Applied Data Science and Smart Systems

Proceedings of 2nd International Conference on Applied Data Science

and Smart Systems 2023 (ADSSS 2023) 15-16 Dec, 2023, Rajpura,
India

Edited by

Jaiteg Singh

S B Goyal

Rajesh Kumar Kaushal

Naveen Kumar

Sukhjit Singh Sehra

Boca Raton London New York

CRC Press is an imprint of the

Taylor & Francis Group, an informa business
First edition published 2025
by CRC Press
4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

and by CRC Press

2385 NW Executive Center Drive, Suite 320, Boca Raton FL 33431

© 2025 selection and editorial matter, Jaiteg Singh, S B Goyal, Rajesh Kumar Kaushal, Naveen Kumar and
Sukhjit Singh Sehra; individual chapters, the contributors

CRC Press is an imprint of Informa UK Limited

The right of Jaiteg Singh, S B Goyal, Rajesh Kumar Kaushal, Naveen Kumar and Sukhjit Singh Sehra to be
identified as the authors of the editorial material, and of the authors for their individual chapters, has been
asserted in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988.

The Open Access version of this book, available at www.taylorfrancis.com, has been made available under a Creative
Commons [Attribution-Non Commercial-No Derivatives (CC-BY-NC-ND)] 4.0 license.

Any third party material in this book is not included in the OA Creative Commons license, unless indicated otherwise
in a credit line to the material. Please direct any permissions enquiries to the original rightsholder.

For permission to photocopy or use material electronically from this work, access www.copyright.com
or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923,
978-750-8400. For works that are not available on CCC please contact [email protected]

Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are used
only for identification and explanation without intent to infringe.

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library

ISBN: 9781032748146 (pbk)

ISBN: 9781003471059 (ebk)

DOI: 10.1201/9781003471059

Typeset in Sabon LT Std

by HBK Digital
 Contents

Preface xx

Chapter 1 AI-driven global talent prediction: Anticipating international graduate admissions 1

Sachin Bhoite, Vikas Magar and C. H. Patil

Chapter 2 English accent detection using hidden Markov model (HMM) 10

Babu Sallagundla, Kavya Sree Gogineni and Rishitha Chiluvuri

Chapter 3 Study of exascale computing: Advancements, challenges, and future directions 17

Neha Sharma, Sadhana Tiwari, Mahendra Singh Thakur, Reena Disawal and Rupali Pathak

Chapter 4 Production of electricity from urine 28

Abhijeet Saxena, Mamatha Sandhu, S. N. Panda and Kailash Panda

Chapter 5 Deep learning-based finger vein recognition and security: A review 34

Manpreet Kaur, Amandeep Verma and Puneet Jai Kaur

Chapter 6 Development of an analytical model of drain current for junctionless GAA MOSFET
including source/drain resistance 43
Amrita Kumari, Jhuma Saha, Ashish Saini and Amit Kumar

Chapter 7 Crop recommendation using machine learning 49

Paramveer Kaur and Brahmaleen Kaur Sidhu

Chapter 8 Environment and sustainability development: A ChatGPT perspective 54

Priyanka Bhaskar and Neha Seth

Chapter 9 GAI in healthcare system: Transforming research in medicine and care for patients 63
Mahesh A., Angelin Rosy M., Vinodh Kumar M., Deepika P., Sakthidevi I. and
Sathish C.

Chapter 10 Fuzzy L-R analysis of queue network with priority 71

Aarti Saini, Deepak Gupta, A. K. Tripathi and Vandana Saini

Chapter 11 Blood bank mobile application of IoT-based android studio for COVID-19 76
Basetty Mallikarjuna, Sandeep Bhatia, Neha Goel, and Bharat Bhushan Naib

Chapter 12 Selection of effective parameters for optimizing software testing effort estimation 82
Vikas Chahar and Pradeep Kumar Bhatia

Chapter 13 Automated detection of conjunctivitis using convolutional neural network 91

Rajesh K. Bawa and Apeksha Koul

Chapter 14 An overview of wireless sensor networks applications, challenges and security attacks 98
N. Sharmila Banu, B.Vidhya and N.Mahendran

Chapter 15 Internet of health things-enabled monitoring of vital signs in hospitals of the future 108
Amit Sundas, Sumit Badotra, Gurpreet Singh and Amit Verma

Chapter 16 Artificial intelligence-based learning techniques for accurate prediction and classification
of colorectal cancer 114
Yogesh Kumar, Shapali Bansal, Ankush Jariyal and Apeksha Koul

Chapter 17 SLODS: Real-time smart lane detection and object detection system 120
Tanuja Satish Dhope, Pranav Chippalkatti, Sulakshana Patil, Vijaya Gopalrao Rajeshwarkar
and Jyoti Ramesh Gangane
Chapter 18 Computational task off-loading using deep Q-learning in mobile edge computing 129
Tanuja Satish Dhope, Tanmay Dikshit, Unnati Gupta and Kumar Kartik

Chapter 19 A comprehensive analysis of driver drowsiness detection techniques 134

Aaditya Chopra, Naveen Kumar and Rajesh Kumar Kaushal

Chapter 20 Issues with existing solutions for grievance redressal systems and mitigation approach using
blockchain network 140
Harish Kumar, Rajesh Kumar Kaushal and Naveen Kumar

Chapter 21 A systematic approach to implement hyperledger fabric for remote patient monitoring 147
Shilpi Garg, Rajesh Kumar Kaushal and Naveen Kumar

Chapter 22 Developing spell check and transliteration tools for Indian regional language – Kannada 152
Chandrika Prasad, Jagadish S. Kallimani, Geetha Reddy and Dhanashekar K.

Chapter 23 Real-time identification of traffic actors using YOLOv7 162

Pavan Kumar Polagani, Lakshmi Priyanka Siddi and Vani Pujitha M.

Chapter 24 Revolutionizing cybersecurity: An in-depth analysis of DNA encryption algorithms in

blockchain systems 172
A. U. Nwosu, S. B. Goyal, Anand Singh Rajawat, Baharu Bin Kemat and Wan Md Afnan
Bin Wan Mahmood

Chapter 25 Exploring recession indicators: Analyzing social network platforms and newspapers textual
datasets 180
Nikita Mandlik, Kanishk Barhanpurkar, Harshad Bhandwaldar, S. B. Goyal, Anand Singh
Rajawat and Surabhi Rane

Chapter 26 NIRF rankings’ effects on private engineering colleges for improving India’s educational
system looked at using computational approaches 187
Ankita Mitra, Subir Gupta, P. K. Dutta, S. B. Goyal, Wan Md. Afnan Bin Wan Mahmood
and Baharu Bin Kemat

Chapter 27 Analysis of soil moisture using Raspberry Pi based on IoT 194

Basetty Mallikarjuna, Sandeep Bhatia, Amit Kumar Goel, Devraj Gautam, Bharat Bhushan
Naib and Surender Kumar

Chapter 28 Drowsiness detection in drivers: A machine learning approach using hough circle
classification algorithm for eye retina images 202
J. Viji Gripsy, N. A. Sheela Selvakumari, S. Sahul Hameed and M. Jamila Begam

Chapter 29 Optimizing congestion collision using effective rate control with data aggregation algorithm
in wireless sensor network 209
K. Deepa, C. Arunpriya and M. Sasikala

Chapter 30 DDoS attack detection methods, challenges and opportunities: A survey 215
Jaspreet Kaur and Gurjit Singh Bhathal

Chapter 31 A review of privacy-preserving machine learning algorithms and systems 220

Utsav Mehta, Jay Vekariya, Meet Mehta, Hargeet Kaur and Yogesh Kumar

Chapter 32 Optimization techniques for wireless body area network routing protocols: Analysis and
comparison 226
Swati Goel, Kalpna Guleria and Surya Narayan Panda

Chapter 33 Securing the boundless network: A comprehensive analysis of threats and exploits in
software defined network 236
Shruti Keshari, Sunil Kumar, Pankaj Kumar Sharma and Sarvesh Tanwar
Chapter 34 A bibliometric analyses on emerging trends in communication disorder 246
Muskan Chawla, Surya Narayan Panda and Vikas Khullar

Chapter 35 Enhancing latency performance in fog computing through intelligent resource allocation
and Cuckoo search optimization 256
Meena Rani, Kalpna Guleria and Surya Narayan Panda

Chapter 36 Pediatric thyroid ultrasound image classification using deep learning: A review 264
Jatinder Kumar, Surya Narayan Panda and Devi Dayal

Chapter 37 Hybrid security of EMI using edge-based steganography and three-layered cryptography 278
Divya Sharma and Chander Prabha

Chapter 38 Efficient lung cancer detection in CT scans through GLCM analysis and hybrid classification 291
Shazia Shamas, Surya Narayan Panda and Ishu Sharma

Chapter 39 Newton Raphson method for root convergence of higher degree polynomials using big
number libraries 298
Taniya Hasija, K. R. Ramkumar, Bhupendra Singh, Amanpreet Kaur and Sudesh Kumar
Mittal

Chapter 40 The influence of compact modalities on complexity theory 307

Lalit Sharma, Surbhi Bhati, Mudita Uppal and Deepali Gupta
Chapter 41 Designing a hyperledger fabric-based workflow management system: A prototype solution
to enhance organizational efficiency 313
Arjun Senthil K. S., Thiruvaazhi Uloli and Sanjay V. M.

Chapter 42 Exploring Image Segmentation Approaches for Medical Image Analysis 322
Rupali Pathak, Hemant Makwana and Neha Sharma

Chapter 43 Design and performance analysis of electric shock absorbers 328

Jenish R. P. and Surbhi Gupta

Chapter 44 Integrating metaverse and blockchain for transparent and secure logistics management 334
A.U. Nwosu, S.B. Goyal, Anand Singh Rajawat, Baharu Bin Kemat and Wan Md Afnan
Bin Wan Mahmood

Chapter 45 A systematic study of multiple cardiac diseases by using algorithms of machine learning 343
Prachi Pundhir and Dhowmya Bhatt

Chapter 46 Forecasting mobile prices: Harnessing the power of machine learning algorithms 348
Parveen Badoni, Rahul Kumar, Parvez Rahi, Ajay Pal Singh Yadav and Siroj Kumar Singh

Chapter 47 Deep learning-based chronic kidney disease (CKD) prediction 363

J. Angel Ida Chellam, M. Preethi, R. Rajalakshmi and E. Bharathraj

Chapter 48 Cattle identification using muzzle images 370

J. Anitha, R. Avanthika, B. Kavipriya and S. Vishnupriya

Chapter 49 Simulation-based evaluating AODV routing protocol using wireless networks 378
Bhupal Arya, Dr. Jogendra Kumar, Dr. Parag Jain, Preeti Saroj, Mrinalinee Singh and Yogesh
Kumar

Chapter 50 Smart agriculture using machine learning algorithms 387

Tript Mann and Jashandeep Kaur

Chapter 51 Cloud computing empowering e-commerce innovation 393

Zinatullah Akrami and Gurjit Singh Bhathal

Chapter 52 Navigating blockchain-based clinical data sharing: An interoperability review 402

Virinder Kumar Singla, Amardeep Singh and Gurjit Singh Bhathal
Chapter 53 Analysis of data backup and recovery strategies in the cloud 409
Sumeet Kaur Sehra and Amanpreet Singh

Chapter 54 Landslide identification using convolutional neural network 416

Suvarna Vani Koneru, Harshitha Badavathula, Prasanna Vadttitya and Sujana Sri Kosaraju

Chapter 55 Retinal vessel segmentation using morphological operations 424

Vishali Shapar and Jyoti Rani

Chapter 56 Liver segmentation using shape prior features with Chan-Vese model 430
Veerpal and Jyoti Rani

Chapter 57 Online video conference analytics: A systematic review 435

Vishruth Raj V. V. and Mohan S. G.

Chapter 58 Sales analysis: Coca-Cola sales analysis using data mining techniques for predictions and
efficient growth in sales 448
Siddique Ibrahim S. P., Pothuri Naga Sai Saketh, Gamidi Sanjay, Bhimavarapu Charan
Tej Reddy, Mesa Ravi Kanth and Selva Kumar S.

Chapter 59 Statistical analysis of consumer attitudes towards virtual influencers in the metaverse 458
Sheetal Soni and Usha Yadav

Chapter 60 Quantum dynamics-aided learning for secure integration of body area networks within the
metaverse cybersecurity framework 467
Anand Singh Rajawat, S. B. Goyal, Jaiteg Singh and Celestine Iwendi

Chapter 61 An optimized approach for development of location-aware-based energy-efficient routing

for FANETs 475
Gaurav Jindal and Navdeep Kaur
Chapter 62 Quantum cloud computing: Integrating quantum algorithms for enhanced scalability and
performance in cloud architectures 482
Anand Singh Rajawat, S. B. Goyal, Sandeep Kautish and Ruchi Mittal

Chapter 63 Integrating AI-enabled post-quantum models in quantum cyber-physical systems

opportunities and challenges 491
S. B. Goyal, Anand Singh Rajawat, Ruchi Mittal and Divya Prakash Shrivastava

Chapter 64 Adaptive resource allocation and optimization in cloud environments: Leveraging machine
learning for efficient computing 499
Anand Singh Rajawat, S. B. Goyal, Manoj Kumar and Varun Malik

Chapter 65 Quantum deep learning on driven trust-based routing framework for IoT in the metaverse
context 509
S. B. Goyal, Anand Singh Rajawat, Jaiteg Singh and Chawki Djeddi

Chapter 66 Advancing network security paradigms integrating quantum computing models for
enhanced protections 517
Anand Singh Rajawat, S. B. Goyal, Chaman Verma and Jaiteg Singh

Chapter 67 Optimizing 5G and beyond networks: A comprehensive study of fog, grid, soft, and
scalable computing models 529
S. B. Goyal, Anand Singh Rajawat, Jaiteg Singh and Tony Jan

Chapter 68 Smart protocol design: Integrating quantum computing models for enhanced efficiency and
security 537
S. B. Goyal, Sugam Sharma, Anand Singh Rajawat and Jaiteg Singh
Chapter 69 Efficient IIoT framework for mitigating Ethereum attacks in industrial applications using
supervised learning with quantum classifiers 544
S. B. Goyal, Anand Singh Rajawat, Ritu Shandilya and Varun Malik

Chapter 70 Quantum computing in the era of IoT: Revolutionizing data processing and security in
connected devices 552
S. B. Goyal, Sardar M. N. Islam, Anand Singh Rajawat and Jaiteg Singh

Chapter 71 A federated learning approach to classify depression using audio dataset 560
Chetna Gupta and Vikas Khullar

Chapter 72 Securing IOT CCTV: Advanced video encryption algorithm for enhanced data protection 565
Kawalpreet Kaur, Amanpreet Kaur, Vidhyotma Gandhi and Bhupendra Singh

Chapter 73 A comprehensive review of federated learning: Methods, applications, and challenges in

privacy-preserving collaborative model training 570
Meenakshi Aggarwal, Vikas Khullar and Nitin Goyal

Chapter 74 Review of techniques for diagnosis of Meibomian gland dysfunction using IR images 576
Deepika Sood, Anshu Singla and Sushil Narang

Chapter 75 The impact of unstable symmetries on software engineering 582

Lalit Sharma, Surbhi Bhati, Mudita Uppal and Deepali Gupta
Chapter 76 Integrating quantum computing models for enhanced efficiency in 5G networking systems 588
Anand Singh Rajawat, S. B. Goyal, Jaiteg Singh and Xiao Shixiao

Chapter 77 Micro-expressions spotting: Unveiling hidden emotions and thoughts 595

Parul Malik and Jaiteg Singh

Chapter 78 Artificial intelligence and machine vision-based assessment of rice seed quality 603
Ridhi Jindal and S. K. Mittal
 List of Figures

Figure 1.1 Graduate admission predictor UI flow 5

Figure 1.2 Train and test F1-score 7
Figure 1.3 Example of XGBoost 7
Figure 2.1 Proposed model diagram 13
Figure 2.2 Model accuracy 14
Figure 2.3 Output when no file is selected 14
Figure 2.4 Output predicted as Indian accent 14
Figure 2.5 Output predicted as Britain 14
Figure 2.6 Output predicted as American 14
Figure 2.7 Spectrogram representation of the model 15
Figure 2.8 Mel-Frequency Cepstral Coefficients 15
Figure 3.1 Exascale computing architecture 19
Figure 3.2 Application of exascale computing 20
Figure 3.3 Benefits of exascale computing 23
Figure 3.4 Challenges of exascale computing 24
Figure 4.1 Block diagram of the proposed system 29
Figure 4.2 Electrolytic cell 29
Figure 4.3 PEM fuel cell 30
Figure 4.4 Schematic representation of the proposed system 31
Figure 5.1 (a) NIR sensor, (b) Finger vein pattern 35
Figure 5.2 Template generation 36
Figure 5.3 Confusion matrix 39
Figure 6.1 Cross-section of cylindrical gate-all-around MOSFET 44
Figure 6.2 GAA JLT in different regions of operation 44
Figure 6.3 Flowchart showing the calculation of IDS 46
Figure 6.4 Drain current validation with the reported data (a) Singh et al., (2011); (b) Moon et al.,
(2013). (Parameters used: (a) L = 160 nm; ND = 6.7 × 1018 cm-3; (b) L = 150 nm; Width of
NW = 18 nm) 46
Figure 6.5 Drain current model validation with the experimental data (Choi et al., 2011). (Parameters
used: L = 50 nm, EOT = 13 nm and ND = 2×1019 cm-3) 46
Figure 6.6 Model validation with the reported data (a) Hu et al. (2014); (b) Wang et al. (2014) 47
Figure 6.7 Comparison of simulated (Lou et al., 2012) and output characteristics of the model 47
Figure 6.8 Variation of the drain current with different drain voltages with and without S/D resistance
(Parameters used: L = 50 nm, ND = 2×1019 cm-3, EOT = 13 nm, Rsd = 15 kΩ) 47
Figure 6.9 Drain current variation with gate overdrive voltage with and without S/D resistance
(Parameters used: L = 50 nm, ND = 2×1019 cm-3, EOT = 13 nm, RSD = 15 kΩ) 47
Figure 7.1 Flow chart of proposed methodology 51
Figure 8.1 Power usage for training large language models (LLMs) based on AI in 2023 (in megawatt
hours) 56
Figure 8.2 Global share of organizations taking steps to reduce carbon emissions from AI use in 2022 57
Figure 8.3 Emissions when training AI-based large language models (LLMs) in 2022 (in CO2
equivalent tons) 57
Figure 8.4 Machine learning (ML) platform emissions in tons of CO2 equivalent in 2022 57
Figure 8.5 Power consumption when training AI based large language models (LLMs) 57
Figure 8.6 Global share of organizations taking action in reducing carbon emissions from their AI use
in 2022 58
Figure 8.7 Types of sustainability activities in which respondents’ organizations using AI in 2022 59
Figure 9.1 Overview of GAI 65
Figure 9.2 Design architecture for GAI 66
Figure 9.3 Decision support system for patients 69
Figure 10.1 Priority queue network model 73
Figure 11.1 The use case diagram of the blood bank 78
Figure 11.2 The sensors connected to the android studio 79
Figure 11.3 Snapshot of the API of android studio 79
Figure 11.4 GUI of new app 79
Figure 11.5 Package explorer 80
Figure 12.1 Software testing stage in SDLC 83
Figure 12.2 Factors governing software testing effort 88
Figure 13.1 AI to detect and classify eye disease 92
Figure 13.2 Proposed system to detect and classify pink as well as healthy eye 93
Figure 13.3 Sample of pink and healthy eyes 93
Figure 13.4 Number of images in the dataset 93
Figure 13.5 Resized dimension of images 93
Figure 13.6 Enhanced eye images 94
Figure 13.7 Augmented images. (a) Healthy eye. (b) Pink eye 94
Figure 13.8 Architecture of CNN model 94
Figure 13.9 Learning curves of CNN model 95
Figure 14.1 Diagram of a typical WSN 99
Figure 14.2 (a) Making centralized decisions design. (b) Distributed determination design 100
Figure 14.3 Sensor nodes architecture 102
Figure 14.4 Applications areas of WSN 102
Figure 14.5 Framework for mobile sinks In WSN 105
Figure 14.6 Multiple nodes request for transmission channel at the same time 105
Figure 14.7 WSN with compromised access point and replicated mobile sink 105
Figure 14.8 Mobile sink revoke compromised node and broadcast control messages to the network 105
Figure 15.1 Provider-controlled electronic health records (EHR) and patient-controlled, mobile,
wearable sensor-based personal health records (PHR) are two types of electronic health
records (EHR) 111
Figure 15.2 Acquisition, storage, processing, and display are shown at the bottom of this schematic of
the Internet of Health Things (IoHT) 111
Figure 16.1 System design for CRC detection 117
Figure 17.1 Statistics for road accidents (Annual report, 2020) 121
Figure 17.2 SLODS 121
Figure 17.3 Lane detection flowchart 122
Figure 17.4 Original image 123
Figure 17.5 Frame gray scaling 123
Figure 17.6 Denoising 123
Figure 17.7 Edge detection 123
Figure 17.8 Hough transform 123
Figure 17.9 Edge detection for each edge detector 124
Figure 17.10 Flowchart for object detection 125
Figure 17.11 Frame capture 125
Figure 17.12 Frame gray scaling 125
Figure 17.13 Extracting ROI 125
Figure 17.14 ROI masking 125
Figure 17.15 Final output 126
Figure 17.16 Real time lane detection 126
Figure 17.17 Real time object detection 127
Figure 18.1 Block diagram of edge computing model 131
Figure 18.2 Number of tasks in queue with respect to time (s) 132
Figure 18.3 Server utilization with respect to time (ms) 132
Figure 18.4 CPU utilization of different CPU’s and number of tasks in queue during a single window
execution, time (s) vs. number of tasks 132
Figure 18.5 Episodes vs. number of tasks for local computation and edge server off-loading using
Q-learning algorithm 132
Figure 18.6 Number of tasks vs. average delay (s) for local and off-loading using Q-learning algorithm 133
Figure 20.1 Prisma flow diagram for literature study 141
Figure 20.2 Structure of grievances redressal system 141
Figure 20.3 Publication trends in existing solutions on grievances redressal system 144
Figure 20.4 Structure of the blockchain-based proposed system 145
Figure 21.1 Transaction flow for hyperledger fabric 148
Figure 21.2 System architecture for hyperledger fabric 149
Figure 21.3 Spec.yaml file for network 149
Figure 21.4 Smart contracts for entities 150
Figure 21.5 Screenshot for creating a patient 150
Figure 21.6 Screenshot for query a patient 150
Figure 21.7 Result for query a patient 150
Figure 21.8 (a) Blocks per min, (b) Transactions per organization 150
Figure 22.1 Chronological order of face shield development 155
Figure 22.2 Input word not present in Bloom filter 156
Figure 22.3 Input word present in Bloom filter 156
Figure 22.4 Input misspelled word to Symspell give suggestion word 157
Figure 22.5 Performance analysis 157
Figure 22.6 User interface for spell check 158
Figure 22.7 Identifying wrong words (underlined in red) 158
Figure 22.8 Selecting the wrong words with options 158
Figure 22.9 Suggesting the correct word for the misspelled word 158
Figure 22.10 Pipeline for the process 158
Figure 22.11 Character BERT embedding 159
Figure 22.12 (a) Mapping vowels to its IPA representation 159
Figure 22.12 (b) Mapping consonants to its IPA representation 159
Figure 22.13 Mapping consonants to its IPA 160
Figure 23.1 A sample image from the dataset 164
Figure 23.2 Architecture diagram of YOLOv7 164
Figure 23.3 Different results curves 168
Figure 23.4 Precision-recall curve 168
Figure 23.5 Confusion matrix 169
Figure 23.6 (a) & (b) represents the sample input images predicted using model 170
Figure 23.7 (a) & (b) represents the predicted images generated by the model for the sample input
images 170
Figure 24.1 Cryptography mechanism of DNA encryption 173
Figure 24.2 The layered diagram of a blockchain 174
Figure 25.1 Data flow diagram 181
Figure 25.2 (a) Tweets collected over time for recession 182
Figure 25.2 (b) Data collected for recession comments (Reddit) 182
Figure 25.2 (c) News articles collected over time for recession 182
Figure 25.3 Memory-usage of each data source 183
Figure 25.4 (a) Number of words for each tweet 183
Figure 25.4 (b) Number of words for each Reddit comment 184
Figure 25.4 (c) Number of words for every NYTimes headline 184
Figure 25.5 (a) Word-cloud for the Twitter data 184
Figure 25.5 (b) Word-cloud for the Reddit data 184
Figure 25.5 (c) Word-cloud for the New York Times data (Khattar et al., 2020) 184
Figure 25.6 (a) Sentiment score of the tweets 185
Figure 25.6 (b) Sentiment score of the Reddit posts 185
Figure 25.6 (c) Sentiment score of the Reddit posts 185
Figure 26.1 Methodology 191
Figure 26.2 Confidence intervals 192
Figure 26.3 Distribution curve 192
Figure 26.4 Histogram 192
Figure 26.5 Power F distribution curve 192
Figure 27.1 Flowchart of soil moisture detection monitoring system 196
Figure 27.2 The proposed system’s component parts 196
Figure 27.3 Circuit details of Raspberry Pi 197
Figure 27.4 ESP 8266 with sensors 197
Figure 27.5 Node MCU (ESP 8266) 197
Figure 27.6 Test result 198
Figure 27.7 Field chart with irrigation of data 198
Figure 27.8 Chart with population of different countries 198
Figure 27.9 Different output representations as per the given data 199
Figure 28.1 Image mining process 203
Figure 28.2 External eye structure 203
Figure 28.3 System architecture 204
Figure 28.4 Pre-processing using noise reduction, sharpness, contrast, brightness 206
Figure 28.5 Pre-processing result using noise reduction, sharpness, contrast, brightness 206
Figure 28.6 Segmentation comparisons for right eye image threshold 1.5, 1.6, 1.8 and image threshold
1.7 using E_GRUNS algorithm 207
Figure 28.7 Performance analysis of classifiers 207
Figure 28.8 Mean absolute error 207
Figure 28.9 Execution time 208
Figure 29.1 Throughput 212
Figure 29.2 Packet loss 212
Figure 29.3 E2E delay 213
Figure 29.4 Data transfer rate 213
Figure 30.1 Architecture of cloud computing 216
Figure 30.2 DDoS attack 217
Figure 30.3 Botnet-based DDoS 218
Figure 31.1 Schematic representation of distributed system 221
Figure 31.2 Schematic representation of outsourced system 221
Figure 32.1 Organization of paper 227
Figure 32.2 Systematic review structure 228
Figure 33.1 Paper’s roadmap 237
Figure 33.2 SDN architecture 237
Figure 33.3 Work flow of SDN 238
Figure 33.4 Flow of analysis 239
Figure 33.5 Analysis of attacks 244
Figure 34.1 Different communication disorders with their prevalence 247
Figure 34.2 Significant rise in publications since 2002 247
Figure 34.3 Flowchart depicting pre-selection of the articles 248
Figure 34.4 Significant rise in cumulative frequency of the keywords 250
Figure 34.5 Cluster formation using co-occurrence of keywords 250
Figure 34.6 Significant works by numerous authors 251
Figure 34.7 Publications by subject area 251
Figure 34.8 Publications contributed by various countries 252
Figure 35.1 The suggested methods for allocating the task in the fog node 260
Figure 35.2 A flowchart of the proposal’s steps 261
Figure 36.1 Endocrine system 265
Figure 36.2 Thyroid gland 265
Figure 36.3 Phases of medical image processing 266
Figure 36.4 Reasons for medical image segmentation 266
Figure 36.5 Problems in image segmentation 267
Figure 36.6 Image segmentation techniques 267
Figure 36.7 Hierarchical relationships of AI, ML, DL and CNN 268
Figure 36.8 Evolution of machine learning technology 269
Figure 36.9 Comparison of ML and DL 270
Figure 36.10 Types of neural network (a) Traditional neural network (b) CNN (c) FCN 270
Figure 36.11 Restricted Boltzmann machines (RBMs) 271
Figure 36.12 Autoencoder architectures with vector 271
Figure 37.1 Commonly used data storage devices 279
Figure 37.2 Common types of EMI 279
Figure 37.3 Commonly used smart medical devices that generate EHR 279
Figure 37.4 Process of normalization of cover image Lena 283
Figure 37.5 Few of the 5856 X-ray images that act as input for the PHM 283
Figure 37.6 Output images achieved after implementing the proposed PHM method 285
Figure 38.1 Proposed block diagram 294
Figure 38.2 Pre-processed image 295
Figure 38.3 FFA segmentation 295
Figure 38.4 SIFT feature extraction 296
Figure 38.5 SURF feature extraction 296
Figure 38.6 Graphical comparison between SIFT and SURF 296
Figure 39.1 Flow chart of the procedure followed to implement 3 experiments using primitive and big
number data types and their accuracy and correctness evaluation 301
Figure 39.2 Coefficient bits and accurate root bits relationship 304
Figure 39.3 Graph of time complexities of root convergence experiments using primitive and big
number data types 304
Figure 40.1 (a) Analysis of online algorithms 309
Figure 40.1 (b) Schematic representation of correlation between heuristic approaches and large-scale
procedures 309
Figure 40.2 Illustration of the phenomenon where distance increases as complexity decreases 310
Figure 40.3 Comparison of median latency between the methodology and other methodologies 310
Figure 40.4 Comparison of expected complexity between KamMone and other algorithms 310
Figure 40.5 Relationship between the expected block size of KamMone and latency 311
Figure 41.1 Case study workflow design 314
Figure 41.2 Prototype application basic design 319
Figure 42.1 The segmentation outcomes 323
Figure 42.2 Left part represents the MR image with noise and right side represents image without noise 324
Figure 42.3 Real-world medical images are used to test the procedure. Column 1 contains the original
images and contours. Column 2 has the final outlines. Column 3 contains photos that have
been adjusted for bias. Column 4 contains the estimated bias fields 324
Figure 42.4 Alternative approaches are compared. Column 1 displays the original pictures and initial
outlines. The C-V model is shown in column 2. There are columns 3: Li’s method. The
LGDF model may be found in column 4. Column 5 – In this scenario, the method should
be followed 325
Figure 42.5 Correct segmentation of CT image 325
Figure 42.6 An example of interactive image segmentation 325
Figure 43.1 Electric shock absorbers 329
Figure 43.2 Valve before rotation 330
Figure 43.3 Valve after rotation 330
Figure 43.4 Ride frequency estimation 330
Figure 43.5 Controller 330
Figure 43.6 Circuit diagram 331
Figure 43.7 Damping 331
Figure 44.1 The architecture and layers of the metaverse 335
Figure 44.2 The transaction process of blockchain 337
Figure 44.3 The architecture of blockchain-based logistics management metaverse system 338
Figure 44.4 Simulations results 340
Figure 44.5 Latency comparison 340
Figure 44.6 Throughput comparison 340
Figure 45.1 2022 leading cause of death 344
Figure 45.2 Number of deaths by cause, World 2019 344
Figure 45.3 Share of total disease burden by cause, World 2019 344
Figure 45.4 Utilizing a phonogram (Baghel et al., 2020) signals from the current CVD classes.
(a) Aortic stenosis (AS), (b) Mitral regurgitation (MR), (c) Mitral stenosis (MS),
(d) Mitral valve prolapse (MVP) and (e) Normal 345
Figure 46.1 Data set values without company columns 350
Figure 46.2 Data set values with company columns 350
Figure 46.3 Exploring the data values 351
Figure 46.4 Example of data set by using head() function 352
Figure 46.5 Data description of dataset that we have taken 352
Figure 46.6 Boxplot diagram for outliers and missing values 353
Figure 46.7 Heatmap for comparing and choosing the attribute to classify our model 353
Figure 46.8 Count plot is used to count the occurrence of the observation 354
Figure 46.9 Count plot is used to count the occurrence of the RAM 355
Figure 46.10 Count plot is used to count the occurrence of the primary camera 356
Figure 46.11 This pairplot is used to understand the best set of feature to explain a relationship between
two or more variables to perform cluster separation 356
Figure 46.12 Predicting the model after applying the decision tree 358
Figure 46.13 Attributes vs. accuracy of five attributes 359
Figure 46.14 Attributes vs. accuracy after adding RAM feature 360
Figure 47.1 Causes of CKD 364
Figure 47.2 Block diagram of the proposed approach for CKD prediction 366
Figure 47.3 Architecture of proposed DNN mode 366
Figure 47.4 Performance of the proposed DN 368
Figure 47.5 Class analysis of CKD dataset 368
Figure 47.6 Comparative analysis 368
Figure 48.1 Output of muzzle classification using YOLOv5 model 372
Figure 48.2 Summary of CNN model 373
Figure 48.3 Flow diagram of proposed system 374
Figure 48.4 Plot of accuracy 375
Figure 48.5 Plot of loss 375
Figure 48.6 Register page 375
Figure 48.7 Dashboard for details 375
Figure 48.8 Matched result 376
Figure 48.9 Not matched result 376
Figure 48.10 Feedback responses 376
Figure 49.1 Wireless networks scenario for routing protocols 380
Figure 49.2 Simulation view 380
Figure 49.3 Total byte sent 381
Figure 49.4 Total packet sent 381
Figure 49.5 First packet sent 382
Figure 49.6 Last packet sent 382
Figure 49.7 Average jitter 382
Figure 49.8 First packet received 383
Figure 49.9 Total byte received 383
Figure 49.10 Total packet received 383
Figure 49.11 Last packet received 384
Figure 49.12 Average end to end delay(s) 384
Figure 49.13 Throughput (bits/s) 384
Figure 50.1 Flow chart for crop recommendation 389
Figure 50.2 Flow chart for fertilizer recommendation 389
Figure 50.3 Flow chart for irrigation scheduling 390
Figure 50.4 Flow chart for soil moisture prediction 391
Figure 51.1 Shows global ecommerce retail sales reached $5.7 trillion in 2022. This share is expected
to increase by 10% in 2023 and reach $6.3 trillion. 395
Figure 51.2 Comparison of traditional e-commerce and cloud e-commerce 395
Figure 52.1 Blockchain structure 403
Figure 52.2 Blockchain applications in healthcare 403
Figure 52.3 Identified research gaps 407
Figure 53.1 Fundamental IoT architecture 410
Figure 53.2 Important blocks of the proposed model 411
Figure 53.3 Components of the centralized backup system 411
Figure 53.4 Challenge-response-authentication process time overhead 413
Figure 53.5 Reference sequence segmentation error comparison 413
Figure 53.6 Comparison with native system throughput 414
Figure 53.7 Comparison of data backup management 414
Figure 53.8 Comparison of off-server copy management 414
Figure 53.9 Comparison of storage device management 415
Figure 54.1 Resized landslide image 418
Figure 54.2 Transfer learning architecture 419
Figure 54.3 Accuracy of the model 420
Figure 54.4 Loss of the model 421
Figure 54.5 Select the input image 421
Figure 54.6 Input image 421
Figure 54.7 Output 422
Figure 55.1 Normal gray scale images 426
Figure 55.2 Abnormal gray scale images 426
Figure 55.3 Illumination corrected normal images (a) and corresponding low frequency components (b) 427
Figure 55.4 Illumination corrected abnormal images (a) and corresponding low frequency component (b) 427
Figure 55.5 Illumination corrected and, adaptive histogram equalized normal images 427
Figure 55.6 Illumination corrected and adaptive histogram equalized abnormal images 428
Figure 55.7 Clique function treated normal images 428
Figure 55.8 Clique function treated abnormal images 428
Figure 56.1 The comparison of existing and the proposed methods. (a) Represents the original images.
(b) Represents the ground truth images. (c) Represents the existing method and
(d) Represents the proposed method. 432
Figure 56.2 Shows the original image belongs to the SLIVER dataset 432
Figure 56.3 Shows the intensity histogram for the whole CT image 433
Figure 56.4 The intensity histogram for the whole CT image 433
Figure 58.1 Flow chart of sales analysis 450
Figure 58.2 Output of elbow method (using clustering techniques) 450
Figure 58.3 Yearly consumption for soft drink 452
Figure 58.4 Output of CAGR 452
Figure 58.5 Relationship between likelihood to recommend and preferred product 453
Figure 58.6 Yearly consumption trends for soft drink brands (with future predictions). Dotted line
shows the future sales of soft drinks. 453
Figure 58.7 Seasonal analysis for soft drink brands with respective to their region 454
Figure 58.8 [Left] Total sales over the year [Right] year-over-year sales growth rate 455
Figure 58.9 Model performance matrices including accuracy, precision, recall 456
Figure 59.1 Flow chart of methodology adopted 461
Figure 59.2 Matrix scatterplot 462
Figure 60.1 Process flow 469
Figure 61.1 Different UAV network level communications 476
Figure 61.2 FANET network in smart city modeling 476
Figure 61.3 Proposed system model 478
Figure 61.4 FANET network deployment 479
Figure 61.5 Back propagation training 479
Figure 61.6 Energy consumption 479
Figure 61.7 Latency in the data gathering among nodes 479
Figure 61.8 End delay (sec) 480
Figure 61.9 Energy consumption per round 480
Figure 62.1 Integrating quantum model for enhanced scalability and performance in cloud architectures 484
Figure 62.2 Proposed model flow chart 486
Figure 63.1 Integrating AI-enabled post-quantum models in quantum cyber-physical systems
opportunities and challenges 496
Figure 64.1 Proposed machine learning model 502
Figure 64.2 Comparative analysis different machine learning model 506
Figure 65.1 Details flow 512
Figure 66.1 Chronological order of face shield development 522
Figure 69.1 Proposed model 546
Figure 71.1 Federated learning model for depression detection using audio data 561
Figure 71.2 Training and validation accuracy and loss results for depression detection using Bi-LSTM,
CNN and LSTM algorithms 562
Figure 71.3 Client training results for depression detection using IID data 562
Figure 71.4 Client validation results for depression detection using IID data 563
Figure 71.5 Server validation results for depression detection using IID data 563
Figure 72.1 Time taken to perform encryption and decryption without compression 568
Figure 72.2 Time taken to perform encryption and decryption with compression using proposed OVEA 568
Figure 72.3 Time taken to perform encryption 568
Figure 73.1 Federated learning framework 571
Figure 73.2 Federated learning taxonomy 572
Figure 74.1 Graph for prevalence of MGD based on ethnicity 577
Figure 74.2 (a) Detailed view of MG’s in upper lid 577
Figure 74.2 (b) Detailed view of MG’s in lower lid 577
Figure 74.3 Malfunction of MG’s 577
Figure 75.1 Illustration of the new optimal models 584
Figure 75.2 Relationship between the median energy of Bolye and the hit ratio 584
Figure 75.3 Relationship between the average block size of Bolye and the popularity of reinforcement
learning 585
Figure 75.4 Median sampling rate comparison of Bolye with other approaches 585
Figure 75.5 (a) Relationship between interrupt rate growth and decreasing popularity of link-level
acknowledgments 586
Figure 75.5 (b) Average energy of the approach in relation to instruction rate 586
Figure 78.1 ANN architecture 605
Figure 78.2 Flow chart of the process 605
Figure 78.3 Test image for grade-2 Ponni rice 606
Figure 78.4 Test image for grade-2 Matta rice 606
Figure 78.5 ANN framework 606
Figure 78.6 Grading of rice grains according to varieties using ANN classifier 607
Figure 78.7 Regression plots 607
 List of Tables

Table 3.1 Technological overview of exascale system 18

Table 3.2 Represents the market growth of EC 26
Table 4.1 Comparison of electrolysis 31
Table 4.2 Comparison of efficiency of PEM fuel cell 31
Table 5.1 Comparison of different biometric modalities 35
Table 5.2 Datasets for the finger vein. 38
Table 5.3 Recent articles related to the CNN and template security of FVR 39
Table 7.1 Feature description 51
Table 7.2 Accuracy of algorithms 51
Table 8.1 Effective steps for reducing environmental impact of ChatGPT 59
Table 10.1 Crisp values 73
Table 10.2 Fuzzy particular values 74
Table 10.3 L-R Fuzzy Values 74
Table 11.1 Frequency of occurring in different blood groups (Fahim et al., 2016). 77
Table 11.2 Description of package explorer 80
Table 12.1 Comparative analysis of existing studies 86
Table 12.2 Categorization of factors effecting the software testing effort 87
Table 13.1 Layered architecture of proposed CNN model 94
Table 13.2 Performance metrics 95
Table 13.3 Hyper-parameters of CNN model 95
Table 13.4 Evaluation of CNN model 95
Table 13.5 Performance summary. 95
Table 13.6 Class-wise performance metrics for eye classification. 95
Table 15.1 Typical hospital vitals used for patient monitoring 109
Table 15.2 Methods, frameworks, and systems for assessing risk 110
Table 16.1 Cases and deaths in the US 2020 due to CRC 115
Table 16.2 Comparative analysis of CRC 116
Table 17.1 Area of interest vertices 122
Table 17.2 Hough transform parameters 122
Table 17.3 Canny algorithm parameters 125
Table 17.4 Parameters for Robert operator 126
Table 17.5 Parameters for Prewitt operator 126
Table 17.6 Parameters for Sobel operator 126
Table 17.7 Parameters for Canny operator 126
Table 17.8 Speed analysis 127
Table 18.1 Parameters for deep Q-learning algorithm 131
Table 19.1 List of various work done on driver drowsiness detection. 138
Table 20.1 Literature review for grievances redressal system. 142
Table 21.1 Process to build up the minifab network. 149
Table 22.1 Dataset details. 157
Table 22.2 Performance analysis. 157
Table 22.3 Dataset split up. 158
Table 22.4 Sample transliterations of Kannada words 160
Table 23.1 Overview of confusion matrix. 169
Table 24.1 Analysis of different types of DNA-based schemes. 175
Table 24.2 Analysis of DNA-based encryption solutions 175
Table 24.3 Analysis of blockchain system-based on DNA encryption algorithm. 175
Table 24.4 Case Studies on the integration of blockchain system with DNA based encryption algorithm 177
Table 24.5 Comparative analysis of existing conventional system and proposed solution. 177
Table 25.1 Comparative analysis of different studies associated with Twitter, Reddit, and the
New York Times. 181
Table 25.2 Comparison of number of records and memory usage for various datasets. 183
Table 26.1 Comparative analysis of different learning outcomes. 190
Table 26.2 H hypothesis. 191
Table 28.1 Performance analysis of classification. 207
Table 32.1 Inclusion and exclusion criteria. 228
Table 32.2 Comparative analysis of routing protocols using optimization techniques 232
Table 33.1 Attacks in application plane 240
Table 33.2 Attacks in control plane 240
Table 33.3 Attacks in data plane 240
Table 33.4 Categorization of attacks, affecting multiple planes 241
Table 33.5 Attack vectors with their tools and techniques 242
Table 33.6 SDN component with their threats 243
Table 34.1 Metadata about the research contributions 252
Table 34.2 CiteScore per year 253
Table 34.3 Source normalized impact per paper 253
Table 34.4 SCImago journal rank (SJR) 253
Table 34.5 Meta-perspectives of the study 254
Table 35.1 Literature survey summary 258
Table 35.2 Parameters of the fog environment 262
Table 35.3 Final outcome 262
Table 36.1 Thyroid diseases and their symptoms 265
Table 36.2 DL segmentation performance parameter 274
Table 37.1 Literature review of studied research articles 281
Table 37.2 The stego-encryption algorithm for the PHM 284
Table 37.3 Comparison based on the size of the data set and programming language used in previous
research 285
Table 37.4 Encryption time for the proposed hybrid method 286
Table 37.5 Proposed method decryption time 286
Table 37.6 Comparison based result table for the proposed hybrid method with the research article
that led to this research 287
Table 39.1 Newton Raphson method to find first real root 300
Table 39.2 Polynomial root finding with primitive data types 302
Table 39.3 Polynomial with primitive co-efficient and big number constant 303
Table 39.4 Polynomial root convergence with big numbers 305
Table 41.1 Comparison of different existing solutions. 318
Table 44.1 Comparative analysis of existing work with limitations 338
Table 44.2 Tools and specifications 339
Table 44.3 Latency results between the proposed system and the existing system 340
Table 44.4 Throughput comparison between the baseline system and the proposed system 340
Table 44.5 Comparative analysis of existing metaverse-based systems 341
Table 45.1 Comparisons of previous studies on classification of cardiac diseases. 346
Table 46.1 Summary of parameters 358
Table 46.2 Top five attributes with accuracy. 359
Table 46.3 Bottom seven attributes with accuracy 359
Table 47.1 CKD dataset description. 367
Table 47.2 Performance of the proposed DNN. 368
Table 47.3 Comparative analysis. 368
Table 48.1 Descriptive statistics of feedback responses. 376
Table 49.1 Parameters list (simulation setup). 379
Table 49.2 Performance metrics. 381
Table 50.1 Features description for dataset of crop recommendation. 388
Table 50.2 Feature description for dataset of fertilizer recommendation. 389
Table 50.3 Features description for dataset of irrigation scheduling. 390
Table 50.4 Feature description for dataset of soil moisture prediction. 390
Table 50.5 Classification metrics for crop recommendation system. 391
Table 50.6 Classification report for fertilizer recommendation system. 391
Table 50.7 Regression metrics for soil moisture prediction. 391
Table 50.8 Classification report for irrigation scheduling. 391
Table 52.1 Frequency of identified problem areas. 406
Table 55.1 Average value of sensitivity, specificity and accuracy for different normal retinal images. 428
Table 55.2 Average values of vessel-to-vessel free area for proposed and existing methods. 428
Table 56.1 Results for SLIVER 433
Table 56.2 Results for IRCAD 433
Table 57.1 Summary of state-of-the-art methods 441
Table 57.2 Summary of limitation, strength and open problems 443
Table 59.1 Descriptive statistics 462
Table 59.2 Collinearity diagnostics 462
Table 59.3 Model summary 463
Table 59.4 ANOVA 463
Table 59.5 Coefficients 463
Table 60.1 Comparative analysis 468
Table 60.2 Data set description 470
Table 60.3 Simulation Parameters for Assisted Learning in Quantum Dynamics (S(BAN)+ Q(MV)) 471
Table 60.4 Analytical results on the utilization of quantum dynamics to enhance learning in a
multi-agent environment 472
Table 61.1 Performance comparison 480
Table 62.1 Comparative analysis 484
Table 62.2 A comprehensive overview of the impact of each variable on the main performance metrics 488
Table 63.1 Comparative analysis 492
Table 63.2 Datasets relevant to quantum cyber-physical systems (QCPS) that incorporate AI-enabled
post-quantum model integration 494
Table 63.3 Simulation parameter 496
Table 63.4 Results analysis 497
Table 64.1 Comparative analysis 501
Table 64.2 simulation parameters utilized in a cloud system that employs machine learning techniques to
achieve optimal resource allocation and optimization. 505
Table 64.3 Results analysis 506
Table 65.1 A comprehensive compilation of critical information from each study 511
Table 65.2 Simulation parameters and quantum parameters. 514
Table 65.3 Results analysis 515
Table 66.1 Comparative table. 519
Table 66.1 Encryption and data protection 525
Table 66.2 A comparative analysis of the performance of QML algorithms and standard (classical)
machine. 525
Table 66.3 Comparative analysis of classical and quantum search algorithms. 526
Table 66.4 A visual representation of the potential superiority of quantum algorithms. 526
Table 66.5 Measurements of electronic components. 527
Table 67.1 Optimizing 5G and beyond networks. 531
Table 67.2 Overall impact on 5G networks. 534
Table 67.3 Impact on 5G and beyond networks. 534
Table 68.1 Comparative analysis. 538
Table 68.2 Comparison of key performance metrics. 540
Table 68.3 Results analysis. 541
Table 68.4 Results analysis traditional protocol design and QC/PQC-enhanced protocol design. 541
Table 69.1 Comparative analysis. 545
Table 69.2 An illustrative dataset for a robust IIoT architecture. 546
Table 69.3 A generic parameters. 549
Table 69.4 An examination of the results pertaining to a proficient IIoT framework. 549
Table 70.1 Comparative analysis between traditional IoT system and quantum-enhanced IoT system. 556
Table 70.2 Results analysis between traditional IoT system and quantum-enhanced IoT system. 557
Table 71.1 Training and validation results of DL models for depression detection 561
Table 71.2 Training and validation results of FL models for depression detection using IID data 562
Table 72.1 Key findings in the literature 567
Table 72.2 A comparison results of encryption speed time (in frame/seconds) for different videos
without compression 567
Table 72.3 A comparison results of encryption speed time (in frame/seconds) for different videos with
compression 568
Table 72.4 Comparison of time taken by different encryption algorithms 568
Table 73.1 Types of FL architectures with applications 572
Table 73.2 Strategies to handle challenges 573
Table 74.1 Ethnicity-based prevalence of MGD (Hassanzadeh et al., 2021) 577
Table 74.2 Different techniques for diagnosis of Meibomian gland dysfunction 579
Table 76.1 Literature review on quantum computing and network security 590
Table 76.2 Test scenario 592
Table 76.3 Results analysis for quantum AI in 5G networks 592
Table 77.1 Few of the available published research reports on MEs analysis of recent years 596
Table 77.2 Spotting techniques employed in facial MEs for pre-processing 596
Table 77.3 Few of the existing techniques employed for spotting a facial MEs 598
Table 78.1 Grade of rice grains 605
Table 78.2 Analysis of rice seeds in a random sample 607
Table 78.3 Results of a random sample regarding the size of the rice seeds with percentage values 608
Table 78.4 Results of a random sample in terms of the size of the rice seeds with percentage values by
Human inspector 608
Table 78.5 Result analysis 608
 Preface: Second International Conference on Applied Data
Science and Smart Systems

The Second International Conference on Applied Data Science and Smart Systems (ADSSS-2023) was held
on 15-16 December 2023 at Chitkara University, Punjab, India. This multidisciplinary conference focussed
on innovation and progressive practices in science, technology, and management. The conference successfully
brought together researchers, academicians, and practitioners across different domains such as artificial intel-
ligence and machine learning, software engineering, automation, data science, business computing, data com-
munication, and computer networks. The presenters shared their most recent research works that are critical
to contemporary business and societal landscape and encouraged the participants to devise solutions for real-
world challenges.
ADSSS-2023 featured an extensive selection of tracks, each delving into critical facets of applied data science
and smart systems. “Machine Learning Principles, Smart Solutions, and Business Strategies” provided insights
into the synergy between ML principles, innovative solutions, and strategic business applications. The track on
“AI and Deep Learning” explored the latest advancements and applications in artificial intelligence and deep
learning technologies. Addressing contemporary challenges, “Data Science Techniques for Handling Epidemic,
Pandemic” showcased the role of data science in managing health crises. “Deep Intelligence for Interdisciplinary
Research” facilitated discussions on the integration of deep intelligence across diverse research domains. The
track focusing on “Software Engineering and Automation” explored methodologies that enhance efficiency and
automate processes in the realm of software development. “Data Communication and Computer Networks”
shed light on evolving communication technologies, while “Computing in Business and Learning” addressed the
intersection of computing technologies with business strategies and educational practices. Finally, “Engineering
Mathematics and Physics” provided a platform for exploring the application of mathematical and physical prin-
ciples in various engineering disciplines. This diverse array of tracks collectively contributed to a comprehensive
exploration of applied data science, fostering interdisciplinary collaboration and offering valuable insights for
future research and advancements in the field.
ADSSS-2023 was honored to host eminent scientists and researchers from across the globe, whose insightful
keynote addresses enhanced participants’ knowledge. The keynotes covered diverse topics such as AI-Powered
Quantum Cryptography, the Role of Large Language Models in Scientific Research, and the Application of Deep
Gaussian Processes in Radio Map Construction and Localization. Beyond knowledge dissemination, the confer-
ence served as a dynamic platform for networking and collaboration among researchers, fostering the exchange
of ideas that may shape future research endeavours. We trust that every participant found the ADSSS-2023
experience to be both enriching and productive.
 Editors

Dr. Jaiteg Singh is an accomplished academician with 19 years of experience in

research, development, and teaching in computer science and engineering. Possesses
expertise in diverse fields like Neuromarketing, Navigation systems, Software
Engineering, Business Intelligence, and Data Mining. He has supervised thirteen Ph.D.
thesis and eighteen M.Tech thesis. He has generated significant consultancy revenue,
organized international conferences, published extensively in renowned journals
indexed at reputed databases like Scopus and Web of Science. He has authored/
edited six books, secured four US copyrights, and even filed for over 50 patents.
He is serving as reviewer for numerous journals from reputed publishers like
Elsevier, Springer, Wiley, MDPI and IEEE.
Email: [email protected]

Dr. S B Goyal, a distinguished personality in the realm of Computer Science &

Engineering, earned his Ph.D. from Banasthali University, Rajasthan, India, in 2012.
With over two decades of rich experience spanning both national and international
levels, he has made significant contributions to the academic and administrative sec-
tors of numerous institutions.
Dr. Goyal’s unparalleled curiosity and dedication to staying abreast of the latest IT
developments have established him as an authority in Industry 4.0 technologies. His
expertise encompasses a wide range of cutting-edge fields, including Big Data, Data
Science, Artificial Intelligence, Blockchain, and Cloud Computing.
Dr. Goyal actively participates in panel discussions, sharing his knowledge on
Industry 4.0 technologies in both academic and industry platforms. His edito-
rial prowess is recognized through his contributions as a reviewer, guest editor, and co-editor for numerous
International Journals and Scopus books published by prestigious organizations like IEEE, Inderscience, IGI
Global, and Springer.
An esteemed IEEE Senior member since 2013, Dr. Goyal has an impressive record of contributions to Scopus/
SCI journals and conferences as an editor. His innovative spirit is further evidenced by his possession of over 10
international patents and copyrights from countries including Australia, Germany, Japan, and India.
Throughout his illustrious career, Dr. Goyal has been the recipient of numerous academic excellence awards
at both national and international levels. Currently, he holds the esteemed position of Director at the Faculty
of Information Technology, City University, Malaysia. His commitment to emerging technologies continues to
inspire and pave the way for future innovations and technological advancements.
Email: [email protected]

Dr. Rajesh Kumar Kaushal is a highly accomplished and dedicated researcher and
educator, boasting a robust background in Computer Science and Engineering, and
accumulating an extensive 19 years of experience in academia since 2004. Holding a
Ph.D. in Computer Science and Engineering from Chitkara University, Punjab, India,
he currently serves as a professor in the Department of Computer Applications and
actively contributing to research endeavors.
Dr. Kaushal’s prolific research output is evident through his involvement in more
than 60 patent filings, with one of them being published in the World Intellectual
Property Organization (WIPO) and 15 being granted. Additionally, he has taken
on the role of Project Manager in a DST-funded project titled “Smart and Portable
Intensive Care Unit,” supported by the Millennium Alliance (FICCI, USAID, UKAID, Facebook, World Bank,
and TDB DST). Furthermore, he served as a Co-Principal Investigator in another DST-funded project named
“Remote Vital Information and Surveillance System for Elderly and Disabled Persons,” under the Technology
Intervention of Disabled and Elderly (TIDE) scheme of the Ministry of Science & Technology, Government of
India. Presently he is working on another DST funded project named “Smart Ergonomic Portable Commode
Chair” under the DST TIDE scheme. He is also actively contributing the research community in the area of
Blockchain and Internet of Things and have published more than 55 research papers and all of them are either
indexed in SCOPUS or SCI.
Additionally, he serves as a visiting professor at Kasetsart University in Sakon Nakhon, Thailand, and actively
participates as a reviewer for the peer-reviewed journal “Technology, Knowledge & Learning.” Recognizing
his exceptional contributions to education, Dr. Kaushal has been honoured with the Teacher Excellence Award
twice, receiving the accolade in both 2017 and 2019 from Chitkara University, Punjab, India. In 2017, he
earned the Teacher Excellence Award in the category of “Most Enterprising,” and in 2019, the recognition was
bestowed upon him in the category of “Most Enterprising & Emerging Leader.”
Email: [email protected]

Dr Naveen Kumar is a distinguished academician and researcher with a Ph.D. in

Computer Science and Engineering, boasting over 25 years of rich experience in
both teaching and research domains. His journey includes notable contributions
as a research fellow at PGIMER, Chandigarh, where he actively contributed to a
Department of Science & Technology (DST) funded project titled “Closed Loop
Anaesthesia Delivery System.” Currently serving as an Associate Professor at
Chitkara University Research and Innovation Network (CURIN) within the Research
Department of Chitkara University, Punjab, Dr Naveen Kumar continues to be at
the forefront of cutting-edge research initiatives. His dedication to academic excel-
lence is evident through his extensive engagement in research activities, reflected in
his impressive track record. Dr. Naveen Kumar’s expertise extends to intellectual
property rights, with a remarkable portfolio comprising over 70 filed patents, of
which more than 30 have been granted. Furthermore, his contributions to scholarly
literature are substantial, with over 50 research articles published across various
prestigious journals and conferences. His commitment to knowledge dissemination is highlighted by his author-
ship of the book “Workshop Practice,” published by Abhishek Publication. He is actively engaged in many Govt.
funded research projects and presently he is working on a research project as a Principal Investigator “Smart
Ergonomic Portable Commode chair”. This initiative, supported by the Technology Intervention for Disabled
and Elderly (TIDE) scheme under the Ministry of Science & Technology, Government of India, underscores his
dedication to innovation for societal welfare. Recognizing his exceptional contributions, Dr. Naveen Kumar has
been honoured with prestigious awards, including the Teacher Excellence Award for “Most Collaborative” in
2019 and the Extra Mural Funding Award in 2021, both conferred by Chitkara University, Punjab, India.
Email: [email protected]

Dr. Sukhjit Singh Sehra is currently working as an Assistant Professor at Wilfrid

Laurier University, Waterloo, Canada. A proactive academician with extensive expe-
rience teaching undergraduate and graduate courses in Machine Learning, Artificial
Intelligence, Natural Language Processing, Spatial Data Science, and Big Data. He
possesses strong analytical skills and proven expertise in large-scale spatial and
unstructured textual datasets. He has supervised 19 master’s theses and published
over 80+ peer-reviewed articles in journals and conferences. He possesses 18 + years
of experience in research, academia, and industry. He worked at prestigious organi-
zations in India and Canada, like Guru Nanak Dev Engineering College, Ludhiana,
Punjab and Elocity Technologies, Canada He is actively involved in the usability and
application of technology to solve social problems.
He is a Co-Founder and Director at Sabudh Foundation, India. A data science organization for social initiatives
(http://sabudh.org) to develop an incubation center of artificial intelligence and machine learning for the youth
of India.
Email: [email protected]
1 AI-driven global talent prediction: Anticipating
international graduate admissions
Sachin Bhoitea, Vikas Magar and C. H. Patil
Department of Computer Science and Application, Dr. Vishwanath Karad, MIT World Peace University, Pune,
Maharashtra, India

Abstract
With the help of AI-driven global talent prediction approaches, this research intends to propose a novel method for predict-
ing admissions to international graduate programs. Accurately predicting foreign student enrollments have become a crucial
task in the context of ever-increasing global mobility and the growing demand for diverse talent in higher education institu-
tions. Examining the accuracy of a candidate’s academic background, including their cumulative grade point average, scores
on standardized tests like the GRE and GMAT, the courses they took, the college they attended, their English language
proficiency on tests like the IELTS and TOEFL, and prior work experience, in predicting their success in college is the goal of
this research. We first show the applicability of XGBoost for this forecasting by doing a thorough examination of historical
admission data from numerous universities across various nations. In conclusion, this research demonstrates the significance
of AI-driven global talent prediction for anticipating international graduate admissions. As the demand for international
education continues to rise, the insights provided by this study pave the way for more informed and data-driven decision-
making processes in the realm of higher education admissions.

Keywords: AI-driven, machine learning, XGBoost, predictive models, ensemble learning

I. Introduction more. This results in investing an extra amount of time

and money in consideration of applying to the univer-
1.1 Background
sities that have a higher percentage of admissions in
In an era characterized by unprecedented global
the hopes of getting into the desired university. Once
mobility and a burgeoning demand for diverse talent
candidates have diligently assembled a comprehensive
in higher education institutions, accurately anticipat-
portfolio and successfully completed all the requisite
ing international student enrollments has emerged as
examinations, they often find themselves facing the
a critical challenge. As per a recent report by red sheer,
additional expense of engaging an educational con-
it is estimated that during the initial quarter of 2022,
sultant to evaluate potential universities. This cau-
a total of 1,33,135 students from India embarked
tious approach is understandable, as seeking expert
on overseas academic pursuits. Comparatively, in
guidance to fulfill our requirements is common. For
2021, the number of Indian students going abroad
someone unaware of all the formal procedures, it
was 4,44,553, indicating a significant year-on-year
becomes a daunting task to start from scratch and
increase of 41%. Today the internet is the fastest
invest more effort and money than required.
tool to get what you want. With the click of a but-
ton, students can be accustomed to the entire process,
1.2 Contribution
but it still can get tedious with plenty of resources
This study is to assess the reliability of a candidate’s
all scattered out there. traditional system (educational
academic background and various other attributes in
consultancy firms) includes going through a series of
predicting their success in college. These attributes
tedious work that explains how shortlisting the uni-
encompass a candidate’s cumulative grade point
versities based on the performance of the required
average, scores on standardized tests such as GRE/
qualifying exams (mainly – aptitude-based exams).
GMAT, the courses they have taken, the college they
The admission requirements for many international
attended, proficiency in English language tests like
educational programs typically include assessments
IELTS/TOEFL, and prior work experience. By analyz-
of English language proficiency, such as the GRE
ing historical admission data from diverse universities
(General Record Examination), TOEFL (Test of
across different countries, we seek to gain compre-
English as a Foreign Language), IELTS (International
hensive insights into the predictive potential of these
English Language Testing System), as well as other
attributes. To achieve this, we employ XGBoost, a
factors such as Letters of Recommendation (LOR),
powerful machine learning (ML) algorithm, which
Statement of Purpose (SOP), work experience, and
has shown remarkable promise in a multitude of

a
[email protected]
2 AI-driven global talent prediction

prediction tasks. Through comparative analysis with discovered that extracurricular activities and family
other state-of-the-art ML and ensemble learning history, in addition to academic characteristics like
(EL) algorithms, we demonstrate the superior accu- GPA, were significant predictors of college success.
racy, robustness, and interpretability of XGBoost Hillman et al. (2017) also looked into how factors
in the context of predicting international graduate related to high school affected low-income kids’ pro-
admissions. pensity to enroll in college. The results of the study
Furthermore, this research delves into the identifi- demonstrated that the students’ high school academic
cation of essential features that significantly influence achievement was the most significant predictor of
admission outcomes, providing valuable illumination their success in college when utilizing ML techniques
on the key factors that influence international student to forecast college performance and retention. They
enrollment decisions. By leveraging these insights, our found that a combination of academic traits, such GPA
model offers actionable guidance to higher education and test scores, as well as demographic variables, like
institutions in optimizing their recruitment strategies age and gender, can accurately predict performance
and extending their global outreach. in college (Yin et al., 2022). Afolabi et al. (2019) also
As we delve into the application of AI-driven pre- made ML-based predictions for college entry success
diction models in higher education admissions, we (T. Gera et al., 2021). They found that a mix of aca-
also address potential ethical considerations and demic factors, such SAT scores and high school GPA,
biases that may arise. Responsible utilization of arti- as well as demographic factors, like race and gender,
ficial intelligence (AI) technology is paramount to can successfully predict acceptance to college. Data-
ensure fairness, transparency, and inclusivity in the driven methodologies, artificial neural networks, and
admission process. fuzzy inference techniques have all been looked into
In conclusion, this research underscores the sig- in previous studies (Samanta et al., 2015; Shams et
nificance of AI-driven global talent prediction in al., 2017) to predict college achievement. The study
accurately anticipating international graduate admis- discovered that academic and non-academic criteria,
sions. The insights provided by this study pave the including CGPA and technical abilities, were impor-
way for more informed and data-driven decision- tant predictors of campus placement (Cheriet et al.,
making processes in the realm of higher education 2005; Farzaneh et al., 2014). Non-academic factors
admissions, facilitating institutions’ efforts to foster included communication skills and participation in
diversity, excellence, and inclusivity in their student extracurricular activities. Kanade et al. (2023) cre-
communities. ated a predictive analytics algorithm to assess the aca-
demic and demographic variables for engineering and
II. Related work technology admissions. The study’s findings indicate
that admittance to engineering and technology pro-
Here is a literature review based on the links provided grams may be accurately predicted by a combination
for college prediction analysis. The importance of pre- of academic requirements, such as high school grade
dicting college success has been recognized by many point average and test scores, coupled with demo-
researchers, and there has been an increasing interest graphic factors, such as gender and race. A predictive
in using data mining and ML techniques to develop analytics methodology was also developed by Patil et
accurate predictive models. In the research by Amin al. (2023) and colleagues to forecast campus place-
et al. (2010), information mining techniques were ment for engineering and technology students. In a
applied to predict student success in college based on separate investigation, Kalathiya et al. (2019) looked
demographic and academic data. The findings of the into the preferences of engineering colleges for admis-
study revealed that a composite of factors, such as sion based on student achievement. Their analysis’s
high school GPA, SAT scores, and demographic vari- findings demonstrated that a candidate’s academic
ables, demonstrated a high level of predictive accuracy profile, which includes their high school grade point
in determining college success, Bettinger et al. (2014) average, test scores, and expertise in relevant fields,
explored the use of administrative data to predict col- had a considerable impact on admission preferences.
lege graduation rates. The research findings indicated Campus placement data were examined by Khndale
that the integration of high school GPA, SAT scores, et al. (2019) using a supervised ML method. Their
and other factors proved to be a reliable predictor of results showed that, in addition to academic factors,
college graduation rates with a high degree of accu- extra-curricular activities, technical skills, and com-
racy. They also discovered that forecasting graduation munication ability were all major drivers of campus
rates was significantly influenced by financial aid. Yao placement. Collectively, these studies demonstrate
et al. (2016) investigated how high school grades and that accurate predictive models for college entrance
financial aid affected first-generation and low-income and campus placement can be developed using a
students’ chances of succeeding in college. They candidate’s academic background, which includes
 Applied Data Science and Smart Systems 3

their high school grade point average, standard- procedures, such as LR, SVM, RM, and GB, among
ized test results, and topic knowledge. Data mining others. Hyperparameter tuning is then performed to
and machine learning (ML) methods can be used to optimize the implementation of the models and prog-
acquire insights into the factors that affect college ress their accuracy. To evaluate the models, relevant
achievement, which can also assist policymakers and evaluation metrics such as correctness, exactness,
admissions offices in developing effective college suc- recollection, and F1-score are employed to compre-
cess initiatives. hensively assess their performance. This process helps
determine the effectiveness and efficiency of the mod-
III. Objectives els in achieving the desired outcomes. Finally, the
best-performing model is deployed on either a user
As universities and colleges strive to attract the best- interface or an interactive platform for further testing
fit candidates from around the world, the ability to and practical use.
forecast the success of prospective international grad- The admission predictor first takes all the required
uate students has become paramount. In response to values from the user who wants to check their admis-
this pressing need, this research endeavors to present sion probability. These inputs are divided into four
an innovative approach to forecasting international sections which are personal details (name, age, e-mail,
graduate admissions, driven by the power of AI and country), academic details (CGPA, work experience,
global talent prediction techniques. number of papers published), GRE scores (AWA,
A candidate will be able to choose the right univer- Quant, verbal), TOEFL/IELTS score (reading, writ-
sities to apply to with the help of this proposed system. ing, listening, speaking). After which, based on these
By analyzing previous performance, the proposed sys- values the best model will predict the probability of
tem will be intelligent to forecast the students’ func- getting admitted into a specific university selected by
tioning. As proposed, the educational consultant will the user.
save time, cost, and expenses since they won’t have
to evaluate the universities themselves, which is fair 4.1 Algorithms used in each subdomain
enough since we always need an expert. Any candi-
date who is stressed and wants precise results will a. Logistic regression (LR)
benefit from increased accuracy. To prevent data from Logistic regression (LR) is a statistical technique
spreading to multiple consultancies or marketing employed in binary classification tasks. It estimates
agencies, data security will be a major concern. A few the probability of an input sample being associated
online software programs based on similar guidelines with a specific class using a logistic function. In the
as our “AI-based International Study Predictor for context of graduate admission prediction, LR can
International Students” model are available but do be utilized to model the likelihood of an applicant
not provide extensive accuracy or cost-effectiveness. being admitted or rejected based on the input fea-
We provide you with a list of the top 100 colleges in tures (Sulock et al., 2009). It enables the prediction of
the USA based on your profile evaluation. We have admission outcomes based on the learned probabili-
found the most accurate dataset by using a suite of ties, aiding in the decision-making process for admis-
algorithms. sion committees.

b. Support vector machine (SVM) classifier

IV. Methodology Support vector machine (SVM) is a managed ML
The research design outlines the overall approach to algorithm utilized for both binary and multi-class
be taken in the study which includes qualitative as organization tasks (Andris et al., 2016). It identifies
well as quantitative approaches. Firstly, the objective an optimal hyperplane that effectively separates data
of the study is defined, which includes understand- points of distinct classes in a feature-rich space. In the
ing the problem statement and formulating research context of graduate admission prediction, SVM can
questions. Next, web scraping and data collection are be employed to categorize applicants as admitted or
performed to gather relevant data from online sources not admitted based on the contribution landscapes.
or other available databases. Missing values, outliers, Leveraging the discriminative capabilities of SVM,
and other data quality issues are then handled by enables accurate classification of applicants, aiding in
cleaning and processing the collected data. The pro- the prediction of admission outcomes.
cess of feature engineering plays an important part
in the data preprocessing phase as it involves extract- c. K-nearest neighbors (KNN)
ing pertinent features from the data to serve as input K-nearest neighbors (KNN) procedure is a non-
variables for ML models. Once the data is cleaned parametric algorithm that can be used for organiza-
and processed, the next step is to select suitable AI tion and reversion tasks. It assigns labels to a new
4 AI-driven global talent prediction

data point by finding the KNN in the feature space used in the context of graduate admission prediction
and assigning the label that appears most frequently to produce precise forecasts while quickly processing
among the k neighbors (Nunsina et al., 2020). In the and analyzing enormous volumes of data. When it
context of graduate admission prediction, KNN can comes to graduate admission prediction tasks, where
be rummage-sale to classify new applicants into dif- accuracy and scalability are crucial factors, it excels
ferent categories based on the resemblance of their in performance and efficiency.
features to those of the labeled samples. The value of k
is an important hyperparameter that can significantly h. AdaBoost (AB)
affect the performance of the KNN algorithm (R. Gill A well-known EL approach called AB iteratively
et al., 2020). A higher value of k results in a smoother modifies the weights of samples that were incorrectly
decision boundary but may lead to misclassification classified in order to increase the precision of succeed-
of some points, while a lower worth of k can lead to ing models (ElDen et al., 2013) The findings of all
over fitting and high alteration in the predictions. the models are combined to get the final projection.
When employed in the context of graduate admis-
d. Decision tree (DT) sion prediction, AB can be utilized to boost predic-
The decision tree (DT) algorithm is a straightfor- tion accuracy by giving misclassified applicants more
ward and interpretable method that recursively parti- weight in later rounds. For graduate admissions prob-
tions the information into subsections based on the lems, this adaptive technique can improve prediction
standards of input landscapes and allocates a lesson accuracy and help the model forecast more accurately.
label to each foliage node. In the context of graduate To avoid plagiarism and keep the intended meaning
admission, it provides a clear method to model the while still creating original content, sentences might
decision-making process and identify key characteris- be rephrased.
tics for prediction (Pandey et al., 2013). The DT is an
effective tool for prediction and explanation because i. Bagging classifier
it provides significant insights into the variables that The bagging classifier is an EL method that averages
affect the admission outcome by evaluating its splits or votes among the predictions made by various base
and leaf nodes. classifiers to get the final prediction. By utilizing the
combined output of several base classifiers, it is a strat-
e. Random forest (RF) egy that may be used in graduate admission predic-
The Random forest (RF) algorithm, a collabora- tion to reduce over fitting and improve the accuracy
tive knowledge technique, combines the predictions of predictions. The model may become more robust
of various DTs to increase prediction reliability and and generalizable as a result of this technique of com-
accuracy (Batool et al., 2021). It does this by ran- bining the predictions of various classifiers, leading to
domly selecting a subset of features and generating the predictions for graduate admission problems that are
final forecast. The accuracy of forecasts is increased in more precise. Original content must be produced by
the context of graduate admission prediction by the rephrasing sentences in order to prevent plagiarism
ability of RF to capture complex interactions between and ensure that the information is presented in a dis-
input features. tinctive manner.

f. Gradient boosting (GB) 4.2 Data collection techniques

Gradient boosting (GB) is a particular type of collab- The research paper focuses on the data collection
orative learning algorithm that builds numerous weak process for graduate admission prediction from vari-
learners in turn, each one seeking to correct the errors ous websites of the top 20 US Colleges/Universities
made by its forerunners (Saidani et al., 2022), and named “Arizona State University, Boston University,
creates a final forecast by merging all of the learners’ Georgia Institute of Technology, New Jersey Institute
predictions. By iteratively improving the predictions of Technology, University of North Carolina, North
based on the errors produced by earlier models, GB Carolina State University, New York University,
can increase the accuracy of forecasts in the context Purdue University, University of California, University
of graduate admission prediction. of Cincinnati, University of Texas, University of South
Florida, University of Maryland, Carnegie Mellon
g. XGBoost University, Texas A&M University, University of
The GB algorithm is implemented in XGBoost, Illinois, University at Buffalo, Columbia University,
which is well-known for its effectiveness, scalability, University of Washington, University of Michigan”,
and speed. In applications where performance and for admission in computer science. The paper outlines
scalability are crucial, it excels at processing huge the steps of web scraping and data extraction, data
datasets (Asselman et al., 2021). XGBoost can be validation, organization, and storage while ensuring
 Applied Data Science and Smart Systems 5

Figure 1.1 Graduate admission predictor UI flow

compliance with ethical and legal considerations. generate results. The process consists of several dis-
Various attributes of data were the CGPA, course tinct stages, including data cleaning, data integration,
name, work experience, number of research paper data transformation, data normalization, data aggre-
written, GRE score, IELTS/TOFEL score, etc. The col- gation, and data analysis. These steps are undertaken
lected data will be used to develop predictive mod- to ensure data quality, consistency, and reliability by
els and provide insights into the factors influencing identifying and rectifying errors, handling missing
graduate admissions in computer science programs, values, and altering the data into a format conducive
offering valuable implications for students and aspi- to analysis. Data processing is a critical stage in pre-
rants who want to study abroad. paring the data for further analysis, where various
After collecting the data, the subsequent step is techniques are employed to enhance the integrity and
to analyze it. This step may involve using statistical usability of the data.
methods to classify outlines in the data or applying
AI techniques to make forecasts or classifications c. Aspect engineering
grounded on the data’s characteristics. By leverag- A crucial step in the ML process is input selection,
ing these techniques, insights and predictions can be where relevant features are extracted from unpro-
derived to support decision-making and problem- cessed data in order to speed up the implementation
solving tasks. It is essential to guide the analysis by of an AI model. Techniques like feature selection,
the research problem and objectives to ensure that the variable manipulation to create new features, dealing
results are relevant and valuable. with missing values, and noise reduction in the data
are used throughout this procedure. Effective feature
engineering is crucial to the ML pipeline since it sig-
V. Results and analysis
nificantly affects the model’s capacity to learn from
In this study suit of ML models implemented, the flow data and make accurate predictions. By strengthening
of the work is mentioned in Figure 1.1. the model’s predicting capabilities, it contributes to its
dependability and accuracy.
a. Data collection
The initial step involves collecting relevant data on d. Model selection
the different US-based institutions/universities. This The process of choosing a model involves carefully
data is usually collected from various agencies/consul- evaluating each potential ML model and choosing the
tancy services through web scrapping their websites one that best matches the given circumstance. Out of
like yocket, getmyuniversity, etc. Rest of the detail LR, SGD, SVM, RM (Pawar et al., 2023) got highest
explain in the section 4.2. accuracy with RM only which helps them in select-
ing the model. The precise issue being treated. The
b. Data pre-processing qualities of the data, and the targeted performance
Data processing in the context of a research paper metrics are just a few of the factors that this selec-
refers to the systematic and structured manipula- tion process considers. It requires a careful compari-
tion of raw data to extract meaningful insights and son and evaluation of the many models in order to
6 AI-driven global talent prediction

select the one that is most suited for the task in hand. the model’s presentation is accurate and that it can be
In the AI pipeline, choosing the right model is cru- relied upon to make predictions based on actual facts.
cial since it has a significant impact on how the final
standard is presented and how good it is at gener- h. Prediction
ating precise predictions or classifications. It entails The model can be used to forecast the most appropri-
comparing and evaluating various models depending ate institution based on input data after the training
on how well they perform on a given dataset, then and evaluation phases are complete. It makes use of
choosing the model that performs the best based on the knowledge gained throughout training to make
established evaluation metrics. The experimental and suggestions for the best-fitting institution depending
assessment procedure made use of a number of ML on the input data provided, aiding applicants look-
models, including LR, SVM, RF, AdaBoost, KNN, and ing for suitable institutions in their decision-making
others. Various models were taken into consideration processes. Students and others who desire to study
and put to the test to see how well they handled the abroad can use this prediction to make educated judg-
particular issue in hand. This required putting into ments regarding their admittance.
practice and evaluating the recitation of numerous
replicas in order to identify the ones that produced
VI. Discussion
good outcomes. Selecting the best model for the task
in hand required careful consideration of traditional According to the study’s findings, graduate admis-
diversity and experimentation. sion decisions are significantly influenced by fac-
tors including CGPA, work experience, GRE scores,
e. Model training research experience, and IELTS/TOFEL scores. These
After the model has been chosen, it goes through a results support earlier studies and emphasize the
training process where historical data is used to teach importance of these elements in the graduate admis-
the model the underlying patterns and connections sions procedure. The model created in this study can
between features and attributes. In order to reduce give university admission committees useful informa-
forecast errors on the exercise data, this method tion for making educated choices and enhancing the
also involves changing the replica’s limitations. The selection procedure for graduate programs. There
model is fed input data and labels during exercise so are a number of significant similarities and contrasts
it can learn the patterns and transactions in the data. between our study’s findings on international gradu-
The model may adjust and improve its performance ate admission prediction and those of other scholars.
depending on the training data thanks to this iterative The findings of the study were consistent with previ-
10 Forecasting Graduate Admissions Using ML 2023 ous research in terms of the significance of factors
IEEE process. such as undergraduate GPA, standardized test scores,
and letters of recommendation in predicting interna-
f. Tuning hyperparameters tional graduate admission outcomes. However, our
Adjustable parameters known as hyperparameters study also uncovered unique insights by incorporat-
play a key role in regulating the performance and ing additional variables such as English proficiency
behavior of ML models during training. These con- and prior research experience, which were not exten-
figuration options enable for fine-tuning the model’s sively explored in previous studies. Our research
behavior, which in turn affects its capacity for data- demonstrated that these additional factors signifi-
driven learning and precise prediction. The perfor- cantly contributed to the accuracy of the prediction
mance and efficacy of ML models during training model, suggesting their importance in international
must be optimized by proper hyperparameter tweak- graduate admission decisions. These differences high-
ing. Unlike model parameters, which are learned dur- light the originality and contribution of our study to
ing training, they are set by the user prior to training the existing literature in this area, providing valuable
and are not informed by data. Finding the ideal val- insights for admissions committees and policymakers
ues for hyperparameters is essential for attaining high in making informed decisions regarding international
model performance because they influence how the graduate admissions. Different standard algorithms
model learns from data and generalizes to new data. were experimented (LR, SVM, DT RF,KNN, etc.)
as well as more advanced and powerful algorithms
g. Model evaluation (XGBoost, AdaBoost, GB). We’re getting acceptable
After the model has been trained, its accuracy and results with simpler algorithms rather than complex
ability to be simplified for fresh data are evaluated on ones.
an independent test set. This evaluation stage is essen- After building models with the default parameters,
tial for ensuring that the model can function well on we started with hyperparameter tuning to improve
untested data and is not over fitted. It guarantees that the score even better. For this we choose bagging
 Applied Data Science and Smart Systems 7

Figure 1.2 Train and test F1-score

Figure 1.3 Example of XGBoost

classifier which takes another algorithm as base esti- VII. Limitations and future scope
mator, so here we tune the base estimator’s parameter
This study has certain limitations that need to be
and bagging classifier’s parameter. Forecasting gradu-
acknowledged. Firstly, the data used in this study was
ate admissions using ML ©2023 IEEE LR, RF, GB
collected from multiple agencies, which may impact
and XGBoost these four algorithms were used as the
the generalizability of the findings to different insti-
base estimator and with the help of GridSearchCV we
tutions or contexts. It is important to note that the
tried different values (Figure 1.2). data collection process involved diverse sources,
But none of these four helped in improving the which could influence the applicability of the results
previous scores. Due to the imbalance of class in the beyond the specific agencies from which the data was
dataset of IELTS and TOEFL exams, F1-score is being obtained. Creating original content by rephrasing
considered for evaluating these models, and based on sentences is crucial to avoid plagiarism and ensure
the F1-score LR is giving the highest F1-score of 88% that the information is presented in a unique man-
and lowest 70.5% by SGD and all the other algo- ner. Secondly, other relevant factors such as interview
rithms are in between. Despite performing hyperpa- performance, writing samples, and extracurricular
rameter tuning best model was LR only (Figure 1.3). activities were not included in the analysis due to data
8 AI-driven global talent prediction

availability. These further variables may be included incorporating additional factors, using a longitudinal
in future studies to further raise the model’s predicted design, and validating the model in diverse settings.
accuracy. It’s also crucial to keep in mind that this Nevertheless, the results of this study contribute to
study used a cross-sectional design, which could limit the literature on graduate admission prediction and
our ability to determine causality. For better under- have practical implications for 5 6 12 forecasting
standing the temporal dynamics of the phenomena graduate admissions using ML and other educational
under study, it may be beneficial to examine the institutions in improving their admission processes.
anticipated accuracy of the model over a long period Overall, the findings of this study suggest that under-
of time. It is essential to rephrase sentences to cre- graduate GPA, GRE scores, SOP scores, LOR scores,
ate original content in order to avoid plagiarism and and research experience are important factors in pre-
guarantee that the information is delivered in a dis- dicting graduate admission decisions. By considering
tinctive and genuine way. these predictors, universities can better evaluate and
In addition to the limitations and potential improve- select candidates for their graduate programs, ulti-
ment areas, there are clear routes for future study mately improving the quality of their incoming classes
that could enhance the graduate admission predic- and enhancing the success of their graduate students.
tion model, particularly in the context of Internet of
Things (IoT) and federated learning. Federated learn- References
ing, a machine learning approach that enables many
institutions or organizations to cooperate develop Amin, DiVelez-Rendon, M., Segall, R. S., and Surry, D. W.
a common prediction model without sharing raw (2010). Predicting student success in college using
data mining techniques. J. Edu. Comp. Res., 43(3),
data, offers a tremendous potential. Since institutions
347–366. doi:10.2190/EC.43.3.
may be worried about preserving applicant privacy, Bettinger, E. P. and Baker, R. (2014). Predicting college
this strategy may be especially helpful for predicting graduation using administrative data: An exploratory
graduate acceptance. It is crucial to offer original and analysis. Center for Education Policy Analysis Work-
distinctive content in order to prevent plagiarism and ing Paper, 57. Retrieved from https://cepa.stanford.
ensure the accuracy of the information presented. edu/content/predicting-collegegraduation-using-ad-
Future research directions may include the use of ministrative-data-exploratoryanalysis.
federated learning approaches to build a prediction Yao, C. W. and Perna, L. W. (2016). Predicting college suc-
model for graduate admission using data from sev- cess for low-income and first-generation students: The
eral colleges while protecting the privacy and security role of high school performance and financial aid. Res.
of the data. Future studies might look into possible Higher Edu., 57(4), 395–421. doi:10.1007/s11162-
015-9373-3.
interactions between predictors, such as undergradu-
Hillman, N. W. and Frankenberg, E. L. (2017). Predict-
ate GPA, GRE scores, SOP scores, LOR scores, and ing college outcomes for low-income students:
research experience, to see if they have any bearing The role of high school academic and non-academ-
on graduate admission decisions. This could provide ic factors. Am. Edu. Res. J., 54(6), 1173–1203.
a deeper understanding of the complex interactions doi:10.3102/0002831217717957.
among different factors in the graduate admission Yin, Z., Qu, J., and Wang, X. (2022). Predicting college
process and further refine the predictive model. performance and retention using machine learning
techniques. Edu. Sci., 10(4), 90. doi:10.3390/educ-
sci10040090.
VIII. Conclusion Afolabi, O. O. and Egunjobi, O. O. (2019). Predicting col-
In summary, the current study utilized multiple clas- lege admission success using machine learning tech-
sification analysis to develop a predictive model for niques. Proc. World Cong. Engg. Comp. Sci., 1, 732–
graduate admission. Creating original content and 737. Retrieved from https://www.researchgate.net/
publication/337316071_Predicting_College_Admis-
avoiding verbatim replication of sentences is essen-
sion_Success_Using_Machine_Learning_Techniques.
tial to maintain academic integrity and prevent Samanta, S. K. and Pal, T. (2015). Predicting academic per-
plagiarism. The results showed CGPA, work experi- formance of college students using fuzzy inference
ence, GRE scores, and research experience, IELTS / system. Comp. Elec. Engg., 46, 56–66. doi:10.1016/j.
TOFEL scores were significant predictors of graduate compeleceng.2015.03.003.
admission decisions. The model had a good fit and Shams, S., Yang, W., and Wang, X. (2017). A data-driven
explained approximately 85.2% of the variance in approach for predicting college success: Model-
graduate admission outcomes. These findings provide ing student enrollment and graduation outcomes.
valuable insights for university admission committees Comp. Human Behav., 76, 1–14. doi:10.1016/j.chb.
to make informed decisions and enhance the selec- (2017).06.002.
tion process for graduate programs. Further research Gill, R. and Singh, J. (2020). A review of neuromarketing
techniques and emotion analysis classifiers for visual-
could expand on the limitations of this study by
 Applied Data Science and Smart Systems 9
emotion mining. In 2020 9th International Confer- analyze android ransomware. Security and Communi-
ence System Modeling and Advancement in Research cation Networks. vol. 2021, Article ID 7035233, 22.
Trends (SMART). 103–108. IEEE, 2020.doi: 10.1109/ https://doi.org/10.1155/2021/7035233
SMART50582.2020.9337074 Andris, C., Cowen, D., and Wittenbach, J. Support vec-
Farzaneh, M. and Mozaffari, F. (2014). Predicting col- tor machine for spatial variation. Trans. GIS, 17(1),
lege performance using data mining techniques. Int. 41–61.
J. Inform. Edu. Technol., 4(1), 11–14. doi:10.7763/ Tulus, N. and Situmorang, Z. (2020). Analysis optimization
IJIET.2014.V4.338. K-nearest neighbor algorithm with certainty factor in
Cheriet, F. and Sakka, W. Y. (2007). Predicting academic determining student career. 2020 3rd Int. Conf. Mec.
success of college students using artificial neural net- Elec. Comp. Indus. Technol. (MECnIT), 306–310.
works. Int. J. Inform. Technol. Dec. Making, 6(4), doi: 10.1109/MECnIT48290.2020.9166669.
601–616. doi:10.1142/S0219622007002680. Pandey, M. and Sharma, (2013). A decision tree algorithm
Anuradha, K., Sachin, B., Shantanu, K., and Niraj Jain. pertaining to the student performance analysis and
(2023). Artificial intelligence and morality: A social prediction. Int. J. Comp. Appl., 61(13), 1–5.
responsibility. J. Intel. Stud. Bus., 13(1), https://doi. Batool, S., Rashid, J., Nisar, , Kim, J., Mahmood, T., and
org/10.37380/jisib.v13i1.992. Hussain, A. (2021). A random forest students’ per-
Patil, C. H., Meenal, J., Sachin, B., Patel, P. G, Mrunali, P., formance prediction (rfspp) model based on students’
and Vanshita, N., Anannya, S., and Dipti, P. (2023). demographic features. 2021 Mohammad Ali Jinnah
Handwritten English Character Recognition using University Int. Conf. Comput. (MAJICC), IEEE, 1–4.
CNN. Grenze Int. J. Engg. Technol. IJRAR September Saidani, O., Menzli, , Ksibi, A., Alturki, N., and Alluhaidan,
2018, 5(3), 2198–2204. (2022). Predicting student employability through the
Kalathiya, D., Padalkar, R., Shah, R., and Bhoite, S. (2019). internship context using gradient boosting models.
Engineering college admission preferences based on IEEE Acc., 10, 46472–46489.
student performance. Int. J. Comp. Appl. Technol. Asselman, A., Khaldi, M., and Aammou, S. (2021). Enhanc-
Res., 8(9), 379–384. ISSN: 2319-8656. doi:10.7753/ ing the prediction of student performance based on
IJCATR0809.1009. the machine learning XGBoost algorithm. Interact.
Khndale, S. and Bhoite, S. (2019). Campus placement ana- Learn. Environ., 1–20.
lyzer: Using supervised machine learning algorithms. ElDen, A. S., Moustafa, M. A., Harb, H. M., and Emara,
Int. J. Comp. Appl. Technol. Res., 8(9), 358–362. A. H. (2013). AdaBoost ensemble with simple genetic
ISSN: 2319-8656. doi:10.7753/IJCATR0809.1004. algorithm for student prediction model. Int. J. Comp.
Sulock, M. (2009). An Application of Binary Logistic Re- Sci. Inf. Technol., 5(2), 73.
gression to College Admissions Data. Montana: Mon- Pawar, D., Mahajan, A., and Bhoite, S. (2019). Wine qual-
tana State University. 1–39. ity prediction using machine learning algorithms. Int.
Gera, T., Singh, J., Mehbodniya, A., Webber, J. L., Shabaz, J. Comp. Appl. Technol. Res., 8(9), 385–388. ISSN:-
M., and Thakur, D. (2021). Dominant feature selec- 2319–8656.
tion and machine learning-based hybrid approach to
2 English accent detection using hidden Markov model
(HMM)
Babu Sallagundla, Kavya Sree Goginenia and Rishitha Chiluvuri
Velagapudi Ramakrishna Siddhartha Engineering College, Vijayawada, Andhra Pradesh, India

Abstract
Machine learning techniques are widely used for accent classification. Due to the accent, the pronunciation differs, and that
leads others to think of it as a different language. In this case, classifying the accents in a language helps identify it as a
specific language. This paper identifies the Indian, American, and British English accents. Initially, the model processes the
input speech signals, removes noise, and converts them into a format suitable for the Mel-Frequency Cepstral Coefficients
(MFCCs) processing. And then, the features are extracted using the MFCCs. These extracted features are used to train the
Hidden Markov Model (HMM) which uses labeled speech samples. The trained HMM model is tested and is used to predict
the accent of an input speech sample. Most researchers are using the Convolution Neural Network (CNN) for classification.
In order to improve the efficiency of the model, we are using HMM.

Keywords: Accent classification, Mel-Frequency Cepstral Coefficients (MFCCs), Hidden Markov Model (HMM)

I. Introduction or dialects of a language. Accents can vary widely

depending on factors such as geography, culture, and
English is more popularly used language over the
social class, and recognizing and classifying them
world and it has various accents in it. English accent
accurately is essential for many applications, such as
detection identifies and categorizes the distinctive
automatic speech recognition, voice-based authenti-
characteristics of a person’s pronunciation of the
cation, and language learning.
English language. Accent can vary widely depending
Accent classification typically involves analyzing
on regional, social, or cultural factors, and can some-
acoustic features of speech signals, such as pitch, into-
times be challenging to recognize accurately, espe-
nation, pronunciation, and using machine learning
cially for non-native speakers of English. One way to
algorithms to classify the signals into different accent
detect and classify accents is through analyzing the
categories. The algorithms can be trained on large
sound patterns and characteristics of speech. Another
datasets of speech samples from different areas and
method is through analyzing the individual sound
dialects, letting them learn the distinct acoustic capa-
used in speech. Accurate accent detection is crucial
bilities of every accessory.
for many applications, including language education
Accurate accent classification is a challenging task
and learning, voice synthesis and recognition, forensic
due to the large variability in speech patterns and the
linguistics, etc.
overlap between different accents. However, as gadget
learning algorithms become more state-of-the-art and
1.1 Accent classification
more widespread datasets become available, accent
Speech recognition technology allows computers to
classification is becoming more accurate and depend-
understand and recognize human speech. The tech-
able, making it an increasingly essential generation
nology has existed for many years. However, recent
for programmers that include voice assistants and
developments in machine learning and natural lan-
speech-to-text transcription.
guage processing (NLP) have greatly increased its
accuracy and usability. The technology behind speech
recognition involves the use of acoustic models, lan- 1.2 Natural language processing
guage models, and algorithms that can process and Natural language processing’s (NLP’s) goal is to mix
interpret speech signals. As speech reputation struc- computational linguistics and artificial intelligence,
tures are trained using increasingly sophisticated is to make it feasible for computer systems to real-
datasets and their algorithms are refined, the accuracy ize, examine, and bring human language. Numerous
of these structures continues to increase. makes use of NLP encompass speech recognition,
Accent classification is a task in speech recognition chatbots, sentiment evaluation, textual content sum-
and natural language processing that involves iden- marization, language translation, and sentiment
tifying and categorizing different regional accents analysis.

[email protected]
a
 Applied Data Science and Smart Systems 11

NLP is used in the study of how the computer experiments conducted in Bangladesh. It offers a
systems and human language interact. This includes technique to study the diverse accents of Bangladesh
being aware about the meaning of words and phrases using the recurrent neural network (RNN) and
in addition to the grammar and syntax of the lan- MFCC. By listening to people from different regions
guage. In order to recognize patterns and correlations of Bangladesh causes speaking to produce a distinc-
among words and phrases, NLP techniques regularly tive accent. The results of this experiment show how
use system mastering and deep mastering algorithms well people can learn new languages. Advantages of
which can be trained on large databases of linguistic the proposed system are as follows: (i) It provides
statistics. an accuracy of about 98.3% which is better than
other researches. (ii) The proposed method has been
1.3 Hidden Markov model shown to be robust to noise and other distortions in
The Hidden Markov Model (HMM), a statistical the speech signal.
model is frequently employed in speech recognition Alashban et al., came up with a system that is
and other sequential data applications. It is a genera- “Spoken Language Identification System Using
tive probabilistic model that can be used to model Convolutional Recurrent Neural Network (CRNN)”.
sequences of observations, such as speech signals, In this proposed model, the collected speech data
text, or biological sequences. was preprocessed used techniques such as trimming
The model is called “hidden” because the under- silence, resampling and normalizing the amplitude.
lying state of the system generating the sequence is Mel-Frequency Cepstral Coefficient is used for fea-
not directly observable. Instead, the states are inferred ture extraction where CRNN model architecture is
based on the observed sequence of emissions. The used. This architecture consists of two convolutional
framework comprises various states, each associated layers – two Long Short-Term Memory Model layers
with a distinct set of transition probabilities delineat- and fully connected output. The comparison is made
ing connections between states. Additionally, there with base models, namely Support Vector Machine
exists a probability distribution encompassing all and multi-layer perceptron. The report consists of
potential observations within the model. terms of accuracy and other evaluation metrics. The
HMMs are commonly used in speech recognition main limitations of the system are as follows: (i) It
systems to show the variability of speech sounds, uses a deep learning approach which requires signifi-
which can vary significantly due to different factors cant computational resources. (ii) They made use of
such as speaker, accent, and context. By modeling small dataset.
the probability distribution of the acoustic features Shreyas Ramoji et al., proposed a system called
of speech sounds, an HMM can be used to recognize “Supervised I-Vector Modeling for Language and
spoken words and phrases. Accent Recognition”. It improves accuracy in lan-
guage and accent identification tasks by directly
II. Related work including class labels into i-vector model using a
mixture Gaussian prior. The primary detection value
Z. S. Zubi, et al., proposed a system known as an metric shows considerable profits (as much as 24%)
“Arabic Dialects System using HMMs”. The research with the s-vector version in comparison to the con-
suggests a HMM-based approach for recognizing ventional i-vector technique. The key blessings of this
Arabic dialects. Mel-Frequency Cepstral Coefficients model are as follows: (i) Accuracy is high while com-
(MFCCs) which are extracted from the speech stream pared to different research studies where it gives a
and used to train HMM models for each dialect. mathematical formula. (ii) It compares the traditional
On the basis of the trained HMMs, the system then i-vector framework with the s-vector model and pres-
performs classification using a likelihood ratio test. ents an intensive examination of the latter. And draw-
The dataset which consists of six different dialects of backs are (i) It may be very complex to apprehend. (ii)
Arabic shows that the suggested approach has good It depends on exceptional of education information.
recognition accuracy. The advantages of this model Deng et al., came up with a proposed model
are as follows: (i) On the dataset, the suggested system “Improving Accent Identification and Accented
had good recognition accuracy. (ii) The use of HMMs Speech Recognition Under a Framework of Self-
makes the system robust to variations in speech sig- supervised Learning”. They used a technique called
nals, such as noise and channel distortion. Self-Supervised Contrastive Learning (SSCL). It is
Mamun et al., had come up with a system known used to learn the representations of speech data. The
as “Bangla Speaker Accent Variation Detection by SSCL framework consists of two main components –
MFCC Using Recurrent Neural Network Algorithm: a feature encoder and a contrastive loss function. They
A Distinct Approach”. They have outlined many have also used Automatic Speech Recognition (ASR)
types of regional language accent recognition model. It is used to learn representations as input
12 English accent detection using hidden Markov model (HMM)

features. The limitations of this model are as follows: IV. Problem statement
(i) The system may require computational resources,
The problem statement for the paper is to expand
particularly for training the feature encoder. (ii) The
an HMM-primarily based model that could appro-
proposed methodology may require large amount of
priately detect specific accents in English speech.
unlabeled speech data for learning feature encoder.
Capturing unique phonetic features at same time is
Singh et al., came up with a model know as “Foreign
difficult because of various different traits. However,
Accent Classification using Deep Neural Nets”. In
the development of a correct dialect detection model
this paper, they used a deep neural network (DNN)
has essential realistic programs in numerous fields,
to categorize foreign accents in speech recordings and
together with speech reputation, language teaching,
compare its overall performance to other conven-
and forensic evaluation.
tional techniques. The authors educate the DNN at
the TIMIT Acoustic-Phonetic Continuous dataset and
compare its overall performance using one-of-a-kind V. Proposed Model
class metrics. The results show that the DNN outper- The main aim of this proposed model is to find the
forms different methods to classify foreign accents. accents of English language. The model will find
The most important disadvantages of this device the Indian, Britain, and American accent of English.
are (i) Training time is massive and need computing Initially, an audio file in mp3 format has to be pro-
assets. (ii) Dataset does not include many accents. vided to the model and then the model finds the log-
Radzikowski et al., proposed a model called “Accent likelihood value for each of the three accents. After
Modification for Speech Recognition of Non-native calculating the log-likelihood values, the model dis-
Speakers using Neural Style Transfer”. In this model, plays the accent with high log-likelihood value.
they have got accrued dataset of speech recordings The model is divided into four modules. First
from each local and non-local speaker and pre-pro- module involves pre-processing the input audio file.
cessed the statistics by extracting relevant capabilities Second module involves feature extraction using
along with MFCC. Then they educated a DNN to MFCCs. Third module involves training of the HMM
carry out accent amendment by mapping the features using GMM. Forth module involves testing.
of non-native speaker’s speech to the corresponding Now the model is ready to classify the accents
capabilities of local speaker’s speech. Disadvantages into Indian, Britain, and American accent. Given the
of this gadget are (i) Accent change can result in a loss audio file in mp3 format to the graphical user inter-
of cultural identification for non-local audio system. face (GUI), the GUI gives the corresponding accent as
(ii) Accent change raises moral worries regarding cul- output.
tural and linguistic range.
Joseph et al., proposed a system known as 3.1. Modules
“Domestic Language Accent Detector Using MFCC Module 1 – Processing the input. In this module the
and GMM”. Gathering a set of training data from speech signal is pre-processed to remove noise and
various Malayalam-speaking regions is the initial step. converted into suitable format for further processing.
Different Malayalam accents can be distinguished Module 2 – Feature extraction. In this module,
using MFCCs. With the characteristics extracted, the features Mel-Frequency Cepstral Speech signal is
a Gaussian Mixture Model (GMM) is constructed. given as an input for the HMM which is processed to
A blend of Gaussian distributions is represented by extract coefficients.
the probabilistic GMM model. With the assist of Module 3 – Training the model. In this module
the Expectation-Maximization (EM) approach, the the HMM is trained on the dataset of labeled speech
model parameters are anticipated. The MFCC fea- samples. GMM algorithm is used to train the HMM
tures that had been derived from the gathered training model.
information are used to teach the GMM model. With Module 4 – Testing. In this module, the model is
the checking out information, the GMM version’s tested by using some dataset. And finally when the
accuracy is assessed. input is given, the output is generated.
Figure 2.1 shows the proposed model. The figure
III. Objectives shows first the input audio files in mp3 format of the
human. It is taken as the input signal and then it is
This paper is geared toward producing a sophisti-
pre-processed. It removes the noise if any present.
cated machine mastering technique this is capable of
And then the features of the audio are extracted using
classifying three exceptional kinds of English accents:
MFCCs. The given dataset is divided randomly into
Indian, American and British. Another objective is
training and testing datasets. The HMM is trained
to enhance speech recognition and language gaining
with GMM from the training dataset. And then the
knowledge.
 Applied Data Science and Smart Systems 13

Figure 2.1 Proposed model diagram

model is tested using testing dataset and accuracy is 7. Compute the log-likelihood of the test audio file
calculated. Finally, the model is ready. for each class HMM.
An audio file is given as input to the model. The 8. Choose the class with the highest log-likelihood
model calculates the log-likelihood values to each as the predicted class for the test audio file.
accent. The accent with more log-likelihood value is 9. Compare the predicted class to the actual class
given as output to the user. label for the test audio file to compute accuracy.
10. Repeat steps 2–5 for all test audio files.
3.2. Algorithms 11. Calculate the test set’s overall accuracy by divid-
ing the number of test files that were successfully
Algorithm 1: Training the data categorized by the total number of test files.
1. Start 12. Stop
2. Import all the required packages.
3. Set the number of classes and HMM states. Algorithm 3: GUI
4. Define the file paths to the data.
5. Define the function to extract features using MF- 1. Start
CCs from audio files. 2. Import all the required packages.
6. Define the function to pre-process the data by 3. Create a window with the required title.
computing the mean MFCCs for each audio file 4. Add a label asking the user to choose an audio
in a directory. file.
7. Define the training and testing ratios. 5. Now add the button correspondingly.
8. Split the data into training and testing sets for 6. Define a function that takes the input file from
each class. the user and shows the English accent in that file.
9. Train a Gaussian HMM for each class on the 7. In the function defined, pass the input audio file
training data using the HMMlearn library. to the model that is built earlier.
10. Stop 8. Display the English accent to the user.
9. Stop
Algorithm 2: Testing the data
VI. Result and analysis
1. Start
To examine the effectiveness of the proposed model
2. Import all the required packages.
in figuring out accents of the English language,
3. Create a sample data set from the test data set.
numerous audio files in mp3 format has been
4. Pre-process each file in the dataset that is split-
provided as input. The model successfully com-
ted.
puted the likelihood values for each of the three
5. Load the pre-trained HMM models for each
accents: Indian, British, and American accent. The
class.
log-likelihood values were then compared and the
6. For each test audio file, extract its MFCC fea-
accent with the highest log-likelihood value was
tures.
14 English accent detection using hidden Markov model (HMM)

determined as the expected dialect for the input

audio file.
The model consisted of four distinct modules, each
serving a crucial purpose in the accent identification
process. The first module focused on pre-processing
the audio files, ensuring optimal data quality for sub-
sequent analysis. The second module involved feature
extraction, utilizing MFCCs to capture the distinctive
characteristics inherent to each accent. In the third
module, HMM was trained using GMM which enables Figure 2.3 Output when no file is selected
the model to learn and differentiate the accent patterns
effectively. Finally, the fourth module encompassed test-
ing, where the trained model was deployed to predict
the accent based on the log-likelihood values obtained.
Through extensive evaluation and experimentation,
the model yielded highly promising results, showcas-
ing its robustness and accuracy in identifying accents.
By employing an intuitive GUI, users could effort-
lessly provide audio files in mp3 format and obtain
the corresponding accent as the output. The poten-
tial applications of the model span various domains, Figure 2.4 Output predicted as Indian accent
including language learning, speech recognition, and
accent-related research, thereby contributing to a
comprehensive understanding and appreciation of
English language accents.
The research’s findings make a substantial contri-
bution to the field of accent analysis and recognition.
With the increasing need for effective communication
across diverse linguistic backgrounds, our model offers
a reliable solution for automated accent identifica-
tion. Future work in this area could explore expand-
ing the repertoire of recognized accents and further
refining the model’s performance. Overall, the pro- Figure 2.5 Output predicted as Britain
posed model stands as a valuable tool with immense
potential for both academic and industry, fostering
advancements in language-related studies and facili-
tating enhanced intercultural communication.
Figure 2.2 demonstrates the suggested model’s accu-
racy following execution in the Jupyter notebook.
Figure 2.3 shows the above output when no file is
selected. The output of the system indicates that no
audio file was provided for analysis. This serves as an
informative response, prompting the user to provide
a valid audio file in mp3 format.
Figure 2.4 shows the categorized output when a Figure 2.6 Output predicted as American
valid audio file is chosen and it will show the actual
accent of the speaker. This output highlights the mod- outcome demonstrates the model’s proficiency in suc-
el’s capability to correctly identify and distinguish the cessfully identifying and distinguishing the distinct
specific traits of the Indian dialect from other English characteristics associated with the British accent in
accents, such as the British and American dialects. spoken English.
Figure 2.5 shows the classified output as Britain Figure 2.6 a significant output has been provided,
when a valid British accent audio file is selected. This when the model predicts the selected audio file as the
American dialect. This result shows the model’s effec-
tiveness in correctly identifying and differentiating the
unique characteristics and speech patterns associated
Figure 2.2 Model accuracy with the American English dialect.
 Applied Data Science and Smart Systems 15

Figure 2.7 Spectrogram representation of the model

Figure 2.8 Mel-Frequency Cepstral Coefficients

Figure 2.7 shows the spectrogram representation visualizing and understanding the acoustic proper-
of the proposed model. The spectrogram presents a ties of different accents. It allows for a comprehensive
visual representation of the audio signals showing the analysis of the frequency bands and spectral charac-
frequency and depth components over time. This rep- teristics that contribute to accent variations.
resentation plays a crucial role in accent identification
as it offers valuable insights into the precise acoustic VII. Conclusion
patterns function of different accents. The spectro-
gram output represents a significant leap forward in In conclusion, the use of HMMs for English accent
the area of accent detection, contributing to improved detection is explored. Promising results are achieved
language understanding, cross-cultural communica- by utilizing HMMs to model the acoustic characteris-
tion, and the broader study of linguistic variations tics of different accent. Through the training process,
within English accents. a unique patterns and transitions present in various
Figure 2.8 provides a visual representation of the English accents is captured, thereby distinguishing
extracted MFCC features showing the distribution between them effectively.
and patterns of the coefficients for each audio sample. By leveraging HMMs, we have demonstrated the
The graph of MFCCs serves as a powerful tool for potential of this approach for accent detection. The
16 English accent detection using hidden Markov model (HMM)

HMM framework provides a robust and flexible Elizabeth, N., Steedman, M., and Goldwater, S. (2020). The
method for modeling temporal dependencies and cap- role of context in neural pitch accent detection in Eng-
turing the variability in speech signals. It has proven lish. arXiv preprint arXiv:2004.14846. Doi - https://
to be particularly suitable for accent classification doi.org/10.48550/arXiv.2004.14846
Al-Jumaili, Zaid, Tarek Bassiouny, Ahmad Alanezi, Wasiq
tasks due to its ability to handle sequential data.
Khan, Dhiya Al-Jumeily, and Abir Jaafar Hussain.
Although the work has yielded encouraging results,
(2022). Classification of Spoken English Accents Using
there is still ample room for improvement and fur- Deep Learning and Speech Analysis. In International
ther exploration in the field of English accent detec- Conference on Intelligent Computing Methodolo-
tion using HMMs. After uploading the audio files to gies. ICIC 2022. Lecture Notes in Computer Science.
the GUI, it predicts the accent. Our future work is to 13395, 277–287. Cham: Springer International Pub-
convert it into web application. It’s also been trying to lishing, 2022. doi: https://doi.org/10.1007/978-3-031-
improve the accuracy thereby to identify the language 13832-4_24
spoken in the audio file. China. (2022). Proceedings, Part III. Cham: Springer Inter-
national Publishing, 2022.
Guntur Radha, K., Krishnan, R., and Mittal, V. K. (2020). A
References system for automatic regional accent classification. In
Zubi, Z. S. and Idris, E. J. Arabic Dialects System using 2020 IEEE 17th India Council International Confer-
Hidden Markov Models (HMMs). WSEAS TRANS- ence (INDICON). 1–5.
ACTIONS ON COMPUTERS. 21. 304–315. doi: Veranika, M. et al. (2022). Language accent detection with
10.37394/23205.2022.21.37. CNN using sparse data from a crowd-sourced speech
Mamun, R. K., Abujar, S., Islam, R., Badruzzaman, K. B. archive. Math., 10(16), 2913.
M., and Hasan, M. (2020). Bangla speaker accent Sami, M. and Habbash, M. (2021). Study of the influence of
variation detection by MFCC using recurrent neural Arabic mother tongue on the English language using
network algorithm: A distinct approach. In: Saini, H., a hybrid artificial intelligence method. Interact. Learn.
Sayal, R., Buyya, R., Aliseri, G. (eds). Innovations in Environ., 1–14.
Computer Science and Engineering. Lecture Notes in Keith, G. (2023) . Accent adjustment based on spoken feed-
Networks and Systems, vol 103. Singapore: Springer. back. Tech. Disclos. Comm.
https://doi.org/10.1007/978-981-15-2043-3 59 Nugroho, K., Winarno, E., Zuliarso, E., and Sunardi.
Alashban, A. A., Qamhan, M. A., Meftah, A. H., Alotaibi, (2023). Multi-accent speaker detection using normal-
Y. A. (2022). Spoken language identification system ize feature MFCC neural network method. J. RESTI
using convolutional recurrent neural network. Appl. (Rekayasa Sistem Dan Teknologi Informasi), 7(4),
Sci., 12, 9181. https://doi.org/10.3390/app12189181. 832–836.
Ramoji, S. and Ganapathy, S. (2020). Supervised I-vector Lan, Y., Xie, T., and Lee, A. (2023). Portraying accent ste-
modeling for language and accent recognition. Comp. reotyping by second language speakers. PLoS ONE,
Speech Lang., 60, 101030. ISSN 0885-2308. https:// 18(6), e0287172.
doi.org/10.1016/j.csl.2019.101030. Klumpp, Philipp, Pooja Chitkara, Leda Sarı, Prashant Serai,
Keqi, D., Cao, S., and Ma, L. (2021). Improving accent Jilong Wu, Irina-Elena Veliche, Rongqing Huang, and
identification and accented speech recognition under a Qing He. (2023). Synthetic Cross-accent Data Aug-
framework of self-supervised learning. arXiv preprint mentation for Automatic Speech Recognition. arXiv
arXiv:2109.07349. preprint vol. arXiv:2303.00802 1–5. Doi: https://doi.
Utkarsh, S. et al. (2020). Foreign accent classification using org/10.48550/arXiv.2303.00802
deep neural nets. J. Intel. Fuzzy Sys., 38(5), 6347–6352. Dylan, W., Dev, S., and Nag, A. (2023). Hilbert-Huang-
Radzikowski, K., Wang, L., Yoshie, O. et al. (2021). Accent transform based features for accent classification of
modification for speech recognition of non-native non-native English speakers. 2023 34th Irish Sig. Sys.
speakers using neural style transfer. J. Audio Speech Conf. (ISSC). IEEE.
Music Proc., 11. https://doi.org/10.1186/s13636- 021- Margot, M. and Carson-Berndsen, J. (2023). Investigating
00199-3. Phoneme Similarity with Artificially Accented Speech.
Joseph, A. P. (2020). Domestic language accent detector us- In vol. Proceedings of the 20th SIGMORPHON
ing MFCC and GMM. Int. J. Appl. Engg. Res., 15(8), workshop on Computational Research in Phonetics,
800–803. Phonology, and Morphology, 49–57. doi: 10.18653/
Khanal, S., Johnson, M. T., Soleymanpour, M., and Bozorg, v1/2023.sigmorphon-1.6
N. (2021). Mispronunciation detection and diagno- Zuluaga-Gomez, J. et al. (2023). CommonAccent: Explor-
sis for Mandarin accented English 30 speech. 2021 ing large acoustic pretrained models for accent clas-
Int. Conf. Speech Technol. Human Comp. Dialog. sification based on common voice. arXiv preprint
(SpeD), Bucharest, Romania, 62–67. doi: 10.1109/ arXiv:2305.18283.
SpeD53181.2021.9587408. Carlos, F. and Polzehl, Y. (2023). Domain Adversarial Train-
Arya, R., Singh, J., Kumar, A. (2021). A survey of multi- ing for German Accented Speech Recognition. Ger-
disciplinary domains contributing to affective com- man Acoustics Society, 1413–1416.
puting. Comp. Sci. Rev., 40. https://doi.org/10.1016/j.
cosrev.2021.100399.
3 Study of exascale computing: Advancements, challenges,
and future directions
Neha Sharmaa, Sadhana Tiwari, Mahendra Singh Thakur, Reena Disawal
and Rupali Pathak
Prestige Institute of Engineering Management and Research, Indore, India

Abstract
Exascale computing is the high performance computing system that can measure quintillion calculations per second. It is
capable to perform the calculations of 1018 floating point operations (FLOPS) per second. It is the term given to the next
50–100 times increased speed over very fast super computers used today. High performance computing application helps
to simulate large scale application, machine learning, artificial intelligence, industrial IoT, weather forecasting, healthcare
industries and many more. The increased computational power will enable researchers to tackle more complex problems,
collects and analyze larger data sets, perform simulations with high accuracy and resolutions. Exascale computing has the
power to transform scientific research, spur innovation, and tackle complex issues that were previously computationally
impractical. This paper describes a brief description, architecture and various applications of exascale computing such as
healthcare, microbiome analysis, etc. This paper also presents the future and research aspects of exascale computing.

Keywords: High performance computing, exascale computing, super computers, parallel processing, data analytics, computer
architecture

I. Introduction management. Exascale systems are designed to get

over the drawbacks and difficulties that current HPC
High performance computing (HPC) technology
systems experience, including high power usage,
affects almost every sphere of our life that includes
memory and storage bottlenecks, limited program-
education, communication, entertainment, economy,
mability, and scalability as shown in Table 3.1.
engineering and science, etc. The next stage of HPC
Exascale computer provides extra ordinary power
is known as exascale computing (EC), where com-
and memory so that it can be applied in HPC areas like
puter systems is capable to perform the calculation large scale simulation, machine learning, deep learn-
at least 1018 floating point operations per second (1 ing, and multi physics (Francis et al., 2020; Matthew
exaFLOPS) or a billion (i.e. a quintillion) calculations et al., 2020; Choongseok et al., 2023). EC has done
per second. lots of improvement in scientific, medical, weather
In comparison to existing petascale systems, it rep- forecasting, and artificial intelligence (Francis et al.,
resents a huge increase in computational capability. 2020; Yuhui et al., 2022).
It is thousand times faster than petaflops machines Media governments such as India, US, EU, China,
(Huang et al., 2019; Matthew et al., 2020). EC will Japan, etc., and industries such as IBM, Intel, etc.,
enable simulations and analyze previously unheard together are putting their efforts to build exascale
of complexity and scope, which will revolutionize computers. There is competition between United
scientific research, engineering, and data analytics States and China to become the first nation that has
(Fabrizio et al., 2019; Matthew et al., 2020). an exascale computer. The estimated cost of this exotic
EC is developed with the increasing demands of computer equipment will be in between $ 400 million
scientific and industrial applications. Because these and $ 600 million. Aurora 2021 (A21) (Matthew et
applications requires large computational capability al., 2020) is therefore first US exascale system.
to solve complex problems in areas like astrophysics,
materials science, energy research, climate modeling, Components of EC
and more. These applications generate large amount a. Processor: Homogenous or heterogeneous plat-
of data and requires complex simulation. These sim- forms can be used for designing of exascale sys-
ulations demand extremely high level of processor tems. In heterogeneous, exascale uses CPU and
power, memory capacity, and storage bandwidth. GPUs to improve the performance efficiently.
To achieve exascale processor, significant advance- b. Memory requirement: In order to meet the per-
ment is required in computer architecture, sys- formance requirement, exascale needs high band-
tem design, software development and energy width memory (HBM). HBM stack can contain

[email protected]
n
18 Study of exascale computing: Advancements, challenges, and future directions
Table 3.1 Technological overview of exascale system

Parameter 2009 2018 Swimlane 1 (extrapolation 2018 Swimlane 2 (represent the

of multi-core design) GPU design point)

Power 6 MW ~20 MW Same as SL1

Memory 0.3 PB 32–64 PB Same as SL1
Node performance 125 GF 1.2 TF 10TF
Latency 1–5 µs 0.5–1 µs Same
Memory Latency 150–250 clock cycles 100–200 clock cycles (~50 ns) Same
(~70–100 ns)
Node memory BW 25 GB/s 0.4 TB/s 4–5 TB/s
Storage 15 PB 500–1000 PB Same
System size (nodes) 18,700 1M 100,000
IO 0.2 TB 60 TB/s Same as SL1

up to eight Dynamic Random Access Memory hardware. EC has various technical challenges such as
(DRAM) module which are connected through power consumption, memory management, parallel-
two channels per module. It includes silicon in- ism, fault tolerance, and scalability (John et al., 2011;
terposer base die with a memory controller and Pete et al., 2012; Judicael et al., 2015; Mahendra et
interconnected through-silicon via (TSVs) and al., 2020; Matthew et al. 2020). In literature, authors
microbumps. Double Data Rate (DDR) memo- Matthew et al. (2020), Maxwell et al. (2021), Francis
ries are generally off-chip dual-in line memory et al. (2020), John et al. (2011) have discussed vari-
modules means they are separated from CPU die. ous benefits, opportunities and challenges in EC.
HBM offers low latency and has high through- Fabrizio et al. (2019) reviewed the political and social
put as compared to DDR because it is close to aspects of exascale computing along with history of
the processor die. HPC architecture. Peter et al. (2013) and Martin et
al. (2019) have explained the requirement analysis of
II. Related work exascale based on cases use. Author has described ref-
erence architecture and technology-based architecture
Enormous research is going on HPC technology to of the process project in EC. Martin et al. (2019) have
improve the performance of high speed application. In proposed novel hardware designs and architectures
2018, exascale system was introduced which performs that can deliver exascale performance while maintain-
calculation of 1018 FLOPS (Matthew et al., 2020). ing energy efficiency and reliability. Peter et al. (2013),
Exascale system helps to simulate high speed applica- Martin et al. (2019) authors summarized the differ-
tions such as healthcare industry, industrial IoT, data ent challenges in operating system such as technical,
analytics and many more (Levent Gurel et al., 2018; business and social for exascale system. This includes
Tanmoy et al., 2019; Francis et al., 2020). Tanmoy research on resource management, job scheduling,
et al. (2019) described that how artificial intelligence power management, fault tolerance.
(AI), Big data and HPC helps to discover new drug
with reduce cost and minimize development cost.
III. Architecture of EC
Francis et al. (2020) explored the role of EC in dif-
ferent areas such as microbiome analysis, healthcare In view of the requirement of different industry, the
industry, chemistry and material applications, data architecture of exascale is divided in to three groups:
analysis and optimization applications, energy appli- virtualization, data and computing requirement (Peter
cation, earth and space science applications and many et al., 2013; Martin et al., 2019). The exascale com-
more. Tanmoy et al. (2019) explained how EC tech- puting architecture is shown in figure 3.1.
nique and AI helps to predict the cancer and tumor In virtualization layer, virtualization requirements
response in advance. Exascale computing enables are taken directly from the application basis of our
engineers and researchers to design, optimize, and test user communities-container support that provides
new products and technologies more efficiently and lightweight virtualization method similar to app
quickly. In his paper, L. Gurel et al. (2018) reviewed packages. Advantages of this technique are flexibil-
that contribution of EC in autonomous driving and ity, reliability, ease of deployment and maintenance.
how EC reduces the software complexity with available User applications require to be distributed across a
 Applied Data Science and Smart Systems 19

Figure 3.1 Exascale computing architecture

variety of computer infrastructure, portability and IV. Key technology for EC

collaboration.
The essential technologies needed for EC are dis-
The primary requirement is to manage exascale
cussed in the following section.
data sets or excessive data flow, it is impossible to alter
and manage at a single data center. It is also integrated
A. HPC system
with Meta data management. Depending on the data
Since EC handles large amount of data (1018 FLOPS)
services, data connection or data transfer is big chal-
and computational workloads (Matthew et al., 2020).
lenge. The exascale platform should support large data
Therefore, it includes numerous linked processing
transfer in all infrastructures (Martin et al., 2019).
units, such as CPUs, GPUs, or specialized accelerators
The computing requirement is needed that can sup-
(Thiruvengadam et al., 2017).
port all HPC, cloud computing and speed require-
ment. Aim of current scientific application is huge
B. Parallel processing
data distribution at all computer research centers or
EC strongly relies on parallel processing, which
sites. As a result, degree of parallelism and concur-
involves running numerous computer processes con-
rency is also increased (J. Singh et al., 2009; Jiangang
currently to achieve high throughput. This entails
et al., 2021). These requirements need to be fulfilled
decomposing complicated issues into simpler issues
while designing computing requirement. In continu-
so that numerous processing units can handle them
ation, computing architecture is proposed based on
simultaneously (Matthew et al., 2020).
modularity and scalability. These two approaches are
useful in high degree parallelism and high distribu-
C. High speed processor
tion. It offers flexibility to be used small modules and
EC necessitates the creation of advanced processor
method that exploits various sources of exascale sys-
that can supply the necessary levels of computational
tems efficiently (Martin et al., 2019).
20 Study of exascale computing: Advancements, challenges, and future directions

power. This can entail utilizing heterogeneous archi- V. Emerging applications of EC

tecture, which pair conventional CPUs with special-
There are various applications of EC that is shown in
ized accelerators like GPUs or FPGAs (Thiruvengadam
Figure 3.2 and the detail descriptions are given below:
et al., 2017).
A. Advances in healthcare (accelerating drug discov-
D. Memory
ery with AI, HPC and big data)
To manage the enormous amount of data required for
The current state of drug development is a long,
EC, large-capacity, fast memory and storage devices
expensive process and, to some extent, a shot in the
are needed. Improvements in random access memory
dark. The cost of developing even a single drug is
(RAM), high-speed cache, and storage technologies
high. According to research by the Tufts Center, the
like solid-state drives (SSDs) or non-volatile memory
cost of drug development was found to be more than
are all included in this (Matthew et al., 2020).
$2.5 billion.
There are many different healthcare sectors, such
E. Energy efficiency
as the pharmaceutical industry, and many more are
EC systems use a lot of electricity, so energy effi-
struggling to develop new drugs, and patients are also
ciency is important. Sustainable energy source are
waiting for new drugs to improve their medical con-
essential for EC. To address the power and thermal
dition. AI, cloud computing, IoT and Big data aim to
concerns, this entails creating low-power CPUs, opti-
shorten development time and reduce costs at every
mizing algorithms, and using cutting-edge cooling
step of the new drug development chain, from initial
techniques.
research to clinical trials. Different emerging tech-
nologies help scientists do retrospective analysis on
F. Software model
existing data analytics, also help find new drugs for
It is essential to create software and programming
disease. AI that runs through large amounts of genetic
models that effectively make use of the extreme par-
data to determine the correlation between a particular
allelism and diverse architectures found in exascale
DNA sequence and a disease that will help identify
systems. This includes providing tools for managing
potentially useful drugs. Once this process is com-
and debugging intricate software systems, as well as
plete, AI uses electronic media recording to identify
optimizing algorithms and constructing parallel pro-
potential drug for the target audience and enable the
gramming frameworks.
industry to develop setup and put drugs into trials
(Tanmoy et al., 2019; Francis et al., 2020).
G. Data management and analytics
Traditionally, multiple clinical trial phases are
EC generates large amount of data that need to be
required once the most promising drugs have been
managed and analyzed. To get useful insight from
identified, which becomes time-consuming, demand-
the enormous datasets produced by exascale simula-
ing and costly. Data analytics, IoT and cloud comput-
tions and computations, effective data management
ing already offer benefits here and promise to bring
approaches, including data storage, retrieval, analysis,
more in the future. Wearable and implantable IoT
and visualization, are required.

Figure 3.2 Application of exascale computing

 Applied Data Science and Smart Systems 21

devices collect enormous amounts of patient infor- These models can be expanded upon in order to
mation from sensors and data storage in the cloud. enhance pre-clinical drug testing and accelerate
The cloud provides large storage that is cheaper and cancer patients’ access to drug-based therapies.
requires high computing power that assists in the data 2. RAS (Rat sarcoma virus) pathway issue 2.
analysis process. In short, accelerating drug discovery 3. Planning for a treatment approach.
with AI, HPC and Big data (Francis et al., 2020):
In order to predict treatment response, compli-
• Current processes for drug discovery are time cated, indirect interactions between drug structures
consuming and expensive. and tumor structures are captured using supervised
• Cutting-edge technologies such as artificial intel- mechanical learning techniques to address drug
ligence, HPC and Big data will reshape method of responses. Using the history of past simulations, the
drug discovery (Tanmoy et al., 2019). RAS technique uses multi-tasking to search a large-
• Requires high computing hardware power results scale space to define the scope of a series of simu-
for the ability to model further drug progress be- lations. Machine learning (ML) models are used to
fore moving on to clinical trials. automatically read and compile millions of clinical
records in order to deal with the treatment approach.
B. Dynamic stochastic power grid Direct conclusions about are provided by ML models.
ExaSGD application is used to preserve the integrity Every issue calls for a distinct approach for to inte-
of power grids and address load imbalances. With the grate the learning, yet they are all supported by the
help of this programe, the grid’s real-time response same CANDLE environment.
optimization against probable disruption occurrences Python library, the runtime manager, and a set of
is created using models and algorithms. ExaSGD deep neural networks are all included in the CANDLE
serves power grid operators and planners and is based package. Tensor Flow, PyTorch, and deep neural
on exascale computing. networks that download and represent three issues
Power grids keep the supply and demand for are employed for exascale computing, with a run-
electricity in balance. Attacks on the grid, whether time supervisor organizing the distribution of work
physical or digital, can result in costly power grid throughout the HPC system. Performance features
components being permanently damaged or experi- include semi-automated uncertainty quantification,
encing large-scale blackouts. Load shedding is utilized large-scale search for hyper parameters, and auto-
to prevent generation-load imbalance and maintain matic search for best model performance.
the functionality of the power grid (Francis et al., Exascale challenges are represented in the urgent
2020). requirement to train many related models. Each test
Cyber-enabled control and sensing, plug-in stor- application’s demand results in cutting-edge models
age devices, censored elements, and smart meters that span the speculative space (which is not specific
managed automatically and remotely can all have an to the idea of an accurate medicine).
impact on how the electrical grid behaves. To avoid
generation and load shedding at the moment, load D. Microbiome analysis
shedding is employed. Using simulations, the ExaSGD Microbial species are important part of our ecosys-
tool offers additional ideal configurations for resolv- tem. They are influencing various domains such as
ing generation-load imbalance. This method enhances agricultural production, pharmaceutical and also
the electricity grid’s ability to recover from various used to make oils, medicines and other products. To
risks (Francis et al., 2020). study and gather information about the microbe’s
genome, sequence methods are used. In genome
C. Deep learning (DL) enabled the precise cure for sequencing, Metagenomics data are larger and more
cancer plentiful results in increased cost of computation. As
Project “CANDLE application” was started by the a solution, the ExaBiome application develops data
DOE and NCI (National Cancer Institute) of the NIH integration tools with high computing power (Francis
(National Institutes of Health). The goal of this proj- et al., 2020).
ect is to develop CANDLE (Cancer Learning Area), Metagenomics is a domain that explores functional
an amazing and in-depth learning environment for and structural details of the microbiome. Metagenome
exascale programs. Three key challenges are being integration, protein synthesis and signature-based
addressed by the CANDLE programe (Tanmoy et al., methods are three major computational problems
2019; Francis et al., 2020): faced in bioinformatics domain. ExaBiome attempts
to provide measurable tools for above stated prob-
1. Find a solution for the drug response issue and lems. Metagenome integration means capturing raw
create models for predicted drug responses. data sequences and generates long gene sequences
22 Study of exascale computing: Advancements, challenges, and future directions

and signature-based methods enable comparable and composition. The cornerstone for comprehending
effective metagenome analysis (Francis et al., 2020). engineering structures, materials, and energy science
MetaHipMer, a well-known metagenome compiler is structural strengths and heterogeneities, or con-
created by the ExaBiome team, scales thousands of formational mutations in macromolecules. Single-
computers in contemporary petascale-class architec- particle imaging (SPI) and X-ray scattering variation,
ture. Additionally, a sizable ecological database has which are non-crystalline based diffractive imaging
been created. To take advantage of the chance for techniques, may see and analyses these structural het-
enhanced node compatibility with memory structures, erogeneity and variations. This characteristic encour-
including GPUs, work is being done on measurable ages interest in the creation of X-ray free-electron
upgrades across nodes and node level improvements. lasers. Effective data processing, fragmentation pat-
With other collaborators, MetaHipMer exhibits terns, and reconstruction of 3D electron cones, how-
competitiveness. The second long-term compiler is ever, enable the visualization of structural changes
also being developed and has a significantly larger over time (Francis et al., 2020).
computer density, making it well suited to exascale The problem with ExaFEL is to devise an auto-
systems even though MetaHipMer is made for short matic analysis pipeline for single-part imaging using
reading data (Illumina) and is meant for long-term different techniques. This requires the reconstruction
data. HipMCL, the second code from ExaBiome, of a 3D cell structure from 2D separating images.
offers a way to measure proteins. The structure of This conversion is done by new Multi-Tiered Iterative
protein families in the billions of proteins may be seen Phasing (M-TIP) algorithm.
thanks to HipMCL, which has thousands of nodes. Diffraction images from distinct particles are gath-
These codes are based on typical compound patterns ered in SPI. The production of molecules (or atoms)
with flexible character unit (DNA or protein) algo- and cohesive areas (or comparable particles) under
rithm alignment, minimal layout, calculation, and specific operating circumstances is also assessed using
analysis of fixed-length strands, as well as a range these diffraction images. Since the shapes and condi-
of graphs and small matrix techniques. Metagenome tions of the particles in the image are unknown and
integration is core of the ExaBiome complicated chal- heavily contaminated by sound, determining prop-
lenge, but that capability will make it simpler for new erties using the SPI test is challenging. Additionally,
bioinformatics problems to emerge (Francis et al., the quantity of accessible particles typically places
2020). a cap on the number of viable images. To determine
the form, areas, and molecular structure from a single
E. Analysis of data for free electron laser particle’s data obtained utilizing structural barriers
X-ray diffraction is used by the Linac Coherent simultaneously, the M-TIP algorithm uses a duplicate
Light Source (LCLS) at the Stanford Linear guessing framework. Additionally, it aids in the com-
Accelerator Centre (SLAC) to model individual prehensive information extraction from single-parti-
atoms and molecules for crucial scientific activities. cle diffraction.
The representation of molecular structure revealed A quick response is necessary to direct the test,
by X-ray fragmentation in close to real time will ensure that enough data is gathered, and modify
need for previously unheard-of computer compres- the sample concentration to obtain a single particle
sion scales and bandwidth data techniques. Data rate. Together, exascale computing power and HPC
detector measurements in light sources have sub- processes can handle the analysis of the expanding
stantially increased; after LCLS-II-HE development data explosion. As a result, researchers will be able to
is complete, LCLS will grow its data by three orders analyses data quickly, respond quickly to test-quality
in magnitude by 2025. The ExaFEL programme data, and simultaneously decide on a three-dimen-
uses exascale computation to accelerate the process sional sample design.
of reconstructing molecular structures from X-ray
diffraction data from weeks to minutes (Francis et F. Autonomous car
al., 2020). Self-driving vehicle will generate and use a variety
Users of LCLS demand an integrated approach to of data to analyze various parameters such as loca-
data processing and scientific interpretation which tion, road condition, and passenger safety. To man-
calls for in-depth computer analysis. Exascale pro- age all the data, you need HPC (Levent Gurel et al.,
cessing capacity will be needed to meet demand for 2018).
real-time analysis of the data explosion which will The car is equipped with sensors, embedded com-
take about 10 minutes (Francis et al., 2020). puters, cameras, high-precision GPS and satellite,
Because of its high repetition rate and brightness, wireless network, 5G connectors to connect to the
LCLS can map individual molecules’ inherent fluc- internet. Autonomous car will exchange data with the
tuation in relation to flexibility and ascertain their management and control system and will sync with
 Applied Data Science and Smart Systems 23

a large database that continuously provides real-time astronomy, materials research, and computational
information such as weather, traffic conditions, emer- biology (Francis et al., 2020).
gency alerts, etc.
Autonomous car will generate a large amount of B. Accelerated innovation
data and will send more than four terabytes of data EC enables engineers and scientists to swiftly and
per hour to the cloud. Exascale high performance efficiently build, optimize, and test new products and
computing and Big data are therefore capable of deliv- technologies. It enables rapid innovation in fields
ering the computing power required to use predictive including aerospace, automobile design, energy sys-
decision support systems to evaluate large amounts tems, and material research by allowing for the study
of data. of a broad design space. EC aids in the identification
of optimal designs, resulting in improved products
VI. Benefits of EC and solutions, by modeling and analyzing compli-
cated systems.
The speed of EC is 50 to 100 times faster than latest
supercomputer. Therefore, this kind of HPC applica- C. Advances in data analytics and AI
tion helps to simulate large scale application, ML and EC enables the processing and analysis of enormous
AI, etc. (Matthew et al., 2020; Maxwell et al., 2021). datasets in real-time, opening up new opportunities in
It is fast, and cost effective. As a result, intelligent data analytics and AI. It makes possible for DL and
storage capacity, computing power can be applied machine learning models to be more accurate and
in industries like health care, chemical, National effective, which advances fields like genomics, person-
Security, reducing pollution, and many more (Francis alized medicine, social network analysis, autonomous
et al., 2020). EC helps to minimize health issues, and systems, and recommendation systems (Tanmoy et al.,
proves the better quality of life by optimizing the 2019; Francis et al., 2020). EC facilitates the extrac-
transportation facilities. In short, EC has a number of tion of useful insights from enormous amounts of
advantages that could revolutionize fields including data, fostering innovation and decision-making.
engineering, society, and scientific study is shown in
Figure 3.3. Some advantages of exascale computing D. Cross-disciplinary collaboration
are: EC fosters cross-disciplinary cooperation among
scholars. Exascale systems’ computational capacity
A. Scientific discovery and resources can be used by scientists, engineers, and
EC enables scientists and researchers to run simula- subject-matter specialists to tackle challenging issues
tions and models at a scale and resolution that have that call for interdisciplinary solutions (R. Arya et
never been possible before. This may result in fresh al., 2021). Through information exchange and inte-
scientific understandings, discoveries, and a better grated problem-solving, this partnership may result
comprehension of intricate processes. Exascale simu- in advances in areas like fusion energy, drug devel-
lations can facilitate discoveries and speed up scien- opment, urban design, and computational social
tific development in areas including climate modeling, sciences.

Figure 3.3 Benefits of exascale computing

24 Study of exascale computing: Advancements, challenges, and future directions

E. Precision and realism et al., 2015; Mahendra et al., 2020; Matthew et al.,
EC allows for simulations and modeling with a level of 2020) are shown in Figure 3.4. These challenges are
accuracy and realism never before possible. Exascale as discussed in the following sections.
simulations deliver more precise results by including
complex interconnections and finer-grained details. A. Technical challenges
This improves decision-making processes, which Exascale system has identified four key challenges:
helps in better forecasts, and encourages the creation increased number of faults, power requirement mini-
of trustworthy and durable systems and technologies. mization, memory management and parallelism at
node level. These challenges are directly related to
F. Economic and social impact exascale OS/R (operating system and runtime soft-
EC holds the promise of fostering both societal and ware) layer. Hardware complexity, resources chal-
economic improvement. By quickening the pace of lenges within OS, programming model, design issues
product development cycles, enhancing efficiency, and are few more to handle.
cutting costs, it encourages innovation and supports
industries. By offering strong tools for modeling, i) Resilience
analysis, and optimization, EC also helps to address As the numbers of components are increasing
major issues like climate change, healthcare, and sus- on chip, the numbers of faults are also increases.
tainable energy. These faults cannot be protected by other error
detection and correction technique. Timely prop-
G. Advances in computation agation fault notification across large network in
EC promotes improvements in computational meth- limited bandwidth scenario is very difficult.
ods and algorithms. To efficiently utilize the pro- ii) Power management
cessing capacity of exascale computers, researchers It is one of the critical challenges of exascale
investigate novel algorithms, optimization techniques, system. It requires 20–30 MW to run any appli-
and parallel programming paradigms. Beyond exas- cation. Resources can be change at any time to
cale computing, these developments help other com- adopt power requirement.
puter platforms and allow for further development in iii) Memory hierarchy
HPC. New memory technology emphasizes on reduc-
ing the power cost while data are transferring be-
VII. Challenges in EC tween different nodes. In case of exascale system,
the OS provide more support to runtime and ap-
Exascale are facing different technical and social chal- plication management and as the complexity re-
lenges (John et al., 2011; Pete et al., 2012; Judicael duces the OS overheads.

Figure 3.4 Challenges of exascale computing

 Applied Data Science and Smart Systems 25

iv) Parallelism ating and runtime systems have been designed to

Exascale system performs the calculation of 1018 impose strict barriers between nodes.
FLOPS. In order to achieve these calculations, • The node operating and runtime systems for an
application performs billions of calculation with exascale system will need to be right-sized in or-
in a second. This situation is significant challenge der to meet the needs of the application while
for developers. Performance cannot be increased minimizing overheads. Legacy operating and run-
through additional clock scaling but additional time systems do not emphasize restructuring to
parallelism is required in order to support next match the needs of an application.
generation of systems. But OS again faces dif-
ferent challenges like efficient and scalable syn- B. Business and social challenges
chronization, scheduling, scalable resource man- In spite of having technical excellence, exascale sys-
agement, global consistency, coordination, and tem is facing different challenges in business like lack
control. of transparency from vendors, sustainability and por-
v) Hardware-related challenges tability, preservation of existing code base and so on.
The hardware resources are heterogeneous in
nature. It includes multiple types of memory VIII. Future and research aspects of EC
and different processing element. Different type
EC has the ability to significantly advance technologi-
of component may have different performance
cal innovation, scientific discoveries, and societal con-
characteristics though they have capabilities to
cerns. The following are some crucial elements of EC’s
perform same function. Therefore, allocation of
future and research potential (Tanmoy et al., 2019;
resources becomes complicated to perform the
Francis et al., 2020; Maxwell et al., 2021; Yuhui Don
calculation. Handler establishment process for
et al., 2022). The market growth of EC is represented
hardware event is more difficult in order to sup-
in the table 3.2.
port responsiveness to faults, application moni-
toring system, and energy management and so on.
vi) OS/R structural challenges A. Simulation and modeling
The new operating system for exascale system is EC will make it possible for scientists and research-
facing different challenges like misalignment of ers to run simulations and models at previously
requirements, user-space resource management unheard-of scales and resolutions. This encompasses
and parallel OS services. disciplines like computational biology, astrophysics,
a. Misalignment of requirements: materials science, and quantum mechanics, among
The interference of OS should be minimized others. New scientific insights and discoveries will be
while at the same time OS should provide neces- made as a result of the ability to mimic and examine
sary support to all other application. complicated phenomena in greater detail.
b. User-space resource management
Generally OS directly manages and controls B. Data analytics and AI
all resources to different application software. EC will revolutionize these fields by making it pos-
However different programming models and ap- sible to handle and analyses enormous datasets in real
plication requires different resources results in time. Applications in fields such as genetics, person-
inefficiency. In the future application software alized medicine, social network analysis, intelligent
and runtime will require increased control over cities, and autonomous systems are included in this.
resources like core, memory, power and so on. Exascale computing and AI techniques have the poten-
c. Parallel OS services tial to revolutionize industries and spur innovation.
OS performs parallel processing; effective sup-
port and development for this interface is diffi- C. Multi-disciplinary research collaboration
cult. EC will promote cross-disciplinary research coopera-
vii) Legacy OS/R issues tion. EC systems will enable scientists, engineers, and
• The node operating and runtime systems for an subject matter experts to solve complicated issues
exascale system will need to be highly parallel, that call for interdisciplinary solutions. This partner-
with minimal synchronization. Legacy operating ship may result in innovations in industries like fusion
and runtime systems tend to be monolithic, fre- energy, medication development, materials design,
quently assuming mutual exclusion for large por- and urban planning.
tions of the code.
• The node operating and runtime systems for an D. Machine learning and DL
exascale system will need to support tight interac- Exascale computing will make it possible to train
tion across sets of nodes (enclaves). Legacy oper- and use deep learning and machine learning models
26 Study of exascale computing: Advancements, challenges, and future directions
Table 3.2 Represents the market growth of EC

Criteria Details

Reference year 2020

Forecast period 2021–2022
Revenue forecast in year 2028 USD 50.3 billion
Growth rate CAGR of 6.3% for the year 2021–2028
Regions covered North America, Europe, South America, Asia Pacific, Middle East and Africa
Profiles of significant players Advanced micro devices (US), Intel (US), HPE (US), IBM (US) , Lenovo (China),
Nvidia’s (Japan), NEC Corporation
Portion covered Through computation, devices, type, deployment, size, price, organization and server

that are more sophisticated and complicated. This IX. Market analysis of EC
will open up new opportunities in fields includ-
Compound annual growth of EC is going to 6.3%
ing speech and image recognition, natural language
throughout the course of the forecast and it is expected
processing, robotics, autonomous cars, and recom-
that market growth will reach USD 50.3 billion by
mendation engines. Researchers will investigate new
2028. This growth is driven by the increasing demand
architectures and algorithms to take use of exascale
for HPC across industries such as healthcare, finance,
capabilities for more precise and effective machine
energy, weather forecasting, and scientific research.
learning.
EC is being actively embraced by numerous sectors
to solve challenging computational issues and gain a
E. Computational fluid dynamics
competitive edge. For the instance, it provides sophis-
Engineers will be able to model and optimize fluid
ticated simulations for drug discovery, genomics, and
flow in unprecedented detail thanks to exascale com-
personalized treatment in the healthcare industry. It
puting’s enormous impact on computational fluid
supports high-frequency trading, risk modeling, and
dynamics (CFD) simulations. This has uses in envi-
portfolio optimization in the financial sector. EC is also
ronmental engineering, energy systems, automotive
used by energy corporations for seismic imaging, res-
design, and aerospace. Higher resolution and more
ervoir modeling, and energy production optimization.
accurate simulation and analysis of complicated flow
EC is strategically important, and governments around
dynamics can result in better designs and more effec-
the world are actively promoting its development. To
tive systems.
speed up exascale computing research and deployment,
numerous nations, including the United States, China,
F. Quantum computing
Japan, and European nations, have started national ini-
EC has the potential to be extremely important for
tiatives and funding programmes. These programmes
the growth and development of quantum computing.
seek to promote governmental, academic, and com-
Exascale systems can be a great resource for expe-
mercial cooperation in order to progress technology
diting quantum research and applications because
and preserve competitiveness. Several companies and
they can provide the enormous processing capacity
organizations are leading the exascale computing such
that quantum simulators and quantum algorithms
as Hewlett Packard Enterprise, IBM, Intel, NVIDIA,
demand. This includes creating quantum-enabled
AMD, and Cray. Universities, research organizations,
algorithms for application in real-world situations,
and national laboratories all contribute significantly to
optimizing quantum algorithms, and simulating
the development of exascale computer systems.
quantum systems.

G. Hardware and software innovation X. Conclusion

Ongoing research and development of new hardware EC is the new frontier of HPC technique. It has capa-
architectures, memory technologies, interconnects, bility to achieve performance of ExaFLOPS in terms
and software frameworks will be necessary for exas- of power and cost constraint. High computational
cale computing in the future. To fully utilize the capa- capability is able to tackle various challenges such
bilities of exascale systems, research will concentrate as scientific, medical, various social aspects and engi-
on enhancing energy efficiency, fault tolerance, scal- neering. It includes various research areas such as sys-
ability, and programmability. tem architecture, software and programming models,
 Applied Data Science and Smart Systems 27

performance optimization, energy efficiency, resilience Gürel, Levent (2018). Towards Exascale Computing for Au-
and fault tolerance, Big data analytics, and applica- tonomous Driving. In 2018 International Workshop
tion-specific research. To fully utilize the capabilities of on Computing, Electromagnetics, and Machine Intel-
exascale systems, researchers are concentrating on cre- ligence (CEMi), 17–18. IEEE, 2018. doi: 10.1109/
CEMI.2018.8610529.
ating innovative hardware architectures, implementing
Beckman, P., Brightwell, R., de Supinski, B. R. et al. (2012).
efficient algorithms, investigating new programming
Exascale operating systems and runtime software
paradigms, and optimizing system performance.EC report. US Department of Engineering. https://doi.
has a bright future ahead of it. It will promote multi- org/10.2172/1471119.
disciplinary research collaboration, improve scientific Zounmevo, Judicael A., Swann Perarnau, Kamil Iskra, Ka-
discovery, enable advances in AI and data analytics, zutomo Yoshii, Roberto Gioiosa, Brian C. Van Es-
revolutionize fields like computational fluid dynamics sen, Maya B. Gokhale, and Edgar A. Leon. (2015).
and quantum computing, and ignite innovation across A container-based approach to OS specialization for
sectors. However, there are obstacles in the way of exascale computing. In 2015 IEEE International Con-
EC full potential. Power consumption, memory con- ference on Cloud Engineering, 359–364. IEEE, 2015.
straints, communication bottlenecks, and the com- doi: 10.1109/IC2E.2015.78.
Gao, Jiangang, Fang Zheng, Fengbin Qi, Yajun Ding, Hon-
plexity of programming for massively parallel systems
gliang Li, Hongsheng Lu, Wangquan He et al. (2021).
are issues that researchers and engineers must over-
Sunway supercomputer architecture towards exascale
come. Moreover, to overcome technical obstacles and computing: analysis and practice. Science China In-
assure the successful deployment of these potent sys- formation Sciences, 64(4): 141101. doi: https://doi.
tems, the development of EC necessitates close coop- org/10.1007/s11432-020-3104-7
eration between academics, industry, and government. Arya, R., Singh, J., and Kumar, A. (2021). A survey of multi-
disciplinary domains contributing to affective comput-
ing. Comp. Sci. Rev., 1–9. https://doi.org/10.1016/j.
References
cosrev.2021.100399
Huang, H., Li-Qian, Z., YuTong, L. et al. (2019). An effi- Chang, C. et al. (2023). Simulations in the era of exascale
cient real-time data collection framework on petascale computing. Nat. Rev. Mat., 8, 309–313. https://doi.
systems. Neurocomput., 361(7), 100–107. https://doi. org/10.1038/s41578-023-00540-6.
org/10.1016/j.neucom.2019.06.039. Vijayaraghavan, Thiruvengadam, Yasuko Eckert, Gabriel
Matthew, N. O., Sadiku, Awada, E., and Musain, S. M. H. Loh, Michael J. Schulte, Mike Ignatowski, Brad-
(2020). Exascale computing (Supercomputers): ford M. Beckmann, William C. Brantley et al. (2017).
An overview of challenges and benefits. J. Engg. Design and Analysis of an APU for Exascale Comput-
Appl. Sci., 15(9), 2094–2096. DOI: 10.36478/jeas- ing. In 2017 IEEE International Symposium on High
ci.2020.2094.2096. Performance Computer Architecture (HPCA), 85–96.
Gagliardi, F., Moreto, M., and Mateo Valero, M. O. (2019). IEEE, 2017. doi: 10.1109/HPCA.2017.42.
The international race towards Exascale in Europe. Shalf, J., Dosanjh, S., and Morrison, J. (2011). Exascale
CCF Trans. HPC, 1, 3–13. https://doi.org/10.1007/ computing technology challenges. High Perform.
s42514-019-00002-y. Comp. Comput. Sci. – VECPAR 2010: 9th Int.
Kogge, P. and Shalf, J. (2013). Exascale computing trends: Conf., 6449, 1–25. https://link.springer.com/chap-
Adjusting to the new normal or computer architec- ter/10.1007/978-3-642-19328-6_1.
ture. Comput. Sci. Engg., 15(6), 16–26. DOI: 10.1109/ Singh, Jaiteg, and Kawaljeet Singh. (2009). Statistically
MCSE.2013.95. Analyzing the Impact of AutomatedETL Testing on
Bobák, M., Hluchy, L., Belloum, A.S., Cushing, R., Meizner, the Data Quality of a DataWarehouse. International
J., Nowakowski, P., Tran, V., Habala, O., Maassen, J., Journal of Computer and electrical engineering. 1(4)
Somosköi, B. and Graziani, M. (2019). Reference ex- 488–495. DOI:10.7763/IJCEE.2009.V1.74
ascale architecture. In 2019 15th International Con- Don, Y. et al. (2022). Exascale image processing for next-
ference on eScience (eScience) 479–487. IEEE. doi: generation beamlines in advanced light sources. Nat.
10.1109/eScience.2019.00063. Rev. Phy., 4, 427–428. https://www.nature.com/ar-
Alexander, Francis, Ann Almgren, John Bell, Amitava Bhat- ticles/s42254-022-00465-z.
tacharjee, Jacqueline Chen, Phil Colella, David Dan- Verma, Mahendra K., Roshan Samuel, Soumyadeep Chat-
iel et al. (2020). Exascale applications: skin in the terjee, Shashwat Bhattacharya, and Ali Asad. (2020).
game. Philosophical Transactions of the Royal Soci- Challenges in fluid flow simulations using exascale
ety A 378(2166): 20190056. 1–31 doi: https://doi. computing. SN Computer Science., 1(3): 178 pp. 1-14.
org/10.1098/rsta.2019.0056. doi: https://doi.org/10.1007/s42979-020-00184-1
Bhattacharya, Tanmoy, Thomas Brettin, James H. Doro- Zimmerman, M. I. et al. (2021). SARS-CoV-2 simulations
show, Yvonne A. Evrard, Emily J. Greenspan, Amy L. go exascale to predict dramatic spike opening and
Gryshuk, Thuc T. Hoang et al. (2019). AI meets exas- cryptic pockets across the proteome. Nat. Chem., 13,
cale computing: advancing cancer research with large- 651–659. https://www.nature.com/articles/s41557-
scale high performance computing. Frontiers in oncol- 021-00707-0.
ogy, 9, 984. DOI: 10.3389/fonc.2019.00984.
4 Production of electricity from urine
Abhijeet Saxena1,a, Mamatha Sandhu2, S. N. Panda3 and
Kailash Panda4
1
Utkal University, Odisha, India
2,3
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India
4
Laxmi Narayan College, Odisha, India

Abstract
The research work explores the possibility of utilizing urine, the most abundant waste on earth as an unconventional, yet,
plausible alternative to generate electricity. A groundbreaking two-phase method has been introduced that utilizes a urea
electrolytic cell to convert urine into electricity. In the initial phase, urea-rich water is broken down to extract Hydrogen,
which serves as the primary input for electricity generation in the subsequent phase. Our paper comprehensively examines
the technology employed in the urine powered generator, elucidates the intricacies of the process, and assesses the overall
efficiency of the integrated system, positioning it as a promising advancement for the future. Furthermore, a comparative
analysis is conducted against existing energy sources, shedding light on the environmental, economic, and technological
advantages of our approach.

Keywords: Clean energy, hydrogen economy, PEM fuel cell, electrolysis, urine, waste-to-energy, unconventional energy sources

I. Introduction most abundant waste (T. O. Ajiboye et al., 2022) on

earth, for generating electricity. The process compris-
“To Evolve is to Sustain.” This world, and everything
ing of two feasible phases converts the input into
within, is evolving at a considerable pace. With every
the desired output. Here, presented concisely, are the
passing minute, we are becoming more and more
methods used along with the schematic diagrams to
equipped for the future which patently, is uncertain
help comprehend well. We shall also showcase the
through researches, inventions and ideas. However,
efficiency of our final input/output integrated system
over the same horizontal space, Homo sapiens have
along with possible comparisons. It shall, besides,
been responsible for aggravating own surround-
cover detailed reactions occurring in the processes to
ings to such great extents that many of our natural
help to understand the working of the system thor-
resources that once were abundantly present are on
oughly. In the end, while concluding, we assign a new
the verge of dying out. The world has a coherent
identity entirely to urine, projecting it as a cheap-yet-
identity lagged in replacing those nearing extinction
not-so-cheap waste. At the end of the discussion, it’s
resources with something more meaningful that can
seen that how this new area of the invention will be
compensate their presence once they are gone. For the
characteristic of the time when power will be cheaper
future generation, we must reserve to make the pro-
and available to every household in the cleanest pos-
cess eventually a prosperous one instead of despair
sible way.
one. If we imagine the world without resources, for
a country like India where most of the power (elec-
tricity) around 57% is sourced by coal, sooner or II. Technology
later, will eventually run out, time has come for us to The technology to generate electricity from urine con-
think about alternate sources (Sataksig et al., 2017). sists of two major steps: Step 1 – Extraction of hydro-
Questions seem to be piling up with time, seeking the gen from urea contained in urine. Step 2 – Generation
right answer. The way forward is to acknowledge the of electricity from hydrogen. The technologies used
problems and work collectively to find remedies to are as follows: Electrolytic cell – The working prin-
each problem. This research aims to attempt and pro- ciple of this cell is no different from any other cell of
pose a solution to one of the many existing problems. the same kind. However, the electrolytic cell for the
Suggestions and analysis indicate, the urgent need to desired purpose needs to be engineered such that it
find clean energy sources before it becomes an acute is efficient and feasible to produce on a mass scale.
issue (Bhashyam et al., 2020). Alternate source to Proton-Exchange Membrane (PEM) (hydrogen) fuel
generate electricity is the problem that needs to be cell – PEM fuel cell is an electro-chemical cell that
prioritized. This research is about using urine, the

[email protected]
a
 Applied Data Science and Smart Systems 29

takes hydrogen as input, reacts with the oxygen in the Equation (1) defines the oxidation of urea at the
air to generate electricity. anode of the electrolytic cell. Equation (2) defines
the oxidation of Ni(OH)2 to NiOOH and the current
III. Proposed methodology produced during the electrolysis process. Equation (4)
demonstrates that a remarkably low potential of less
This section details the complete process into 2 phases; than 1.23 V is required for the electrolysis of water,
the extraction of hydrogen from urine (S. Yeasmin et and theoretically 70% hydrogen (Amanda K. et al.,
al., 2022) and its conversion into electricity. Phase 1 2015) is obtained. This implies that during the nitrate
– Urine is abundantly available. The prime element remediation of wastewater, nitrogen is generated from
of urine is urea, from which hydrogen [H], carbon the anode while hydrogen, a valuable constituent for
[C], nitrogen [N] and oxygen [O] can be extracted. the imminent hydrogen economy (Kar et al., 2022),
Regardless of technological advancements, there is is liberated at the cathode. In simple words, pure
still no technology that can convert urea to hydro- hydrogen (H2) can be collected at the cathode while at
gen. This proposed process could sustain not only anode the nitrogen can be collected along with traces
hydrogen resources but also do the de-nitrification of of oxygen as well as hydrogen (S. A. Grigoriev et al.,
urea-abundant water which is generally discharged 2006). Phase 2 – Owing to the flammable property
into rivers. The proposed system block diagram is as of hydrogen, the gas extracted in phase 1 need to be
shown in Figure 4.1. The electrolytic cell designed (T stored in a cylinder with safety valves on the inlet and
Gera et al., 2021) would use the proposed electro- outlet tubes for not allowing its reverse flow as dis-
chemical process (Kumar et al., 2018) for extracting cussed in (Langmi et al., 2022). The hydrogen from
hydrogen from urea (Amanda K. et al., 2015; Jinqi Li. the outlet tube of the cylinder is released into the PEM
et al., 2022) as shown in Figure 4.2. fuel cell as seen in Figure 4.3. In the PEM fuel cell,
Using above-described electrolytic cell along with the hydrogen reacts with the oxygen from the air to
inexpensive transition metal nickel, the electro-chem- release energy while forming water. Detailed working
ical oxidation of human urine, is represented in the of the PEM cell is as mentioned below (Tolga Taner et
following four equations: al., 2018), where the hydrogen molecule gets oxidized

(1)

(2)

(3)

(4) Figure 4.2 Electrolytic cell

Figure 4.1 Block diagram of the proposed system

30 Production of electricity from urine

Figure 4.3 PEM fuel cell

and loses two electrons, as it passes through the mem- relatively little heat, and without producing any light.
brane. Thus, two ions of hydrogen are generated as Due to these characteristics, the reaction is not clas-
oxidation half-reaction at the anode as represented in sified as combustion. In PEM fuel cell as discussed
the Equation (5). in Yun Wang et al. (2022), considering the energy-
producing step only, i.e., by omitting other parts of
(5) the energy picture, then, electricity produced by it is
more environmentally friendly, than that produced by
The hydrogen ions H+, combines with oxygen (O2) coal-fired or nuclear power plant. The process does
while passing through the proton exchange mem- not emit any greenhouse gas or pollutants, or radioac-
brane, producing two electrons of water as reduc- tive waste. With hydrogen as the fuel in PEM fuel cell,
tion half-reaction at the cathode as represented in the the only chemical product released is water. The water
Equation (6). released by the fuel cell can be an added benefit for
the astronauts in the space station/shuttle who oth-
erwise must rely on moisture/water from respiration,
(6)
sweat and urine (Nehir Atasay et al., 2023), For per
mole of water formed, the overall reaction releases
The overall cell equation, as is with the gal-
286 kJ of energy. However, rather than being liber-
vanic cells, is given as total amount of half-reaction
ated in the form of heat, 40–60% of this energy is
equations:
converted to electric energy by the fuel cell. The con-
version proportion is much higher compared to 20%
 (7) or less usable in case of internal combustion engine
for generating electricity from fossil fuels (Singla et
Two electrons (2e−) with two Hydrogen atoms al., 2021). The electricity produced from the phase
(2H+) thus cancel as represented in Equation (8): 2 of the entire system can be stored in a battery for
further use as per the requirements (as depicted in
(8) Figure 4.4).

The electrons flowing from the anode to the cath- IV. Comparison of results analysis
ode of a fuel cell move through an external circuit
to do work which is the whole point of the device. Table 4.1 shows the comparison of energy consump-
Thus, in a fuel cell, a movement of electron occurs tion between electrolysis (Panigrahy et al., 2022) of
from H2 to O2. This flow occurs with no flame, with water and urea, using Ni anodes, under lab conditions,
 Applied Data Science and Smart Systems 31

Figure 4.4 Schematic representation of the proposed system

Table 4.1 Comparison of electrolysis Table 4.2 Comparison of efficiency of PEM fuel cell

Electrolysis Energy (Wh/g) H2 cost (INR/kg) Efficiency of PEM Wattage (in KWh) H2 (in kg)
fuel cell (in %)
Urea 37.5 187.5
Water 53.6 268.0 100 33.33 1
60 20 1
60 1 0.05

based on cost of energy @ INR 5 per kWh. The com- follows: Urine is 95% water implies that 1 l of urine
parison is on two parameters: (a) Wattage per gram contains 0.95 l water. Further, 1 l of water weighs
of hydrogen (b) Cost of producing 1 kg of hydrogen. 1 kg implies 950 ml of water would weigh 950 g.
Illustration of unit economics of the system – For During electrolysis, 2 moles of water liberate 2 moles
evaluation of the energy unit (in terms of units of of hydrogen and 1 mole of oxygen. The molar mass
electricity consumed) economics of the system, i.e., of water being 18.015 g implies that 950 g of water
amount of hydrogen needed to produce electricity (J. is equivalent to 52.73 moles of water. Thus, 52.73
Singh et al., 2019) and subsequently, checking urine moles of hydrogen would be produced from 950 g
and energy required for the production of hydrogen of water (or 1 l of urine). Since, mass = molar mass *
(quantity that will produce 1 kWh of electricity). number of moles, and molar mass of hydrogen is 2.02
The total quantity of hydrogen to produce 1 kWh of g (approx.), the mass (of 52.73 moles of hydrogen) =
electricity is estimated in following two steps: One 52.73 * 2.02 = 106.5 g (rounded to the nearest tenth).
kg of hydrogen carries 33.33 kWh of energy and the Thus, 1 l of urine produces 106.5 g of hydrogen.
efficiency of the PEM cell is 60%. As discussed by Therefore, 50 g of hydrogen, consequently, would
O. Bilgin et al. (2015), many procedures are imple- require 0.470 l or 470 ml of urine, (say nearly half-
mented in hydrogen calculations. a-liter). Hence, to conclude, in order to get 1 KWh of
Hence, to generate 1 KWh of electricity, we energy as the output from the system, we need 1.875
need 0.05 kg or 50 g of hydrogen gas as shown in kWh and 450 ml of urine as input. The efficiency of
Table 4.2. The quantity of hydrogen is 50 g. Energy the system is 53.33%, which is higher than other
and urine required for producing 50 g of hydrogen, sources of power generation. Besides, 1 kg of hydro-
referring to the information in Table 4.1, we see that gen when used in fuel cell could drive vehicles up to
37.5 kWh of energy is required for the production 97–100 km (Oldenbroek et al., 2020). For a vehicle
of 1 kg of hydrogen. Hence, to generate hydrogen of running petrol or diesel, to cover the same distance,
50 g, it would need 1.8 kWh of input energy. The considering an ideal mileage 22 km/l, would consume
quantity of urine to produce 50 g of hydrogen is as around 4–4.5 l fossil fuel, This @ INR 72/l would cost
32 Production of electricity from urine

about INR 288–324. This would be INR100 more this landscape is a constraint that requires dem-
than the price for 1 kg of hydrogen. Hence, compared onstrating clear advantages.
to other sources of energy or of hydrogen itself, this • Energy return on investment (EROI): Evaluat-
PEM fuel cell process (I. Schimidhalter et al., 2021; ing the energy return on investment, considering
K. Ondrejicka et al., 2022) is economical and is made all energy inputs and outputs, is a constraint in
feasible on mass scale. When supplemented with determining the practicality and sustainability of
other renewable sources of energy (A. U. Rehman the urine-to-electricity system. A positive EROI is
et al., 2017; M. Sandhu et al., 2022; Rehman et necessary for long-term viability.
al., 2022), the system could become self-sustaining,
thereby decrease the dependency on paid sources of VI. Conclusion
electricity. Such system would save a considerable
amount of money. The article demonstrated the concept of producing
electricity from urine. Few comparisons theoreti-
cally prove the process to be not only plausible but
V. Challenges, pitfalls and constraints
economical too. With further research for making
Challenges this process commercially feasible, it would create a
• Technological feasibility: Implementing the pro- significant impact on the present and future demand
posed urine-to-electricity process efficiently and and supply scenarios of energy, paving ways for new
cost-effectively on a large scale is a complex chal- advancements in the field of energy and automobile. In
lenge, as laboratory conditions may differ signifi- a country like India where the majority of the popula-
cantly from real-world applications. tion face the perils of vehicle pollution, where most of
• Safety concerns: Handling and storing flammable the group housings and industrial setups use fossil fuel
hydrogen gas safely is a crucial challenge. Ad- electricity backups, this development has the potential
equate safety measures, such as pressure relief of providing a clean-energy and also counter the men-
valves and leak detection systems, must be in ace of human wastes. Sooner or later, mankind will
place to mitigate potential risks. approach times in near future, when the fossil fuels
• Economic viability: While the paper suggests eco- of the world would get exhausted thereby demand-
nomic feasibility, the true cost-effectiveness of ing a new source of energy for domestic, industrial
the process depends on factors like infrastructure and transportation uses. This proposed development
costs, energy efficiency, and market dynamics. A could be a proactive step in that direction. Last but
comprehensive economic analysis is essential to not least, we look forward to contributing something
address this challenge. very useful out of something considered useless. After
all, there is no such thing as waste in this ecosystem.
Pitfalls
• Long-term durability: Ensuring that the electro- References
lytic cells, PEM fuel cells, and other components
can withstand continuous operation over an ex- Sataksig. (2017). Eureka Green, (When will the Earth run
tended period is a potential pitfall. Unexpected out of Fossil Fuel) https://earthbuddies.net/when-will-
we-run-out-of-fossil-fuel/ [Online Resource].
wear and tear could affect the system’s reliability.
Adithya, B., James, H., and Charles, D. (2020). The clean
• Public acceptance: Convincing the public to em- energy imperative. Renew. Energy Fin., 2, 14–15.
brace the idea of using urine for electricity genera- Ajiboye, T. O., Ogunbiyi, O. D., Omotola, E. O., Adeyemi, W.
tion can be challenging due to the societal stigma J., Agboola, O. O., and Onwudiwe, D. C. (2022). Urine:
associated with waste materials. Overcoming this Useless or useful waste?, Results Engg., 16, 100522,
psychological barrier and fostering acceptance is a https://doi.org/10.1016/j.rineng.2022.100522.
potential pitfall in the adoption of the technology. Yeasmin, S., Ammanath, G., Onder, A., Yan, E., Yildiz, U.
H., Palaniappan, A., and Liedberg, B. (2022). Cur-
Constraints rent trends and challenges in point-of-care urinalysis
• Urine collection and processing: Building the nec- of biomarkers in trace amounts. TrAC Trend Anal.
essary infrastructure for collecting, transporting, Chem., 157, 116786.
Kumar, Kaushik, Zindani, Divya, and Davim. (2018). Elec-
and processing urine is a significant constraint. It
trochemical process advanced machining and manu-
requires substantial investment and adherence to facturing processes. Springer International Publishing,
sanitation and hygiene standards. 1, 105–122.
• Competition with existing technologies: The pro- Luther, A. K., Desloover, J., Fennell, D. E., and Rabaey, K.
posed urine-based energy generation system must (2015). Electrochemically driven extraction and re-
contend with established clean energy technolo- covery of ammonia from human urine. Water Res.,
gies, such as solar and wind power. Competing in 87, 367–377.
 Applied Data Science and Smart Systems 33
Li, J., Zhang, J., and Yang, J.-H. (2022). Research progress Bilgin, O. (2015). Evaluation of hydrogen energy produc-
and applications of nickel-based catalysts for electro- tion of mining waste waters and pools. Int. Conf.
oxidation of urea. Int. J. Hyd. Energy, 47(12), 7693– Renew. Energy Res. Appl. (ICRERA), Palermo, Italy,
7712. https://doi.org/10.1016/j.ijhydene.2021.12.099. 557–561. doi: 10.1109/ICRERA.2015.7418475.
Sanjay K. K., Harichandan, S., and Roy, B. (2022). Biblio- Vincent, O., Smink, G., Salet, T., and van Wijk, Ad J. M.
metric analysis of the research on hydrogen economy: (2020). Fuel cell electric vehicle as a power plant:
An analysis of current findings and roadmap ahead. Techno-economic scenario analysis of a renewable in-
Int. J. Hyd. Energy, 47(20), 10803–10824. tegrated transportation and energy system for smart
Gera, Tanya, Jaiteg Singh, Abolfazl Mehbodniya, Julian L. cities in two climates. Appl. Sci. 10(1), 143. https://doi.
Webber, Mohammad Shabaz, and Deepak Thakur. org/10.3390/app10010143.
(2021). Dominant feature selection and machine Ondrejička, K., Putala, R., and Mikle, D. (2022). Fuel
learning-based hybrid approach to analyze android cells as backup power supply for production pro-
ransomware. Security and Communication Networks. cesses. Cybernet. Informat. (K&I), 1–6. doi: 10.1109/
vol. 2021, Article ID 7035233, 22 pages, 2021. https:// KI55792.2022.9925936.
doi.org/10.1155/2021/7035233 Schmidhalter, I., Aguirre, P. A., and Eva Aimo, C. (2021).
Grigoriev, S. A., Porembsky, V. I., Fateev, V. N. (2006). Pure Phenomenological modeling and optimization of the
hydrogen production by PEM electrolysis for hydro- sizing and operation of a proton exchange membrane
gen energy. Int. J. Hyd. Energy, 171–175. fuel cell. XIX Workshop Inform. Proc. Con. (RPIC),
Henrietta, L. W., Engelbrecht, N., Modisha, P. M., and 1–6. doi: 10.1109/RPIC53795.2021.9648520.
Bessarabov, D. (2022). Electrochemical power sourc- Sandhu, M. and Thakur, T. (2022). Harmonic reduc-
es: Fundamentals, systems, and applications. Hyd. tion in a microgrid using modified asymmetrical in-
Storage Elsevier, 455–486. verter for hybrid renewable applications. IEEE Int.
Taner, T. (2018). Introductory chapter: An overview of PEM Conf. Power Elec. Smart Grid Renew. Energy (PES-
fuel cell technology proton exchange membrane fuel GRE), Trivandrum, India, 1–6. doi: 10.1109/PES-
cell. Intech Open, 1–5. GRE52268.2022.9715962.
Wang, Y., Pang, Y., Xu, H., Martinez, A., and Chen, K. S. Singh, Jaiteg, Saravjeet Singh, Sukhjit Singh, and Hardeep
(2022). PEM fuel cell and electrolysis cell technologies Singh. (2019). Evaluating the performance of map
and hydrogen infrastructure development – A review. matching algorithms for navigation systems: an em-
Energy Environ. Sci., 6, 1–8. pirical study. Spatial Information Research. 27: 63–
Atasay, N., Atmanli, A., and Yilmaz, N. (2023). Liq- 74. doi: https://doi.org/10.1007/s41324-018-0214-y
uid cooling flow field design and thermal analysis Rehman, A. U., Zeb, S., Khan, H. U., Shah, S. S. U., and
of proton exchange membrane fuel cells for space Ullah, A. (2017). Design and operation of microgrid
applications. Int. J Energy Res., 16. https://doi. with renewable energy sources and energy storage sys-
org/10.1155/2023/7533993. tem: A case study. IEEE 3rd Int. Conf. Engg. Tech-
Singla, M. K., Nijhawan, P., and Oberoi, A. S. (2021). Hy- nol. Soc. Sci. (ICETSS), Bangkok, Thailand, 1–6. doi:
drogen fuel and fuel cell technology for cleaner future: 10.1109/ICETSS.2017.8324151.
A review. Environ. Sci. Pollut. Res., 28, 15607–15626. Abdul, R., Radulescu, M., Cismas,, L. M., Cismas,, C.-M.,
https://doi.org/10.1007/s11356-020-12231-8. Chandio, A. A., and Simoni, S. (2022). Renewable en-
Bharati, P., Narayan, K., and Ramachandra Rao, B. (2022). ergy, urbanization, fossil fuel consumption, and eco-
Green hydrogen production by water electrolysis: A nomic growth dilemma in Romania: Examining the
renewable energy perspective. Mat. Today: Proc., 67, short- and long-term impact. Energies, 15(19), 7180.
1310–1314. https://doi.org/10.3390/en15197180.
5 Deep learning-based finger vein recognition and security:
A review
Manpreet Kaura, Amandeep Verma and Puneet Jai Kaur
Information Technology, UIET, Punjab University, Chandigarh, India

Abstract
The recognition system implies development that passes through the various stages. The finger vein recognition (FVR) is the
lead over the other biological modalities like finger print, face iris, etc., this paper reviews the worked done in the area of
FVR. The pre-processing is needed to enhance the images for better results. The feature extraction module provides the col-
lection of the best features in the finger vein images, which is used for template generation. The template-based schemes are
in fact the best and most appropriate for security purpose because it only preserved the scrambled information rather than
the original features of the human beings. According to this review the convolutional neural network (CNN)-based models
are best for FVR but still there have been some challenges faced by it so those would be improved in the experimental work
of this review.

Keywords: FVR, CNN, template, security, datasets

I. Introduction these modalities are captured in the visible light. It

has the unique characteristic pattern available on the
With the digitization in every field, the biometric
outer skin so that, it is easy for the intruders to forge
identification and authentication after the password-
and duplicate or collect the patterns from the sensor
based security mechanism, is the one of the most
surface with the help of silicon (Sarkar and Singh,
widely used techniques for the human authentication
2020). In the information technology field, there is
on the digital system. The various biometric modali-
need to maintain the balance among the security, reli-
ties are used to recognize the human beings based
ability and cost of the digital systems. While using the
on the physiological and behavioral characteristics
iris biometric, becomes a higher capability sensor is
(Shaheed et al., 2021) such as fingerprint (Jacob et
needed to capture the inner part of the human eye
al., 2021), face (Malik et al., 2018), hand (Park and
(Linsangan et al., 2019, Agarwal and Jalal, 2021).
Kim, 2013), iris (Czajka, Bowyer, and Flynn, 2019),
This becomes a very costly and time-consuming pro-
signature (Tanwar, Obaidat, and Tyagi, 2019), finger
cess. So, there is need to maintain the quality of every
vein (Z. Liu et al., 2010), walking pattern (Casale,
aspect related to the digital system development. The
Pujol, and Radeva, 2012), brain waves (Campisi et
comparison is shown in the Table 5.1.
al., 2014), etc. The biometric technology has gained
the popularity in recent years due to unique feature
of every human being, even a twin has the different II. Finger vein recognition
biological patterns. The traditional identification and The biometric recognition is divided into two catego-
authentication methods are token- and password- ries: (a) Extrinsic – This is based on the visibility of
based and have the various security risks while trans- the biological information. In this the information is
ferring the information over communication channels easily available to the capturing devices that include
or store the data on the server, the information related faces, fingerprints, iris etc. (b) Intrinsic – This is less
to the human assets in digital form is always at major easily visible to the capturing devices than the extrin-
risks of being leaked out. With the development of sic. The special devices are used to collect the inner
the biological system, the risk of the loss of the digi- patterns of the vessels like palm vein, finger vein, etc.
tal information has gradually increased in the recent The recognition is the process of identifying some-
years. But there is dire need to maintain the security thing, which is available in the database through the
of the biological information of an individual it is of capturing devices. This information had previously
great importance because if the system recognizes been stored (Jain and Kumar, 2012).
an individual with his/her fingerprint then there are The first FVR system was introduced by the Japanese
only ten options to identify his/her at the terminals. scientist for medical purpose in 2002 (https://cir.nii.
Various security risks are from all physiological ac.jp/articles). After that, the Japan-based company
modalities, namely fingerprint, face, hand, because Hitachi had manufactured the devices the FVR for the

[email protected]
a
 Applied Data Science and Smart Systems 35
Table 5.1 Comparison of different biometric modalities

Modality Data type Devices Contact Cost Security

Face Images Camera N-Contact Low Normal

Finger Print Images Optical Sensors Contact Low Good
Finger Vein Images NIR Sensors N-Contact Low Superior
Iris Images Special Camera N-Contact High Normal
Signature Images Electronic Tablet& Pad N-Contact Low Normal

layers of network which is able to provide the higher

accuracy during the recognition process. Deep learn-
ing is a successor of machine learning which includes
multiple layers with learning algorithms. This allows
the deep learning method to learn hierarchical feature
from the data. Therefore, deep learning has substi-
tuted the conventional feature extraction approach in
several areas involving speech, computer vision, and
natural language processing (Yin, Zhang, and Liu,
2021). The pre-trained models provide better accu-
Figure 5.1 (a) NIR sensor, (b) Finger vein pattern racy than the conventional methods so the researchers
use these deep learning models in the biometric areas.
The researchers have brought deep learning in to the
purpose of identification and authentication. This is biometric field, Due to its strong capability with fea-
much more popular in Japan because it preserves very ture representation. The new network models and the
private information of the person which is different in pre-trained models are developed/tested on different
every human being (Uhl, 2020). datasets.
From the perspectives of security, finger vein is The deep learning methods were applied the finger
far better than all other physiological modalities veins to recognize the finger vein. But it did not pro-
(Shaheed, 2018) (Table 5.1). The FVR rely the vascu- vide good results because the system was trained and
lar pattern which is situated inside the skin. Due to the tested over the non-publicly available datasets (Radzi
inside pattern, the forgery of the blood vessels is more and Bakhteri, 2016; R. Arya et. al., 2021). While the
difficult than of other which is the unique in every features were extracted from the raw images, some
human being. The NIR sensors are used to capture information gets lost due to various factors. So the
the inner information related to the vessels of human author improved the robustness by applying the fully
body. This is a big benefit of using the finger vein as an connected CNN model to recover the lost pattern
identification and authentication because the vessel of the finger vein and to achieve the good verifica-
pattern can be collected only when the finger is under tion performance (Zoph and Le, 2017). The CNN
NIR lights and only detected while the blood flows in models are complex and not a light weight as other
the human body. So the risk of the duplicity is mini- techniques. The researchers have developed the light
mal as compared to other modalities (Hong, Lee, and weight models for vein recognition (Hong, Lee, and
Park, 2017; Al-khafaji and Al-tamimi, 2022). Figure Park, 2017). The traditional methods of image seg-
5.1 (a) has shown the finger under NIR sensor and mentation have been combined with the CNN model
collect the information of finger vein. Figure 5.1 (b) to improve the recognition performance. The labels
has shown the extracted feature of finger vein for fur- for the vein pixel information were automatically
ther process. The collected good features are sent to assigned by the system. The CNN model was trained
the template generation module for provides the good and tested to predict the pixel information for the
security to the finger vein information. Then, the vein finger vein. The special module of CNN was devel-
formation is verified for genuine users. oped to find out the missing pixels in the segmented
images (IEEE and IEEE, 2017). CNN and supervised
discrete hashing finger vein identification process has
III. Literature survey of finger vein
proposed to overcome the problem, which decrease
Vein pattern recognition based on deep learning the template size up to 250 bytes and also surges the
The deep learning models are based on the large data- matching speed (P. Sharma et al., 2018). Outmoded
set processing. The tasks are hidden in number of finger vein identification methods can be cracked by
36 Deep learning-based finger vein recognition and security: A review

hackers because the template used by the scheme is of sparse because all the training samples are based
in the form of plain data (Qin and El-yacoubi, 2017). on dictionary matrix. Thus, need was felt to optimize
The author proposed the FVR-DLRP to secure revo- the selection of dictionary data to improve the system
cable and to do efficient finger vein template genera- performance (Fang et al., 2022). Finger vein recog-
tion. Most of the traditional finger vein recognition nition was done with the use of oval PDCNN, this
systems have a shading and misalignment of finger was the advanced version of the PDKs. It provided the
vein problem. Need to pay the much more effort best performance, the first ten layers were from the
and time for extracting the features from the images MobileNet and all other layers from the SqueezeNet
which is a complicated and complex process in deep network could be compressed the network to achieve
CNN (Y. Liu et al., 2018). To improve these problems the higher performance (Li et al., 2023).
the researchers had proposed a robust CNN model The CNN models have improved the recognition
which had the error rate of 0.396 it was collected on performance for FVR. But this advancement has
a good quality dataset (Hong, Lee, and Park, 2017). needed to improve the feature extraction module and
All the publicly available datasets for finger vein have defense against security attacks, and the CNN models
the small images collection. are not lightweights.
Although, the CNN model used for the finger vein,
achieved higher accuracy yet it face the problem Security of FVR system
related to the training process (W. Liu et al., 2017). In the present time, the security of each system is at
The CNN model is successfully applied for the fin- major risk of losing the information and digital assets
ger vein identification process (Simonyan, 2019). The which are protected using the several security mecha-
light weight convolutional neural network (CNN) nisms. Certain things are to be kept in mind before
model was proposed to improve the small training choosing the bio-information in any application to
dataset problem by using the similarity measure net- enhance the security. First of all, it should be clear
work (Qin and El-yacoubi, 2018). The dense net was that, all the phases of technologies which are to be
proposed for finger vein recognition (FVR). This was used in the application must be safe in terms of data
used it remove the noise in the images and the two privacy and protection. Although biometric technol-
or more images were combined for recognition sys- ogy yet it faces various security attacks is considered
tem and also used for feature extraction. The system to be a secure system of real-world market.
has huge computational cost (Song, Kim, and Park,
2019). Although the all-available methods were tested Template protection for finger vein
on publicly available datasets but still these systems The template protection is the process of generating
had failed in the practical uses. Depth based separate the precise or related information from the images by
CNN model was developed to overcome this prob- feature extraction process (Kumar, 2019). This is the
lem. The system is simple but it still has weakness of unique precise information associated with the differ-
recognizing the less defined features of the images ent human beings. The template scheme is divided into
(Tang et al., 2019). The CNN-CO was based on the two categories: (a) Bio-cryptosystem (Kaur, Kumar,
local descriptor for pre-training the ImageNet model. and Singh, 2023). This combines the best feature of
Practically this system was best out of all the con- both the worlds i.e., cryptographic keying methods
ventional systems (Y. U. Lu, 2019). Finger vein-based and biometric schemes and (b) cancellable templates
authentication model was proposed by developing the (Manisha, 2020). In this field, the biometric template
lightweight Siamese network. When images were col- of a person is distorted in such a manner that the
lected the feature information got lost. The GCNet original data is not available to the intruder but still
and multi-scale feature was used during the process to identity recognition can be performed. The template
solve the faced problem. The system has been tested generation in the CNN models for higher accuracy for
over three publicly available datasets but the need FVR can be performed (Yin, Zhang, and Liu, 2021).
was felt to improve the feature extraction module by The template generation is shown in Figure 5.2.
changing the width and cardinality of the CNN net- The security to the templates of finger vein to provide
work (Fang, Ma, and Li, 2023). The parameters for
training and testing the network for data happen to
be complex due to the complex hidden structure of
the neural network. Although the result was 99.98%
but we need a light weight model for FVR (Wang and
Shi, 2022). The double-weighted group sparse rep-
resentation classification was developed to solve the
FVR. It had the lower accuracy than the other mod-
els and also took a long time to solve the coefficient Figure 5.2 Template generation
 Applied Data Science and Smart Systems 37

the password based key derivation function has been state of the art does address various issues related to
used to derive the key named FVR-DLRP. To provide the finger vein pattern recognition for artificial neural
the information still remains even though the pass- network. The question arises as who can suggest good
word is cracked. But this system has lower accuracy matching for prob and gallery images for recognition
in terms of FAR rate, which leads to FAR attack and process and has also been advised to prepare and
it even has lower GAR rate (Y. Liu et al., 2018). generate the light weight model for finger vein (Yin,
Deep CNN with hard mining finger verification Zhang, and Liu, 2021).
scheme was proposed which achieved better perfor-
mance than achieved through commercial finger vein Security attacks
verification systems. This method also accelerates the On the other side, the restricted system is responsible
complete training process. The huge template size for security attacks. The attackers generate fake bio-
requires enormous amount of storage space (Huang metric template and modify it at different levels, the
et al., 2017). The Gabor filter was used, built a fin- finger vein faces various security attacks related prob-
ger vein authentication system based on the light- lems, from time-to-time various security methods and
weight CNN and supervised discrete hashing so as their patches for breaches (Tome and Vanoni, 2014)
to improve the finger vein images. Despite all these are introduced to amend the system.
steps, this method has decreased the template size Transferable deep convolutional network was
and surges the performance of finger vein verification. proposed to handle the presentation attack. The
The connection between training time and recogni- researcher has modified the system by adding seven
tion outcome was not thoroughly measured (Xie and layers to the existing Alex-Net to overcome the over
Kumar, 2019). The fusion based system was devel- fitting problem. The artifacts for the finger vein were
oped for fingerprint and the finger vein biological generated by using two different printers. The modi-
datasets. The feature level fusion was applied to this fied system is able to handle the PAD for finger vein.
system. The system has the higher matching perfor- The transferable deep learning neural network for the
mance and security of the data (Yang et al., 2018). finger vein has still to face video presentation attack
The BDD-based FVR system was based on the deep (Raghavendra et al., 2017). Another mechanism for
CNN. The system was combined with ML-ELM to security of the data template was developed but it is
form a FVR system that provides the protected pri- still facing security breaches during different process
vacy, and also provides the security to the template. like storing the template and matching the templates
In case of tempering by the intruders the template for finger vein. The FVR with the template-based pro-
performs the undoing operation independently. Then tection is able to handle the presentation attack but
the new template version is generated by user specific still has the problem related to the adversarial attack
keys (Yang et al., 2019). The weighted least squares (Ren et al., 2021). The survey provided by the Yimin
regression has been used to improve the template gen- Yin et al. (Yin, Zhang, and Liu, 2021), had suggested
eration. It minimizes the verification errors but this and elaborated all the security breaches in finger vein
was based on an assumption, but the template has CNN methods. The system has developed to handle
the very little distance in the intra-class for the same the impersonation attack and check the system for
image data (Qin, 2019). The cancelable biometric- authentication with minimum enrolment time. The
based scheme for CIRF, and proposed a low-rank MC-CLAHE method is used to process the images
approximation-based cancelable indexing scheme of finger vein prior to the CNN training process but
which was based on CIRF was introduced to solve the the system is only providing security to the database
problem of excessive computational overhead. Low- (Safie, Zarina, and Khalid, 2023). Various templates-
rank approximation of biometric images was used based CNN models are developed. It has been pro-
to speed up the calculation of CIRF and also used vided the moderate defense against security attacks.
the minimum spanning tree representation for low- But the security problem still remains exist in the
rank matrices in the Fourier domain. The researcher template based FVR systems. Table 5.3 has shown the
proved the reliability of the projected method in pro- most recently articles related to the CNN, Template
tecting related biological information (Murakami et and security of FVR.
al., 2019). The template protection scheme has been
proposed to align the images. The IoM hash is used to IV. Datasets
realize the required privacy and security for the FVR
(Kirchgasser et al., 2020). The security is provided Various datasets (Y. Lu et al., 2013) are always freely
to the template to solve the issue of the presentation available to the researchers to train and test the pre-
attack by CNN model but templates still faces adver- trained models for the CNN. The models learn the
sarial sample attack and does not have light weight features from the images in the dataset to train the sys-
feature extraction module (Ren et al., 2021). The tem, then the system based on this training recognizes
38 Deep learning-based finger vein recognition and security: A review

the person on the basis of pre-stored trained data of CIR

that person. The datasets are the most essential part The CIR is the parameter to measure the false color
of the biometric recognition system development. The photograph called color infrared that shows the
most commonly used datasets for the finger vein have reflected electromagnetic wave form an object. It is
been listed in the Table 5.2. used to check the security flaws in the biometric sys-
tems (Ren et al., 2021).
V. Parameters of performance measurement
VI. Analysis and discussion
The performance of any system directly depends on
its error rate, which occurs at various levels of the In this survey of biometric system, we have been
system development. The error rate is calculated for found out that, the FVR system is best over the other
each activity during the whole process. The accuracy biometric modalities (Table 5.1). The survey has pro-
of the system, mainly depends on the performance: vided related to the CNN models and template-based
capturing a quality image, short time interval between security and attacks.
the enrolment and verification phase, robustness of
the recognition system, and the environmental factors • Most of the researchers has directly apply the
are temperature, humidity, illumination conditions of CNN pre-trained models for the FVR, no chang-
the sensor surface (Unar, Chaw, and Abbasi, 2014). es did make to the CNN models, due to this the
To evaluate the performance of the biometric system, extracted features from the vein images have the
all criteria have been defined in the standard way in poor information for human recognition. while
the series of ISO/IEC 19795 (Draft et al., 2006). The recognize the person, the system has denied for
accuracy has been defined in terms of error rate as: the same user who is enrolled earlier, which is
the greatest issue related to the feature extraction
Confusion matrix module. So, need to provide the best feature ex-
This is the calculated formula in the Table form for traction method by altering the pre trained net-
the classification algorithm. In this the actual classes work models. Some adjustments should be made
for training have been defined at the y-axis of the table for better results.
and the predicted classes have been defined at the • One of another issue related to the feature extrac-
x-axis of the table. When any one to train the CNN tion module is the mismatching of the gallery and
model for two classes, the 2×2 confusion matrix is the probe images. This problem is addressed by
generated. From this matrix, the other values can be only few researchers according to the FVR survey
calculated (Muthusamy and Rakkimuthu, 2022). A in this review. The image has stored at the time of
confusion matrix for analysis of result is shown in identification is not matches with the image while
Figure 5.3. The FAR and FRR are defined as: recognition. So those need to apply the feature ex-
traction module effectively to solve this problem.
• Most of the FVR solutions are based on the pre-
trained models which has template protection
for biological information. But it is still facing
the various issues related to the accuracy and se-
curity. The CNN-based template generation has

Table 5.2 Datasets for the finger vein.

Datasets No. of images No. of related to each person Resolution

FV-USM 5904 4 (both M, I) 640×480

HKPU 6264 2 (left M, R) 513×256
IDIAP 440 4 665×250
MMCBNU-6000 6000 6 (both M, I, R) 640×480
SCUT 10800 6 (both M, I, R) 640×480
SDUMLA-HMT 3816 6 (both M, I, R) 320×240
THU-FVFDT1, 2 440, 2440 1 (left I) for each 200×100 for both
UTFV 1440 6 (both M, I, R) 672×380
M=Middle finger, I=Index finger, R=Ring finger
 Applied Data Science and Smart Systems 39

vides the security only to the database which is

stored on the server.
• The pre-trained model of CNN is complex in na-
ture and difficult to modify. Most of the research-
ers addressed the issue related to the models that
are not light weight, in the terms of parameters,
filters and functionality. This is the interesting re-
search area of AI for biometric recognition.
• The quality of images becomes degraded, due to
the posture of the image while capturing, illumi-
Figure 5.3 Confusion matrix
nation of the light from the sensors (NIR). Only
the few researchers have performed the experi-
ment on the FV-USM dataset. It has the low-qual-
been provided by the few researchers, but it yet ity images which are suitable to build the more
has flaws, this is the major research area in FVR. robust system. The researcher had performed the
• When the attack of the adversarial sample is ig- experimental work on this dataset but system still
nored, the system produces the incorrect recogni- faces the security attack, when the samples (ad-
tion. Ren et al. (2021) has provided the cancel- versarial) are ignored.
able template-based protection to the FVR with • Although, the all-available CNN based FVR has
CNN. This system has solved the issue related to the higher accuracy but most of the systems is still
the presentation attacks. But some of the CNN facing the issues related to the security, template
based FVR has still faces the video presentation generation, FAR, GAR and accuracy of recogni-
attack problem. Some of the researchers has pro- tion.

Table 5.3 Recent articles related to the CNN and template security of FVR

Method Datasets Performance Improvements/ Journal Ref.

parameters need to improve

Deep CNN DS1, DS2 & DS3 EER - DS1=0.42%, Improvement IEEE International (Huang et
DS2= 1.41% & DS3= of vein pattern Conference al., 2017)
2.14% matching on Identity,
Security and
Behavior Analysis
(IEEE-2017)
Deep learning FV_NET64 GAR=91.2% Enhancement the Soft Computing (Y. Liu et
& random FAR=0.3% revocability of the (Springer-2018) al., 2018)
projection template
EP-DFT FVC2002 DB2, EER=0.45% Enhancement of Pattern Recognition (Yang et al.,
FVC2004 DB2, non-revocability of (Elesvier-2018) 2018)
FV-HMTD templates
CNN and Two session EER=0.0887 Reduced the Pattern (Xie and
supervised databases template size Recognition Letters Kumar,
discrete hashing (Elesvier-2019) 2019)
BDD-ML-ELM SDUMLA, CIR=93.09%, 98.70%, New non invertible IEEE Transactions (Yang et al.,
MMCBNU_6000 98.61%, respectively templates for on Industrial 2019)
& UTFVP (datasets) finger vein Informatics
(IEEE-2019)
Weighted HKPU & FV-USM EER=1.28, 1.43, Improvement in MDPI (Qin, 2019)
least square respectively (datasets) enrolment template (Information-2019)
regression for finger vein
Correlation- N Genuine template Fast and secure Pattern (Murakami
invariant identification=164.7 biometric Recognition Letters et al., 2019)
random identification 7 (Elesvier-2019)
filtering No leakage of
information of the
template
40 Deep learning-based finger vein recognition and security: A review

Method Datasets Performance Improvements/ Journal Ref.

parameters need to improve

Self-attention FV-USM, Recall is best over Need to improve Infrared Physic (Fang, Ma,
mechanism MMCBNU_6000, SDUMLA= 0.9944 the feature and Technology and Li,
(SAC) Siamese SDUMLA-HMT F1 Score is best over extraction module (Elsevier Dec-2020) 2023)
network FV_USM=0.9925 by changing
EER is best over the width and
MMCBNU_6000= cardinality of the
0.0012 network models
RSA for SDMLA, Scheme A is best over The feature Knowledge Based (Ren et al.,
template MMCBNU_6000, HKPU=99.03% extraction method System (Elsevier 2021)
protection HKPU, FV-USM and is not lightweight May-2021)
using CNN Scheme B is best over and the system is
FV_USM=99.18% not able to handle
the adversarial
sample attack and
presentation attack
Survey on SDUMLA-HMT, Performance summary Need to generate the Computer Vision (Yin,
ANN for FV-USM, HKPU, of all CNN methods light weight models, and Pattern Zhang, and
finger vein MMCBNU_6000, for finger vein solve the problem Recognition Liu, 2021)
UTFVP, THU- of mismatching (Springer
FVFDT, SCUT, of gallery and Aug-2022)
IDIAP prob image,
dynamic finger vein
extraction
Double PolyU [], FV-USM, The system has best The sparse International (Fang et al.,
weighted SDUML-HMT performance over coefficient takes the Journal of Machine 2022)
group sparse FV_USM=97.0-95.88- long time, so that Learning and
representation 88.72% over three need to improve it Cybernetics
classification different variants of and improve the (Springer
models accuracy of the May-2022)
system
Multimodal CASIA-WebFace, 99.98% Model should be Sensors (MDPI (Wang and
approach SDUMLA-FV, lightweight Aug-2022) Shi, 2022)
based on CNN FV-USM
(RESNET,
AlexNet,
VGG-19)
MC-CLAHE FV-USM AUC=0.78–0.91 Security only International (Safie,
(CNN- provided to the Journal of Online Zarina,
AlexNet) database and Biomedical and Khalid,
Engineering (iJOE 2023)
2023)
N=Data not available.

VII. Conclusion related issues. Most commonly used datasets have

been explored for FVR. At the end the analysis and
Finger vein recognition methods have been explored discussion has been provided based on the articles
in the present review based on the deep learning added in this review. In future, the system will
methods for template generation which provides develop in such a way for which the feature extrac-
more security than other traditional methods for tion module would be appropriate for template
biometric information. The template generation generation and the system would not face the issue
directly depends on the feature extraction module. related to the security attacks. Deep learning models
So, there is need to improve the feature extraction will be used for it.
module and apply it effectively to the template
generation module. The feature extraction related
problem has also been addressed. The templates References
have also been checked for security attacks based Rohit, A., and Jalal, A. S. (2021). Presentation attack detec-
on the efficiency of the feature of the biometric tion system for fake iris: A review. Multimedia Tools
information but the FVR still faces various security Appl., 80, 15193–15214.
 Applied Data Science and Smart Systems 41
Al-khafaji, Ruaa S, S. and Al-tamimi, M. S. H. (2022). Vein of the 2nd International Conference on Communi-
biometric recognition methods and systems: A review. cations and Cyber Physical Engineering. 405–411.
Adv. Sci. Technol. Res. J., 16(1), 36–46. Springer Singapore.
Patrizio, C. and La Rocca, D. (2014). Brain waves for Changyan, L., Dong, S., Li, W., and Zou, K. (2023). Fin-
automatic biometric-based user recognition. IEEE ger vein recognition based on oval parameter-depen-
Trans. Inform. Foren. Sec., 9(5), 782–800. https://doi. dent convolutional neural networks. Arabian J. Sci.
org/10.1109/TIFS.2014.2308640. Engg., 1–16. https://doi.org/10.1007/s13369-023-
Pierluigi, C., Pujol, O., and Radeva, P. (2012). Personal- 07818-5.
ization and user verification in wearable systems Linsangan, N. B., Flores, P. R., Poligratis, H. A. T., Victa,
using biometric walking patterns. Pers Ubiquit A. S., and Villaverde, J. (2019). Real-time iris recogni-
Comput., 563–580. https://doi.org/10.1007/s00779- tion system for non-ideal iris images. Proc. 2019 11th
011-0415-z. Int. Conf. Comp. Automat. Engg., 32–36. https://doi.
Adam, C., Bowyer, K. W., and Flynn, P. J. (2019). Domain- org/10.1145/3313991.3314002.
specific human-inspired binarized statistical image Wenjie, L., Li, W., Sun, L., Zhang, L., and Chen, P. (2017).
features for iris recognition. 2019 IEEE Winter Conf. Finger vein recognition based on deep learning. IEEE
Appl. Comp. Vision (WACV), 959–967. https://doi. Conf. Indust. Elec. Appl., 205–210.
org/10.1109/WACV.2019.00107. Yi, L., Ling, J., Liu, Z., Shen, J., and Gao, C. (2018). Fin-
Draft, Working, Committee Draft, Final Committee Draft, ger vein secure biometric template generation based
Final Draft, International Standard, International on deep learning. Soft Comput., 22(7), 2257–2265.
Standard, Information Tech-, Information Tech-, In- https://doi.org/10.1007/s00500-017-2487-9.
formation Tech-, and Information Tech. (2006). Bio- Zhi, L., Yin, Y., Wang, H., Song, S., and Li, Q. (2010).
metric standards – An update. Elesvier. Finger vein recognition with manifold learning. J.
Chunxin, F., Ma, H., and Li, J. (2023). A finger vein au- Netw. Comp. Appl., 33(3), 275–282. https://doi.
thentication method based on the lightweight Siamese org/10.1016/j.jnca.2009.12.006.
network with the self-attention mechanism. Infrared Lu, Y. U. (2019). Exploring competitive features using
Phy. Technol., 128, 104483. https://doi.org/10.1016/j. deep convolutional neural network for finger vein
infrared.2022.104483. recognition. IEEE Acc., 35113–35123. https://doi.
Chunxin, F., Ma, H., Yang, Z., and Tian, W. (2022). A fin- org/10.1109/ACCESS.2019.2902429.
ger vein recognition method based on double weight- Yu, L., Xie, S. J., Wang, Z., and Park, D, S. (2013). An
ed group sparse representation classification. Int. J. available database for the research of finger vein
Mac. Learn. Cybernet., 13(9), 2725–2744. https://doi. recognition. 2013 6th Int. Cong. Image Signal Proc.
org/10.1007/s13042-022-01558-y. (CISP 2013), 410–415. https://doi.org/10.1109/
Gil, H. H., Lee, M. B., and Park, K. R. (2017). Convolution- CISP.2013.6744030.
al neural network-based finger vein recognition using Fiqri, M., Azis, A., Nasrun, M., Setianingsih, C., and Mur-
NIR image sensors. Sensors, https://doi.org/10.3390/ ti, M. A. (2018). Face recognition in night day using
s17061297. method Eigenface. 2018 Int. Conf Signals Sys. (IC-
Houjun, H., Liu, S., Zheng, H., Ni, L., Zhang, Y., and Li, W. SigSys), 103–108. https://doi.org/10.1109/ICSIG-
(2017). Deep vein: Novel finger vein verification meth- SYS.2018.8372646.
ods based on deep convolutional neural networks. Manisha. (2020). Cancelable biometrics : A comprehensive
2017 IEEE Int. Conf. Iden. Sec. Behav. Anal. (ISBA), survey. Artif. Intel. Rev., 53(5), 3403–3446. https://
5, 1–8. https://doi.org/10.1109/ISBA.2017.7947683. doi.org/10.1007/s10462-019-09767-8.
IEEE, Student Member and Fellow IEEE. (2017). Efficient Takao, M., Ohki, T., Kaga, Y., Fujio, M., and Takahashi,
processing of deep neural networks : A tutorial and K. (2019). Cancelable indexing based on low-rank ap-
survey. Proc. IEEE, 105, 2295–2329. proximation of correlation-invariant random filtering
Jeena, J. I., Betty, P., Darney, P. E., Raja, S., and Robinson, for fast and secure biometric identification. Pattern
Y. H. (2021). Biometric template security using DNA Recogn. Lett., 126, 11–20. https://doi.org/10.1016/j.
codec based transformation. Multimedia Tools Appl., patrec.2018.04.005.
5, 7547–7566. Dharmalingam, M. and Rakkimuthu, P. (2022). Trilat-
Jain, A. K. and Kumar, A. (2012). Biometric Recognition : An eral filterative hermitian feature transformed deep
Overview. https://doi.org/10.1007/978-94-007-3892-8. perceptive fuzzy neural network for finger vein veri-
Prabhjot, K., Kumar, N., and Singh, M. (2023). Biometric fication. Exp. Sys. Appl., 196, 116678. https://doi.
cryptosystems: A comprehensive survey. Multimedia org/10.1016/j.eswa.2022.116678.
Tools Appl., 2022, 16635–16690. Gitae, P. and Kim, S. (2013). Hand biometric recognition
Simon, K., Kauba, C., Lai, Y., Zhe, J., and Uhl, A. (2020). based on fused hand geometry and vascular patterns.
Finger vein template protection based on alignment-ro- Sensors, 13, 2895–2910. https://doi.org/10.3390/
bust feature description and index-of-maximum hash- s130302895.
ing. IEEE Trans. Biomet. Behav. Ident. Sci., 2(4), 337– Huafeng, Q. (2019). A template generation and improve-
349. https://doi.org/10.1109/TBIOM.2020.2981673. ment approach for finger-vein recognition. Informa-
Gunjan, Vinit Kumar, Puja S. Prasad, and Saurabh Mukher- tion, 1, 1–19. https://doi.org/10.3390/info10040145.
jee. (2019). Biometric template protection scheme- Sharma, P. and Singh, J. (2018). Machine learning based ef-
cancelable biometrics. In ICCCE 2019: Proceedings fort estimation using standardization. 2018 Int. Conf.
42 Deep learning-based finger vein recognition and security: A review
Comput., Power Comm. Technol. (GUCON), doi: Jong, S., M. I. N., Kim, W. A. N., and Park, K. R. (2019).
10.1109/GUCON.2018.8674908. Finger-vein recognition based on deep denseNet using
Huafeng, Q. and El-yacoubi, M. A. (2017). Deep represen- composite image. IEEE Acc., 7, 66845–66863. https://
tation-based feature extraction and recovering for doi.org/10.1109/ACCESS.2019.2918503.
finger-vein verification. IEEE Trans. Inform. Foren. Su, T., Zhou, S., Kang, W., Wu, Q. and Deng, F. (2019).
Sec., 12(8), 1816–1829. https://doi.org/10.1109/ Finger vein verification using a Siamese CNN. IET
TIFS.2017.2689724. Biomet., 8, 306–135. https://doi.org/10.1049/iet-
Huafeng, Q. and El-Yacoubi, M. A. (2018). Deep represen- bmt.2018.5245.
tation for finger-vein image-quality assessment. IEEE Tanwar, Sudeep, Mohammad S. Obaidat, Sudhanshu Tyagi,
Trans. Cir. Sys. Video Technol., 28(8), 1677–1693. and Neeraj Kumar. (2019). Online signature-based
https://doi.org/10.1109/TCSVT.2017.2684826. biometric recognition. Biometric-based physical and
Ahmad, R. S. and Bakhteri, R. (2016). Finger-vein biomet- cybersecurity systems. 255–285. doi: https://doi.
ric identification using convolutional neural network. org/10.1007/978-3-319-98734-7_10
Turkish J. Elec. Engg. Comp. Sci., 24(3), 1863–1878. Pedro, T. and Vanoni, M. (2014). On the vulnerability of
https://doi.org/10.3906/elk-1311-43. finger vein recognition to spoofing. 2014 Int. Conf.
Raghavendra, R., Venkatesh, S., Raja, K. B., and Busch, Biomet. Special Interest Group (BIOSIG), 1–10. Ge-
C. (2017). Transferable deep convolutional neural sellschaft für Informatik e.V. - GI.
network features for fingervein presentation attack Andreas, U. (2020). Advances in Computer Vision and Pat-
detection. Proc. 2017 5th Int. Workshop Biomet. tern Recognition Handbook of Vascular Biometrics.
Foren., IWBF 2017, 1–5. https://doi.org/10.1109/ Unar, J. A., Chaw, W., and Abbasi, A. (2014). A review of
IWBF.2017.7935108. biometric technology along with trends and pros-
Arya, R., Singh, J., and Kumar, A. (2021). A survey of pects. Pattern Recogn., 47(8), 2673–2688. https://doi.
multidisciplinary domains contributing to affective org/10.1016/j.patcog.2014.01.016.
computing. Comp. Sci. Rev., 40, 100399. https://doi. Yang, W. and Shi, D. (2022). Convolutional neural network
org/10.1016/j.cosrev.2021.100399. approach based on multimodal biometric system with
Hengyi, R., Sun, L., Guo, J., Han, C., and Wu, F. (2021). fusion of face and finger vein features. Sensors, 1–15.
Finger vein recognition system with template protec- Cihui, X. and Kumar, A. (2019). Finger vein identification
tion based on convolutional neural network. Knowl. using convolutional neural network and supervised
Based Sys., 227, 107159. https://doi.org/10.1016/j. discrete hashing. Pattern Recogn. Lett., 119, 148–156.
knosys.2021.107159. https://doi.org/10.1016/j.patrec.2017.12.001.
Safie, S. I., Zarina, P., and Khalid, M. (2023). Practical Wencheng, Y., Wang, S., Hu, J., Zheng, G., and Valli, C.
consideration in using pre-trained convolutional neu- (2018). A fingerprint and finger-vein based cancelable
ral network (CNN) for finger vein biometric. Int. J. multi-biometric system. Pattern Recogn., 78, 242–251.
Biomet. Engg., 19(02), 163–175. https://doi.org/10.1016/j.patcog.2018.01.026.
Arpita, S. and Singh, B. K. (2020). A review on performance, Wencheng, Y., Wang, S., Hu, J., Zheng, G., Yang, J., and Val-
security and various biometric template protection li, C. (2019). Securing deep learning based edge finger
schemes for biometric authentication systems. Multi- vein biometrics with binary decision diagram. IEEE
media Tools Appl., 27721–27776. Trans. Indust. Inform., 15(7), 4244–4253. https://doi.
Kashif, S. (2018). A systematic review of finger vein recogni- org/10.1109/TII.2019.2900665.
tion techniques. Information. https://doi.org/10.3390/ Yin, Yimin, Renye Zhang, Pengfei Liu, Wanxia Deng, Sil-
info9090213. iang He, Chen Li, and Jinghua Zhang. (2022). Arti-
Kashif, S., Mao, A., Qureshi, I., Kumar, M., Abbas, Q., Ul- ficial neural networks for finger vein recognition: a
lah, I., and Zhang, X. (2021). A Systematic Review on survey. 1–83. arXiv preprint arXiv:2208.13341.
Physiological ‑ Based Biometric Recognition Systems : Barret, Z. and Le, Q. V. (2017). Neural architecture search
Current and Future Trends. Archives of Computa- with reinforcement learning. Int. Conf. Learn. Repre-
tional Methods in Engineering. Netherlands: Springer. sent., 1–16.
https://doi.org/10.1007/s11831-021-09560-3. https://cir.nii.ac.jp/articles
Karen, S. (2019). Darts: Differentiable architecture search. https://www.flexlines.be/showcases/nir-vein- detector
Int. Conf. Learn. Represent., 1–13. arXiv.
6 Development of an analytical model of drain current
for junctionless GAA MOSFET including source/drain
resistance
Amrita Kumari1,a, Jhuma Saha2, Ashish Saini1 and Amit Kumar1
1
Quantum School of Technology, Quantum University, Roorkee, India
2
Indian Institute of Technology, Gandhinagar, Gujarat, India

Abstract
Fabrication of devices in deca nanometer regime suffers from several limitations as the devices are being scaled so that the
speed and transistor density can be increased. This has led to a series of innovative techniques by the industry as well as aca-
demia. Depletion regions formed in association with the p-n junctions is one of the restrictive factors in scaling short channel
devices in case of junction-based (JB) metal-oxide-semiconductor field-effect transistors (MOSFETs). This has led to several
short channel effects (SCEs). Recently, novel MOSFET structures have been developed that are devoid of p-n junctions and
have also been successfully fabricated. These devices are named “junctionless transistors (JLTs)”. MOSFETs employing
gate-all-around (GAA) architecture have been reported as an ultimate structure in silicon integrated circuits (ICs). In this
paper, we have developed an analytical drain current model for short channel GAA JLT, including source (S)/drain (D) series
resistance, which is also one of the important parameters when devices with short channel are fabricated. We have obtained
the potential distribution profile using Poisson’s equation. It was then used for obtaining the model for drain current. The
validation of the model has been obtained with both the simulation as well as experimental results. We have further analyzed
the effect of S/D resistance on the drain current for different device parameters.

Keywords: GAA, junctionless, short channel

I. Introduction effectively controls the electrostatic potential inside

the channel is effectively under the control of the gate.
Limitations imposed on the fabrication techniques of
Incorporating GAA in JL devices can further enhance
devices as they are scaled in the nanometer regime,
the device characteristics (Colinge et al., 2010; Duarte
in accord with Moore’s law, led to the innovation of
et al., 2011; Yu, 2014). The current research therefore
alternative device structures. One of the challenging
focuses on the GAA JLTs.
factors that need to be overcome in short channel (SC)
Numerous reports exist in literature on modeling
junction-based (JB) devices is the fabrication of sharp
of drain current of long channel GAA JLT (Duarte et
and abrupt junction between the channel and source/
al., 2011; Yu, 2014). However, the modeling of short
drain (S/D) region. A lot of short channel effects
channel devices has been reported by only few of
(SCEs) are associated with creation of these junctions.
them (Hu et al., 2014; Jiang et al., 2014; Sehra et al.,
Such challenges led to the evolution of junctionless
2020; Raut and Nanda, 2022; Smaani et al., 2022). In
(JL) architecture in which the concentration of dop-
this paper, we have developed a drain current model
ants is uniform all over the S/D and channel region
of short channel GAA JLT, which is a recent area of
(Colinge et al., 2010). As a result, no exorbitant high-
research (Chaujar et al., 2023; Kumar et al., 2023;
speed annealing techniques are required. This lessens
Kumari et al., 2023; Smaani et al., 2023). We have
the demand on fabrication processes and the thermal
incorporated S/D series resistance in our model. The
budget (Colinge, 2007; Colinge et al., 2010). This
model is based on the previous compact model of
permits one to fabricate SC devices. Simple architec-
junction-based GAA MOSFETs (Tsormpatzoglou et
ture (no p-n junction), no concentration gradient, low
al., 2009). The S/D series resistance is a vital element
leakage and improved short channel characteristics
in modeling of SC devices as in such devices; the series
are some of the advantages of junctionless transistors
resistance becomes a considerable portion of the total
(JLTs).
resistance and hence needs to be considered in the
Compared to other multi-gate architectures,
device modeling. The validation of the model has been
Metal-oxide-semiconductor field-effect transistors
obtained with both the experimental as well as simu-
(MOSFETs) with Gate-all-around (GAA) struc-
lation data. The effect of changing series resistance on
ture provide better immunity to SCEs, as the gate
drain current characteristics has also been obtained.

a
[email protected]
44 Development of an analytical model of drain current for junctionless GAA MOSFET

II. Theoretical details voltage (VFB) of the device, a complete neutral chan-
nel is created and we reach flat band condition. The
GAA structures offer superior short-channel charac-
conduction and valence bands become flat and now
teristics owing to the excellent control the gate offers
we can say that the device is turned ON. An accumu-
over the channel in such structures. Due to the absence
lation layer is created at the surface on further increas-
of junctions in JLTs, there is no requirement of doping
ing the gate voltage and the negative charge carriers
concentration gradient and hence the problems asso-
get accumulated resulting in the flow of surface cur-
ciated with the junctions are eliminated. Such devices
rent along with the bulk current. Figure 6.2 depicts
are also reported to deliver improved driving current
the energy-band diagram of GAA JLT in different
and sub-threshold properties when combined with
regions of operation along with the device schematic.
GAA architecture. Figure 6.1 depicts the cross-section
For an n-type semiconductor, in the cylindrical
of such JL GAA MOSFET.
coordinate, the Poisson’s equation can be written as
Device physics
JLTs are characterized as devices that are strongly and
(1)
evenly doped throughout. This means the type of the
dopants and their concentration is same all over the
junction. For n-type devices, p+ polysilicon is used as
where, φ signifies the potential,
the gate material and n+ polysilicon is used for p-type
r represents the radial direction,
devices. This results in a difference of approximately
V is the applied voltage,
1 eV in the work function which causes the channel
Nd represents the concentration of dopant throughout
to deplete. To bring the channel out of depletion, gate
the source, drain and channel,
bias must be applied. The working principle of GAA
VT is thermal voltage, and
JLT is as follows.
εSi is the permittivity of Si.
When no gate voltage (VG) is applied, the channel
Neglecting depletion charge density and considering
is fully depleted and a negligible amount of current
only the mobile carrier’s density, the above equation
flows through the region between the S and D. In such
can be simplified (Trevisoli et al., 2012) as
a situation, the transistor is said to be in OFF con-
dition. This is the sub-threshold region of operation.
The valence band is completely filled while the con- (2)
duction band is empty. When the voltage applied at
the gate equals the device’s threshold voltage (VTH),
bulk current starts flowing along a thin neutral path,
which is a non-depleted region formed near the center
of the channel. The path gets widened on increasing
the gate voltage. This increases the current flowing
through it. This has been reflected in the energy band
diagram where it can be seen that the concentration of
positive charges in the valence band is reduced. When
the applied voltage at the gate is same as the flat-band

Figure 6.1 Cross-section of cylindrical gate-all-around

MOSFET Figure 6.2 GAA JLT in different regions of operation
 Applied Data Science and Smart Systems 45

Boundary conditions applied for the GAA MOSFET

can be expressed as (Jimenez et al., 2004) 

(3)

The change due to shortening of channel is reflected

in the threshold voltage as ΔVTH, that can be expressed
as (Tsormpatzoglou et al., 2009)

(4)

where VTH,L and VTH,S are the long channel (Duarte et

al., 2011) and short channel (Chiang, 2012) threshold
voltages of GAA JLT, respectively.
The expression for drain current taking into
account the drift-diffusion model is given by

(5)

In the above equation, qi = Qi(V)/(4εsikT/qtsi), char-

acterizes the normalized sheet charge density, Qi sig-
nifies the inversion charge density, and V varies from
source to drain voltage.
On integrating the above equation, the drain cur-
rent expression attained is as follows:

(6)

The expression for mobility can be represented as

(Gaubert et al., 2010)

(7)

where, α accounts for the Coulomb scattering, θ1

represents the scattering due to phonons and θ2 rep-
resents the scattering due to surface roughness respec-
tively. VGS represents the gate to source voltage, µ0 is
the low field mobility. Further, we have used a factor
FCLM (Tsormpatzoglou et al., 2009) to consider the
effect of channel length modulation. FCLM can be
expressed as

(8)

where, the exponents A = 0.9 – (λ0/L) and B = 0.8(1+

(λ0/L)).

(9)
46 Development of an analytical model of drain current for junctionless GAA MOSFET

Figure 6.4 Drain current validation with the reported

data (a) Singh et al., (2011); (b) Moon et al., (2013).
(Parameters used: (a) L = 160 nm; ND = 6.7 × 1018 cm-3;
(b) L = 150 nm; Width of NW = 18 nm)

Figure 6.5 Drain current model validation with the ex-

Figure 6.3 Flowchart showing the calculation of IDS perimental data (Choi et al., 2011). (Parameters used:
L = 50 nm, EOT = 13 nm and ND = 2×1019 cm-3)

as experimental data. The verification has been done

for channel lengths that range from 20 nm to 160 nm. result (Choi et al., 2011). The gate length, doping
In our computations, Vsc and Qsc have the range from concentration and effective oxide thickness (EOT)
0.2 to 1.3. The validation with experimental data was taken to be 50 nm, 2×1019 cm-3 and 13 nm,
(Singh et al., 2011) and (Moon et al., 2013) has been respectively.
depicted in Figure 6.4 (a) and (b). In this paper, the extracted value of RSD = 15 KΩ
We have incorporated the effect of source-drain has been used for the computation of drain current.
series resistance (RSD) to validate the model accuracy The results obtained are in line with the experimental
with the experimental data. The source-drain series values with deviation from the reported one below
resistances used in the calculations were extracted 5%.
utilizing (Kim et al., 2013). The figure shows that the Simulation results have also been used to validate
model is in accord with the experimental records with the model.
deviation below 5%. The drain current characteristics of GAA JLT are
Figure 6.5 illustrates the graph of drain current with presented in Figure 6.6 (a) and (b). The current values
VDS with device parameters taken from experimental for different drain voltages having 20 nm and 30 nm
 Applied Data Science and Smart Systems 47

Figure 6.8 Variation of the drain current with different

drain voltages with and without S/D resistance (Param-
eters used: L = 50 nm, ND = 2×1019 cm-3, EOT = 13 nm,
Rsd = 15 kΩ)

Figure 6.6 Model validation with the reported data (a)

Hu et al. (2014); (b) Wang et al. (2014)

Figure 6.9 Drain current variation with gate overdrive

voltage with and without S/D resistance (Parameters
used: L = 50 nm, ND = 2×1019 cm-3, EOT = 13 nm,
RSD = 15 kΩ)

that the drain current decreases on inclusion of RSD,

due to the reduction in terminal voltages.
Figure 6.7 Comparison of simulated (Lou et al., 2012) The drain current variation with gate overdrive
and output characteristics of the model voltage (VGT = VGS – VTH) for different values of
S/D resistance has also been plotted and depicted in
Figure 6.9.
channel length, respectively have been depicted. As
shown in the figure, good accuracy of the proposed V. Conclusion
model has been obtained with the reported simulation
data (Hu et al., 2014; Wang et al., 2014). We have proposed a model of drain current for short
We have obtained one more graph for drain current channel JL GAA NW n- MOSFET. We have incor-
(Figure 6.7) to show the model validation with the porated S/D series resistance in our work. The vali-
simulated data (Lou et al., 2012). The channel length, dation has been obtained with experimental and
concentration of dopant and oxide thickness is taken simulated data. We have also presented algorithm, for
to be 40 nm, 2×1019 cm-3 and 2 nm, respectively. A the computation of the drain current which is based
good agreement of the calculated results has been on multi-iterative technique. These iterations are nec-
obtained with the simulated one, that further sup- essary because upon including S/D series resistance,
ports the accuracy of the model. the drain current transforms into the transcendental
equation, resulting in a number of coupled equations
involving the mobility, threshold voltage, and other
IV. Results and discussion
variables. The effect of RSD on the drain current has
Figure 6.8 depicts the change in drain current with also been investigated. The proposed model may be
and without incorporating RSD for different drain helpful to the research community in predicting the
voltages. It can be comprehended, from the figure, performance parameters of the devices and circuits
48 Development of an analytical model of drain current for junctionless GAA MOSFET

before going into final fabrication. This may save parameters of nanoscale strained silicon MOSFET-
time as well as resources. We have not taken quantum based CMOS inverters. Microelec. J., 55, 8–18.
mechanical effects into account which becomes sig- Subindu, K., Kumari, A., and Das, M. K. (2017). Model-
nificant in ultra scaled devices. ing gate-all-around Si/SiGe MOSFETs and circuits for
digital applications. J. Comput. Elec., 16, 47–60.
Amrita, K., Saini, A., Kumar, A., Kumar, V., and Kumar,
References M. (2023). Recent developments and challenges in
strained junctionless MOSFETs: A review. 2023 Int.
Rishu, C. and Yirak, M. G. (2023). Sensitivity investigation
Conf. Comput. Intel. Sustain. Engg. Sol. (CISES),
of junctionless gate-all-around silicon nanowire field-
118–122.
effect transistor-based hydrogen gas sensor. Silicon,
Haijun, L., Zhang, L., Zhu, Y., Lin, X., Yang, S., He, J., and
15(1), 609–621.
Chan, M. (2012). A junctionless nanowire transistor
Te-Kuang, C. (2012). A new quasi-2-D threshold voltage
with a dual-material gate. IEEE Trans. Elec. Dev.,
model for short-channel junctionless cylindrical sur-
59(7), 1829–1836.
rounding gate (JLCSG) MOSFETs. IEEE Trans. Elec.
Dong-Il, M., Choi, S.-J., Duarte, J. P., and Choi, Y.-K.
Dev., 59(11), 3127–3129.
(2013). Investigation of silicon nanowire gate-all-
Sung-Jin, C., Moon, D., Kim, S., Ahn, J.-H., Lee, J.-S., Kim,
around junctionless transistors built on a bulk sub-
J.-Y., and Choi, Y.-K. (2011). Nonvolatile memory by
strate. IEEE Trans. Elec. Dev., 60(4), 1355–1360.
all-around-gate junctionless transistor composed of
Sehra, S. S., Singh, J., Rai, H. S., and Anand, S. S. (2020).
silicon nanowire on bulk substrate. IEEE Elec. Dev.
Extending processing toolbox for assessing the logi-
Lett., 32(5), 602–604.
cal consistency of OpenStreetMap data. Trans. GIS,
Jean-Pierre, C. (2007). Multi-gate SOI MOSFETs. Micro-
24(1), 44–71. https://doi.org/10.1111/tgis.12587.
elec. Engg., 84(9–10), 2071–2076.
Pratikhya, R. and Nanda, U. (2022). A charge-based analyt-
Jean-Pierre, C., Lee, C.-W., Afzalian, A., Akhavan, N. D.,
ical model for gate all around junction-less field effect
Yan, R., Ferain, I., Razavi, P. et al. (2010). Nanow-
transistor including interface traps. ECS J. Solid State
ire transistors without junctions. Nat. Nanotechnol.,
Sci. Technol., 11(5), 051006.
5(3), 225–229.
Pushpapraj, S., Singh, N., Miao, J., Park, W.-T., and Kwong,
Duarte, J. P., Choi, S.-J., Moon, D., and Choi, Y.-K. (2011).
D.-L. (2011). Gate-all-around junctionless nanowire
A nonpiecewise model for long-channel junctionless
MOSFET with improved low-frequency noise behav-
cylindrical nanowire FETs. IEEE Elec. Dev. Lett.,
ior. IEEE Elec. Dev. Lett., 32(12), 1752–1754.
33(2), 155–157.
Billel, S., Rahi, S. B., and Labiod, S. (2022). Analytical com-
Philippe, G., Teramoto, A., and Ohmi, T. (2010). Modelling
pact model of nanowire junctionless gate-all-around
of the hole mobility in p-channel MOS transistors fab-
MOSFET implemented in verilog-A for circuit simula-
ricated on (1 1 0) oriented silicon wafers. Solid-State
tion. Silicon, 14(16), 10967–10976.
Elec., 54(4), 420–426.
Billel, S., Nafa, F., Upadhyay, A. K., Labiod, S., Rahi, S. B.,
Guangxi, H., Xiang, P., Ding, Z., Liu, R., Wang, L., and
Benlatreche, M. S., Akroum, H., Lakhdara, M., and
Tang, T.-A. (2014). Analytical models for electric po-
Yadav, R. (2023). Compact modeling of junction-
tential, threshold voltage, and subthreshold swing of
less gate-all-around MOSFET for circuit simulation:
junctionless surrounding-gate transistors. IEEE Trans.
Scope and challenges. Device Circuit Co-Design Is-
Elec. Dev., 61(3), 688–695.
sues in FETs, 57–78. CRC Press.
Chunsheng, J., Liang, R., Wang, J., and Xu, J. (2014). Ana-
Trevisoli, R. D., Doria, R. T., de Souza, M., Das, S., Ferain,
lytical short-channel behavior models of junctionless
I., and Pavanello, M. A. (2012). Surface-potential-
cylindrical surrounding-gate MOSFETs. 2014 Int.
based drain current analytical model for triple-gate
Symp. Next-Gen. Elec. (ISNE), 1–2. IEEE.
junctionless nanowire transistors. IEEE Trans. Elec.
David, J., Iniguez, B., Sune, J., Marsal, L. F., Pallares, J.,
Dev., 59(12), 3510–3518.
Roig, J., and Flores, D. (2004). Continuous analytic
Tsormpatzoglou, A., Tassis, D. H., Dimitriadis, C. A.,
IV model for surrounding-gate MOSFETs. IEEE Elec.
Ghibaudo, G., Pananakakis, G., and Clerc, R. (2009).
Dev. Lett., 25(8), 571–573.
A compact drain current model of short-channel cy-
Ye-Ram, K., Lee, S.-H., Sohn, C.-W., Choi, D.-Y., Sagong,
lindrical gate-all-around MOSFETs. Semicond. Sci.
H.-C., Kim, S., Jeong, E.-Y. et al. (2013). Simple S/D
Technol., 24(7), 075017.
series resistance extraction method optimized for
Juncheng, W., Du, G., Wei, K., Zhao, K., Zeng, L., Zhang,
nanowire FETs. IEEE Elec. Dev. Lett., 34(7), 828–
X., and Liu, X. (2014). Mixed-mode analysis of differ-
830.
ent mode silicon nanowire transistors-based inverter.
Alok, K., Gupta, T. K., Shrivastava, B. P., and Gupta, A.
IEEE Trans. Nanotechnol., 13(2), 362–367.
(2023). Impact of temperature variation on noise
Yu, Y. S. (2014). A unified analytical current model for
parameters and HCI degradation of recessed source/
N-and P-type accumulation-mode (junctionless) sur-
drain junctionless gate all around MOSFETs. Micro-
rounding-gate nanowire FETs. IEEE Trans. Elec. Dev.,
elec. J., 134, 105720.
61(8), 3007–3010.
Subindu, K., Kumari, A., and Das, M. K. (2016). Develop-
ment of a simulator for analyzing some performance
7 Crop recommendation using machine learning
Paramveer Kaura and Brahmaleen Kaur Sidhu
Department of Computer Science and Engineering, Punjabi University, Patiala, Punjab, India

Abstract
Agriculture serves as the cornerstone of India’s economic expansion, constituting the primary income source for a significant
proportion of its populace, encompassing both those directly engaged in agricultural activities and those who depend on it
indirectly for their livelihoods. Therefore, it is essential for farmers to make the correct option possible when cultivating any
crop so that the farmer can make maximum profit from the agriculture field. To make the agriculture sector profitable, one
of the technologies that may be used in this age of rapid technological improvement is known as machine learning (ML). In
this research paper, various ML algorithms, such as logistic regression (LR), decision trees (DT), LightGBM, and random
forest (RF), have been utilized to analyze a dataset. The primary objective is to predict the most suitable crop based on soil
attributes such as (nitrogen, phosphorous, potassium) NPK content, humidity, temperature, soil pH level, and rainfall. Out of
this random forest and LightGBM comes with great accuracy whereas decision tree and logistic regression have less accuracy.
In addition, ML algorithms will likely find applications in a variety of agricultural subfields in the near future, including the
diagnosis of plant diseases, the selection of soil types, and the forecasting of retail pricing.

Keywords: Machine learning, crop recommendation, decision tree, random forest, logistic regression, LightGBM

I. Introduction solved by using the advanced technologies to improve

the agriculture sector. The farmer cannot predict the
The importance of agriculture to the Indian economy
weather but the technology can predict the climate
and to human life cannot be ignored. It serves as one
like rainfall, humidity and temperature based on past
of the main professions that are necessary for human
data and help famers in real manner.
existence. India’s population relies heavily on income
Precision agriculture (PA) is a method of farm man-
from agriculture and related industries. Around 82%
agement that takes into account the specific condi-
of farmers are classified as small and marginal, under-
tions of particular fields and crops via the collection,
scoring the central importance of agriculture as the
organization, and analysis of data (António Monteiro,
main source of income for 70% of rural households
2021). It is observed that in recent years precision
(Chavva, n.d.). The financial health of the agricultural
agriculture technologies have helped the farmers as
sector is tightly intertwined with the success of every
well as the environment by suggesting the required
harvest, which, in turn, is influenced by a wide range
amount of fertilizers, pesticides and water for crops.
of variables, including weather patterns, soil qual-
Farmers have got more profit by these technologies in
ity, fertilizer use, and market prices. Due to climate
terms of money also.
change it has become difficult for farmers to choose
In the realm of computer technology, the most
appropriate crop for particular season. Price of crops
recent advancements encompass block chain, inter-
given by the government to farmers also the effective
net of things (IoT), deep learning, machine learning
factor to grow any crop. According to National Crime
(ML) and cloud computing, which are useful for solv-
Records Bureau (NCRB) statistics, the farmer suicides
ing difficult problems in various fields like health,
rate has remained high between 2014 and 2020. In
biochemistry, agriculture, cybercrime, robotics, bank-
2014, 56 farmers have committed suicide, and by
ing, meteorology, medicine and robotics (Vishal
2020, the number has risen to 5,500 (Affarirs, n.d.).
Meshram, 2021). However in this paper, the focus is
The choice of crop to be sown depends on avail-
only on ML in agriculture. ML algorithms (Sharma et
ability of resources like soil, water, seed, manure, fer-
al., 2018) which are support vector machines, Naïve
tilizer and market profit. However, climate conditions
Bayes, neural network, decision tree (DT), K-Nearest
and soil properties are consider as natural parameters
Neighbor, XG-boost (e-Xtreme Gradient Boosting),
for the success of any crop grown by farmer. Due to
multi-variate linear regression, linear regression (LR),
variations in soil, water, and air quality throughout
chi-square automatic interaction detection (CHAID)
the year, it is difficult to predict how best to use dif-
and sliding window non-linear regression are helpful
ferent types of fertilizer and what crop to be grown.
at various stages of crop grown cycle and provides
The rate of agricultural output is falling continuously
maximum accuracy. In this proposed system out of
in this situation (Pande et al., 2021). The issues can be

a
[email protected]
50 Crop recommendation using machine learning

this four, ML models are deployed to make accurate farmers’ crop management issues like crop selection,
crop recommendations. yield, and profit. Researchers employ decision trees,
Naive Byes, SVM, LR, RF, and Xgboost. Pradeepa
II. Related work Bandara et al. developed a crop recommendation sys-
tem for Sri Lanka (Bandara et al., 2020). The study
Pudumalar et al. (2016) addressed precision agricul- provides a theoretical as well as a conceptual plat-
ture. This study proposes an ensemble model with form for a recommendation system using Arduino
majority voting technique utilizing RT, Naïve Bayes, microcontrollers, ML approaches such as Naive
CHAID, and K-nearest neighbor (KNN) as learn- Bayes (multi-nomial) and SVM and unsupervised ML
ers to effectively and correctly suggest a crop for algorithms that are K-Means Clustering and Natural
site-specific parameters. The study (Kanaga Suba Language Processing (NLP) (sentiment analysis).
Raja et al., 2017) analyses historical data to predict Avinash Kumar et al. (2019) addressed crop selection
a farmer’s crop output and price. Sliding window and disease issues. SVMs classification model, deci-
non-linear regression predicts agricultural output sion tree model, and logistic regression model were
depending on rainfall, temperature, market prices, used to create this recommendation system.
land area, and crop yield. Zeel Doshi et al. (2018)
developed a soil dataset-based crop recommendation
III. Objectives
system for only four crops. The ensemble model uses
random forest (RF), Naive Bayes, and linear support The aim of the proposed work is to implement ML
vector machines base learners. The majority voting algorithms for developing crop recommendation sys-
technique is employed in the combination approach tem and it is based on chemical properties of soil and
because it is the most accurate. The author uses Big weather condition and to evaluate the proposed system.
Data analytics and ML to create an AgroConsultant,
an system that assists Indian farmers choose the best IV. Background techniques
crop based on sowing season, farm location, soil
properties, and environmental factors like tempera- A. Logistic regression
ture and rainfall (Doshi et al., 2018). Rainfall predic- It is a ML algorithm primarily applied to classifica-
tor, another approach created by academics, predicts tion problems, operates on the foundation of predic-
annual precipitation. The system uses DT, KNN, RF, tive analysis rooted in probability (Rymarczyk et al.,
and neural networks. 2019). Notably, the LR model adopts a more intricate
Shilpa Mangesh Pande et al. (2021) provide farm- cost function than linear regression. This cost func-
ers a simple yield projection tool. Farmers utilize a tion, often referred to as the “Sigmoid function” or
smartphone app to connect to the internet. GPS “logistic function,” replaces the linear function. Due
locates users. Enter location and soil type. ML sys- to the fundamental premise of logistic regression, the
tems can identify the most profitable crops and pre- output range of the cost function is inherently con-
dict agricultural yields for user-selected crops. SVM, fined to the interval [0, 1]. This constraint stems from
RF, MLR, ANN, and KNN are used to predict agri- the nature of logistic regression, where it inherently
cultural production. The research (A et al., 2021) models the probability of an event occurring, ren-
suggests a way to assist farmers pick crops by con- dering linear functions unsuitable for capturing its
sidering planting time, soil qualities including type of nuances and characteristics.
soil, pH value, and nutrient content, meteorological
aspects like rainfall, temperature, and state location. B. Decision tree
The suggested system has been developed using linear When it comes to representing models for use in
regression as well as neural network. Another study data classification, decision trees are among the most
(Gosai et al., 2021) forecasts the best crop based on popular approaches (Jijo and Abdulazeez, 2021). DT
N, P, K, pH of soil, humidity, temperature, and rain- stand as versatile assets in numerous domains, span-
fall. Decision trees, SVM, Nave Bayes, Support vec- ning machine learning, image processing, and pattern
tor machine, LR, RF, and XGBoost were utilized to recognition. Their core role revolves around the task
develop suggested system, and the maximum accu- of classification and, as a result, they find wide appli-
racy was of XGBoost. Distribution analysis, majority cation as classifiers within the field of data mining.
voting, correlation analysis and ensembling are used These decision trees are architecturally composed of
to create 22 crop recommendations (Kulkarni et al., interconnected nodes and branches, each node rep-
2018). A three-level technique solves crop recommen- resenting a collection of attributes within discrete
dations. Chhikara et al. (2022) propose a ML-based classification categories. Each branch within the tree
crop recommender system that can accurately fore- signifies a potential value associated with the respec-
cast the yield of 22 different crop types, addressing tive node. Decision trees earn considerable favor for
 Applied Data Science and Smart Systems 51

their proficiency in swiftly and accurately managing

large datasets.

C. Random forest
In the field of ML, random forest, a supervised learn-
ing technique, has demonstrated considerable success.
It’s versatile and can handle various tasks like classifi-
cation and prediction. What sets random forest apart
is that it operates like a team of decision trees collabo-
rating to solve problems. Rather than rely on a one
decision tree, it combines the results of many trees,
each trained on different parts of the data. This coop-
erative approach enhances accuracy. Essentially, the
more trees in this “forest,” the better the performance,
and it’s less prone to errors (Dabiri et al., 2022; Gera
et al., 2021).

D. LightGBM
Tree-based learning algorithms are at the foundation
of LightGBM, a gradient boosting framework (Tang
et al., 2020). It has the many benefits because of its
decentralized and efficient design such as increased
efficiency and accelerated training time. It uses less
memory and allows for multi-GPU and distributed
learning. LightGBM has ability to process large Figure 7.1 Flow chart of proposed methodology
amount of data as well as gives enhanced precision.

Table 7.1 Feature description

V. Methodology
The proposed methodology has been shown in the Attribute Attribute description
Figure 7.1 to implement an accurate crop recommen-
Nitrogen Proportion of nitrogen amount in soil
dation system. The detailed process of developing rec-
Phosphorus Ratio of phosphorus content in soil
ommender system has been provided which includes
various stages data collection, data pre-processing, Potassium Ratio of potassium amount in soil
model development and training and testing. Rainfall Rainfall (mm)
In the first stage dataset used for proposed approach Humidity Relative humidity (%)
was retrieved from Github repository https://github. pH level Soil’s pH value
com/gabbygab1233/Crop-Recommender/blob/ main/
Temperature Temperature (degree celsius)
Crop_recommendation.csv. The dataset is combina-
tion of rainfall, climate and fertilizers dataset which
is collected from Indian data available for agriculture
sector. Table 7.1 shows the description of features Table 7.2 Accuracy of algorithms
that has been used. It has 8 features and 2200 records.
The preparation of the collected raw data to make Algorithm/classifier Predicted accuracy
it appropriate for use in building the ML model is the
LightGBM 0.99
next step. This is the initial and most crucial step in
Decision tree 0.98
the creation of an ML model. The data that is avail-
able on different sources is not clean always. Data Random forest 0.99
must be cleaned and formatted before any action can Logistic regression 0.95
be performed on it. Pre-processing the data is neces-
sary for this purpose.

crop recommendation system, including LR, DT, RF,

VI. Results
and LightGBM. The accuracy of mentioned algo-
Table 7.2 displays the performance of machine learn- rithms has been assessed using evaluation parameters
ing approaches employed in the development of a such as precision, recall, F1-score, and support.
52 Crop recommendation using machine learning

Precision is imperative to explore innovative solutions. In

It indicates the percentage of accurate predictions this research paper, a crop recommendation system
executed by the model. It specifically assesses the clas- powered by ML technologies is introduced. The sys-
sifier’s capacity to correctly discern instances as either tem is designed to help farmers in making accurate
positive or negative. The precision parameter is deter- decision when choosing the crop varieties based on
mined by dividing the accurately predicted positive the intricate interplay of soil attributes and weather
counts by the sum of all predictions, encompassing conditions.
both correct and incorrect classifications.
VIII. Conclusion
The research paper introduces a crop recommenda-
tion system designed for farmers to make optimal crop
Recall
choices through predictive analytics. This system takes
It denotes the percentage of positive cases success-
into account critical parameters, such as soil charac-
fully identified by the model. It quantifies a classifier’s
teristics (NPK content, pH, and humidity), along with
ability to effectively locate all instances that belong to
meteorological factors (temperature and rainfall). To
the designated target class. The recall metric is com-
develop this system, ML approaches, including LR,
puted by dividing the count of correct predictions for
DT, RF along with LightGBM, are employed. This
the target class by the total number of predictions
algorithms have been applied on collected dataset.
made, encompassing both accurate and inaccurate
Dataset was normally distributed with negligible out-
classifications.
liers and this is analyzed after pre-processing steps.
LightGBM and RF exhibit the highest accuracy among
various algorithms. Looking forward, enhancing the
system’s performance is achievable through the con-
F1-score tinuous updating of datasets, ensuring its alignment
It signifies the proportion of positive predictions that with evolving agricultural conditions. Moreover, the
are accurate. This metric is computed from the combi- system’s scope can be expanded to include crop dis-
nation of both precision and recall, and it falls within ease detection which can assist farmers in maximizing
a range from 1.0 – indicating the best performance benefits in the agriculture sector.
to 0.0 – representing the worst. Given that F1-scores
encompass both precision and recall in their compu-
References
tation, they are considered more informative than
simple accuracy measurements. When assessing and Chavva, Mr Konda Reddy. India at a glance. n.d. https://
comparing classifier models, it is advisable to use the www.fao.org/india/fao-in-india/india-at-a-glance/en/
weighted average of F1-scores rather than relying (accessed june 15, 2023).
Gera. T., Singh, J., Mehbodniya, A., Webber, J. L., Shabaz,
solely on overall accuracy.
M., and Thakur, D. (2021). Dominant feature selec-
tion and machine learning-based hybrid approach
to analyze android ransomware. Sec. Comm. Netw.,
2021, 1–22.
Affarirs, M. O. H. NCRB on data portal. [Online]. Avail-
Support able: https://data.gov.in/search?title=suicide.
It refers to the frequency of actual occurrences of a Pande, Shilpa Mangesh, Prem Kumar Ramesh, ANMOL
specific class within the dataset. It serves as an indi- ANMOL, B. R. Aishwarya, KARUNA ROHILLA,
cator of how evenly or unevenly the training data is and KUMAR SHAURYA. (2021). Crop recommender
distributed. Disparities in support could potentially system using machine learning approach. In 2021 5th
suggest that the reported scores by the classifier may international conference on computing methodologies
not be as reliable as desired. It is important to note and communication (ICCMC), 1066–1071. IEEE.
that support remains consistent across various mod- António Monteiro, S. S. P. G. (2021). Precision agriculture
for crop and livestock. MDPI/Animals, 1–18.
els; its primary role is to aid in the analysis of the
Meshram, Vishal, Kailas Patil, Vidula Meshram, Dinesh
testing methodology. Hanchate, and S. D. Ramkteke. (2021). Machine
learning in agriculture domain: A state-of-art survey.
VII. Discussion Artificial Intelligence in the Life Sciences, 1, 1–11. doi:
https://doi.org/10.1016/j.ailsci.2021.100010
The agriculture sector plays pivotal role in driv- Sharma, Pinkashia, and Jaiteg Singh. (2018). Machine learn-
ing a nation’s economic development. To enhance ing based effort estimation using standardization. In
the profitability and productivity of this sector, it 2018 International Conference on Computing, Power
 Applied Data Science and Smart Systems 53
and Communication Technologies (GUCON), 716– Kulkarni, Nidhi H., G. N. Srinivasan, B. M. Sagar, and N. K.
720. IEEE. Cauvery. (2018). Improving crop productivity through
Pudumalar, S., E. Ramanujam, R. Harine Rajashree, C. Ka- a crop recommendation system using ensembling tech-
vya, T. Kiruthika, and J. Nisha. (2017). Crop recom- nique. In 2018 3rd International Conference on Com-
mendation system for precision agriculture. In 2016 putational Systems and Information Technology for
Eighth International Conference on Advanced Com- Sustainable Solutions (CSITSS), 114–119. IEEE. doi:
puting (ICoAC), 32–36. IEEE. 10.1109/CSITSS.2018.8768790
Jijo, B. T. and Abdulazeez, A. M. (2021). Classification Priyadharshini, A., Swapneel Chakraborty, Aayush Ku-
based on decision tree algorithm for machine learning. mar, and Omen Rajendra Pooniwala. (2021). Intel-
J. Appl. Sci. Technol. Trends, 02(1), 20–28. ligent crop recommendation system using machine
Rymarczyk, T., Kozlowski, E., Klosowski, G., and Niderla, learning. In 2021 5th international conference on
K. (2019). Logistic regression for machine learning in computing methodologies and communication
process tomography. Sensors, 1–19. (ICCMC), 843–848. IEEE. doi: 10.1109/ICC-
Dabiri, Hamed, Visar Farhangi, Mohammad Javad Moradi, MC51019.2021.9418375
Mehdi Zadehmohamad, and Moses Karakouzian. Gosai, D., Raval, C., Nayak, R., Jayswal, H., and Patel, A.
(2022). Applications of Decision Tree and Random (2021). Crop recommendation system using machine
Forest as Tree-Based Machine Learning Techniques learning. Int. J. Sci. Res. Comp. Sci. Engg. Inform.
for Analyzing the Ultimate Strain of Spliced and Technol., 554–569.
Non-Spliced Reinforcement Bars. Applied Sciences. Ray, Rakesh Kumar, Saneev Kumar Das, and Sujata
12(10), 4851. 1–13 doi: https://doi.org/10.3390/ Chakravarty. (2022). Smart Crop Recommender
app12104851. System-A Machine Learning Approach. In 2022 12th
Tang, Mingzhu, Qi Zhao, Steven X. Ding, Huawei Wu, International Conference on Cloud Computing, Data
Linlin Li, Wen Long, and Bin Huang. (2020). An im- Science & Engineering (Confluence), 494–499. IEEE.
proved lightGBM algorithm for online fault detection doi: 10.1109/Confluence52989.2022.9734173
of wind turbine gearboxes. Energies, 13(4): 807. 1–16. Chhikara, Sonam, and Nidhi Kundu. (2022). Machine
doi: https://doi.org/10.3390/en13040807. Learning based Smart Crop Recommender and
Raja, S. Kanaga Suba, R. Rishi, E. Sundaresan, and V. Yield Predictor. In 2022 International Conference
Srijit. (2017). Demand based crop recommender on Computing, Communication, and Intelligent Sys-
system for farmers. In 2017 IEEE Technological tems (ICCCIS), 474–478. IEEE. doi: 10.1109/ICC-
Innovations in ICT for Agriculture and Rural De- CIS56430.2022.10037678
velopment (TIAR), 194–199. IEEE. doi: 10.1109/ Bandara, P., Weerasooriya, T., R. T. H., Nanayakkara, W., D.
TIAR.2017.8273714. M. A. C., and P. M. G. P. (2020). Crop recommenda-
Doshi, Zeel, Subhash Nadkarni, Rashi Agrawal, and Neepa tion system. Int. J. Comp. Appl., 175(22), 22–25.
Shah. (2018). AgroConsultant: intelligent crop rec- Kumar, Avinash, Sobhangi Sarkar, and Chittaranjan Prad-
ommendation system using machine learning algo- han. (2019). Recommendation system for crop iden-
rithms. In 2018 Fourth International Conference on tification and pest control technique in agriculture.
Computing Communication Control and Automa- In 2019 International Conference on Communica-
tion (ICCUBEA), 1–6. IEEE. doi: 10.1109/ICCU- tion and Signal Processing (ICCSP), 0185–0189.
BEA.2018.8697349. IEEE, 2019.doi: 10.1109/ICCSP.2019.8698099
8 Environment and sustainability development: A ChatGPT
perspective
Priyanka Bhaskar1 and Neha Seth2,a
1
Assitant Professor, School of Commerce and Management, Central University of Rajasthan, Kishangarh, Ajmer, India
2
Associate Professor, Symbiosis Institute of Business Management, Noida Symbiosis International (Deemed Univer-
sity), Pune, India

Abstract
Artificial Intelligence (AI) and sustainability are two sides of same coin. AI is a reliable ally in the fight for sustainability,
leading us to a brighter future. AI illuminates renewable energy, resource management, and eco-friendly decision-making by
analyzing large datasets. However, the energy usage and carbon footprint of AI models and AI sustainability are increasingly
under review. This research paper examines the environmental implications of AI models, focusing on ChatGPT, and empha-
sizes the necessity for sustainable AI development. Recent studies show that AI model creation and use significantly impact
the global carbon footprint due to energy, water, and carbon emissions. With its massive computational needs, ChatGPT con-
tributes to environmental issues. To tackle this dilemma, sustainable AI development must be promoted. Model compression,
quantization, and knowledge distillation improve AI energy efficiency. The use of renewable energy and the establishment
and enforcement of AI model energy efficiency requirements are equally crucial. ChatGPT and comparable models can be
environmentally friendly by using sustainable AI development methods. In this line, the objective of the present study is to
analyze the impact of the use of AI tools, specifically ChatGPT, on sustainability and environmental protection by analyzing
existing reports and studies on the environmental impact of artificial intelligence models.
Academicians, developers, politicians, institutions and organizations must work together to create rules and frameworks
for energy-efficient AI algorithms, renewable energy use, and responsible deployment. This study article concludes that AI
models’ energy usage and carbon footprint must be understood and reduced. By promoting sustainable practices, the AI
community may encourage a more environmentally sensitive and responsible approach to AI development, leading to a
greener future that meets global sustainability goals.

Keywords: Artificial intelligence, AI-language models, ChatGPT, environmental impact, carbon footprint, water footprint,
greenhouse gas emissions, energy consumption, sustainability, mitigation strategies

I. Introduction tons that each human emits annually. Similarly, it also

produces a considerable water footprint, mostly due
It cannot be denied that artificial intelligence (AI) is
to the training process, which uses a lot of energy and
already transforming the world and will continue to
turns it into heat, necessitating surprisingly sufficient
do so. Even while AI has the potential to be bene-
freshwater to keep equipment cool and sustain tem-
ficial, society may suffer due to it. ChatGPT, a siz-
peratures (Patterson, 2022).
able language model created by Open AI, is one such
Therefore, the objective of the present study is to
prominent AI model which has gained widespread
analyze the impact of the use of AI tools, specifically
application. The worldwide market size for AI is
ChatGPT, on sustainability and environmental pro-
US$ 2,00,000 million in 2023 and it is expected to
tection by analyzing existing reports and studies on
grow by almost 9 times to reach US$ 1,85,000 by
the environmental impact of artificial intelligence
2030 (statista.com). Due to its outstanding Natural
models.
Language Processing (NLP) abilities, ChatGPT has
The content collected for this study is secondary
drawn much attention and, without a doubt, made
in nature and was gathered from different sources,
everyone’s life simpler. But ChatGPT has a price,
such as digital articles, papers, research papers from
just as anything worthwhile has a price. The energy
reputed journals, government websites, statista.com,
required for building and training this AI system
etc. This paper will outline a comprehensive overview
results in substantial negative environmental costs,
of the present state of knowledge in this field so far.
such as generating a substantial carbon and water
The effects of artificial intelligence are explored on
footprint that are frequently disregarded.
the environment, with special reference to ChatGPT.
According to data and calculations, ChatGPT gen-
The study also investigates environmental problems
erates 8.4 metric tons of carbon dioxide per year to
and continues to increase the energy efficiency of AI
run the data centers, more than double the 4 metric

[email protected]
a
 Applied Data Science and Smart Systems 55

models by addressing such challenges. Suggestions for significantly impacts business sustainability. Ray
the adoption of sustainable practices and the usage of (2023) asked ChatGPT how it will play a significant
renewable energy sources in all fields are part of the role in agricultural science and technology in future
study. This article addresses the environmental effect and got responses which may lead to the sustainable
of ChatGPT and offers suggestions for sustainable AI development of the farming sector.
development, but it also has certain restrictions and After refereeing a number of research papers and
opens up avenues for future exploration. articles published on ChatGPT and its application in
The study highlights the need to take into account various fields, it was found that individually many
the ecological implications of AI systems and the papers talk about the application of ChatGPT in vari-
demand for sustainable practices in creating and ous areas for sustainable development and extend-
implementing them. The last part includes the conclu- ing similar work, in this study, authors are trying to
sion and future prospects of the study. analyze the impact of the use of AI tools, specifically
ChatGPT, on sustainability and environmental pro-
II. Review of literature tection by analyzing existing reports and studies on
the environmental impact of artificial intelligence
In the contemporaneous literature available in the models.
field of information technology, numerous studies are
available that provide information about the chatbots,
III. ChatGPT and potential areas of concern
language models and IT platforms. This section pres-
ents the evolvement of research based on ChatGPT Kain (2023) stated that ChatGPT is an advanced
and its relation to sustainable development. language model developed by OpenAI and released
Zhu et al. (2023) raised concerns about environ- in November 2022. The acronym “ChatGPT” com-
mental issues due to the introduction of another natu- bines the terms “Chat”, which refers to the chatbot
ral language processing model, ChatGPT. They have functionality of the framework, and “GPT” stands for
used ten real-world examples to study the impact of “Generative Pre-trained Transformer” and it is a type
ChatGPT and its impact on the environment. Another of Large Language Model (LLM). ChatGPT is based
study by Khowaja et al. (2023) focused on an aspect on the core GPT models from OpenAI, GPT-3.5 and
of large language models which were ignored, includ- GPT-4, which provide conversational interaction
ing sustainability, privacy, digital divide and ethics (Lund et al., 2023).
(SPADE) and based on primary data and visualization, To generate intelligent and captivating text-based
they suggested that not only ChatGPT other models replies to user input, the latest AI conversation tool
should also undergo SPADE analysis. George (2023) takes advantage of the most recent advancements in
raised the issue of water consumption by ChatGPT. It machine learning as well as natural language process-
was found that the water consumption by AI models ing (NLP) (Bhaskar, 2022).
is relatively less than in other industries but it is still OpenAI launched in November 2022, ChatGPT
a matter of concern, and it should be further reduced revolutionized how people interact globally by pro-
by taking appropriate measures like improving energy ducing replies to common writing jobs in seconds.
efficiency, utilizing renewable energy sources, opti- Despite the fact that ChatGPT’s “outputs may be
mizing algorithms and implementing strategies to inaccurate, untruthful, and otherwise misleading at
conserve water. times”, as stated in its FAQs, the model’s speed and
Biswas (2023) integrated with ChatGPT to get adaptability make it widely applicable for simple
responses on the effect of ChatGPT on global warm- writing tasks, including cover letters, and many more
ing and analyzed the replies received. Biswas con- uncountable things.
cluded that ChatGPT can be used in various ways to ChatGPT does not directly impact the environ-
aid climate research, including model “parameteriza- ment because it is an AI language model. However,
tion, data analysis and interpretation, scenario gen- the infrastructure and data centers needed to support
eration, and model evaluation”. Sohail et al. (2023) ChatGPT, as well as the technology that enables it, may
reviewed 100 Scopus papers on ChatGPT and found have an impact on the environment. Such as training
that ChatGPT has applications in various fields like and operating complex language models like GPT-3
healthcare, marketing and financial services, software require a substantial amount of computer power,
engineering, academic and scientific writing, research which is why ChatGPT uses a lot of energy. If the
and education, environmental science, and natural energy required to power data centers and computer
language processing and its potential to address real- infrastructure comes from non-renewable sources, it
world problems. Vrontis et al. (2023) analyzed the may cause carbon emissions and environmental dam-
role of ChatGPT and skilled employees in business age (Teubner, 2023). Also, the energy needed to run
sustainability and found that leadership motivation AI models results in the emission of carbon dioxide
56 Environment and sustainability development: A ChatGPT perspective

and other greenhouse gases, which fuel global warm- ChatGPT’s 1 million users sent one request each day,
ing. The carbon footprint of ChatGPT and other AI ChatGPT would get the same number of requests per
systems depends on factors such as the energy source, day as BLOOM did at that time. At least based on the
cooling requirements, and hardware efficiency (An et volume of discussion about ChatGPT in traditional
al., 2023; Khattar et al., 2020). and social media platforms, ChatGPT handles far
ChatGPT needs a lot of processing and storage more daily requests than similar services. Although
power, which is often provided by big data centers. there is a great deal of ambiguity around this estimate
Massive amounts of water are used to cool these data since it is founded on some dubious presumptions,
centers, which leads to a water footprint. Apart from compared to in-depth analyses of BLOOM’s carbon
that, they also need various other materials like metals footprint, a comparable linguistic model, it seems
and minerals to build and maintain them (Qin, 2023). plausible.
As AI technology develops quickly and becomes obso- Figure 8.1 (Statista.com) shows the energy con-
lete, it may cause an increase in e-waste as outmoded sumed by AI models during training is significant, with
gear is discarded. E-waste poses environmental risks both GPT-3, the first version of the current edition of
due to improper disposal, as it often contains hazard- OpenAI’s popular ChatGPT, and Gopher requiring
ous and toxic elements (Khan, 2023). well over a thousand-megawatt hours of electricity.
Because this is solely for the training model, the over-
IV. ChatGPT’s: Creating carbon footprint all energy consumption of GPT-3 and other LLMs is
expected to be substantially greater. GPT-3, the great-
The phrase “carbon footprint” denotes the total est energy user, consumed nearly the equivalent of
quantity of “carbon dioxide (CO2)” pollutants gener- 200 Germans in 2022. While not enormous, it repre-
ated by a particular person or a company (such as a sents a significant usage of energy.
nation, business, building, etc.). Both immediate emis- While it is undeniable that training LLMs requires
sions from the energy generation that drives consumer a significant amount of energy, the energy sav-
products and services and further emissions from the ings are anticipated to be significant. Any AI model
burning of fossil fuels for industry, transportation, that improves operations by a fraction of a second
and heating make up this total. Additionally, meth- might save hours of shipping, liters of gasoline, or
ane, nitrous oxide, and chlorofluorocarbons (CFCs) hundreds of computations. Each consumes energy,
emissions are frequently taken into consideration and the total amount of energy saved by an LLM
when discussing a carbon footprint idea (An et al., may considerably outpace its energy cost. Mobile
2023; Euronews, 2023). phone carriers are an excellent example, with one-
Kain (2023) stated that the carbon footprint of third expecting AI to lower power usage by 10 per
creating ChatGPT isn’t public information, but if cent to 15 per cent. Given how much of the world
understood correctly, it is based on a GPT-3 variation. relies on mobile phones, this would be a significant
Estimates show that training GPT-3 consumed 1,287 energy saver. The CO2 emissions from training LLMs
MWh and generated 552 tons of CO2. are also significant, with GPT-3 emitting over 500
Mclean’s (2023) research paper stated that it is tons of CO2. This, too, might be drastically altered
most certainly considerably greater than that of GPT- depending on the sorts of energy generation that
3. The energy expenditures would increase if it had to cause the emissions. Most data center operators,
be rebuilt frequently in order to refresh its knowledge. for example, would want to have nuclear energy, a
The amount of carbon dioxide that ChatGPT is esti-
mated to produce annually is 8.4 tons, which is more
than twice as much as the annual emissions of a single
person. i.e., 4 tons.
ChatGPT’s daily emissions of 23.04 kg of CO2 per
day would add up to 414.72 kg CO2 during the course
of 18 days and on the contrary, Big Science Large
Open-science Open-access Multilingual Language
Model (BLOOM) (Luccioni, 2022) released 360
kg of CO2 during the course of 18 days. The differ-
ence between the two emission estimates can be due
to many things, including the varying carbon inten-
sities of the electricity (Jiafu, 2023) generated by Figure 8.1 Power usage for training large language
BLOOM and ChatGPT. It’s also crucial to remember models (LLMs) based on AI in 2023 (in megawatt
that BLOOM handled 230,768 requests in total over hours)
18 days, or 12,820 on average every day. If 1.2% of Source: Statista
 Applied Data Science and Smart Systems 57

notably low-emission energy generator, play a big well over a thousand-megawatt hours of electricity.
role. Because this is solely for the training model, the over-
Figure 8.2 (Statista.com) shows how the develop- all energy consumption of GPT-3 and other large lan-
ing world is making the most aggressive efforts to guage models (LLMs) is expected to be substantially
reduce emissions from AI models used in organiza- greater.
tions. This covers enormous regions in India, Africa, Figure 8.4 (Statista.com) shows CO2 emissions
Latin America, and the Middle East, including a siz- from AI are significant compared to the average
able chunk of the planet and most of its inhabitants. human emission in 2022. GPT-3 training, not includ-
The proportion of organizations tackling emissions in ing the model that is now operating, produced more
those regions is approximately double that of North than a hundred individuals in a year. Training Gopher
America. North American and European organiza- was the equivalent of seventeen Americans’ emissions.
tions are taking fewer steps, which might be due to While the figure may appear considerable, it must be
the fact that the energy utilized to power this technol- seen from the perspective of potential emission reduc-
ogy on those continents is often greener than in the tions through more efficient business strategies.
developing world. It is observed from Figure 8.5 (Statista.com) that
Figure 8.3 (Statista.com) shows the energy con- the power consumption while training AI-based LLM
sumed by AI models during training is significant, with
both GPT-3, the first version of the current edition of
OpenAI’s popular ChatGPT, and Gopher requiring

Figure 8.2 Global share of organizations taking steps

to reduce carbon emissions from AI use in 2022
Figure 8.4 Machine learning (ML) platform emissions
Source: *Includes Hong Kong and Taiwan **Includes India,
Latin America, Middle East, North Africa, and Sub-Saharan in tons of CO2 equivalent in 2022
Africa. Source: Statista

Figure 8.3 Emissions when training AI-based large lan- Figure 8.5 Power consumption when training AI based
guage models (LLMs) in 2022 (in CO2 equivalent tons) large language models (LLMs)
58 Environment and sustainability development: A ChatGPT perspective

bottle, making the overall water footprint for infer-

ence significant given its billions of users. Therefore,
the water footprint of chat GPT and other AI systems
has no doubt become an increasingly significant envi-
ronmental issue as the use of AI continues to grow
(Mclean, 2023).
According to Yadav (2023), the water footprint of
AI refers to the quantity of water used to generate
electricity and provide cooling for data centers that
run AI models. The water footprint has two compo-
nents: direct water consumption and indirect water
consumption.
AI’s water footprint uses a lot of fresh water, which
Figure 8.6 Global share of organizations taking action adds to the problem of water shortage. sourced from
in reducing carbon emissions from their AI use in 2022
natural sources, such as lakes and rivers for manu-
facturing and processing of hardware and semicon-
is maximum in the case of ChatGPT as compared to ductors, as well as for cooling AI infrastructure. So,
Goopher, BLOOM and OPT. this makes the world’s water shortage problem worse.
Figure 8.6 (Statista.com) demonstrates the global Furthermore, water shortage disproportionately
share of organizations taking action to reduce carbon affects vulnerable groups whose survival depends
emissions from their AI use during 2022, the data is on scarce water resources. By allocating water away
displayed according to region. From the figure, it is from regions that need it the most, the water require-
visible that contribution from developing markets ments of AI can worsen already existing inequalities
is highest in taking actions, followed by Asia-Pacific (Alam, 2022; Nova, 2023).
region and Greater China (including Taiwan and Large-scale water withdrawals from rivers and
Hong Kong) and Europe, while North America con- streams can alter natural water flows and lower the
tributes the least. quantity of water available for other purposes, includ-
ing agriculture and drinking water (Srivastava et al.,
2022). This may result in decreased water quality, a
V. ChatGPT: Water footprint and environment
fall in aquatic habitats, and biodiversity loss.
Van (2021) stated in his research paper that the The cooling of data centers generates large amounts
amount of freshwater (David, 2023) utilized to train of wastewater, which can contain a range of pollut-
AI models is referred to as their “water footprint” and ants. If this effluent isn’t correctly handled, it may
run these models. This encompasses both the water have an adverse effect on the environment, causing
necessary to produce electricity and cool the servers local water sources to become
utilized by AI models and the water used to create the Due to the high energy requirements for data center
hardware parts of these models. Although AI models cooling, the energy consumption of AI systems is inti-
use very little water directly, the indirect water con- mately related to the water they use. When fossil fuels
sumption related to their development and upkeep are used to provide this energy, greenhouse gas emis-
can be substantial. sions may be produced, which contributes to climate
Researchers at the University of California, Riverside change (Sharma, 2019). In addition, the manufacturing
have identified the water footprint of AI models, and shipping of the hardware for AI systems consumes
namely ChatGPT-3 and 4. According to the study, energy which increases greenhouse gas emissions.
Microsoft utilized over 700,000 gallons of fresh water Apart from its long-term sustainability should also
for GPT-3 training in its data centers, the same amount be taken into account. If the water footprint problem
of water required to make 370 BMW automobiles is not solved, the growing AI sector can put further
or 320 Tesla cars. This is mostly due to the training stress on the water supply. Since both the growth of
procedure, during which large amounts of energy are AI and the availability of water depend on water, it is
lost and converted into heat, requiring an astonishing essential to address the water footprint.
quantity of freshwater to regulate temperatures and
cool down equipment (Mclean, 2023; Yadav, 2023). VI. Environmental impacts of ChatGPT and it’s
Furthermore, when ChatGPT is utilized for feasibility
activities like replying to queries or producing text,
Additionally, the model consumes a lot of water to Kain (2023) stated that from the consumers’ perspec-
make its inferences. The amount of water used in a tive, it is obvious that there is very little space for
basic chat of 20–50 answers is similar to a 500 ml action in terms of minimizing environmental effects;
 Applied Data Science and Smart Systems 59

nevertheless, providers have a number of options for Figure 8.7 (Statista.com) shows that when AI tools
reducing their digital footprint. were employed, organizations in 2022 were primarily
Table 8.1 represents the steps that are effective in concerned with reducing their physical influence on
reducing the environmental impact of Chat GPT: the environment. This is most likely owing to such
These steps are effective in reducing the environ- enhanced efficiency simply translating to improved
mental impact of ChatGPT. However, their effective- corporate growth and expenses. Organizations priori-
ness depends on the specific circumstances of the data tize ethical product sourcing since it may sometimes
center where ChatGPT is located. Some solutions result in direct cost increases to manufacturing and
may be more feasible than others based on the loca- supply lines.
tion of the data center, the type of hardware used, and
other factors.
For example, optimizing the location of a data
center may not be feasible in all cases. Data centers
may be located in areas with limited access to cooler
climates or water sources. In these cases, alternative
cooling methods or improving energy efficiency may
be more feasible solutions (Mclean, 2023).

1. Improving the organizations environmental im-

pact (e.g., improving energy efficiency, optimiz-
ing transportation)
2. Evaluating sustainability efforts (e.g., bench-
marking)
3. Improving the organization’s governance (e.g.,
regulatory compliance, risk management)
4. Improving the organization’s social impact (e.g., Figure 8.7 Types of sustainability activities in which
sourcing ethical products). respondents’ organizations using AI in 2022

Table 8.1 Effective steps for reducing environmental impact of ChatGPT

Choose required information The cost of training a model may be greatly decreased by utilizing just the necessary
data or by successfully adapting current models for a new purpose, making AI more
viable
Invest in green energy In order to decrease CO2 emissions, efforts are needed to increase the use of renewable
energy sources in data centers. Therefore, it is advisable and essential to rely on cloud
service providers to make sure that electricity is delivered from renewable energy
Reduce unnecessary Unnecessary computations should be reduced in order to lower the overall workload of
computations ChatGPT. By improving the model’s data processing methods and algorithms, this can
be accomplished. Less energy and water will be needed to power ChatGPT by reducing
the amount of computing that is not required
Monitoring and analyzing It’s crucial to routinely track and evaluate ChatGPT’s water usage. Data center
water consumption operators may use this to streamline their processes and find places where water usage
can be decreased
Advocate for greater The creation and maintenance of measurements and standards for assessing the energy
transparency efficiency of creating and implementing ML models is one approach to resolving this
problem (Henderson et al., 2020)
Optimize data center location Locate and promote areas with the potential to host greater amounts of renewable
energy and data centers with reduced carbon footprints
Prolonging life of AI models Increasing the longevity of AI technology and infrastructure through upkeep,
maintenance, and updates can cut down on the production of electronic waste. To
minimize environmental impact, it is also crucial to recycle and properly dispose of old
AI technology
Encourage environmental Environmental awareness is critical because it will help promote responsible AI
awareness industry practices that will open the door for “greener” AI. One tactic for doing this is
to emphasize the limitations of language models and to lessen the excitement around
novel, eye-catching AI systems like ChatGPT. We may actively support new lines of
inquiry that do not simply rely on creating more complicated ones (Zhu, 2023)
60 Environment and sustainability development: A ChatGPT perspective

VII. Conclusion a future in which AI innovations like ChatGPT con-

tribute to a sustainable and environmentally friendly
In conclusion, our study has highlighted the impor-
society by putting the suggestions made in this study
tance of sustainability in AI development and shed
into practice.
light on ChatGPT’s negative environmental effects.
Unquestionably, ChatGPT’s ability to transform sev-
eral sectors and enhance people’s lives is one of its VIII. Recommendations and suggestions
strongest assets. However, the massive amounts of It is drawn from the reports by McKinsey & Co.
computing power needed to develop and maintain and extracted from statista.com that organizations,
ChatGPT have negative environmental effects. nowadays, are trying to improve their environmental
The findings of this research highlight the exces- impact, organization governance, and social impact
sive energy consumption associated with training and while evaluating sustainability efforts. This study
running ChatGPT models. The carbon footprint gen- addresses the environmental effect of ChatGPT and
erated by these processes raises concerns about the offers suggestions for sustainable AI development, but
contribution of AI development to climate change and it also has certain restrictions and opens up avenues
environmental degradation. Moreover, the extensive for future exploration.
data requirements for training ChatGPT raise ethical Some limitations of this study involve the lack of
questions regarding privacy, data collection, and the primary data in this study. As this study relies solely
potential exploitation of personal information. on secondary data, it may be limited by the avail-
To address these issues, the paper proposes several ability and quality of the existing studies and reports.
recommendations for sustainable AI development. Therefore, more comprehensive and accurate empiri-
In the first place, there is a requirement for greater cal data on the environmental impact of AI models
accountability and openness within the AI community. like ChatGPT is necessary to validate the findings and
When training and using AI models like ChatGPT, recommendations.
developers and organizations should be transparent Due to differences in methodologies, reporting
about the energy used and carbon emissions pro- standards, or particular purposes of the research
duced. This transparency will enable researchers, examined, secondary data sources may show biases
policymakers, and the public to make informed deci- or inconsistencies. These restrictions may impact the
sions and encourage the adoption of energy-efficient generalizability and overall reliability of the findings.
practices. Secondary data sources could lack precise contextual
The research also promotes the creation and appli- information regarding the AI models under analysis,
cation of renewable energy sources for the purpose of such as geographical differences, data center locations
powering AI infrastructure. The environmental effect or particular hardware configurations. This restric-
of AI systems may be considerably reduced by switch- tion may affect the findings’ accuracy and suitability
ing from fossil fuels to clean energy. To facilitate this for use in practical situations.
shift, cooperation between energy suppliers, politi-
cians, and AI developers is essential.
The study also stresses the significance of optimiz- IX. Future prospect of the study
ing AI models to lower computing demands without Future research could build upon the existing study by
sacrificing performance. AI systems may reduce their addressing the limitations and exploring the primary
energy consumption and environmental impact using data. In order to strengthen the research, future stud-
methods like model compression, knowledge distilla- ies should take into account primary data collection
tion, and effective algorithms. techniques, such as direct measurements, experiments,
Finally, the study emphasizes the need for contin- interviews and surveys. This strategy would give more
ued AI research and innovation. The development of accurate and thorough details on the energy usage,
sustainable AI solutions can be aided by encouraging carbon footprint, and overall environmental effects
multidisciplinary partnerships between environmen- of AI models. Conducting a holistic assessment of
tal scientists and professionals in artificial intelligence. the environmental impact of AI systems, considering
These initiatives may result in the developing strong, all stages of their lifecycle, including manufacturing,
effective, and environmentally responsible AI models. operation, and disposal, to understand the complete
In conclusion, while ChatGPT and similar AI sustainability picture. While the study emphasized
models hold immense potential, their environmen- ChatGPT’s environmental effect, comparing that
tal impact cannot be ignored. To secure a peaceful impact to that of other AI models or to more con-
coexistence between technological progress and the ventional human procedures can offer a wider per-
preservation of our planet, sustainability in AI devel- spective. Comparative research between various AI
opment is crucial. We can create the conditions for architectures or environmental impact testing of AI
 Applied Data Science and Smart Systems 61

models against human performance would provide nov. J., 1(2), 97–104. doi: https://doi.org/10.5281/
important insights into the relative sustainability of zenodo.7855594.
AI development. Collaborative working with busi- Jiafu, A. N., Wenzhi, D. I. N. G., and Chen, L. I. N. (2023).
ness stakeholders, AI developers, and environmen- Correspondence: ChatGPT: Tackle the growing car-
bon footprint of generative AI. Nature, 615(7953),
tal specialists would promote a multidisciplinary
586. doi:https://doi.org/10.1038/d41586-023-00843-
approach to comprehending and reducing the envi-
2.
ronmental effects of AI models. These collaborations Tanushree, K. (2023). Is ChatGPT harmful to the environ-
might make it easier for people to acquire informa- ment. Sigma Earth. https://sigmaearth.com/is-chatgpt-
tion, share expertise, and create sustainable practices harmful-to-the-environment/.
and norms. Mehreen, K. and Chaudhry, M. N. (2023). Artificial intel-
As time goes on carrying out longitudinal studies ligence and the future of impact assessment. Available
over a lengthy period of time, researchers will be able at SSRN 4519498. doi: http://dx.doi.org/10.2139/
to monitor changes in the environmental effect of AI ssrn.4519498.
models. In order to lessen the environmental impact Ali, K. S., Khuwaja, P., and Dev, K. (2023). ChatGPT needs
of AI models, this would assist to detect trends, tech- SPADE (Sustainability, PrivAcy, Digital divide, and
Ethics) evaluation: A review. arXiv preprint arX-
nological improvements, and viable mitigation tech-
iv:2305.03123. doi: https://doi.org/10.48550/arX-
niques. Adding case studies and field research to
iv.2305.03123.
secondary data analysis will offer insightful informa- Alexandra Sasha, L., Viguier, S., and Ligozat, A.-L. (2022).
tion on the precise environmental effects of AI mod- Estimating the carbon footprint of bloom, a 176b
els in various industries or applications. These studies parameter language model. arXiv preprint arX-
could concentrate on actual situations and take into iv:2211.02001.
account things like the setup of data centers, power Lund, B. D., Wang, T., Mannuru, N. R., Nie, B., Shimray, S.,
sources, and energy-saving techniques. and Wang, Z. (2023). ChatGPT and a new academic
reality: Artificial Intelligence-written research papers
and the ethics of the large language models in schol-
References arly publishing. J. Assoc. Inform. Sci. Technol., 74(5),
Gulzar, A., Ihsanullah, I., Naushad, M., and Sillanpää, M. 570–581. doi:https://doi.org/10.1002/asi.24750.
(2022). Applications of artificial intelligence in water Khattar, N., Singh, J., and Sidhu, J. (2020). An energy ef-
treatment for optimization and automation of adsorp- ficient and adaptive threshold VM consolidation
tion processes: Recent advances and prospects. Chem. framework for cloud environment. Wireless Personal
Engg. J. 427, 130011. doi:https://doi.org/10.1016/j. Comm., 113, 349–367.
cej.2021.130011. Sophie, M. (2023). The environmental impact of ChatGPT:
An, J., Ding, W., and Lin, C. (2023). ChatGPT: Tackle A call for sustainable practices in AI development.
the growing carbon footprint of generative AI. Na- Earth.org. https://earth.org/environmental-impact-
ture, 615(7953), 586. doi:https://doi.org/10.1038/ chatgpt/#:~:text=The/Environmental/Impact/of/Data/
d41586-023-00843-2. Centres&text=According/to/estimates/C/ChatGPT/
Priyanka, B. and Sharma, K. D. (2022). A critical insight emits.
into the role of artificial intelligence (AI) in tour- Kannan, N. (2023). AI-enabled water management sys-
ism and hospitality industries. Pacific Business Rev. tems: An analysis of system components and interde-
Int., 15(3), 76–85. https://www.researchgate.net/ pendencies for Wwater conservation. Eigenpub Rev.
publication/371110624_A_Critical_Insight_into_the_ Sci. Technol., 7(1), 105–124. https://studies.eigenpub.
Role_of_Artificial_Intelligence_AI_in_Tourism_and_ com/index.php/erst.
Hospitality_Industries. David, P., Gonzalez, J., Hölzle, U., Le, Q., Liang, C., Mun-
Biswas, S. S. (2023). Potential use of chatGPT in global guia, L.-M., Rothchild, D., So, D. R., Texier, M., and
warming. Ann. Biomed. Engg., 51(6), 1126–1127. Dean, J. (2022). The carbon footprint of machine
doi:https://doi.org/10.1007/s10439-023-03171-8. learning training will plateau, then shrink. Computer,
David, D. (2023). AI creates new environmental injus- 55(7), 18–28. doi: 10.1109/MC.2022.3148714.
tices, but there’s a fix. News. https://news.ucr.edu/ Chengwei, Q., Zhang, A, Zhang, Z., Chen, J., Yasunaga,
articles/2023/07/12/ai-creates-new- environmental- M., and Yang, D. (2023). Is ChatGPT a general-
injustices-theres-fix. purpose natural language processing task solv-
Euronews. (2023). Chat: What is the carbon footprint er?. arXiv preprint arXiv:2302.06476. doi:https://doi.
of generative AI models? Euronews. https://www. org/10.48550/arXiv.2302.0647.
euronews.com/next/2023/05/24/chatgpt-what- Partha Pratim, R. (2023). AI-Assisted Sustainable Farming:
is-the-carbon-footprint-of-generative-ai-mod- Harnessing the Power of ChatGPT in Modern Agri-
els#:~:text=Researchers%20estimated%20that%20 cultural Sciences and Technology. ACS Agricultural
creating%20GPT. Science & Technology. 460–462. Doi: https://doi.
George, A. S., George, A. S. H., and Martin, A. S. G. (2023). org/10.1021/acsagscitech.3c00145
The environmental impact of AI: A case study of wa- Preeti, S. and Priyanka, P. (2019). Climate change and sus-
ter consumption by Chat GPT. Partn. Univ. Int. In- tainable development: Special context to Paris agree-
62 Environment and sustainability development: A ChatGPT perspective
ment. Proc. Int. Conf. Sustain. Comput. Sci. Technol. Aimee, V. W. (2021). Sustainable AI: AI for sustainabil-
Manag. (SUSCOM). doi:http://dx.doi.org/10.2139/ ity and the sustainability of AI. Spring Link, 1(3),
ssrn.3356829. 213–218. doi:https://doi.org/10.1007/s43681-021-
Sohail, S. S., Farhat, F., Himeur, Y., Nadeem, M., Madsen, 00043-6.
D. Ø., Singh, Y., Atalla, S., and Mansoor, W. (2023). Demetris, V., Chaudhuri, R., and Chatterjee, S. (2023).
Decoding ChatGPT: A taxonomy of existing research, Role of ChatGPT and skilled workers for business
current challenges, and possible future directions. J. sustainability: Leadership motivation as the mod-
King Saud Univ. Comp. Inform. Sci.. 101675. doi: erator. Sustainability, 15(16), 12196. doi: https://doi.
https://doi.org/10.1016/j.jksuci.2023.101675. org/10.3390/su151612196.
Aman, S., Jain, S., Maity, R., and Desai, V. R. (2022). De- Pooja, Y. (2023). Explained: What is the water footprint
mystifying artificial intelligence amidst sustainable ag- of AI and how AI tools are raising environmental
ricultural water management. Curr. Dir. Water Scar. concerns. IndiaTimes. https://www.indiatimes.com/
Res., 7, 17–35. doi:https://doi.org/10.1016/B978-0- explainers/news/explained-what-is-the-water-foot-
323-91910-4.00002-9. print-of-ai-and-how-ai-tools-are-raising-environmen-
Timm, T., Flath, C. M., Weinhardt, C., van der Aalst, W., tal-concerns-603277.html.
and Hinz, O. (2023). Welcome to the era of chat- Jun-Jie, Z., Jiang, J., Yang, M., and Ren, Z. J. (2023).
gpt. The prospects of large language models. Busin. ChatGPT and environmental research. Envi-
Inform. Sys. Engg., 65(2), 95–101. doi: https://doi. ron. Sci. Technol. doi:https://pubs.acs.org/action/
org/10.1007/s12599-023-00795-x. showCitFormats?doi=10.1021/acs.est.3c01818.
9 GAI in healthcare system: Transforming research in
medicine and care for patients
Mahesh A.1,a, Angelin Rosy M.2, Vinodh Kumar M.3, Deepika P.4,
Sakthidevi I.5 and Sathish C.6
1,4
Sri Sairam College of Engineering, Karnataka, India
2,6
Er. Perumal Manimekalai College of Engineering, Tamilnadu, India
3
P.S.V College of Engineering and Technology, Tamilnadu, India
5
Adhiyamaan College of Engineering, Tamilnadu, India

Abstract
GAI also known as generative artificial intelligence, represents a category of artificial intelligence (AI) that possesses the
capability to produce novel content, encompassing images, written text, and music. Although it remains in its emerging
phases of advancement, this technology holds the promise of revolutionizing numerous sectors, ranging from healthcare and
finance to entertainment. The subject of GAI is rapidly developing and holds the capacity to transform the field of health-
care. The adoption of GAI technology has revolutionized the healthcare industry, transforming the way patients are treated
and medical research is conducted. This article explores the many potential applications of GAI in healthcare, including its
ability to improve diagnostic accuracy, optimize treatment, accelerate drug discovery, and enhance medical image analysis.
GAI, as demonstrated by advanced neural network algorithms like variational autoencoders (VAEs) and generative adver-
sarial networks (GANs) enables healthcare practitioners, medical analyst, technologist and scientists to generate realistic
and high-fidelity medical data. Using this technology, medical professionals can improve diagnosis accuracy by combining
varied information about patients, allowing for more robust and individualized treatment strategies. Furthermore, GAI aids
in the generation of realistic medical images, allowing medical practitioners to better grasp and interpret difficult illnesses.
In the field of drug exploration, GAI speeds up the process for determining possible compounds and molecules, saving time
and money over traditional methods. It investigates how GAI encourages interaction among human experts and artificially
intelligent machines, allowing medical practitioners to make better decisions. This complementary partnership takes use of
the capacity of artificial intelligence to analyze large datasets, detect trends, and recommend viable treatment paths, while
human knowledge provides the context-sensitive knowledge required for informed decision-making. Ethical concerns and
obstacles related with the application of GAI to medical procedures are also addressed, with an emphasis on the importance
of responsible application, data protection, and transparency. The healthcare sector aspires ready to bring in a new era of
distinctive effective and cost-effective treatment for patients and research in medicine by adopting the revolutionary potential
of GAI and managing its ethical consequences.

Keywords: Generative artificial intelligence, healthcare, generative adversarial networks, variational autoencoders, ethical

I. Introduction and implications for ethics. Employing an extensive

dataset for machine learning (ML) model training can
In recent years, generative artificial intelligence (GAI)
result in the emergence of a robust GAI system. The
has been gaining significant traction. It is not unex-
model acquires the data’s patterns and framework,
pected that healthcare and GAI are becoming increas-
and it can subsequently be used to produce new
ingly popular. Artificial intelligence (AI) has swiftly
data with comparable features. There are numerous
altered several industries, including the medical field. In
approaches to building a GAI system, but in the fol-
healthcare, one subset of AI which is GAI has emerged
lowing sections some of those that are most common
as an immersive changer (Mondal et al., 2023).
are discussed.
Generative artificial intelligence machines are capa-
ble of producing latest information, images, or even
whole instances of art. The use of this technology a. Generative adversarial networks (GANs)
has enormous potential in healthcare for improving A sort of neural network in which two models com-
diagnostics, drug discovery, patient care, and medical pete against after other. The generator and discrimi-
research. This article investigates the possible applica- nator are two neural networks that compete against
tions and positive aspects of creative AI in healthcare, each other to generate and identify real and fake data.
as well as, the problems related to implementation The two models interact with each other, and the

a
[email protected]
64 GAI in healthcare system: Transforming research in medicine and care for patients

generator improves over time at producing data that in the upcoming years thanks to the ongoing develop-
is accurate. ment of this technology (Singh et al., 2019; Jovanović
et al., 2022; Samant et al., 2022).
b. Variational autoencoders (VAEs)
VAEs are a type of neural network that can compress II. Evolution of GAI
data into a smaller, hidden space, and then decom-
press it back to the original data. The latent space is a The advent of GAI signifies resulted in substantial
space with fewer dimensions that captures the data’s advances in ML and AI. In the following sections,
basic characteristics. After learning to represent data a timeline of how generative AI has progressed is
in the space known as the latent space, the VAE can discussed.
be used to produce new data by sampling points from
the latent space and decoding these again into the a. Beginning principles (1950–1990s)
data that was originally collected space. During the early years of AI research, the underlying
principles of GAI were established. In domains such
c. Recurrent neural networks (RNNs) as natural language processing and music creation,
RNNs are a type of neural network system that is researchers investigated rule-based systems and sym-
capable of processing sequential data. As a result, bolic representations to generate content.
they are well-suited for generating text, music, and
other sorts of data with a periodic order. b. The rebirth of neural networks in the 2000s
The various approaches needed to build a GAI sys- The “deep learning revolution,” or the resurrection of
tem will vary depending on the application. GANs, artificial neural networks, was critical in the creation
for instance, are frequently used to produce images, of generative artificial intelligence. Neural networks
but VAEs are frequently used to generate text. The with deep learning revealed the ability to learn data
following are some of the steps involved in developing hierarchies, allowing for the development of higher-
a GAI system: level and more intricate outputs (Davies et al., 2021).

i. Data collection: The initial step involves gather- c. Variational autoencoders (VAEs) (2013)
ing an extensive array of data pertinent to the VAEs pioneered a probabilistic approach to genera-
intended application. For instance, if the goal is tive modeling. To construct a latent space represen-
to create cat photographs, the initial task entails tation of data, they incorporated aspects from both
amassing a dataset comprising images of cats. generative and recognition models. VAEs enabled
ii. Dataset pre-processing: Before the data can be seamless interpolation and manipulation of latent
employed for model training, it might necessitate space data points.
pre-processing. This step could encompass tasks
such as data cleansing, noise reduction, and nor- d. Generative adversarial networks (GANs) (2014)
malization. GANs, suggested by Ian Goodfellow and colleagues,
iii. Algorithm choice: Multiple models exist for con- represented a significant development in generative
structing a GAI system. The selection of an ap- AI. GANs are made up of the discriminator and gen-
propriate model hinges on the specific applica- erator, two distinct neural networks that compete
tion and the quantity of available data. with one another in a manner akin to a game. While
iv. Model training: The algorithm is subjected to a the discriminator seeks to distinguish between genu-
training process using a dataset. The duration of ine and produced data, the generation process aims to
this process can vary based on the dataset’s size provide data that is as realistic as possible.
and the desired precision of the model, some-
times spanning a considerable timeframe. e. Visualizing the future era (2014–current)
v. Generate fresh data: Following the completion During this period, GAN gained prominence due
of model training, it becomes feasible to employ to their ability to create high-resolution images that
the model for generating novel data. The ap- closely resemble authentic photographs. Renowned
proach employed in this process is contingent on GAN architectures like deep convolutional GAN
the specific framework being used and dictates (DCGAN), StyleGAN, and BigGAN elevated the
the manner in which the new data is crafted. caliber and diversity of the generated images to new
heights (Guo et al., 2022).
While GAI systems are currently in their nascent
stages of research, they hold immense potential to f. Text and language making (2015–current)
revolutionize various industries we may anticipate Progress in the realm of natural language process-
seeing more cutting-edge and significant uses of GAI ing and deep learning has yielded the creation of text
 Applied Data Science and Smart Systems 65

and language generative frameworks. Innovations produce offensive content, or reproduce biases exist-
like long- short- term memory (LSTM) networks and ing in training data sparked debate over ethical imple-
transformers have empowered the generation of logi- mentation and mitigating techniques.
cally connected and contextually fitting textual con-
tent (Sathish et al., 2023). j. Ongoing exploration and advancement (current and
beyond)
g. Music and audio production (2016–current) Researchers are working on ways to improve the
Generative algorithms have been used to compose quality of generated content by making it more
music and synthesize audio. Melodies, harmonies, realistic, diverse, and controllable. Hybrid models
and even full music recordings have been generated that combine different GAI techniques are being
using recurrent neural networks along with differ- developed to improve the performance of models.
ent sequence-to-sequence algorithms. WaveGAN and Creative applications of GAI are being explored
other approaches have also showed promise in pro- in areas such as art, music, and video games.
ducing realistic signals for audio. Interdisciplinary collaborations between researchers
from different fields are helping to advance the state
h. Applications in healthcare and science (2010–cur- of GAI research.
rent) Advances in deep learning architectures, computa-
GAI has been used in healthcare since 2010s for a tional power, and the availability of massive datasets
variety of purposes, including image interpretation, have driven the development of GAI. As technologi-
drug development, and personalized medicine. GANs cal advancements continue, AI with generative capa-
and VAEs have found use in creating artificial medical bilities has the potential to impact a wide range of
images to improve diagnostic accuracy and expand persistence, from entertainment and art to health-
small datasets (Figure 9.1) (Rebecca Perkins et al., care and scientific research in this system (Cai et al.,
2022). 2019).

i. Concerns about ethics and racism (2010–current) III. Construction

Concerns regarding ethical issues and biases arose as
AI that generates gained prominence. The potential Designing and training a neural network architecture
for AI-generated content to propagate disinformation, to produce data that corresponds to the dataset being

Figure 9.1 Overview of GAI

66 GAI in healthcare system: Transforming research in medicine and care for patients

studied is the first step in building a generative AI sys- Autoregressive algorithms generate information
tem. In the following sections, the stages below out- throughout a sequential manner, projecting the next
line the general technique for developing a generative component based on prior components.
artificial intelligence system.
d. Functions of loss
a. Select a generative approach GANs – The generator and discriminator networks
Variational autoencoders, GANs, and autoregressive are adversarial trained. The discriminator strives to
models such as transformers are examples of genera- accurately classify both real and generated data, while
tive models that can be used. As per an individual the generator aims to diminish the discriminator’s
wish, he/she can choose the model that best fits the ability to distinguish genuine from generated data.
data which has to be developed (Figure 9.2). VAEs – To guarantee space of latent information is
well-structured, the model is trained using a combina-
b. Collection of data and pre-processing tion of reconstruction loss (how well the generated data
A wide and representative dataset of the type of data matches the original input) and a regularization term.
(as per wish) is compiled to generate (e.g., photo- Autoregressive models optimize the expected prob-
graphs, document, audio, etc.) (Hajarolasvadi et al., ability distribution over the next element using nega-
2019). tive log-likelihood loss.
The data is pre-processed to ensure that it remains
consistent and in the correct format for the mathemat- e. System training
ical framework of choice. This could include scaling The representation’s parameters are prepared infor-
photos, standardizing the values of pixels, represent- mally. The model is feed with real data (for GANs,
ing text, and so on. this is the discriminator’s input; for VAEs, this is the
input for encoding and reconstruction). Fake data
c. Design of architecture samples are generated and feed to the model.
The design of GANs comprises of a generating net- The loss for both the real and generated data is
work and a discriminator network. The genera- calculated and use back propagation to update the
tor generates samples of data, and a discriminator model’s parameters (Walczak et al., 2018).
attempts to differentiate between genuine and pro-
duced samples. f. The iteration process and refinement
The design for VAEs consists of an encoder net- Continuous training of a computational framework
work, a decoder network, and a latent space in involves multiple rounds of iterative adjustments
between. The encoder converts input data to a lower- to enhance its performance. To avoid over fitting,
dimensional latent space, from which the decoder keep an eye on the model’s results on the validation
produces data. information.

Figure 9.2 Design architecture for GAI

 Applied Data Science and Smart Systems 67

g. Assessing and further refinement fitting. Nearly 10% of the data is used to gauge the
To analyze the quality and diversity of the gener- effectiveness of the model.
ated samples, domain-specific metrics or judgment
by humans are used. To improve outcomes, tweak c. Pre-processing of data
hyper parameters, model architectural design, as well Pictures are resized to a common resolution. Scale
as training procedures as appropriate (Alam et al., pixel values to a standard range, such as 0 to 1. To
2018). boost dataset diversity, supplement data with rota-
tions, flips, and other transformations.
h. Development / The next generation
Following the training session, the generator’s results d. Selection of appropriate model
can be employed to produce new data samples by A convolutional neural network (CNN) architecture
supplying random deep space points (for VAEs) or is choose that is appropriate for image categorization
randomly generated noise vectors (for GANs). when selecting a model.

i. Concerns about ethics and intolerance e. System training

The generative model is checked that it does not Binary cross-entropy loss is used to train the CNN
unintentionally spread biases from the training using the training set. Implement early halting based
data. This can be accomplished by carefully select- on the validation set’s results. During training, keep
ing training data and employing bias-mitigation an eye on measures like accuracy, precision, recall,
approaches such as data augmentation and adver- and score.
sarial training.
Implement ways for dealing with potential ethical f. Testing
issues, such as the creation of sensitive or inappro- Metrics are calculated such as the area under the ROC
priate content. This can be accomplished by remov- curve (AUC-ROC), accuracy, sensitivity, and specific-
ing sensitive content from the training data via filters, ity to assess how well the training model performed
creating standards for how the model should be used, on the test set.
and educating users about the possible risks of utiliz-
ing generative AI. g. Interpretability
Interpretability approaches like Grad-CAM is used to
j. The deployment see which X-ray regions are influencing the model’s
The trained generative model is used in the applica- decisions.
tions such as content generation, data enhancement,
innovative artwork generation, and others. h. Evaluation of bias
Building a GAI system necessitates a thorough Any bias in the predictions made by the AI model is
understanding of both the selected model and the spe- determined and corrected, especially with regard to
cific domain of application. Furthermore, consistent racial or gender-based variables.
tracking, assessment, and continuous enhancement
are required to get the outcomes that are desired. i. Medical validation
To evaluate the sensitivity, specificity, and usability
IV. Experimental methods of the AI model in clinical settings, one should work
closely with radiologists.
The experimental data methodology described is dis-
cussed in the following sections to create a GAI model j. Deployment of system
to aid radiologists in identifying lung nodules in chest In a safe, healthcare-compliant setting, such as a hos-
X-rays. pital’s picture archiving and communication system
(PACS), the AI model is deployed.
a. Dataset
A dataset of chest X-ray pictures is applied which is k. Monitoring and maintenance
tagged as either “nodule” or “no nodule” with their The model should be continuously monitored for per-
related annotations. formance, and it should be updated when new data
become available.
b. Split of data
The dataset is separated into three groups: About l. Ethical considerations
80% of the data used to train the AI model is in the The use of the AI model complies with ethical stan-
training set. A 10% of the data is used in the valida- dards and privacy laws are ensured, particularly with
tion set to adjust the hyper parameters and avoid over regard to patient data.
68 GAI in healthcare system: Transforming research in medicine and care for patients

Creating generative AI models for healthcare that Chatbots for personalized medical care
are both efficient and secure by adhering to these thor- Medical chatbots can be developed by healthcare
ough material and methodology standards, which will institutions to give patients with tailored medical
also help to improve patient care and outcomes. It is information and suggestions. Babylon healthcare, for
kept in mind that successful development of health- illustration, has created a chatbot that uses GAI to ask
care AI necessitates a multi-disciplinary approach and patients about the symptoms they are experiencing
collaboration with healthcare professionals. and provide individualized medical recommendations.

V. GGAI applications in the healthcare industry Treatment of patients

Personalized treatment plans for patients can be cre-
Drug discovery ated using GAI. To generate a personalized plan,
Because of the time-consuming and costly tradi- the algorithm can assess a patient’s medical history,
tional drug-development approach, many drugs take genetic information, lifestyle choices, and other
decades to produce. Generative AI can speed up the aspects. For example, the program can analyze a
process by creating novel drug compounds that have patient’s tumor DNA and identify the genetic abnor-
the potential to be transformed into new medica- malities that are causing the disease. It can then pro-
tions. To speed the drug discovery process, pharma- vide a customized, precise therapy strategy addressing
ceutical professionals can simply employ GAI. The specific genetic alterations. Additionally, GAI can
software program can produce new compounds sim- assist doctors and healthcare practitioners in predict-
ilar to existing medications by learning from a big ing patient outcomes.
collection of chemical structures and their properties.
Scientists can then put these new compounds to the Imaging in medicine
test in laboratory conditions and assess their poten- Medical imaging, such as MRIs, CT scans, and PET
tial as new medications. Identifying potential drug scans, are important components of patient care
candidates and assessing their efficacy and safety are because they assist quickly spot critical injuries and ill-
critical elements in the time-consuming and costly nesses. Here, GAI can help healthcare practitioners by
drug development process. Generative AI can speed providing faster responses and streamlining the imag-
up the process by discovering potential medication ing process. Furthermore, generative AI algorithms
candidates from a vast collection of substances and can reduce image noise. It can also reduce scan times
their properties. when used with ML. It can detect problems in patient
Another application of AI in drug development scans without the need for intervention from humans.
is the creation of virtual substances. AI algorithms The anticipated result of these increased capabilities is
can generate virtual molecules and investigate them faster patient care, which is a vital touch point when-
in silicon (in a computer simulation rather than a ever time is of the essence.
laboratory). As a result, the time and money spent
on researching new medications is dramatically Medical investigation and research
reduced. Scientists can utilize generative AI to cre- Scientists can employ GAI to accelerate medical
ate novel compounds in order to find new medica- research. A massive dataset of scientific literature can
tions. The program can learn from a large database be utilized to train the methodology which can subse-
of chemical structures and attributes. It can then quently uncover patterns relevant to certain research
design novel chemicals that are suited to a specific fields. This can help academics generate new research
target. topics and perspectives.

Disease diagnosis Individualized treatment plans

Generative AI has the potential to transform disease By analyzing the massive amounts of patient data and
diagnosis by examining extensive collections of medi- providing treatment recommendations based on that
cal images to detect patterns associated with particular data, GAI can construct tailored plans for treatment.
conditions. For instance, dermatologists can employ
generative AI to diagnose skin cancer by scrutinizing Simulation in medicine
a vast dataset of skin images. AI can recognize indica- Healthcare workers can use GAI to create medical
tive patterns, thereby assisting healthcare profession- simulation to aid with practical knowledge.
als in rendering more precise and timely diagnoses.
Furthermore, generative AI can be applied to analyze Documentation of clinical trials
various other forms of medical imagery, including CT Medical documentation is accomplished by record-
scans, X-rays, and MRIs, to diagnose a diverse array ing and summarizing physician–patient consulta-
of diseases (Wang et al., 2018). tions. This immediately consolidates paperwork by
 Applied Data Science and Smart Systems 69

Figure 9.3 Decision support system for patients

capturing data, producing electronic health records, personalized patient care. However, the challenges
and reducing complex medical terminology for and ethical considerations inherent in integrating
patient comprehension (Figure 9.3). generative AI within healthcare must be thoroughly
deliberated upon. With ongoing exploration in addi-
VI. Challenges generative AI in healthcare tion improvement, potential for generative artifi-
cial intelligence to reshape healthcare and amplify
Although AI that regenerates has enormous potential patient well-being in the coming years remains
in healthcare, several problems must be overcome. substantial.

• Interpretation and trust: The created content can

References
be difficult to interpret at occasions. The inability
to comprehend the algorithm’s decision-making Mondal, S., Das, S., and Vrana, V. G. (2023). How to bell
procedure will have a consequence on confidence. the cat? A theoretical review of generative artificial
• Gathering huge datasets to use as training might intelligence towards digital disruption in all walks
be difficult, limiting effectiveness in particular ar- of life. Technologies, 11, 44. https://doi.org/10.3390/
technologies11020044.
eas.
Jovanović, M. and Campbell, M. (2022). Generative arti-
• Transparency is critical for resolving biases and ficial intelligence: Trends and prospects. Computer,
mistakes and creating confidence between physi- 55(10), 107–112. doi: 10.1109/MC.2022.3192720.
cians and patients. Samant, R. M., Bachute, M. R., Gite, S., and Kotecha, K.
• Privacy, security, and algorithmic bias raise ethi- (2022). Framework for deep learning-based language
cal considerations, needing careful attention to models using multi-task learning in natural language
avoid inequities in healthcare results. understanding: A systematic literature review and
future directions. IEEE Acc., 10, 17078–17097. doi:
10.1109/ACCESS.2022.3149798.
VII. Conclusion Sathish, C. Mahesh, A., Karpagam, N. S., Vasugi, R., In-
Through enhancements in diagnostics, the accel- dumathi, J. and Kanchana, T. Intelligent email auto-
eration of drug development, the customization of mation analysis driving through natural language
therapies, and the facilitation of medical research, processing (NLP). 2023 Second Int. Conf. Elec.
Renew. Sys. (ICEARS), 1612–1616, doi: 10.1109/
generative artificial intelligence stands poised to
ICEARS56392.2023.10085351.
revolutionize the healthcare system. By harnessing Alex, D., Veličković, P., Buesing, L., Blackwell, S., Zheng,
the capabilities of GAI, healthcare professionals can D., Tomašev, N., Tanburn, R. et al. (2021). Advanc-
potentially attain heightened accuracy in diagno- ing mathematics by guiding human intuition with AI.
ses, pioneer novel pharmaceuticals, and administer Nature, 600(7887), 70–74.
70 GAI in healthcare system: Transforming research in medicine and care for patients
Guo, S., Wang, Y., and Yang, W. (2022). A study on the Cai, Q., Wang, H., Li, Z., and Liu, X. (2019). A survey on
collision of artificial intelligence and art based on multimodal data-driven smart healthcare systems:
generative adversarial networks (GAN). 2022 Int. Approaches and applications. IEEE Acc., 7, 133583–
Conf. 3D Immer. Interac. Multi-sens. Exp. (ICDI- 133599. doi: 10.1109/ACCESS.2019.2941419.
IME), Madrid, Spain, 27–31, doi: 10.1109/ICDI- Hajarolasvadi, N. and Demirel, H. (2020). Deep facial emo-
IME56946.2022.00014. tion recognition in video using eigenframes. J. IET
Perkins, R., Jeronimo, J., Hammer, A., Novetsky, A., Guido, Image Proc., 14 (14), 3536–3546. doi:10.1049/iet-
R., del Pino, M., Louwers, J., Marcus, J., Resende, C., ipr.2019.1566.
Smith, K., Egemen, D., Befano, B., Smith, D., Antani, Singh, Jaiteg, and Nandini Modi. (2019). Use of informa-
S., de Sanjose, S., and Schiffman, M. (2022). Compari- tion modelling techniques to understand research
son of accuracy and reproducibility of colposcopic trends in eye gaze estimation methods: An automated
impression based on a single image versus a two-min- review. Heliyon, 5, 1–12.
ute time series of colposcopic images. Gynecol. On- Steven, W. and Velanovich, V. (2018). Improving progno-
col., 167(1), 89–95. ISSN 0090-8258, https://doi.org/ sis and reducing decision regret for pancreatic cancer
10.1016/j.ygyno.2022.08.001. treatment using artificial neural networks. Dec. Sup-
Alam, F., Ofli, F., and Imran, M. (2018). Processing social port Sys., 106, 110–118.
media images by combining human and machine com- Wang, Q., Shen, L., and Shi, Y. (2020). Recognition-driven
puting during crises. Int. J. Human Comp. Interac., compressed image generation using semantic-prior in-
34(6), 311–2258. doi:10.1080/10447318.2018.1427 formation. J. IEEE Sig. Proc. Lett., 8(9), 1150–1154.
831. doi:10.1109/LSP.2020.3004967.
10 Fuzzy L-R analysis of queue network with priority
Aarti Saini1,a, Deepak Gupta2, A. K. Tripathi3 and Vandana Saini4
Maharishi Markandeshwar Engineering College (Deemed to be University) Mullana, Haryana, India
1,2,3

1
Govt College for Women, Shahzadpur (Ambala), Haryana, India
4
Govt College, Naraingarh (Ambala), Haryana, India

Abstract
This paper is the fuzzy analysis of a queue network model with the assumptions of pre-emptive priority discipline on biserial
subsystems and general arrival is on parallel subsystems. It is presupposed that service time and the interval between two
succeeding arrivals follow the Poisson distribution. Both arrivals and service costs are fuzzy in nature. Performance of the
purposed model evaluated by using L-R fuzzy numbers. L-R method is more flexible and simplest method for fuzzy analysis
as compared to other existing methods. Fuzzy triangular number and all classical formulae used to calculate fuzzy queue
characteristics. A numerical calculation well illustrated the results.

Keywords: Fuzzy number, priority, parallel channels, biserial server, L-R method

I. Introduction analyzed fuzzy queue models with DSW algorithm

and n-policy queues in finite, infinite capacity. Ritha
Mathematical analysis of waiting lines in queuing
and Josephine (2017) evaluated priority queue model
theory is an important aspect because it provides a
by fuzzy L-R method, L-R technique was applied by
way to shorten queue lengths and waiting time. In the
Mukeba et al. (2015, 2016) to measure performance
present networking systems, queuing models are very
of queues and single server retrial queues in uncer-
useful to improve the efficiency of any service organi-
tainty, The L-R technique was used by Saini and
zations. Sometimes in queuing model, customers are
Gupta, and A. K. Tripathi (2022) to study feedback
served on priority base. An efficient priority queues
and the varied behavior of servers in a probabilistic
have great significance in providing quality of service
and fuzzy environment.
to different class of customers. Priority queue applica-
In the present paper, using L-R triangular fuzzy
tions can be found in communication networks, hos-
integers and fuzzy arithmetic operations, we are
pitals, service industries, banks, inventory controls,
attempting to analyze the performance indicators
transportations, check-in-counters at airports, etc. In
of the purposed priority queue model in a fuzzy
literature most of the work was on analysis of prior-
environment.
ity queues. But in priority queuing model, input data
is uncertain to remove this uncertainty fuzzy logics
have been used. II. Definitions
Most of the researchers like Prade (1980), Li Fuzzy set
and Lee (1988, 1989), Kao et al. (1999), T. P. Singh If the result of the membership function for a function
et al. (2010, 2015, 2016), Devaraj and Jayalakshmi defined on the universal set X is either
(2012), Gupta D. and Sharma S. (2011, 2013, 2015), or where x is modal value of the
B. Kalpana et al. (2021) extensively studied fuzzy function is said to be fuzzy.
queue characteristics by α-cut Zadeh extension prin-
ciple. Kao et al., developed membership function for Fuzzy triangular number
fuzzy queue characteristics with the use of parametric A number is a fuzzy triangular number
linear programming. B. Kalpana et al. (2018) applied and membership function is defined as
non-linear programming in fuzzy on non-preemptive
priority queues. Selvakumaria and Revathi (2021)
used new ranking method with triangular and trap-
ezoidal numbers to measure the effectiveness of the
non-preemptive priority queue model. Ritha and
Robert (2009), Ritha and Menon (2011), Ning Y.
et al. (2009), Wang et al. (2010), Srinivasan (2013)

a
[email protected]
72 Fuzzy L-R analysis of queue network with priority

Fuzzy L-R number Notations

A number is fuzzy L-R ⇔ three real = fuzzy low and high priority arriving customers,
number m1, f1, h1 > 0 as well as two continuous, posi- i = 1,2 & j = L, H
tive and decreasing functions L and R, from R to [0,1] = fuzzy Priority input rate, i = 1,2 & j = L, H
exist, such that = fuzzy general arrivals, i = 1,2
L (0) = 1, L (1) = 0, L (x) > 0, limx→∞ L(x) = 0 = fuzzy cost of service for low and high priority
R (0) = 1, R (1) = 0, R (x) > 0, limx→∞ R(x) = 0 visitors, i = 1,2 & j = L, H
´
= fuzzy service rate at parallel subsystems
= fuzzy probabilities from i’th server to j’th server
= fuzzy queue length of the system

III. Stochastic mathematical modeling

The proposed model consists of biserial and parallel
A fuzzy number is represented in L-R form as its
severs C1 and C2 both linked to server C3. The sub-
L-R representation is of the form
systems C11 and C12 are in bi-series relation and C21
where m1, f1, g1 are used as modal value, left and right
and C22 are parallel at server C1 and C2 respectively.
spread of , respectively.
Both type of customers with Poisson Mean arrivals
Supp ( ) = (m1 - f1, m1 + g1)
arrived at subsystems C11, C12,
C21 and C22 for availing services with probable con-
L-R fuzzy arithmetic
ditions α12 + α15 = 1, α21 + α25 = 1, α35 = 1, α45 =
Let us consider two L-R fuzzy number
1, where priority is taken only at entry level biserial
& and define L-R fuzzy arithmetic
subsystems C11 and C12. After, that customer move for
operations on them as
next phase service at C3 and finally leave the system
(Figure 10.1).
Let us, define a probability function
(t) at any time t for the
arrivals m1L, m1H, m2L, m2H, m2, m3, from outside in
the system. The model’s continuous solution is derived
from the solutions of differential equations by using
GF and PGF solution methodology as

To solve the above equation, by applying L’ hos- = 1. And find the utilization factors at different servers
pital rule with conditions | Z1|=| Z2|=| Z3|=| in stochastic environment
Z4|=|Z5|=|Z6|=|Z7|=1 and H(Z1, Z2, Z3, Z4, Z5, Z6, Z7)
 Applied Data Science and Smart Systems 73

Table 10.1 Crisp values

m1L = 2 λ1L = 2 µ1L = 9 α12 = 0.4

m1H = 4 λ1H = 4 µ1H = 16 α15 = 0.6
m2L = 3 λ2L = 3 µ2L = 12 α21 = 0.3
m2H = 5 λ2H = 5 µ2H = 15 α25 = 0.7
m2 = 3 λ′1 = 4 µ′1 = 10 α35 = 0.6
m3 = 4 λ′2 = 3 µ′2 = 13 α45 = 0.8
µ3 = 27

Figure 10.1 Priority queue network model

Time- independent solution of the proposed model is

that

Fuzzified model
With the conditions exist if γ1, γ2, γ3, γ4, γ5, γ6, γ7 ≤ 1
Let us represent approximate crisp parameters
IV. Numerical illustration in the form of fuzzy numbers as
For particular crisp values, we get then
Using Table 10.1, utilization factor and queue from stochastic environment results, the fuzzy utiliza-
length is tion factor and queue characteristics can be written as
74 Fuzzy L-R analysis of queue network with priority

VII. Results
Fuzzy lengths of queues • Utilization of first server by high priority cus-
tomers lies between 0.1504 and 0.8271. Utiliza-
tion factor and partial queue length’s maximum
allowed values are 0.3906 and 0.6410. Utiliza-
tion of first server by low priority customers lies
between 0.2061 and 1.5529. The partial queue
lengths and Utilization factor maximum allowed
Average waiting time
values are 1.6738 and 0.6260.
• Utilization of second server by high priority
customers lies between 0.1772 and 0.8717.
Utilization factor and partial queue length’s
maximum allowed values are 0.4167 and
VI. Numerical illustration 0.7144. Utilization of second server by low
Table 10.3 is the fuzzy L-R representations of fuzzy priority customers lies between 0.2695 and
triangular numbers from Table 10.2. 1.6776. Utilization factor and partial queue
Using these numerical values, we get L-R represen- length’s maximum allowed values are 0.7045
tations of traffic intensity at servers are, and 2.3841.
• Utilization of third server lies between 0.2679
and 1.2498. Utilization factor and partial queue
length’s maximum allowed values are 0.5555 and
1.2497.

Table 10.3 L-R Fuzzy Values

Arrival times Service costs Probabilities

Modal values of are 0.3906, = (2,1,1) = (14,1,1) = (0.4,0.2,0.2)

0.4167, 0.5555, 0.4, 0.913, 0.6260, 0.7045 and for = (4,2,2) = (16,2,2) = (0.6,0.2,0.2)
are 0.6410, 0.7144, 1.2497, = (3,1,1) = (15,1,1) = (0.3,0.1,0.1)
0.6, 3.4984, 1.6738, 2.3841, respectively.
= (5,2,2) = (18,2,2) = (0.7,0.1,0.1)
= (4,1,1) = (12,2,2) = (0.6,0.2,0.2)
= (3,2,2) = (15,2,2) = (0.5,0.2,0.2)
= (27,1,1)

Table 10.2 Fuzzy particular values

Customers in queue Arrival times Service costs Probabilities

m1L = 2 = (1,2,3) = (13,14,15) = (0.2,0.4,0.6)

m1H = 4 = (2,4,6) = (14,16,18) = (0.4,0.6,0.8)
m2L = 3 = (2,3,4) = (14,15,16) = (0.2,0.3,0.4)
m2H = 5 = (3,5,7) = (16,18,20) = (0.6,0.7,0.8)
m2 = 3 = (3,4,5) = (10,12,14) = (0.4,0.6,0.8)
m3 = 4 = (1,3,5) = (13,15,17) = (.3,.5,.7)
= (26,27,28)
 Applied Data Science and Smart Systems 75

• Utilization of fourth server lies between .084 and Mittal, Meenu, T. P. Singh, and Deepak Gupta. (2015).
1.2821. Utilization factor and the length of par- Threshold effect on a fuzzy queue model with batch
tial queue maximum potential values are 0.4 and arrival. Arya Bhatta Journal of Mathematics and In-
0.6. formatics. 7(1), 109–118.
Singh, T. P., Mittal, M., and Gupta, D. (2016). Modelling of
• Utilization of fifth server lies between 0.2153
a bulk queue system in triangular fuzzy numbers using
and 1.4258. Utilization factor and partial queue
α-cut. Int. J. IT Engg., 4(9), 72–79.
length’s maximum potential values are 0.6087 Devaraj and Jayalakshmi. (2012). A fuzzy approach to pri-
and 3.4984. ority queues. Int. J. Fuzzy Math. Sys., 2(4), 479–488.
Gupta, D., Sharma, S., and Gulati. (2011). On steady state
VIII. Conclusion behavior of a network queuing model with bi-serial
and parallel channels linked with a common server.
In the present work, based on L-R fuzzy arithmetic Comp. Engg. Intel. Sys., 2(2).
operations, priority queues have been analyzed by Seema, Gupta, D., and Sharma, S. (2013). Analysis of bise-
the L-R technique. This method is used to evaluate rial servers linked to a common server in fuzzy envi-
numerical values of various performances of queues ronment. Int. J. Comp. Sci. Math., 68(6), 26–32.
like traffic intensity and length of queues at different Sharma, S., Gupta, D., and Seema. (2015). Network Aanaly-
servers in fuzzy environment. Fuzzy L-R representa- sis of fuzzy bi-serial and parallel servers with a multi-
stage flow shop model. 21st Int. Cong. Model. Simul.
tion is more informative than basic classical methods
Gold Coast Australia, 697–703.
in stochastic environment. For this numerical calcu-
Kalpana. (2021). Evaluation of performance measures of
lation is used to authenticate the study. While using fuzzy queues with preemptive priority using different
same approximate crisp and fuzzy data, then deter- fuzzy numbers. Adv. Appl. Math. Sci., 20(11), 2975–
mine the outcomes in the event of precise numbers for 2985.
the fraction of both type customers high and low pri- Kalpana and Anusheela. (2018). Analysis of fuzzy non-pre-
ority using the first server are 39.06% and 75.68%, emptive priority queue using non-linear programming.
second server usage by high and low priority custom- Int. J. Math. Trends Technol. (IJMTT), 56(1), 71–80.
ers is 50% and 85.98%, third, fourth and fifth server Selvakumaria and Revathi (2021). Analysis of fuzzy non-
usage are 66.66%, 28.85% and 77.77%, respectively. preemptive priority queuing model with unequal ser-
Accessing these servers while dealing with ambiguous vice rate. Turkish J. Comp. Math. Edu., 12(5), 1457–
1460.
data is 39.06% and 62.60%, 41.67% and 70.45%,
Rita, W. and Robert. (2009). Application of fuzzy set theory
55.55%, 40% and 60.87%, respectively. Thus, from
to retrial queues. Int. J. Algorith. Comput. Math., 2(4),
results we can observe that utilization of 1st and 2nd 9–18.
server by high priority customers is approximate same Ritha, W. and Menon, S. B. (2011). Fuzzy n policy queues
but utilization of servers in stochastic environment by with infinite capacity. J. Phy. Sci., 15, 73–82.
low priority customers is high as compared to fuzzy Ning and Zhao. (2009). Analysis on random fuzzy queu-
environment. The usage of 4th sever in crisp data is ing systems with finite capacity. 9th Int. Conf. Elec.
28.85% and in fuzzy data is 40%. Thus, the study in Busin., 1–7.
future can be extended for more queuing models with Yang, W. and Li. (2010). Fuzzy analysis for the n-policy
batch arrival, priority arrivals on parallel subsystem queues with infinite capacity. Int. J. Inform. Manag.
and biserial servers instead of parallel subsystem. Sci., 21, 41–45.
Srinivasan. (2013). Fuzzy queuing model using DSW algo-
rithm. Int. J. Adv. Res. Math. Comp. Appl., 1(1), 1–6.
References Ritha, W. and Vinnarasi, J. S. (2017). Analysis of priority
queuing models: L-R method. Ann. Pure Appl. Math.,
Prade, (1980). An outline of fuzzy or possibilistic mod-
15(2), 271–276.
els for queuing systems. Wang P. P. and Chang S. K.
Mukeba, J. P., Mabela and Ulungu. (2015). Computing
(eds), Fuzzy Sets. Plenum Press. 147–153. https://doi.
fuzzy queuing performance measures by L-R method.
org/10.1007/978-1-4684-3848-2_13.
J. Fuzzy Sets Valued Anal., 1, 57–67.
Li and Lee. (1988). Analysis of fuzzy queues. Proc. NAFIPS,
Mukeba, J. P. (2016). Application of L-R method to single
158–162.
server fuzzy retrial queue with patient customers. J.
Li and Lee. (1989). Analysis of fuzzy queues. Comp.
Pure Appl. Math. Adv. Appl., 16(1), 43–59.
Math. Appl., 17(7), 1143–1147. http://dx.doi.
Saini, V., Gupta, D., and Tripathi, A. K. (2022). Analysis
org/10.1016/0898-1221(89)90044-8.
of heterogeneous feedback queue model in stochastic
Li, K. and Chen. (1999). Parametric programming to the
and in fuzzy environment using L-R method. Math.
analysis of fuzzy queues. Fuzzy Sets Sys., 107, 93–100.
Stat., 10(5), 918–924.
http:/dx.doi.org/10.1016/S0165- 0114(97)00295-9.
Singh, T. P., Kusum, and Gupta, D. (2010). On network
queue model centrally linked with common feedback
channel. J. Math. Sys. Sci. 6(2), 18–31.
11 Blood bank mobile application of IoT-based android
studio for COVID-19
Basetty Mallikarjuna1, Sandeep Bhatia2,a, Neha Goel3, and
Bharat Bhushan Naib4
1
Department of Information Technology, Institute of Aeronautical Engineering, Dundigal-500043, Tamil Nadu, India
2,4
School of Computing Science and Engineering, Galgotias University Greater Noida, Uttar Pradesh, India
3
Department of Electronics & Communication Engineering, RKGIT, Ghaziabad, India

Abstract
It is impossible to manufacture the blood, as it can be given by the donors. Blood bank retrieval information can be given
through the android studio application, but there is not much work on the integrated environment like IoT sensor connected
with android studio application development. This paper provides the IoT healthcare sensors connected to the android
studio mobile application development blood donors and blood receivers. The mobile application is most useful in an
emergency during the COVID-19 pandemic. This observational study gives the web-based application development and also
android-studio mobile application development for blood bank information retrieval system. The results are carried out in a
real-time environment and updated features of the blood bank mobile application.

Keywords: Internet of things, COVID-19, blood bank, mobile application, android studio

I. Introduction okkadunnadu becomes popular with this concept.

Many the people required the blood of different types
In past years, finding a blood donor or a specific blood
as shown in Table 11.1 (Fahim et al., 2016).
group in an emergency is very difficult as sometimes
The 4 common types of blood groups such as A, B,
may be due to the rare blood group or maybe the
AB, and O, which was invented by Karl Landsteiner
blood group is not available, this problem is increas-
in 1901, on his birthday celebrated as “World Blood
ing day by day. Blood bank application development
Donors Day”. India celebrated “National Blood
through android application with the IoT sensors is
Donations Day” on 1st October (Fahim et al., 2016).
the best solution in the COVID-19 pandemic (Bassam
If a person starts to donate blood at age of 18 every
et al., 2021). Through the blood bank, application
90 days until he reached 60, he/she would have 30
users can easily save their time and effort (Kayode et
gallons of blood and save 600 lives approximately
al., 2019). In the COVID-19 pandemic every two sec-
(Fahim et al., 2016).
onds, some need blood as per the WHO reports (Priya
In this article, develop the blood bank application
et al., 2014).
with IoT healthcare sensors on android studio pro-
People have to stand in a long queue for blood
gramming, that application asks at the time of reg-
requirements and ask the blood in different places.
istration process, donors or receivers’ names, phone
People don’t have money to purchase the blood; it is
numbers, location, and blood group (Altameem et
very difficult to get the blood during the COVID-19
al., 2022). The user login through the credentials
pandemic period (Fahim et al., 2016). In the metro-
and can easily find the receivers or donors’ blood
politan cities, people feel difficult to give or get the
group. The report describes the layout and coding
blood, and also it is problematic for who is coming to
of the application. Blood bank applications can also
give their or donate blood, the perfect mobile appli-
be developed using the Java programming concept
cation is required for donors and receivers in nearby
(Krishna et al., 2019). This project describes the
areas.
application through which users can easily save
There are several types of blood groups, several
time by asking the blood from people in a nearby
hospitals most often blood group type “O”, during
location. Sometimes traffic can cause life. The appli-
the COVID-19 approximately 1 million people are
cation will also reduce the management cost. The
diagnosed every 3–4 hours as per WHO recorded
proposed application provides a real-time, robotic
news (Altameem et al., 2022). The rarest blood group
structure. The following objectives of this study are
was “hh” or Bombay blood group was discovered
as follows:
by Dr. Y. M. Bhende in Bombay 1952. Movies are

[email protected]
a
 Applied Data Science and Smart Systems 77
Table 11.1 Frequency of occurring in different blood blood information management mobile application
groups (Fahim et al., 2016). which has its mobile search engine used to search for
blood donors and receivers from the registered appli-
S. No Approximate frequency of occurring blood type
cation. This study also provides that registered users
1 O +ve: 1 person might be among 3 persons send a notification to donors and receivers. The pro-
2 A +ve: 1 person might be among 3 persons posed application also has certain disadvantages, it
requires an internet connection and manages particu-
3 O -ve: 1 person might be among 15 persons
lar functions required for a large database.
4 A -ve: 1 person might be among 16 persons In this paper section 2 deals with the related work
5 B +ve: 1 person might be among 12 persons existing to differentiate the proposed methodology,
6 B -ve: 1 person might be among 67 persons section 3 deals with the methodology, section 4 pro-
7 AB +ve: 1 person might be among 29 persons vides the implementation, and section 5 deals with the
8 AB -ve: 1 person might be among 167 persons
conclusion followed by references.

Incident Approximate estimation

usage of blood I. Related work
Most of the research work (Kayode et al., 2019) is on
1 Automobile minor Approximately 45 to 50
android-based mobile blood bank application devel-
accident units of blood
opment and information retrieval procedure. The
2 Heart surgery for a Approximately 6 units of
existing research work (Shah et al., 2022) on android
single patient blood
applications and most of the recent work are relevant
3 Organ Approximately 40 units
to the blood bank management web portal applica-
transplantation for of blood
a single patient tion development but not interconnected with the
IoT sensors. The work existing (Priya et al., 2014) on
optimized blood donor information systems, covers
• It shares the blood donors or blood receivers’ re- all blood donors’ information processing approaches
quests for urgent blood in the community of the but not integrated environment on mobile healthcare
city as per the location and can find donors in the application environment. In healthcare, World Health
current city. Organization and Health care Medical Information
• It is required to find the blood at an emergency System (HMS) said, people needed for convenient
time and a shortage of time. mobile blood bank application system (Fahim et al.,
• This proposed application provides real-time in- 2016).
formation about the availability of the blood do- The aim of this proposed work related to the inte-
nors’ nearby location. grated environment with IoT with android studio
• Users can easily request the blood or can give it mobile application provides the real-time environ-
by providing the basic details at the time of reg- ment with mobility with GPS connectivity, those who
istration. needed blood in the COVID-19 pandemic as per medi-
cal emergency, to supply the blood as early as possible.
As per the existing system, people always rush The author’s previous work (Altameem et al.,
into the hospital’s long queue in the blood bank in 2022) automated brain tumor detection worked on
hospitals, it is sometimes impossible to find the spe- web-based application development; this work is a
cific blood group in the given time. To overcome this feature enhancement of the previous work, more than
problem, the given observational study and metrol- 38,000 people needed for the blood every day. Blood
ogy provide a better way to save lives. bank android-based IoT application system reduces
the manual activity and which saves the labor cost
• It saves the time of the users who are looking (Krishna et al., 2019) blood bank record-keeping has
for blood in a long queue in hospitals and blood been carried out manually over the past decades using
bank centers. files allocation, the upcoming technologies are most
• To Searching for the blood of a particular person invented in this field using blockchain (Mallikarjuna
can take a long time, in this period can lose their et al., 2021, Singh et al., 2021). The blood donors’
life. and receivers’ data keep securely, for that required
• The proposed application decreases the manage- blockchain technology (Mallikarjuna, 2022), the web-
ment cost. based blood management system with IoT healthcare
endowment provides the smart home automation
The proposed application has a high featured mobile (Khan et al., 2021), the blood donation activity is an
application to update as per the real-time data of the important objective of the society (Mufaqih et al.,
78 Blood bank mobile application of IoT-based android studio for COVID-19

2020). The blood bank management system devel-

oped with the cloud environment (Arifin et al., 2021).
The web-based blood bank management system is
a very important and crucial issue for quick access
of donor’s information, it monitors blood donation
activities and receivers’ information and prediction
are challenging tasks, organization and time manage-
ment and decision-making are upcoming features in
this area (Narang et al., 2019). The blood bank man-
agement system consists of different modules such as
the patient module, donor module, and blood mod-
ule, the responsibility of this approach has the user
connected with the administrator (Pohandulkar et al.,
2018) approach not suitable for the COVID-19 sce-
nario (Reddy et al., 2016).
The proposed approach overcomes the all disadvan- Figure 11.1 The use case diagram of the blood bank
tages of the existing approach (Prasad et al., 2013), the
currently developed application brings the donors and
appropriate receives gathered into one place (Tatale android device, the android studio device the con-
et al., 2020), and also provides donors and receives nection through the internet and establish between
through the chabot application environment. The afore- the android studio device to healthcare sensors. The
mentioned related work (Sastry et al., 2019; Prasad et application is configured to be connected through
al., 2018) reviewed and updated the proposed and pro- the internet and easily programmable to the android
vide quick response during the COVID-19 pandemic. studio device. The android device gathers the data
The current observational study also provides the web- from the sensors, the user registered with his user’s
based and android mobile application for a better and name and password, the android program executes
more effective integrated platform of information man- inside the device and these data can be recorded in
agement of blood bank on IoT-based healthcare sen- the real-time database and generates alerts to the user.
sors with an android studio application environment. The use case diagram of the application is shown in
(Bhatia et al., 2023) focuses 4G to 8G communica- Figure 11.1. The user interacted with the major blood
tion in IoT and its impact in IoT application. Ganai bank application and registered through the login, the
et al. 2022 highlights the security and privacy issues in application interacts with the IoT sensors and man-
IoT devices as they are vulnerable. Bhatia et al. (2023) ages the modules from pop-up reports, client and
focuses on upgradation in IoT communication from server details, manage profiles, message details, client
3G to 7G and IoT reliability in mobile application. details, conference details, etc.
The IoT healthcare technologies simple and efficient
III. Objectives accessible through the measuring and recording of
the real-time data of the user and connected through
This paper is aimed to design mobile application the information android studio programming. The
which is IoT enabled for blood bank. The main objec- low-cost IoT sensors such as Altimeter, ECG, EMG,
tive is to create a mobile application which can be run Actuators, and Microcontrollers, not only provide the
on android device, the android studio device the con- blood information and also give the various medical
nection through the internet and establish between issues. The current user is easily aware of the symp-
the android studio device to healthcare sensors. The toms during the COVID-19 pandemic as shown in
application is configured to be connected through the Figure 11.2.
internet and easily programmable to the android stu- The API interface is made on android studio plus,
dio device. the user has to register to the app by providing some
basic details like his/her name, mobile number, blood
IV. Methodology group and city. The user can log in to the app by using
the credentials. The user can now enter the app after
This methodology deals with the development of successful login and then the user can see the people
blood application with android studio with IoT sen- requesting blood or maybe they are happy to give you
sors integrated mobile healthcare application devel- the blood or become a donor.
opment, this project implemented with the low-cost You can share the details on different apps by using
sensors which are shown in Figure 11.1, the blood the share button. You can easily call the user by the
bank application program which can be run on icon/feature of calling present in the application.
 Applied Data Science and Smart Systems 79

Figure 11.2 The sensors connected to the android

­studio

The user can also search for the blood in any city by
clicking on the search button at the top and provid-
ing the details like which blood group the user wants
and in which city and it will show you the details if
anybody was there. The user can also become a donor
so the user can save any life by giving the blood group Figure 11.3 Snapshot of the API of android studio
and people can see who is given the blood in that city.
The activity is divided into several parts like login
activity to increase the security of the application.
Then the main activity tells about the people need-
ing the blood and also the activity where the user can
request the blood or become a donor. So, it’s a very
modern compact and the best application to over-
come a serious problem as shown in Figure 11.4.
One of the significant and important of this obser-
vational study is not alert the blood and also provides
the various healthcare issues of the registered users
during the COVID-19 pandemic.

V. Implementation
To create a simple android application project, to set
up the application for the following steps below:
File -> New -> Select New.
Fill in all the entries shown in the above Figure
11.3. Set the name and location of the project. Select
the language in which you want to code. After fill-
ing all the fields click on finish. Once the project is
Figure 11.4 GUI of new app
successfully created the screen will show as shown in
Figure 11.4.
There are some directories and files in the android pandemic. Here are some examples of how the appli-
project which we should be created before start- cation has been used in the fight against COVID-19:
ing our application as shown in Figure 11.5 and the
description of packages as shown in Table 11.2. • Blood donation and distribution in real-time
• Contactless donation
• Inventory management
VI. Applications of our work
• Emergency response
The android studio-based blood bank mobile appli- • Analysis of data to predict demand.
cation was created utilizing IoT-based technology • Remote health monitoring.
to handle COVID-19 difficulties, and it can have a • Donation of post-recovery blood plasma
variety of uses and advantages in the context of the • Community awareness and involvement
80 Blood bank mobile application of IoT-based android studio for COVID-19

• Connecting to health records

• Partnership between the government and NGOs

Healthcare systems can improve their capacity

to successfully address the issues raised by COVID-
19 by utilizing IoT technology via the Blood Bank
Mobile Application. The program can greatly aid in
managing COVID-19 cases overall, particularly those
requiring blood transfusions, by streamlining blood
donation procedures, ensuring safety standards, and
providing real-time data.

VII. Conclusions and future enhancements

During the COVID-19 pandemic, most people needed
blood. Sometimes emergencies and so crucial that
they can cost a life. There is more demand for blood
donors and blood suppliers. Sometimes it is very dif-
ficult to arrange the blood group to save life or need
in operation during a COVID-19 pandemic. To solve
this type of problem, several types of blood banks
are there and many people become blood donors to
save their life. This observational study describes the
application through which users can easily save their
time by searching for blood donors from particular
Figure 11.5 Package explorer geographical regions. This study reduces the time and
more convenient for the users to save money. And
Table 11.2 Description of package explorer also, this study helps users easily track and contact
the donors near them. And also, its great impact on
S.No. Folder, File and Description medical authorities. The feature enhancement of this
1 Src
works the infected people to predict analyze blood
type by using machine learning (ML) and deep learn-
This directory mainly contains the Java source
files. There is the main activity file which has
ing algorithms to extract information and analyze the
an activity class that runs when we launch our data and also transmit the data by using blockchain
app using the app icons technology.
2 Gen A key step in improving healthcare services is the
There is a .R file generated by the compiler.
creation of a blood bank mobile application using
This file mentions all the resources of our IoT-based technology on the android studio platform
project. We cannot update or modify this file to address the problems brought on by COVID-19.
3 Bin This program optimizes the whole supply chain by
This folder has the android packages that were
streamlining the donation and distribution of blood
built during the build process. This folder also while also utilizing IoT capabilities to provide real-
contains everything needed to run the app time monitoring, tracking, and control of blood units.
4 Res/drawable This program is extremely helpful in the COVID-19
pandemic environment, when effective healthcare sys-
This directory contains all the drawable object
files tems are essential.
The program makes sure that the inventory of
5 Res/layout
blood units is constantly monitored and any changes
All the designed layout file is contained in this
in supply and demand are swiftly addressed through
folder
the use of IoT devices. The android-based software
6 Res/values
also guarantees widespread accessibility by enabling
Files that contain strings and color definitions users to quickly locate nearby blood banks, book
are kept in this folder
appointments, and receive notifications all through
7 AndroidManifest.xml their smartphones. The application’s capabilities are
All the fundamental characteristics of the app further improved by the incorporation of IoT technol-
are described in this file ogy by enabling remote monitoring of blood storage
 Applied Data Science and Smart Systems 81

conditions, lowering waste, and increasing overall in Computational Intelligence and Communication
effectiveness. Technology: Proceedings of CICT, 2019: 501–510.
Incorporating these future scopes would not only doi: https://doi.org/10.1007/978-981-15-1275-9_41.
increase the blood bank mobile application’s efficiency Mufaqih, Sukron, Abiyyu Fawwaz Kanz, Sahid Nur Rama-
dhan, and Ahmad Nurul Fajar. (2020). Blood Bank
and efficacy but will also make a major improvement
Information System Based on Cloud In Indonesia.
to healthcare services generally, which is especially
IOP Conf. Series: Journal of Physics: Conf. Series
important during pandemics like COVID-19. 1179(2019) 012028. 1–6.
Singh, Jaiteg, Gaurav Goyal, and Rupali Gill. (2020). Use of
References neurometrics to choose optimal advertisement meth-
od for omnichannel business. Enterprise Information
Al Bassam, N., Hussain, S. A., Al Qaraghuli, A., Khan,
Systems. 14(2): 243–265. doi: https://doi.org/10.1080
J., Sumesh, E. P., and Lavanya, V. (2021). IoT based
/17517575.2019.1640392
wearable device to monitor the signs of quarantined
Sultanul, A. and Taposi, S. (2021). Blood bank mobile ap-
remote patients of COVID-19. Informat. Med. Un-
plication. 1–39.
lock., 24, 100588.
Mahima, N., Nigam, C., and Chaurasia, N. (2019). m-Health:
Aderonke Anthonia, K., Adeniyi, A. E., Ogundokun, R. O.,
community-based android application for medical ser-
and Ochigbo, S. A. (2019). An android based blood bank
vices. Smart Healthcare Sys., 69–81. CRC Press.
information retrieval system. J. Blood Med., 119–125.
Pohandulkar, Surabhi, S., and Khandelwal, C. S. Blood
Aman, S., Shah, D., Shah, D., Chordiya, D., Doshi, N., and
bank app using raspberry PI (2018). 2018 Int. Conf.
Dwivedi, R. (2022). Blood bank management and in-
Comput. Tech. Elec. Mech. Sys. (CTEMS), 355–358.
ventory control database management system. Proce-
Reddy, C. K. K., Anisha, P. R., and Prasad, L. N. (2016). A
dia Comp. Sci., 198, 404–409.
novel approach for detecting the bone cancer and its
Priya, P., V. Saranya, S. Shabana, and Kavitha Subramani.
stage based on mean intensity and tumor size. Recent
(2014). The optimization of blood donor information
Res. Appl. Comp. Sci., 20(1), 162–171.
and management system by Technopedia. Interna-
Narasimha, P. and Munirathnam Naidu, M. (2013). Gain
tional Journal of Innovative Research in Science, En-
ratio as attribute selection measure in elegant decision
gineering and Technology. 3(1), 1–6.
tree to predict precipitation. 2013 8th EUROSIM
Muhammad, F., Cebe, H. I., Rasheed, J., and Kiani, F.
Cong. Model. Simul., 141–150.
(2016). mHealth: Blood donation application using
Tatale, Subhash, and V. Chandra Prakash. (2020). Enhanc-
android smartphone. 2016 Sixth Int. Conf. Dig. In-
ing acceptance test driven development model with
form. Comm. Technol. Appl. (DICTAP), 35–38.
combinatorial logic. International Journal of Ad-
Ayman, A., Mallikarjuna, B., Saudagar, A. K. J., Sharma, M.,
vanced Computer Science and Applications., 11(10),
and Poonia, R. C. (2022). Improvement of automatic
268–278.
glioma brain tumor detection using deep convolution-
Sastry, J. K. R. and Lakshmi Prasad, M. (2019). Testing
al neural networks. J. Comput. Biol., 29(6), 530–544.
embedded system through optimal mining technique
Krishna, P. V., Gurumoorthy, S., Obaidat, M. S., Mallikarjuna,
(OMT) based on multi-input domain. Int. J. Elec.
B., and Arun Kumar Reddy, D. (2019). Healthcare appli-
Comp. Engg., 9(3), 2141–2150.
cation development in mobile and cloud environments.
Prasad, M. L. and Sastry, J. K. R. (2018). Generation of
Internet of things Personal. Healthcare Sys., 93–103.
test cases using combinatorial methods based multi-
Archit, S., Mallikarjuna, B., Murtuza, M., and Tiwari, V.
output domain of an embedded system through the
(2021). Design and implementation of superstick for
process of optimal selection. Int. J. Pure Appl. Math.,
blind people using internet of things. 2021 3rd Int. Conf.
118(20), 181–189.
Adv. Comput. Comm. Con. Netw. (ICAC3N), 691–695.
Sandeep, B., Mallikarjuna, B., Gautam, D., Gupta, U., Ku-
Mallikarjuna, B., Sathish, K., Gitanjali, J., and Venkata
mar, S., and Verma, S. (2023). The Future IoT: The
Krishna, P. (2021). An efficient vote casting system
current generation 5G and next generation 6G and
with aadhar verification through blockchain. Int. J.
7G technologies. 2023 Int. Conf. Dev. Intel. Comput.
Sys. Sys. Engg., 11(3–4), 237–256.
Comm. Technol. (DICCT), 212–217.
Mallikarjuna, B. (2022). Feedback-based resource utili-
Ganai, P. T., Bag, A., Sable, A., Abdullah, K. H., Bhatia, S.,
zation for smart home automation in fog assistance
and Pant, B. (2022). A detailed investigation of imple-
IoT-based cloud. Res. Anthol. Cross-Dis. Des. Appl.
mentation of internet of things (IOT) in cyber security
Automat., 803–824.
in healthcare sector. 2022 2nd Int. Conf. Adv. Com-
Khan, Mohammad Asaduzzaman, Hasibur Rahaman, Iske-
put. Innov. Technol. Engg. (ICACITE), 1571–1575.
daheer Alam, Khayrul Alam, Sumon Mondal, and
Sandeep, B., Goel, N., and Verma, S. (2023). The current
Alimuzzaman Khan. (2021). Development of Applica-
generation 5G and evolution of 6G to 7G technolo-
tion to Find A Nearby Live Blood Donor Using the
gies: The future IoT. Handbook Res. Mac. Learn-En-
Updated Location e-Information. International Jour-
abled IoT Smart Appl. Across Indust., 456–478.
nal of Electrical Engineering and Applied Sciences
Sandeep, B., Goel, N., Ahlawat, V., Naib, B. B., and Singh, K.
(IJEEAS), 4(1), 17–21.
(2023). A comprehensive review of IoT reliability and
Modi, Nandini, and Jaiteg Singh. (2021). A review of various
its measures: Perspective analysis. Handbook Res. Mac.
state of art eye gaze estimation techniques. Advances
Learn-Enabled IoT Smart Appl. Across Indust., 365–384.
12 Selection of effective parameters for optimizing software
testing effort estimation
Vikas Chahara and Pradeep Kumar Bhatia
Guru Jambheshwar University of Science & Technology Hisar, India

Abstract
Software testing holds a significant role within the realm of software development. Its purpose is to bolster and elevate the
reliability and quality of software. This encompassing process involves several key steps, including estimating the required
testing effort, assembling an appropriate test team, formulating effective test cases, carrying out software execution using
these test cases, and meticulously analyzing the outcomes derived from these executions. Thus, precise software testing effort
estimation holds high significance and governs the overall cost of the software development. To support accurate estimation
of software testing effort, the paper presents a detailed analysis and categorization of various factors have great impact on
the software testing effort. The analysis shows that the parameters include various elements such as quality, stability, risk,
resources, etc., that have a significant impact on software testing effort. Thus, the paper contributes towards a platform for
extracting the basic information essential for supporting software testing effort estimation for successful project planning
and execution.

Keywords: Software testing effort, parameters, quality, testing resources

I. Introduction and categorization of parameters that govern the high

quality STE estimation.
In the realm of software development, estimating the
effort required for testing is pivotal for project success. A. Software project
This estimation hinges on a multitude of interconnected A software project is a planned and well-organized
parameters that collectively influence the scope, com- effort to create, develop, and deliver a software prod-
plexity, and precision of testing efforts (Trendowicz, uct or system that meets certain requirements and
Münch, and Jeffery, 2011). Understanding these objectives. It involves various stages such as require-
parameters is essential to ensure effective testing, ments analysis, design, coding, testing, deployment,
maintain project timelines, and deliver high-quality and maintenance (Borade and Khalkar, 2013). The
software products. In other words, efficient software software projects can vary in size, complexity, and
testing effort (STE) estimation is pivotal for project scope, ranging from small applications to large-scale
success (Cibir and Ayyildiz, 2022). This process hinges enterprise system designs (Mohammed et al., 2017;
on understanding various influencing factors such as Singh et al., 2020). A structured framework that out-
project size, complexity, and functionality intricacies lines the various stages and processes involved in the
(Bluemke and Malanowska, 2021). In this process, a creation and maintenance of software is known as
clear requirements and meticulous planning alleviate software development life cycle (SDLC) (Satapathy,
uncertainties, while risk assessment pinpoints need- Acharya, and Rath, 2016). It provides a systematic
ing thorough testing. Obviously, the choice of testing approach to managing and controlling the entire soft-
types, levels, and automation has great impact sup- ware project from inception to completion. The SDLC
ported by the expertise of the testing team, effective ensures that the software is developed efficiently, fol-
communication, and suitable testing tools (Badri, lowing high quality standards, and meets the needs of
Toure, and Lamontagne, 2015). Moreover, the proj- the stakeholders (Figure 12.1).
ect’s schedule and quality goals interact with testing In other words, it is understood that the software
efforts to deliver high quality software projects. development encompasses the entire process of creat-
The paper is aimed to provide a detailed analy- ing software applications. Software testing is an inte-
sis, categorization and discussion of the factors and gral part of this process that ensures the software’s
parameters that govern the effective STE estimation. quality and reliability before it is deployed to users.
To achieve this, the paper first discusses the software Both software development and testing are crucial
project, importance of software testing in the software components of a software project which is the orga-
development life cycle, various types of software test- nized effort to create a specific software product with
ing, concept of STE followed by a detailed summary defined goals and requirements.

[email protected]
a
 Applied Data Science and Smart Systems 83

fort. A skilled and well-staffed testing team can

execute testing tasks more efficiently, reducing
the overall effort required. Conversely, a short-
age of skilled testers can lead to longer testing
periods or a reduction in the comprehensiveness
of testing.
e. Technical debt (TD): Technical debt refers to the
shortcuts or sub-optimal solutions taken during
the development process that might lead to ad-
ditional work in the future. If a software proj-
ect has accumulated significant technical debt,
Figure 12.1 Software testing stage in SDLC
it can increase testing effort. This is because the
presence of technical debt often results in more
complex code, increased likelihood of defects,
and challenges in maintaining and enhancing the
B. Different aspects of testing effort software.
The software testing process, for a software applica- f. Resource availability (RA): The availability of
tion or system involves allocating resources, time and resources, including both human resources and
effort to plan, design, execute and manage the tests. infrastructure, can impact testing effort. A lack
These activities aim to guarantee that the software of necessary tools, testing environments, or
meets quality standards and performs as intended. skilled personnel can lead to increased testing
The extent of the testing effort depends on factors time and effort. On the other hand, having the
such, as the softwares complexity, project require- right resources readily available can streamline
ments and the level of thoroughness needed in test- testing activities and reduce effort.
ing (Kaner et al., 1999). Some of the factors that can
influence software development and testing efforts are
II. Software testing
as follows:
Software testing is an integral part of the software
a. Software complexity (SC): The complexity of a development process discussed in the last section
software system refers to the intricacy and in- and is thus closely related to the software project
terdependency of its components and features. planning and management (Sharma and Kushwaha,
Higher complexity can lead to increased testing 2011). It is the process of evaluating a software prod-
effort as more interactions between components uct or system to identify any discrepancies between
need to be considered, and there’s a higher likeli- the expected behavior and the actual behavior of the
hood of defects due to the intricacies involved. software.
Complex software systems often require more
thorough testing to ensure all possible scenarios A. Software testing life cycle (STLC)
are covered. The software testing life cycle (STLC) is a systemic
b. Software quality (SQ): The desired level of soft- procedure that defines the different phases and tasks
ware quality directly impacts testing effort. If encompassing the testing of software applications
the project requires a high level of quality, more to guarantee their quality, dependability, and opera-
comprehensive testing, including various testing tional capabilities. It establishes an organized method
types (functional, performance, security, etc.), for strategizing, creating, implementing, and docu-
is needed. Striving for higher software quality menting the testing procedures. Acting as a blueprint,
generally leads to increased testing activities and the STLC directs testing teams through the complete
consequently higher testing effort. testing journey, commencing from the initial assess-
c. Schedule pressure (SP): Project timelines and ment of requirements and culminating in the ultimate
deadlines can significantly influence testing ef- deployment. The main goal of software testing is to
fort. When there’s pressure to meet tight sched- ensure that the software meets its intended require-
ules, testing might be rushed or streamlined, po- ments, functions correctly, are reliable and robust
tentially leading to inadequate testing coverage. (Rajamanickam, 2016). Software testing is typically
On the other hand, sufficient time for testing al- categorized into several types, including:
lows for more thorough and meticulous testing
efforts. • Unit testing: Testing individual components or
d. Work force drivers (WFD): The availability and units of code in isolation to ensure their correct-
expertise of the workforce can impact testing ef- ness
84 Selection of effective parameters for optimizing software testing effort estimation

• Integration testing: Testing the interactions be- 1) Background factors

tween different components or modules to ensure • Project requirements: Understanding the soft-
they work together as expected ware’s functional and non-functional require-
• System testing: Testing the complete software ments is fundamental. The complexity of the
application to validate its overall functionality requirements can influence the testing approach
against the defined requirements and the depth of testing required.
• User acceptance testing (UAT): Involving end- • Software complexity: The intricacy of the soft-
users to test the software in a real-world environ- ware’s architecture, design, and interactions
ment to ensure it meets their needs and expecta- among components can impact testing efforts.
tions More complex systems may require more exten-
• Regression testing: Repeating tests to ensure that sive testing.
new code changes do not introduce new defects • Project timeline: The available time for testing
or break existing functionality affects the testing strategy. Tight schedules might
• Performance testing: Evaluating the software’s re- necessitate prioritizing testing activities and em-
sponsiveness, scalability, and resource usage un- ploying automation.
der different conditions • Budget and resources: The budget allocated to
• Security testing: Identifying vulnerabilities and testing, along with the availability of skilled tes-
ensuring that the software is secure from poten- ters and testing tools, influences the testing ap-
tial threats proach and scope.
• Compatibility testing: Compatibility testing
checks the software’s compatibility with differ- 2) Selection criteria
ent devices, browsers, operating systems, and net- • Critical functionality: The core functionalities of
work environments the software that directly impact users’ needs are
• Usability testing: Usability testing evaluates the prioritized for testing
software’s user-friendliness and user experience. • Business impact: Features that have a significant
It ensures that the software is intuitive and easy impact on the organization’s business goals, rev-
to use enue generation, or user satisfaction are given
• Localization and internationalization testing: higher testing priority
These tests verify that the software is adapted to • High-risk areas: Components or functionalities
various languages, cultures, and regions that are historically prone to defects, or those in-
• Accessibility testing: Accessibility testing en- volving complex interactions, require thorough
sures that the software is accessible to users testing
with disabilities by adhering to accessibility • Customer feedback: Inputs from users or custom-
standards. ers regarding key areas of concern guide testing
efforts to address real-world issues
Software testing occurs throughout the software • Legal and regulatory requirements: Features that
development life cycle, with different types of test- must comply with legal or industry-specific regu-
ing being relevant at different stages. It is an itera- lations necessitate testing that validates adher-
tive process, where issues identified during testing are ence.
addressed, and the software is retested to ensure the
fixes didn’t introduce new problems. These concepts relates to software measurement,
evaluation, and estimation. They provide tools and
B. Background factors and selection criteria methodologies to assess various aspects of software
In software testing, background factors and selection development, complexity, quality, and effort estima-
criteria play a crucial role in determining the scope, tion. These approaches are utilized to enhance deci-
approach, and methods used for testing software sion-making and planning in software projects as
applications. These factors and criteria help testing discussed below (Thakore and Upadhyay, 2013)
teams make informed decisions about which testing
techniques and strategies to employ. These factors • Function point: Function points (FP) are a stan-
guide testing teams in making informed decisions dardized unit of measurement used to quantify
that align with project goals and end-user expecta- the functionality provided by a software applica-
tions, ultimately ensuring the software’s quality and tion. They measure the software’s size based on
reliability (Suri et al., 2015). Here’s an overview of the user’s interactions with it, regardless of the
background factors and selection criteria in software underlying technology or implementation. Func-
testing: tion points consider inputs, outputs, inquiries, in-
 Applied Data Science and Smart Systems 85

ternal files, and external interfaces to determine defects, vulnerabilities, and inconsistencies in the
the complexity and size of a software system. software before it’s released to end-users (Sharma and
This metric is often used in software estimation, Kushwaha, 2011; Hidmi and Sakar, 2017; Brar et al.,
project management, and cost analysis. Math- 2022). The effort invested in software testing is influ-
ematically it can be expressed as: enced by several parameters that impact the complex-
ity and scope of the testing process (Bhattacharya,
(1) Srivastava, and Prasad, 2012; Jin and Jin, 2016b). The
relationship between software projects and software
where, β is fixed quotient of the software project, LOC testing can be understood in the following ways:
is total number of lines of codes to execute “n” num-
ber of functions under fρ number of function points. • Quality assurance: Software testing is essential for
ensuring the quality of the software product be-
• Fuzzified OOPS metrics: Object-oriented pro- ing developed within a software project. It helps
gramming systems (OOPS) metrics refer to identify defects, errors, and vulnerabilities in the
measurements used to evaluate the quality and software, allowing developers to address these is-
complexity of object-oriented software. Fuzzi- sues before the software is released to users
fied OOPS metrics involve applying fuzzy logic • Verification and validation: Software testing is
to these metrics to handle imprecise or uncertain a means of verifying that the software is being
data. Fuzzy logic allows for handling vagueness developed correctly (verification) and validating
in software quality attributes by assigning degrees that it meets the user’s needs (validation). It helps
of membership to different categories, providing confirm that the software aligns with the project’s
a more flexible and nuanced understanding of requirements and objectives
software complexity • Risk mitigation: Software projects inherently
• Cosmic function point (CFP)-based factor analy- involve risks, including the risk of defects or er-
sis and selection: Cosmic function points (CFP) rors. Effective testing helps mitigate these risks by
are a variation of traditional function points used catching and addressing issues early in the devel-
to measure the functional size of a software ap- opment process, reducing the chances of critical
plication based on its business functionality. Fac- failures after deployment
tor analysis and selection in the context of CFP • Iterative development: Many modern software
involves identifying and assigning appropriate development methodologies, such as Agile and
complexity factors to account for variations in DevOps, promote iterative and incremental de-
software projects. These factors help in adjust- velopment. Testing is performed throughout these
ing the functional size measurement to reflect the iterations to continuously assess the software’s
software’s unique characteristics progress and maintain its quality
• COCOMO analysis and new OOPS metrics: • Documentation: Software testing generates docu-
COCOMO (constructive cost model) is a soft- mentation about the software’s behavior, test cas-
ware cost estimation model used to predict the es, and results. This documentation is valuable for
effort, cost, and schedule required for software project managers, developers, and stakeholders to
development. It considers various factors like the track progress and make informed decisions
size of the project, development team experience, • Resource allocation: Software projects need to
and complexity. In the context of object-oriented allocate resources, including time and effort, for
programming, COCOMO can be used to esti- testing activities. The scope and depth of testing
mate effort based on new OOPS metrics, which depend on the project’s requirements and priori-
are measurements specific to object-oriented soft- ties
ware. These new metrics might include measures • Feedback loop: Testing provides feedback to the
of class complexity, coupling, cohesion, and other development team about the software’s perfor-
object-oriented design attributes. mance, functionality, and usability. This feedback
loop helps developers improve the software and
enhance user satisfaction
III. Software testing effort
• The discussion shows that the software testing is
Software testing effort refers to the resources, time, a critical aspect of software projects that ensures
and activities required to effectively test a software the quality, reliability, and functionality of the
application or system to ensure its quality, function- software being developed. It supports the overall
ality, and reliability (Nassif et al., 2019; Cibir and success of the project by identifying and address-
Ayyildiz, 2022). It’s an essential phase of the soft- ing issues, mitigating risks, and providing valu-
ware development life cycle that aims to identify able insights for continuous improvement.
86 Selection of effective parameters for optimizing software testing effort estimation
Table 12.1 Comparative analysis of existing studies

Authors Objectives Techniques Key findings Limitations

Badri, M., Toure, F., Predict unit testing Regression analysis Predictive model Limited sample size
and Lamontagne, L. effort levels of classes for testing effort
estimation
Bhattacharya, P., Estimate software test PSO (Particle swarm PSO-based estimation Requires tuning PSO
Srivastava, P. R., and effort optimization) of test effort parameters
Prasad, B.
Bluemke, I. and Review and summarize Survey and review Overview and Lack of original
Malanowska, A. testing effort categorization of research data
estimation techniques
Borade, J. G. and Provide an overview of Review of Overview of software Limited focus on
Khalkar, V. R. effort estimation estimation effort estimation specific estimation
techniques methods techniques
Liao, X. and Naseem, Review COCOMO Review of Overview of Limited focus on
A. models and extensions COCOMO models COCOMO models COCOMO models
and extensions
Satapathy, S. M., Early-stage software Random forest Early-stage effort Limited to use
Acharya, B. P., and effort estimation estimation with use case point-based
Rath, S. K. case points estimation
Sharma, A. and Develop metric suite Requirement Metric suite for early Limited validation of
Kushwaha, D. S. for testing estimation engineering estimation of testing the metric suite
document
Singh, V., Kumar, V., Select influential Fuzzy logic and Parameter selection for Limited to parameter
and Singh, V. B. testing parameters AHP-TOPSIS influencing testing selection
Srivastava, P. R., Estimate test effort Bat algorithm Test effort estimation Limited to bat
Bidwai, A., Khan, A., using bat algorithm based on bat algorithm algorithm
Rathore, K., Sharma,
R., and Yang, X. S.

The comparative analysis of the existing stud- • Risk assessment

ies is given in Table 12.1 to present analysis for the • Testing strategy
findings, techniques and the posed limitations of the • Test environment set-up
research works. The review is used a starting point for • Testing tools and frameworks
presenting a depth analysis of software testing effort • Testing documentation
estimation. • Personnel and skill levels
• Iterations and changes
IV. Factors and parameters governing efficient • Review and collaboration
software testing effort estimation • Data management
• Non-functional testing
The effort estimation process in software testing plays • Project deadlines
a critical role in project planning, resource allocation, • Stakeholder expectations
and budgeting (Liao and Naseem, 2012; Singh, Kumar, • Historical data from the past projects.
and Singh, 2023). An accurate estimation ensures that
testing activities are adequately resourced and aligned Overall, it can be understood that the software test-
with project timeline that involves considering vari- ing effort estimation is a multi-faceted process that
ous factors and parameters that influence the com- considers a range of factors and parameters. Based
plexity and scope of the testing process (Jin and Jin, on the detailed discussion Table 12.2 provides a con-
2016a; Mensah et al., 2016). Some of the key factors cise categorization of various factors and parameters
and parameters that contribute to high quality STE essential for a better understanding of their impact on
estimation. the STE (Figure 12.2).
Analyzing the factors discussed in Table 12.2 helps
• Scope and requirements project managers and testers estimate the testing
• Complexity of the system effort accurately, plan testing activities effectively, and
 Applied Data Science and Smart Systems 87
Table 12.2 Categorization of factors effecting the software testing effort

S. No. Factors Category/Key parameter Description Remarks

Size/complexity Larger and complex projects need STE is directly proportional to

more rigorous testing which is due size and complexity of software
to increased number of interactions, project
features and issues
Scope Any frequent change in the scope of STE is directly proportional to
project leads to an additional testing frequency in scope change
effort
1 Interfaces Software projects that involve STE is governed by the extent
Quality objectives Project characteristics

interaction with the external systems of interaction with external

need a thorough software testing and systems
may add up to the software testing
effort
Clarity of requirements Ambiguous or unclear requirement STE is inversely proportional
of the software project raises to clarity of requirements
misunderstandings among the developer
and management
2 Quality standards When project aims at higher quality STE is directly proportional
standards such as ISO, CMMI it to quality of standards to be
requires rigorous testing effort achieved
Criticalness Projects dealing with critical functions STE is directly proportional
such as aviation or medical device to critical ness of the software
development require higher testing project
efforts
3 Technology stacks The type of technology employed and STE is governed by the
the focused platform govern the extent technology stack
of testing effort. New or unfamiliar
technology need more testing effort
Architecture

Integration complexity When the systems are integrated with STE is governed by the extent
Software

number of third party components the of integration complexity

system confronts more issues and need
more testing effort
4 Depth of testing When comprehensive software testing STE is directly proportional to
strategies such as regression, usability, depth of testing
security are involved, more testing effort is
required in comparison to the basic testing
Testing strategy

effort
Automation Test automation initially require higher STE gradually decreases with
testing effort, however, eventually, the the passage of time
manual software testing effort gets
reduced with the passage of time
Skilled workforce The skilled testers may reduce the STE is directly proportional to
testing phase that otherwise may get the length of testing phase
prolonged when it lacks in skilled
testers
Time constraints Tight software project schedule limits STE is inversely proportional
the testing window and increases the to time constraint
software testing effort
Test data Availability of diverse test data is STE is inversely proportional
5 essential. The generation of test data to the availability of test data
needs additional effort that increase
the overall testing effort. Here, data
Resource Availability

consistency could be time consuming

process when generated
Financial support Industry regulated software projects STE is inversely proportional
need huge investment. In absence of to financial support
financial support high testing effort
is required to meet the compliance
standards
88 Selection of effective parameters for optimizing software testing effort estimation

S. No. Factors Category/Key parameter Description Remarks

Testing configuration Hard and complex software set-up STE is governed by the
configuration can increase software software set-up environment
testing effort during configuration and configuration
Testing tools Availability of suitable tools such as STE is governed by the
hardware, software network resources suitability of testing tools
Testing infrastructure

for the software testing have a high

6 impact on the testing effort
Testing environment The software testing in different STE is governed by the
environment to justify the software infrastructure performance
performance requires a stable and
reliable infrastructure otherwise the
testing effort get increased or decreases
efficiency of testing

Key characteristics of the early design model include:

• Early design emphasis: This model prioritizes the

design phase and encourages in-depth analysis
and planning before moving into the implemen-
tation phase
• Comprehensive design: Design decisions are
made with careful consideration of the system’s
architecture, component interactions, and overall
structure
• Iterative refinement: While emphasizing early
design, the model acknowledges that design de-
Figure 12.2 Factors governing software testing effort cisions can evolve and improve as development
progresses. Iterative cycles of design refinement
are common
allocate resources efficiently to ensure a successful • Reuse and patterns: The model promotes the use
software testing phase. of design patterns and the reusability of existing
components to expedite development and ensure
A. Early design and reuse model (EDRM) proven design practices
The software development involves both code design- • Efficiency and quality: By addressing design intri-
ing and planning well from the inception of the proj- cacies early, the model aims to minimize potential
ect to the successful delivery. The most crucial part defects, reduce costly rework, and enhance soft-
in this is accurate software testing effort estimation, ware performance and maintainability
as it allows manipulating the existing components • Communication and collaboration: Close collab-
in order to reduce the overall development time. The oration between design, development, and other
models tested via reuse enhances the reliability of the stakeholders is crucial to ensure that design deci-
software project while reducing the defects leading sions align with project goals.
to more reliable prediction of software testing effort
estimation. The effort calculation, based on a standardized algo-
rithmic model, is depicted using the given equation.
1) Early design model
The early design model is a software development (2)
approach that focuses on thorough and thoughtful
design in the initial stages of the software develop- where PM signifies the total effort in months.
ment life cycle. The primary goal of this model is to M = {PERS’ RCPX’ RUSE’ PDIF’ PREX’ FCIL’
establish a strong foundation by making informed SCED} Initial calibration sets A = 2.94, Size in K-LOC,
design decisions early on, which in turn leads to and B between 1.1 and 1.24, based on the project’s
enhanced software quality, reduced rework, and originality, development flexibility, risk management
smoother development progression. strategies, and process maturity.
 Applied Data Science and Smart Systems 89

Multipliers: Developer skill, non-functional needs, the adjustments needed based on factors like ASLOC
platform familiarity, and other factors are reflected in and AT, as explained earlier. The COCOMO II model
multipliers. primarily relies on your estimation of the software
project’s size, measured in thousands of Source Lines
• RCPX – product reliability and complexity; of Code (KSLOC), to calculate the required effort in
• RUSE – the reuse required; terms of Person–Months (PM). In essence, the model’s
• PDIF – platform difficulty; effort estimation heavily depends on your assessment
• PREX – personnel experience; of the project’s scale, as quantified by the size of the
• PERS – personnel capability; codebase.
• SCED – required schedule;
• FCIL – the team support facilities. (5)

2) The reuse model Eaf stands for “Effort Adjustment Factor,” which
The reuse model is an approach in software devel- is derived from the cost drivers. The exponent E in
opment that focuses on leveraging existing software the formula is determined by the five scale drivers.
components, modules, or solutions to enhance effi- Estimation techniques, such as expert judgment, his-
ciency, reduce development time, and improve overall torical data analysis, and specialized software tools,
software quality. It centers on the idea that by reusing can aid in determining the effort required for the
well-tested and proven components, developers can reuse model. In essence, the reuse model’s estimation
avoid reinventing the wheel and instead build upon involves evaluating the integration effort, customiza-
established solutions. The reuse model encourages the tion needs, and associated activities when incorporat-
systematic identification, selection, and integration of ing existing components into a new project.
reusable assets to streamline the development process. With a holistic approach, taking into account the
intricacies of the software, the testing strategy, the
a. Reuse model estimation: team’s capabilities, and the project context, the paper
Estimating effort and resources for the reuse model contributes to provide foundation for the accurate
involves considering factors unique to integrating and testing effort estimation. Regular review and adjust-
adapting reusable components. This estimation model ment of estimates based on evolving project dynamics
uses Equation (3) to estimate the effort. further contribute to improved project planning and
successful software delivery.

(3) V. Conclusion
The paper presents a detailed analysis of various fac-
ASLOC stands for “Actual Source Lines of Code,” tors that are critical for accurate estimation of soft-
which represents the total number of lines of ware testing effort that plays a critical role in project
code that have been created for a software proj- planning and management. The paper delved into
ect. AT refers to the “Proportion of Automatically existing research presented by the research commu-
Generated Code,” indicating the fraction of code that nity in predicting software testing effort estimation.
is generated through automated tools or processes. The paper claims important contribution in laying
ATPROD ­represents “Engineers’ Productivity in Code down the foundation and preliminary factor analysis
Integration,” signifying how efficiently developers prior to estimating the testing effort for a software
integrate this code. If the estimation process is based project. The multifaceted nature of software projects
solely on manually written code, the estimated Lines necessitates a comprehensive approach to estimation,
of Code (LOC) can be determined using Equation (4). encompassing parameters such as project complexity,
size, requirements volatility, team expertise, and his-
torical data analysis. By considering these parameters,
(4)
organizations can enhance their ability to create more
reliable and realistic testing effort estimates. As the
ESLOC stands for “Estimated Source Lines of Code,” software development landscape continues to evolve,
which refers to the calculated number of lines of code the parameters influencing testing effort estimation
expected in a software project. The costs associated are subject to change with the passage of time and
with modifying reused code, understanding how to advancement of technology. Therefore, a proactive
integrate it, and making decisions about its reuse stance towards continuous improvement and adap-
are considered in determining the adaption adjust- tation is necessary. The collaboration between devel-
ment multiplier. This multiplier takes into account opment and testing teams, ongoing communication,
90 Selection of effective parameters for optimizing software testing effort estimation

and learning from each estimation cycle’s outcomes review. J. Comput., 3(5), 683–693. Available: http://
are crucial for refining the estimation process over www.cisjournal.org.
time. Altogether, a successful software testing effort Mensah, Solomon, Jacky Keung, Kwabena Ebo Bennin, and
estimation demands a harmonious blend of empiri- Michael Franklin Bosu. (2016). Multi-objective opti-
mization for software testing effort estimation. SEKE,
cal analysis, domain expertise, and technological
1–6. doi: 10.18293/SEKE2016-017
advancements. With fine-tuning of the parameters
Brar, P. S., Shah, B., Singh, J., Ali, F., and Kwak, D. (2022).
discussed in this paper, organizations can pave the Using modified technology acceptance model to eval-
way for more accurate STE estimations, leading to uate the adoption of a proposed IoT-based indoor
better resource allocation, project planning, and ulti- disaster management software tool by rescue work-
mately, the delivery of high-quality software systems. ers. Sensors, 22(5), 1866. https://doi.org/10.3390/
In future, the study can be followed for evaluating the s22051866.
effect of integrating concept of machine learning in Mohammed, N. M., Niazi, M., Alshayeb, M., and Mah-
STE estimation works. mood, S. (2017). Exploring software security ap-
proaches in software development lifecycle: A sys-
tematic mapping study. Comp. Stand. Interf., 50,
References 107–115. doi: 10.1016/J.CSI.2016.10.001.
Badri, M., Toure, F., and Lamontagne, L. (2015). Predict- Nassif, A. B., Azzeh, M., Idri, A., and Abran, A. (2019). Soft-
ing unit testing effort levels of classes: An exploratory ware development effort estimation using regression
study based on multinomial logistic regression model- fuzzy models. Computat. Intel. Neurosci., 2019.
ing. Proc. Comp. Sci., 62, 529–538. Rajamanickam, Leelavathi. (2016). Principles and Goals of
Bhattacharya, P., Srivastava, P. R., and Prasad, B. (2012). Software Testing. International Journal of Advanced
Software test effort estimation using particle swarm Engineering, Management and Science, 2(5): 239455.
optimization. Adv. Intel. Soft Comput., 132, 827–835. 427–430.
doi: 10.1007/978-3-642-27443-5_95/COVER. Satapathy, S. M., Acharya, B. P., and Rath, S. K. (2016). Ear-
Bluemke, I. and Malanowska, A. (2021). Software testing ly stage software effort estimation using random for-
effort estimation and related problems. ACM Comput. est technique based on use case points. IET Software,
Sur. (CSUR), 54(3). doi: 10.1145/3442694. 10(1), 10–17. doi: 10.1049/IET-SEN.2014.0122.
Borade, Jyoti G., and Vikas R. Khalkar. (2013). Software Sharma, A. and Kushwaha, D. S. (2011). A metric suite for
project effort and cost estimation techniques. Interna- early estimation of software testing effort using re-
tional Journal of Advanced Research in Computer Sci- quirement engineering document and its validation.
ence and Software Engineering, 3(8), 730–739. 2011 2nd Int. Conf. Comp. Comm. Technol., 373–
Cibir, E. and Ayyildiz, T. E. (2022). An empirical study on 378. doi: 10.1109/ICCCT.2011.6075150.
software test effort estimation for defense projects. Singh, V., Kumar, V., and Singh, V. B. (2023). A hybrid novel
IEEE Acc., 10, 48082–48087. doi: 10.1109/AC- fuzzy AHP-TOPSIS technique for selecting parame-
CESS.2022.3172326. ter-influencing testing in software development. Dec.
Singh, J., Goyal, G., and Gill, R. (2020). Use of neuro- Anal. J., 6, 100159. doi: 10.1016/J.DAJOUR.2022.
metrics to choose optimal advertisement method for 100159.
omnichannel business. Enterp. Inform. Sys., 14(2), Srivastava, P. R., Bidwai, A., Khan, A., Rathore, K., Shar-
243–265. https://doi.org/10.1080/17517575.2019.16 ma, R., and Yang, X. S. (2014). An empirical study
40392. of test effort estimation based on bat algorithm.
Hidmi, O. and Sakar, B. E. (2017). Software development ef- Int. J. Bio-Ins. Comput., 6(1), 57–70. doi: 10.1504/
fort estimation using ensemble machine learning. Int. IJBIC.2014.059966.
J. Comput. Commun. Instrum. Engg., 4(1), 143–147. Suri, R., Pushpa, and Harsha, S. (2015). Object oriented
Jin, C. and Jin, S. W. (2016). Parameter optimization of software testability (OOSTE) metrics analysis. Int. J.
software reliability growth model with S-shaped test- Comput. Appl. Technol., 4(5), 359–367.
ing-effort function using improved swarm intelligent Thakore, D. and Upadhyay, A. R. (2013). A framework to
optimization. Appl. Soft Comput., 40, 283–291. doi: analyze object-oriented software and quality assur-
10.1016/J.ASOC.2015.11.041. ance. Int. J. Inn. Technol. Explor. Engg. (IJITEE), 5,
Jin, C. and Jin, S.-W. (2016). Parameter optimization of 254–258.
software reliability growth model with S-shaped test- Trendowicz, J., Münch, J., and Jeffery, R. (2011). State of
ing-effort function using improved swarm intelligent the practice in software effort estimation: A survey
optimization. Appl. Soft Comput., 40, 283–291. and literature review. Lecture Notes – Comp. Sci.
Kaner, C., Falk, J., an Nguyen, H. Q. (1999). Testing com- (including subseries Lecture Notes in Artificial Intel-
puter software. John Wiley & Sons. ligence and Lecture Notes in Bioinformatics), 4980
Liao, X. and Naseem, A. (2012). Software models, exten- LNCS, 232–245. doi: 10.1007/978-3-642-22386-
sions and independent models in cocomo suite: A 0_18/COVER.
13 Automated detection of conjunctivitis using convolutional
neural network
Rajesh K. Bawa1 and Apeksha Koul2,a
1
Department of Computer Science, Punjabi University, Patiala, Punjab, India
2
Department of Computer Science and Engineering, Punjabi University, Patiala, Punjab, India

Abstract
Conjunctivitis, commonly referred to as “pink eye,” is a prevalent and contagious eye condition that affects millions world-
wide. Detecting conjunctivitis early and accurately is vital for timely intervention and effective management. Here, a convo-
lutional neural network (CNN) model has been customized to automate the detection of conjunctivitis using eye images. Our
dataset encompasses a diverse array of eye images, which include both healthy and conjunctivitis-affected cases. To tackle
the challenge of limited data, we employ data augmentation techniques to expand the dataset. After pre-processing and aug-
mentation, we curate a collection of 5135 eye images representing both pink-eye pathology and healthy states. Subsequently,
these augmented images undergo classification using the developed CNN model. During execution, the customized CNN
model obtains an impressive accuracy of 88.80%, with a loss of 0.25, and demonstrates precision, recall, and F1 scores of
0.50. The CNN model holds promise as an automated solution for conjunctivitis detection. Its accuracy and efficiency could
substantially support medical professionals in early diagnoses, facilitating timely treatment and curbing transmission rates.

Keywords: Conjunctivitis, pink eye, deep learning, CNN, augmentation

I. Introduction occasionally blurred vision. In a nutshell, conjunc-

tivitis symptoms can vary and consulting a medical
The current monsoon season has resulted in signifi-
professional for accurate diagnosis and appropriate
cant disruptions across various regions of the coun-
treatment is crucial, especially if the symptoms persist
try. The occurrence of floods in various regions of the
or worsen (Rodrigues, 2019).
country has led to a notable escalation in the suscep-
There are many types of conjunctivitis and gener-
tibility to vector-borne diseases (Targhotra, 2023).
ally they are grouped into four main types, depending
Delhi this year saw its worst recorded flood in the
on their causes (McManes, 2022):
last four decades, with unprecedented water levels in
the Yamuna river that caused water-logging in major Viral conjunctivitis: This is usually caused by a virus
parts of the city. As the Yamuna river surpasses the and is highly contagious. It often accompanies com-
danger mark, a sudden rise of eye diseases such as mon cold symptoms and spreads via direct contact
conjunctivitis is also seen in Delhi, Maharashtra, and with the eye secretions of any infected person.
parts of Gujarat (Rawat, 2023).
Bacterial conjunctivitis: Caused by bacteria, this type
can result in a thick discharge that can cause the eye-
A. Background
lids to stick together. It can also be easily transmitted
Conjunctivitis, also referred to as “pink eye,” is an
through direct contact.
inflammation of the conjunctiva which is a delicate
transparent tissue that coats the inner side of the eye- Allergic conjunctivitis: This is triggered by some al-
lid and envelops the white portion of the eye. The dis- lergens like pollen, dust, or pet dander, this type is not
ease is caused by irritants, allergens, bacteria as well contagious. It typically causes itching, redness, and
as viruses, like coronavirus (Roth, 2022). There are excessive tearing.
various symptoms that can be seen such as redness Irritant conjunctivitis: This is typically a result of be-
in the white part of the eye and inner eyelid which ing exposed to irritants such as chemicals, smoke, or
creates a noticeable pink or reddish appearance. The foreign objects. It’s not contagious and typically re-
eye may feel itchy or experience a burning sensation, solves once the irritant is removed.
prompting frequent rubbing, excessive tearing along
with a clear, white, yellow, or green discharge, can be B. Role of AI for the detection and diagnosis of eye
present which may lead to crust formation on the eye- conjunctivitis
lids, especially upon waking. Swelling of the eyelids The conventional techniques involve clinical
might occur, accompanied by sensitivity to light and approaches which are being conducted by medical

a
[email protected]
92 Automated detection of conjunctivitis using convolutional neural network

this purpose. This was done to substantiate the asser-

tion and to attain the targeted accuracy level of 84%.
Likewise, a convolutional neural network (CNN)
model was employed by Erdin and Lalitkumar (2023)
to detect eye diseases. The primary goal of this study
was to classify human eyes into four unique groups:
trachoma, conjunctivitis, cataract, and healthy. The
study’s accuracy percentage was 88.36%. The CNN
model was evaluated and obtained recall of 88.75%,
precision of 89.25%, and F1 score of 88.5%. The
technology exhibited the potential for early diagnosis
Figure 13.1 AI to detect and classify eye disease
of numerous eye illnesses based on the accuracy and
evaluation results.
Verma et al. (2015) focused on the diagnosis and
experts. Various tools such as slit lamp microscope classification of hyperemia, a condition where the
are used by the professionals to carefully observe few white portion of the eye becomes red, using deep
indications such as swelling, redness, and discharge. learning. The paper discussed the subjective and
Although this method worked effectively but it has objective methods of assessing bulbar redness and
a drawback that it introduces the dependence on the highlights the limitations of these methods. Their
experience and interpretation of practitioners (Azari proposed model used deep learning to automatically
and Amir, 2020). This is where artificial intelligence extract features and classify the results, minimizing
(AI) steps in to offer a solution. the dependency on operators for evaluation.
Artificial intelligence (AI) has been as a valuable
tool in the domain of medical diagnostics which
II. Objectives
includes the field of ophthalmology and plays an
important role to detect and diagnose eye condi- Based on the impact of AI techniques in detecting and
tions such as conjunctivitis, as shown in Figure 13.1 diagnosing eye conjunctivitis, the goal of the manu-
(Schmidt-Erfurth et al., 2018). script is to detect and perform binary classification
AI systems have the ability to examine eye pictures between pink eye and healthy eye using modified
or scans with the help of advanced image processing CNN model.
which assist them to identify suspected conjunctivitis The contribution that has been done to conduct the
cases. These algorithms are trained for detecting mild research is as followed:
symptoms such as swelling, redness, and discharge
which allow the models to identify patterns linked 1. Initially, the customized dataset has been cre-
with the disorder (Han, 2022). AI algorithms provide ated which consisted of 265 pink eyes and 130
important insights to medical personnel to establish healthy eyes.
accurate diagnoses by referencing large databases of 2. In the next phase, pre-processing has been
existing cases. This convergence of technology as well performed by resizing the size of images to
as healthcare has the potential to escalate the diag- (224×224) and later is enhanced by using histo-
nosis of conjunctivitis diagnosis, thereby improving gram equalization and unsharp masking.
patient treatment, and potentially reduce the strain on 3. After this, the dataset of images are augmented
medical practitioners (Koul et al., 2023). using three augmentation techniques such as ro-
In fact, the researchers have also contributed in tation, horizontal flipping and vertical flipping
the field of detection and diagnosis of eye conjunc- which results up to 5135 images.
tivitis. Gunay et al. (2015) worked on the diagnosis 4. Following augmentation, an enhanced CNN
of conjunctivitis by analyzing corneal images. The model was meticulously devised. Rigorous eval-
approach entailed the segmentation of the infected uation ensued, encompassing crucial parameters
area within the images to quantify vasculariza- which include precision, F1 score, accuracy, re-
tion and the intensity of redness in pink eyes. Their call, and loss thereby substantiating the model’s
approach successfully detected instances of eye infec- effectiveness in conjunctivitis detection.
tions and accurately identified potentially contagious
patients in 93% of instances. Mukherjee et al. (2021)
III. Methodology
developed a mobile healthcare application (iConDet)
for the purpose of conducting preliminary conjuncti- This section covers the flow to detect and classify the
vitis detection. Deep learning methods were applied pink eye and the healthy eye using proposed CNN
to the conjunctivitis dataset that was compiled for model and the framework is presented in Figure 13.2.
 Applied Data Science and Smart Systems 93

Figure 13.2 Proposed system to detect and classify

pink as well as healthy eye

Figure 13.3 Sample of pink and healthy eyes

Figure 13.5 Resized dimension of images

B. Data pre-processing
The eye data collected are of different sizes which can
hamper the performance of the system, hence their
size have been reduced to (224×224), as shown in
Figure 13.5.
Later, the quality of the resized images have been
enhanced and the process starts with individual histo-
gram equalization on each color channel (blue, green,
and red) to enhance contrast by spreading pixel inten-
Figure 13.4 Number of images in the dataset sity distribution. This enriches visual appeal, making
dark and light areas distinct. Following this, unsharp
masking is applied. A blurred version of the enhanced
A. Dataset image is subtracted from the original, emphasizing
The dataset employed in this research has been care- high-frequency components like edges and details.
fully customized by gathering specific images that These components are then blended back into the
show pink eye from the eye diseases virus dataset vol- image, sharpening it and highlighting features. This
ume 1 (Kaggle, 2020). Moreover, a separate collection blend of techniques enhances contrast and sharpness,
of images depicting healthy eyes has been acquired resulting in an image with heightened visual appeal
from various reputable online sources, as shown in and clear details. The approach is demonstrated by
Figure 13.3. showcasing the original and enhanced images, as
A total of 356 .jiff images illustrating instances of shown in Figure 13.6.
pink eye were amassed for analysis. It is important to
mention that a small subset of these images was con- C. Data augmentation
sidered unsuitable and was manually excluded from It involves applying transformations to the original
the dataset, which ends up with the 265 number of images to diversify the dataset and improve model
images. Furthermore, an independent set of 130 .jpg training. Here, ImageDataGenerator() has been used
images featuring healthy eyes (Singh et al., 2019) was to perform rotation (within a range of -50 to +50
obtained for the purpose of comparison, as shown in degrees), horizontal flipping, and vertical flipping
Figure 13.4. generating 12 augmented images of single image (4
94 Automated detection of conjunctivitis using convolutional neural network
Table 13.1 Layered architecture of proposed CNN model

Layer Output shape Param #

Conv2d (None, 222,222,32) 896

Max_pooling2d (None,111,111,32) 0
Conv2d_1 (None,109,109,64) 18496
Max_pooling2d_1 (None, 54,54,64) 0
Conv2d_2 (none, 52,52,128) 73856
Max_pooling2d_2 (None, 26,26,128) 0
Flatten (None, 86528) 0
Dense (None, 512) 44302848
Figure 13.6 Enhanced eye images Dropout (None,512) 0
Dense_1 (None,1 ) 513

Table 13.1 represents the layered architecture of a

CNN model. Each row corresponds to a layer in the
network, and provides information about the layer
type, output shape, and the number of parameters (or
weights) in each layer.

Convolutional layers: The network starts with

three convolutional layers (`conv2d`, `conv2d_1`,
`conv2d_2`). These layers use 2D convolutions to
Figure 13.7 Augmented images. (a) Healthy eye. (b) extract features from the input data. The number
Pink eye after the layer name (e.g., 32, 64, 128) indicates the
number of filters applied in each layer.
Max pooling layers: Following each convolutional
layer is a max-pooling layer (`max_pooling2d`, `max_
pooling2d_1`, `max_pooling2d_2`). Max pooling
reduces the dimensions of the feature maps, aiding in
Figure 13.8 Architecture of CNN model
retaining important features while reducing compu-
tational load.
Flatten layer: After the convolutional and pooling
layers, there’s a `flatten` layer that reshapes the data
augmented images each simulating different angles from 2D arrays to a 1D vector. This prepares the data
from which the eye might be captured). Hence, in for the subsequent fully connected layers.
total we have 1690 healthy images and 3445 pink
Dense (fully connected) layers: Two fully connected
eye images. This augmented dataset aids learning
layers (`dense`, `dense_1`) follow the flattening layer.
models in learning from a wider range of image
These layers process the flattened data to make final
variations, leading to better generalization and per-
predictions. The number after “Dense” indicates the
formance. Figure 13.7 presents the main augmented
number of neurons in each layer.
images.
Dropout layer: The “dropout” layer is used for regu-
D. CNN larization. It randomly sets a fraction of input units
Convolutional neural networks (CNNs) are a special- to zero during training, reducing the risk of over
ized type of deep learning model designed for process- fitting.
ing visual data like images. They consist of layers that 44396609 total parameters have been generated
automatically learn and extract features from images. which provides the total number of trainable parame-
The core components are convolutional layers, which ters in the network to represent the weights and biases
detect patterns in the input image, and pooling layers, that the model learns during training. These param-
which downsample the data, as shown in Figure 13.8 eters are updated to minimize the loss function during
(Kumar, 2023). training.
 Applied Data Science and Smart Systems 95
Table 13.2 Performance metrics

Metrics Formulae

Accuracy

Loss

Precision

Recall

F1 score

Table 13.3 Hyper-parameters of CNN model

Hyper-parameters Values

Activation ReLu/Sigmoid
Optimizer Adam
Class mode Binary
Figure 13.9 Learning curves of CNN model
Batch_Size 32
Loss Binary cross entropy
Dropout rate 0.5 Table 13.5 Performance summary.
Epochs 10 Model Recall Precision F1 score

CNN 0.50 0.50 0.50

Table 13.4 Evaluation of CNN model

Model Training Testing

Table 13.6 Class-wise performance metrics for eye
Accuracy Loss Accuracy Loss classification.

CNN 0.83 0.35 0.88 0.25 Class Precision Recall F1 Score

Healthy eye 0.33 0.38 0.35

E. Performance metrics Pink eye 0.67 0.64 0.62
In Table 13.2, the metrics provide crucial insights
into the performance of applied learning models and
aid in comprehending the strengths and weaknesses
of models across different aspects of classification Table 13.4 shows that the CNN model achieved
performance (Modi et al., 2021; Kumar et al., 2022; an accuracy of 83% on the training dataset with a
Koul et al., 2023). corresponding loss of 0.35. On the testing dataset,
the model performed even better, with an accuracy
of 88% and a lower loss of 0.25. This indicates that
IV. Results
the CNN model has been able to generalize well from
In this section, the applied CNN model has been the training data to the testing data, demonstrating
evaluated on the basis of various performance metrics its effectiveness in classifying the eye images correctly.
such as accuracy, loss, precision, recall, and F1 score. The performance of the model has been also ana-
The hyper-parameters used to compile the model are lyzed graphically on the basis of their curves as shown
mentioned in Table 13.3: in Figure 13.9.
Table 13.4 provides a summary of the performance Table 13.5 provides a summary of performance
metrics for a trained CNN model on both the training metrics related to precision, recall, and the F1 score
and testing datasets. for a specific model.
96 Automated detection of conjunctivitis using convolutional neural network

Similarly, the model has been also evaluated for consideration. The model’s performance may vary
binary class of the dataset i.e., for healthy eye and with factors like dataset size, diversity, and quality.
pink eye on the basis of precision, recall, and F1 score Over fitting remains a concern, especially if the data-
in Table 13.6. set is small or imbalanced.

V. Discussion VI. Conclusion

Conjunctivitis is a condition that has the potential to The paper highlights the significance of the proposed
affect individuals across all age groups. However, it CNN model for automating the detection of con-
is observed that this phenomenon is more prevalent junctivitis using eye images. The model’s accuracy
among the pediatric population and individuals in of 88.80% and efficiency in distinguishing between
the young adult age group. This can be attributed to healthy and conjunctivitis-affected eyes offer a prom-
the fact that children and young adults often gather ising solution for aiding medical professionals in early
in educational institutions and workplaces, where diagnoses. This advancement in automated detection
they engage in close physical proximity and frequent holds potential to enhance eye health management
bodily contact with one another. The exponential by enabling timely interventions and reducing trans-
proliferation of the contagion is the primary factor mission rates. In future, the work can be extended by
contributing to the nearly two-fold increase in the increasing the dataset and developing an automated
afflicted population within the current calendar year learning model which can perform multi-class clas-
(Hashmi, 2022). sification of different types of conjunctivitis.
According to the available sources, there has been Overall, the paper underscores the valuable role of
a notable increase in conjunctivitis cases, with a rise AI-powered diagnostic tools in improving conjuncti-
of approximately 50–60%. The demographic pri- vitis detection and overall eye health care.
marily impacted by this phenomenon consists pre-
dominantly of individuals in their childhood stage References
of development. According to our research findings,
it has been observed that approximately one out of Azari, Amir A., and Amir Arabi. (2020). Conjunctivitis:
every three children exhibits symptoms of red eyes a systematic review. Journal of ophthalmic & vi-
sion research. 15(3): 372–395. doi: 10.18502/jovr.
or conjunctivitis (PTI, 2023). The observed data indi-
v15i3.7456
cates a significant rise in the incidence of conjunctivi- Manjumdar, Pulak. (2010). Preliminary phytochemical and
tis cases, ranging from 10% to 15%, when compared wound healing activity of zyziphus oenoplia (l) mill.
to previous seasonal patterns. According to expert PhD diss., Rajiv Gandhi University of Health Sciences
analysis, it has been determined that a significant (India), 1–24.
proportion, potentially reaching up to 40%, of those Erdin, Muh. and Patel, L. (2023). Early detection of eye dis-
impacted by the situation in question are children ease using CNN. Int. J. Res. Appl. Sci. Engg. Technol.
(Debroy, 2023). 11(4), 2683–2690.
In this paper, the study aims to contribute to the Melih, G., Goceri, E., and Danisman, T. (2015). Automated
early and accurate detection of conjunctivitis, a con- detection of adenoviral conjunctivitis disease from fa-
tagious eye condition affecting millions worldwide. cial images using machine learning. 2015 IEEE 14th
Int. Conf. Mach. Learn. Appl. (ICMLA), 1204–1209.
The obtained results demonstrate the CNN model’s
Jae-Ho, H. (2022). Artificial intelligence in eye disease: Re-
performance in terms of accuracy, precision, loss, cent developments, applications, and surveys. Diag-
recall, and F1 score. The model showcases a notable nostics, 12(8), 1927.
accuracy level, particularly in distinguishing between Hashmi MF, Gurnani B, Benson S, Price KL. Conjunctivi-
healthy and conjunctivitis-affected eyes. tis (Nursing). (2022). Dec 6. In: StatPearls [Internet].
The CNN model demonstrated a precision, recall, Treasure Island (FL): StatPearls Publishing; 2023 Jan–
and F1 score of 0.50. These results indicate that the . PMID: 33760572.
model strikes a balance between correctly identify- Kaggle. (2020). Eye diseases virus dateset vol 1. October 26.
ing positive instances (true positives) and minimiz- Apeksha, K., Bawa, R. K., and Kumar, Y. (2023). Artificial
ing false negatives. Nonetheless, it’s worth noting intelligence techniques to predict the airway disorders
that an F1 score of 0.50 suggests there is potential illness: A systematic review. Arch. Comput. Method
Engg., 30(2), 831–864.
for enhancing the model’s overall performance since
Ajitesh, K. (2023). Different types of CNN architectures
higher precision, recall, and F1 score values are typi- explained: Examples - Data analytics. Data Analyt.,
cally preferred for more accurate and dependable August 23.
predictions. Likewise, the class-wise metrics indicate Yogesh, K., Koul, A., and Mahajan, S. (2022). A deep learn-
balanced performance for both healthy and conjunc- ing approaches and fastai text classification to predict
tivitis classes. However, certain limitations warrant 25 medical diseases from medical speech utterances,
 Applied Data Science and Smart Systems 97
transcription and intent. Soft Comput., 26(17), 8253– en-in/conditions/eye-discharge/ [Accessed on Nov 15,
8272. 2023]
Singh, Jaiteg, and Nandini Modi. (2019). Use of informa- Erica, R. (2022). What you need to know about conjunc-
tion modelling techniques to understand research tivitis. Healthline. Available at https://www.health-
trends in eye gaze estimation methods: An automated line.com/health/conjunctivitis [Accessed on Oct 15,
review. Heliyon, 5(12), 1–12. 2023]
Debrowski, Adam. (2022) . Types of conjunctivitis: Bacteri- Schmidt-Erfurth, Ursula, Amir Sadeghipour, Bianca S. Ge-
al, viral, allergic and others. Available at https://www. rendas, Sebastian M. Waldstein, and Hrvoje Bogu-
allaboutvision.com/en-in/conditions/conjunctivitis- novic. (2018). Artificial intelligence in retina. Progress
types/ [Accessed on Dec 10, 2023] in retinal and eye research. 67: 1–29.
Prateeti, M., Bhattacharyya, I., Mullick, M., Kumar, R., Modi, Nandini, and Jaiteg Singh. (2021). A review of
Roy, N. D., and Mahmud, M. (2021). i ConDet: An various state of art eye gaze estimation techniques.
intelligent portable healthcare app for the detection Advances in Computational Intelligence and Com-
of conjunctivitis. Int. Conf. Appl. Intel. Informat., 29– munication Technology: Proceedings of CICT 2019.
42. Cham: Springer International Publishing. 501–510. doi: https://doi.org/10.1007/978-981-15-
Nangia, Aarti. (2023). Cases of conjunctivitis, other eye in- 1275-9_41
fection on rise in Delhi: Doctors. Available at https:// Targhotra, Prerna. (2023) . Eye flu cases rise amid floods in
economictimes.indiatimes.com/news/india/cases-of- India; Know causes, symptoms and preventive mea-
conjunctivitis-other-eye-infection-on-rise-in-delhi- sures. Available at https://english.jagran.com/lifestyle/
doctors/articleshow/102085752.cms [Accessed on eye-flu-cases-rise-amid-floods-in-india-know-causes-
Dec 9, 2023] symptoms-and-preventive-measures-10089567 [Ac-
Rabat, Sudeep. (2023). Floods spark eye flu outbreak in cessed on Oct 25, 2023]
Delhi; Know about cause and symptoms. Available at Verma, Sherry, Latika Singh, and Monica Chaudhry. (2019).
https://www.business-standard.com/health/delhi-ncr- Classifying red and healthy eyes using deep learning.
facing-eye-flu-outbreak-after-flood-passes-here-how- International Journal of Advanced Computer Science
to-stay-safe-123072400508_1.html [Accessed on Nov and Applications, 10(7), 525–531. Doi: 10.14569/
10, 2023] ijacsa.2019.0100772.
Rodrigues, Aimee. (2023). Eye discharge: Causes and treat-
ment. Available at https://www.allaboutvision.com/
14 An overview of wireless sensor networks applications,
challenges and security attacks
N. Sharmila Banua, B.Vidhya and N.Mahendran
Sri Ramakrishna College of Arts and Science, SR University, India

Abstract
A wireless sensor network (WSN) is a key technology in the implementation of several applications, including light-duty data
streaming applications and straightforward event/phenomena monitoring systems. The energy-efficiency of wireless sensor
networks is a crucial issue in case of development and implementation. The goal of this effort is to increase the information
processing and routing process of energy efficiency. This research paper’s primary goal is to provide a complete review of
WSN. This article gives a broad overview of the WSN and some of its key features. This study also discusses several WSN
threats, WSN research obstacles, and WSN applications.

Keywords: Wireless sensor network, clustering WSN, mobile sinks in WSN

I. Introduction collected to choose the next course of action. Ahmad

et al. (2020) provides definitions of several context
A wireless sensor network (WSN) is created by the
kinds, such as temporal context, social context, moti-
connecting of various tiny sensor nodes. These net-
vational context, location context, etc.
works are primarily used to gather data on the
In order to meet the requirements outlined in the
environment in which they are implemented. Many
deployment of pervasive computing, WSN have
computer domains, including data transfer, network-
evolved as an appropriate technology for sensing
ing protocols, signal processing, information pro-
events and acquiring data that is typically dispersed
cessing and aggregation, storage, etc., are brought
over numerous sites in a geographic region (Al Qundus
together by WSN (Al Qundus et al., 2022). WSN
et al., 2022). Wireless Sensor Networks are made up
is more adaptable and effective in monitoring the
of cheap, compact, battery-operated computer devices
environment than the large sensors used in earlier
with radio transceivers and sensors that can perceive
times. Also, with no significant infrastructure, fewer
events and interact with one another (Al Qundus et
resources are needed for anything other than environ-
al., 2022). The self-organizing sensor nodes would
ment monitoring.
either pass on the sensed data to a centralized sink
Pervasive computing is the idea that technology
where the data would be analyzed and the presence
should be seamlessly incorporated into every part
of the event inferred, or they would interact with one
of human existence while remaining fully unobtru-
another to cooperatively determine the occurrence of
sive, i.e., without becoming the center of attention.
an event in a dispersed way (Figure 14.1).
Pervasive computing aims to create an intelligent,
A typical WSN is subject to a number of limita-
flexible environment that continuously facilitates
tions. A typical Wireless Sensor Network is subject to
interactions between people and their surround-
a number of limitations (Alghamdi, 2020) , some of
ings by detecting their actions and anticipating their
which are as follows:
needs from their surroundings. This greatly improves
the quality of interaction between people and their
• The WSN is battery-operated, compact in size,
surroundings. It further assumes that this would be
equipped with cheap, low-accuracy sensors
accomplished through the presence of a significant
and short-range radios. They also have limited
number of tiny computing devices with sensing and
computation and memory capacity. As a result,
radio communication capabilities that are widely dis-
these nodes are energy-constrained, have lim-
persed throughout the environment, gathering data
ited processing power, poor sensing precision,
on the environment, gathering data on the actions of
and are vulnerable to hardware and connectiv-
the human subject, and monitoring the interaction
ity issues.
of the human subject with the environment. It also
• Due to problems like channel fading and interfer-
assumes that these computer systems will cooperate
ence, the wireless medium itself is prone to erratic
with one another and be cognizant of the surround-
and unexpected behavior.
ing environment as they evaluate the data they have

[email protected]
a
 Applied Data Science and Smart Systems 99

sensor nodes to maximize coverage area was pro-

posed. Al-Turjman, (2019) explored deterministic
and random (uniform random) node deployments for
large-scale WSNs, taking into account performance
parameters including coverage, energy use, and mes-
sage transmission time. The developed simple energy
model demonstrated the effectiveness of THT as a
node deployment approach for WSN applications.
Jawad et al. (2017) investigation focused on the
cooperative network’s cluster-based coded collabora-
tion with numerous receiving nodes. In this approach,
each member of the receiving cluster relays its sig-
nal copy to the destination while the sending node
sends a packet to the receiving cluster. The destination
Figure 14.1 Diagram of a typical WSN
node decodes the original information bits using code
combining methods. While using the same amount of
In addition, since every node in a network uses the power, the link layer dependability in a cluster-based
same channel for communication, there is a significant network is significantly increased.
likelihood that information will be lost due to packet Coded operation was developed by T. E. Hunter
loss and network congestion. and A. Nosratinia for transmission between two send-
ing nodes and one receiving node. Only one of the
transmitting nodes sends a data block in each time
II. Related work
slot, which consists of N1 bits from its own coded bits
A distributed system for cooperative MIMO trans- and N2 bits from its partner. The receiver then uses
missions developed by Hsin Yi Shen et al. (Alghamdi, code combining to combine the bits it has received
2020) makes use of space-time block coding and code from the two senders (Elappila et al., 2020). The
combining in the transmitting and receiving groups. coded collaboration for the cluster-based network,
In order to estimate the numerous carrier frequency however, lacked clarity.
offsets (CFO) from received mixed pilot signals, an A distributed space time block coding-based coop-
uncorrelated pilot symbol generation method based erative transmission technique was proposed by
on a pseudo noise sequence with iterative updates has Zhu et al. (2012). The performance was examined
been used. Additionally, the evaluation of the mini- on the presumption that nodes cooperate to decode
mum mean square estimator (MMSE) detector for received packets and that error detection occurs at
receiving STBC (Space Time Block Code) coded data the packet level. An optimization approach has been
under several CFO. The system’s projected BER and used to reduce total energy usage based on the perfor-
overall energy usage are compared to those of com- mance analysis. It is clear that adding more nodes to
parable cooperative designs. The suggested strategy a cluster could not increase energy efficiency due to
dramatically raises BER and energy effectiveness. the additional circuit energy that cooperating nodes
A specific plan that combines STBC with coopera- could need. Additionally, the ideal sensor cluster size
tive code combining. The challenges of transmitter changes based on the needed packet error rate (PER).
and receiver diversity in cooperative MIMO systems Even with rigorous throughput and delay constraints,
are addressed by the use of STBC and code combin- considerable energy savings can still be made com-
ing. STBC are deployed in the sending group to take pared to non-cooperative transmission.
advantage of transmitter diversity once the sending
and receiving groups have been established. To create III. Wireless sensor and actuator networks
receiver diversity, the destination combines the signals (WSAN) evolution
from the nodes in the receiving group using error con-
trol code combining. It has been demonstrated that WSAN have developed through time and now include
the system makes use of MIMO diversity benefits to special nodes called actuator nodes. The job of the
deliver dependable and effective gearbox (Díaz et al., WSN is to acquire data about the environment in
2011). which they are placed. In addition, unique nodes
For grid-based WSNs, (Tian et al., 2008) presented known as “actuator nodes” are added to the net-
mathematical methodology for maximizing net- work. These nodes have the potential to actuate in
work lifespan. When the sensor node’s communica- response to certain control components that affect the
tion radius is equal to or greater than its detecting environment in which the sensor nodes are placed.
radius, a technique for deploying the fewest possible In addition to having more processing and memory
100 An overview of wireless sensor networks applications, challenges and security attacks

provided by the sensor nodes. This means that in

order to prevent a delay in the control action, the
interval between the time an action is sensed by
the node mobility and the time when the action
is executed by the actuators must be as short as
is practical.
• As a result, there is an additional restriction
placed on the network latency time (TNLT)
which is crucial in the context of WSAN. TNLT
must not exceed the desired actuation latency
time of the application. This requirement applies
to the time lag between an event’s detection by a
Figure 14.2 (a) Making centralized decisions design. sensor node and the time the same is noted at the
(b) Distributed determination design sink or actuator.

IV. Attacks on WSN

capacities, stronger communication skills, and the
ability to operate on a controlled element, Actuator WSNs are vulnerable to a range of attacks. Secure and
nodes are not energy-constrained (Singh et al., 2019). dependable data transport from the sensing environ-
The following two network architectures (as depicted ment to the base station is necessary for critical appli-
in Figure 14.2), were introduced with the addition of cations (Chaitra and Sivakumar, 2017).
actuator nodes to the network.
Node capture attack: Attackers insert special equip-
• Making centralized decisions (semi-automated). ment into particular network nodes and gather data
• Distributed determination (automated). on sensor communications and security protocols.
By taking the data from sensor communication, the
The sensor nodes transmit the perceived data to a attacker can obtain the cryptographic information.
centralized sink in the event of a centralized decision- With prior knowledge, the attacker begins taking part
making method, where the choice on the control in network activity and is capable of physically cap-
action to be taken by the many actuators is made in turing nodes.
light of the information acquired. The actuators then Wormhole attack: An evil node will pose as the one
receive this information to be put into action by the closest to the base station in order to gather sensed
actuators. information from its neighbors. The real communica-
In the case of a distributed decision-making strat- tions from the nodes won’t reach the authentic base
egy, the sensor nodes relay the information they have station. The malicious node will ignore the messages
gathered to particular actuators, who then converse from its neighbors, bringing down the entire network.
and work together to decide on the precise control Attacks on network energy use: The attacker inserts
action that should be carried out by each of them. The malicious nodes into the system, and these nodes con-
second strategy is more in line with the real concept stantly broadcast connection requests, forward mes-
of WSN since it emphasizes the problem of coopera- sages, and drain the energy of nearby nodes during
tively deducing an event and responding to it. In non- request processing.
real time, simple control action applications, WSAN
Denial of service (DoS) attacks is frequent in net-
are anticipated to greatly speed up the acceptance
works. DoS attacks prevent the legitimate node from
of wireless data acquisition and control systems.
participating in the network’s predefined operations.
Nevertheless, decisions made for the control of actu-
ating action depend on the information gathered by Attacks using replication: The attacker sets up mali-
the sensor network. New limitations (Farhan et al., cious nodes that have the same ID as the real nodes
2017; Ahmed et al., 2019) have been put onto the and are loaded with hacked cryptographic data.
dependability of the network’s data gathering process Instead than capturing a significant number of nodes,
which is as follows: the deployment of numerous replicas makes it simple
to compromise the whole network.
• The control action must be time coherent with
the conditions of the environment because the WSN is essential to C4ISRT systems, which stand for
actuator node(s) must decide whether to con- Surveillance, Command, Control, Communications,
duct the control action based on an estimation Computation, Intelligence, Reconnaissance, and
or re-construction of the event using information Targeting. A very promising sensing method for
 Applied Data Science and Smart Systems 101

military C4ISRT is created by sensor networks’ fast Quick replication

exploitation, self-organization, and error accepting With snapshot replication, data is duplicated pre-
characteristics. Sensor nodes are intended as sensor cisely for the given interval. In contrast to other tech-
blotches that are spread around the area to be detected niques, snapshot replication doesn’t fully account for
in the border monitoring application. The sensor data changes. This type of replication is used because
nodes transmit the data to the network entrance via data changes are most likely to happen rarely, such
the sensor blotch. Also, the entrance is in responsibil- when publishers and subscribers finish their first
ity of transmitting that data to the base station. The synchronizations.
base station connects to database clones of the secure
systems that are only accessible to authorized person- Fusion of replication
nel. Finally, several user interfaces are used to display This type of replication is widely used in server-to-
the information to the military workforce. client configurations because it enables both the sub-
scriber and publisher to make dynamic changes to
A. Replication methods the data. Using merge replication, which consolidates
Attack node replication attack: In a node duplication information from several databases into one, makes it
attack, a copy of the recognized vertex that is present more challenging.
into the network tries to contact nearby nodes that
are within its wireless communication range. Even if Incremental replication based on keys
the attacker is unsuccessful in creating a link with a Only data that has changed since the last update is
valid node, it will continually carry out the same pro- transferred using this technique, sometimes referred
cedure throughout the network. to as key-based incremental data collection. Keys can
Attack on access points: Access points are the infor- be viewed as database elements that lead to data rep-
mation aggregation hubs that send data to the sink. lication. Because each update only involves a few row
An advanced technique called access point replication copies, the costs are very low. The drawback is that
allows an attacker to take control of a section of the this replication mode cannot be used to retrieve data
network. Successful access point replication makes it that has been permanently deleted since the key value
simpler for the attacker to capture the network. is also removed along with the record.
It is difficult to locate these replications in WSNs.
Attack on sink replication: Sink replication is a
Several academics have suggested methods for recog-
sophisticated assault against WSNs. Many WSNs
nizing these dangerous assaults.
simply utilize one sink to collect data. After complet-
ing sink replication, the attacker will have complete
control over the network. Even though WSNs have V. Modes of data acquisition
numerous sinks, sink duplication causes more harm. A WSN’s primary objective is to gather information
about the region in which it is located and then com-
B. Different replication methods municate that information to the sink, which may
Fully replicated tables be far away. The events occurring in the deployment
Whole table replication is the phrase used to describe zone are then determined or reconstructed using
the replication of all data. This includes both new and the information obtained. The deployed nodes nor-
updated information that is replicated from the begin- mally scan and capture data regularly but transfer
ning to the end. Costs are frequently greater since this the gathered data to the actuator, based on the tech-
replication technique needs a lot of processing power nique of data acquisition, which may be one of the
and network bandwidth. But, as will be discussed later following(Ahmed et al., 2018) .
in this article, whole table replication can be helpful
for retrieving data that has been hard destroyed as 1. Periodic data acquisition: In this mode, the nodes
well as data that does not have replication keys. will regularly communicate the information they
have acquired to the sink in addition to periodi-
Replication in a transaction cally collecting the information themselves. The
Using this approach, whenever data is modified, rate at which the acquired data is transferred to
and the user database receives updates after com- the sink would typically be significantly higher
plete first copies of the data are being created from than the frequency of data sampling. This is be-
origin to destination by the data replication soft- cause the node typically uses far less energy for
ware. This is a more effective replication technique sensing and computing than for delivering the in-
since fewer rows are copied when data is changed. formation, but there are rare applications where
Transactional replication is widely used in server-to- the contrary has been demonstrated to be true
server configurations. (Elappila et al., 2018).
102 An overview of wireless sensor networks applications, challenges and security attacks

2. Event-based data acquisition: In this mode, the b) Sensing unit

deployed nodes may scan and gather data on a Typically, the sensing device is coupled to one or
regular basis, but they will only send the data more physical sensors. The sort of sensors de-
when a certain predefined event takes place. sired can be included into the node according on
As the nodes won’t transmit frequently in this the application’s requirements.
mode, there is a tendency for the network life c) Communication unit
to be substantially greater than in periodic data This component keeps the WSN connected. It is
acquisition mode. Applications requiring regu- made up of an antenna-related radio transceiver
lar information about the event region, such as integrated circuit (IC). The radio IC’s communi-
habitat monitoring, are not suited for this mode. cation range may be adjusted, and it mostly de-
3. Query-based data acquisition: In this mode, the pends on the requirements of the application.
deployed nodes may periodically look for and
compile data, but they will only formally request The node’s communication range and power use
and transfer the data. This technique, sometimes are tightly correlated. The node’s power consump-
referred to as interest propagation, involves the tion rises in tandem with the communication range
sink initiating a query that spreads across the Rc. The radio IC typically operates in four modes:
network. The data that nodes have access to will transmit, receive, wait and snooze (Sutagundar and
only be delivered to those that satisfy the query’s Manvi, 2013). Wait mode power utilization is lower
requirements. One of the main criteria of this than the transmit and receive modes which both use
mode is that the query must be disseminated to roughly the same amount of energy during operation.
all nodes in the least amount of time and energy. Sleep mode is the least energy-intensive mode when
4. Hybrid data acquisition: It combines the first compared to the others; it uses a tiny fraction of the
three types. One or more of the aforementioned energy that other modes use.
data collecting modes may be employed by vari-
ous parts of the deployed nodes.

Environmental applications for WSN include coal

mining, tsunami, earthquakes, flood detection, gas
leak detection, forecasting of forest fires, cyclones,
water quality, range of rainfall, volcanic eruption,
and more. The network helps in the implementation
of safety procedures to some extent since it provides
early identification and forecasting of all these natu-
ral disasters. The sensor gathers the information, then
it is transferred to the master station over the net.
This aids in both alerting people to the approaching
disaster and implementing precautionary actions. The
monitoring of forest fires, air pollution, coal mining, Figure 14.3 Sensor nodes architecture
gas leaks, and water quality will all benefit from this.

VI. Single node architecture

In a WSN, the architecture of a sensor node is very
simple and it is subdivided into three primary units:
(i) the processing unit; (ii) the communication unit;
and (iii) the sensing unit. A sensor node’s block dia-
gram is shown in Figure 14.3.

a) Processing unit
This part often acts as the sensor node’s heart-
beat. Its internal microcontroller processes the
data that is sent to it. One or more of the mi-
crocontrollers that are most often used in sensor
nodes include the MSP 430, Intel Strong ARM,
and SA-1100. To store the instruction set, a flash
memory is also linked to this device. Figure 14.4 Applications areas of WSN
 Applied Data Science and Smart Systems 103

VII. Applications areas of WSN VIII. Challenges in WSN

WSNs have many uses now, and their potential appli- The structure of WSN faces a number of difficulties,
cations will expand in the future (Jawad et al., 2017). including topology, design concerns, scalability, net-
These are the few WSN application domains (Figure work longevity, energy utilization, etc. Difficulties,
14.4): network longevity are a crucial factor in the effective
deployment of WSN. Even though all other issues
(a) Habitat observation: As part of the Great Duck have been resolved with no appreciable lifetime
Island project (Ali et al., 2017), (Haseeb et al., enhancement, it is worthless for a WSN application
2020) nodes are placed across the island to (Jayarajanneditors, n.d.;, Jaiswal and Anand, 2020;
track petrel activities. The tiny sensor nodes that Haseeb et al., 2020).This work focuses in particular
were dispersed over the island captured data on on optimizing the clustering architecture in WSNs to
changes in the nest’s pressure, humidity, and tem- lengthen the network lifetime. The main goal of the
perature. Based on the aforementioned variables, clustering algorithms suggested for WSN is to lower
petrel activity may be reliably observed without the network’s energy consumption during each cycle
doing any environmental harm. Large connected of data collecting. Yet, it is discovered in reality that a
sensors would not allow for the same type of ex- round’s total energy reduction alone will not lengthen
perimenting. the network lifetime (Preeth et al., 2018; Shukla and
(b) Precision agriculture: Lack of understanding of Tripathi, 2020). In order to extend the duration of
the correct soil composition is the main prob- WSN, energy balancing is a crucial component along
lem in agriculture (Weng and Lai, 2013; Yu et with energy minimization. While though energy bal-
al., 2013; Al-Turjman, 2019). Water logging is ance and reduction appear to be the same, they are
caused by an overabundance of water in many actually two completely unique facets of the same
agricultural areas, which has a significant im- issue, namely network lifespan. As a result, the opti-
pact on the production. With the appropriate mum clustering method for WSNs must guarantee
micro sensors dispersed across the agricultural both network energy balance and energy reduction.
field, it is possible to measure the temperature,
soil composition, and other nutrient concentra- Energy
tions. Energy competence is the first and frequently most
(c) Coal mining: A major issue in coal mining is significant design problem for a WSN. The three
the frequent fire incidents (Benayache et al., functional areas of communication, sensing, and data
2019; Behera et al., 2020) Click or tap here to processing – each of which requires development –
enter text. which result in the loss of priceless are given power. The battery life can have a signifi-
human lives as well as financial hardship. WSN cant impact on the sensor node’s longevity. As sensor
has been utilized to pinpoint the precise site nodes have limited energy budgets, this restriction is
of fire catastrophes in the early stages, saving typically associated with sensor network approach.
priceless human lives and preventing serious Sensors are often powered by batteries, which should
mishaps. be changed or refilled after they run out.
(d) Determining the path of a forest fire takes more
time when utilizing satellite photos due to im- Limited bandwidth
precise recognition of using high tower sensing In WSN, processing information uses far more energy
locations (Jayarajann editors, n.d.; Zhou et al., than transmitting it. At present wireless communica-
2008; Shahraki et al., 2021). The catastrophic tion is restricted to data rates between 10 and 100
harm to the trees and the wild species that live Kbits per second. As message transfers between sen-
there is unlikely to be stopped unless the message sors are disrupted by bandwidth restrictions, synchro-
reaches the control center as quickly as possible. nization is impossible.
The distributed sensor nodes are effective in de-
tecting forest fires earlier and with greater preci- Node prices
sion. A large number of sensor nodes make form a sensor
(e) Military observation: As of now from two de- network. It follows that the estimation of a distinct
cades, there occurs a significant change to how node is crucial to the sensor network’s overall finan-
battles are fought. Using acoustic or video sen- cial metric. It is obvious that in order for the global
sors, the WSN can deliver trustworthy battlefield metrics to be bearable, the estimation of every sensor
data to the control room. Important human lives node must be considered. Depending on how the sen-
can be saved with prompt intervention (IEEE sor network is used, a significant number of sensors
Staff and IEEE Staff, n.d.). might be scattered randomly across the environment
104 An overview of wireless sensor networks applications, challenges and security attacks

for purposes like weather observation. If the price X. Mobile sinks in WSN: challenges
was reasonable overall for sensor networks, it would
Location Identification: To transmit the detected
be much more reasonable and successful for consum-
data, the sensor nodes require the availability of the
ers who demand careful consideration.
mobile actuator. The broadcast techniques are used
by mobile sinks to relay their position to network
Placement
nodes. Unfortunately, these methods need a signifi-
Node location in WSN could be a simple problem to
cant amount of resources to broadcast position data
tackle. Special strategies are required to position and
from the sink to the network on a regular basis. The
handle a broad spectrum of nodes in a relatively lim-
network requires a lot of message forwarding from
ited environment. In a highly sensitized region, 100 to
each node, which makes it difficult to use resources
10,000 of sensors have also been placed.
efficiently. By using an overhearing method, it takes
less broadcast messages to locate a mobile sink. The
Restrictions on strategy
mobile sink creates beacons with fresh position data
The creation of smaller, more affordable, and more
and transmits them to nearby access points. The access
effective devices is the main goal of wireless sensor
points that are in the mobile sink’s communication
design. The design of WSN will be influenced by a vari-
range receive the beacon and alter their message head-
ety of additional competitions. WSN has encountered
ers to point at the mobile sink. The remaining nodes
limited-restriction hardware and software approach
in the networks locate the mobile sink’s new position
paradigms (Tyagi and Kumar, 2013).
after hearing the changed header. Fewer fixed points
are used in the footprint-based technique. The sensor
IX. Clustering architecture in WSN nodes use fixed positions to determine the communi-
Clustering is the practice of assembling sensor nodes cation channel to the mobile sink. Nodes determine
that are geographically adjacent to one another their logical coordinates and the path to the mobile
(Gao et al., 2010; Gardašević et al., 2020), (Surya sink based on the fixed locations. The overhead
Engineering College & Institute of Electrical and caused by the location identification protocol should
Electronics Engineers, n.d.). A cluster head (CH) is a be maintained and considered to reduce the need for
node that manages a cluster and may start the cluster- retransmission of broadcast messages and extend net-
ing process. Cluster members are the residual nodes in work lifespan.
the cluster. These CM nodes will continually perceive Routing and mobile sink trajectory: The mobile sink
their surroundings and transmit data to the corre- trajectory is essential for the routing of sensed data.
sponding CH nodes. As all member nodes in a cluster In contrast to large-scale WSNs, a mobile sink may
are close to one another, the information they provide travel to each node in the net to collect the sensed
will also be redundant. In the majority of application data. Using special nodes that are fewer than the total
instances, sending this duplicate information to the number of nodes in the network, the random walks-
BS is unnecessary, and it also shortens the network’s based strategy decouples the mobile sink route from
lifespan. In order to create a single piece of infor- the sensor nodes. Movable sinks may roam around
mation, the CH nodes combine the data they have freely and collect data from a desired number of
obtained from the CMs. The BS will only be informed nodes (which act as a sub sinks or access points). By
of this one piece of information. In general, WSNs using position identification methods, sensor nodes
benefit from the clustering architectural paradigm in locate the mobile sink. In order to enhance the dura-
the following ways. tion of WSNs, the trajectory and routing path choices
are crucial. The mobile sink’s trajectory must guaran-
Conserving bandwidth: It occurs when a network is tee sink availability across the network with the least
grouped and nodes are logically segmented and con- amount of routing overhead.
nected to their respective groups. As a result, the clus- Transmission scheduling: The nodes in WSNs transfer
ters may share the same communication bandwidth the detected data to the mobile sink as soon as it enters
without encountering any interference. To prevent their communication range. WSNs are resource-con-
inter-cluster interferences, each cluster will have its strained networks that strive to cut down on energy
own spread code. use while distributing data to a mobile sink. Till the
Scalability: After the first deployment, new sets of mobile sink arrives at the nearest point in the commu-
nodes are included in the network to improve the nication range, the nodes will not be able to send their
accuracy level of the information. The newly added data according to the transmission scheduling scheme.
nodes may be readily accommodated using clustering
strategy as a new cluster or included in the available When more than one node recognizes the
groups during the process. mobile sink within its communication range, data
 Applied Data Science and Smart Systems 105

Figure 14.5 Framework for mobile sinks In WSN

Figure 14.7 WSN with compromised access point and
replicated mobile sink

Figure 14.6 Multiple nodes request for transmission Figure 14.8 Mobile sink revoke compromised node
channel at the same time and broadcast control messages to the network

transmission in WSNs with densely placed sensor Security: Because the sink is mobile, WSNs are vulner-
nodes becomes challenging. Data scheduling tech- able to a variety of attacks. For WSNs with mobile
niques addressed how the mobile sink in densely dis- sinks, conventional security techniques are insuf-
tributed WSNs, as depicted in Figure 14.5, chooses ficient (Abella et al., 2019). A mobile sink may be
the transmission channel. The mobile sink assigns a compromised to access the whole network. If WSNs
channel based on the measure of data to be trans- deploy numerous mobile sinks, seizing one will give
mitted and the node’s remaining power. Effective data the attacker access to a significant chunk of the net-
distribution in WSNs requires a distance and speed work. Two distinct key pools are generated by the
traveled – Finding a reasonable compromise between key management method, one for connecting sensor
efficient data aggregation and the network’s capacity nodes and access points and the other for connecting
to access the mobile sink is a difficult task. Mobile mobile sinks and access points. With the sink’s mobil-
sink must remain inside the sensor nodes’ communi- ity comes an increase in security complexity. Figure
cation range until the nodes have finished transmit- 14.5 depicts the case of an access point being taken
ting. The base station must respond quickly to WSN over and a duplicated mobile sink being brought into
applications in order to manage the regions of inter- the network by an attacker. Researchers need to pay
est. The mobile sink moves along the trajectory at a close attention to WSN security.
controlled speed thanks to transmission scheduling Upkeep of the network: The mobile sink’s privilege
and routing algorithms, forcing the sensor nodes to level must be set before to deployment in order to
transfer any observed data to the sink. To extend the allow the base station to manage the network. Figure
lifespan of the network, a trajectory with a minimal 14.6 illustrates how mobile sinks are utilized for node
trip distance and maximum network coverage must revocation as well as broadcasting private control
be chosen. The suggested model enhances the perfor- messages to the whole network during major security
mance in terms of the speed and distance that mobile threats.
sinks move.
106 An overview of wireless sensor networks applications, challenges and security attacks

The need for mobile sinks in WSNs is rapidly grow- XII. Conclusion
ing. The introduction of mobile sink in WSNs pres-
WSN is a key technology in the implementation of
ents further difficulties for networks. The difficulties
several applications, including light-duty data stream-
of implementing mobile sinks in WSNs are noted.
ing applications and straightforward event/phenom-
WSNs still experience resource depletion when using
ena monitoring systems. The energy-efficiency of
the current techniques (Figures 14.7 and 14.8).
wireless sensor networks is a crucial issue while devel-
oping and running them. The goal of this effort is to
XI. General basic features increase the information processing and routing pro-
A. Sensor network architecture cess’ energy efficiency. The primary goal of this work
Any of the following methods are used by sensors to is to provide a complete review on WSN.
send data to the base station.
References
• In a flat adhoc architecture, sensors work togeth-
Abella, C. S., Bonina, S., Cucuccio, A., D’Angelo, S., Gi-
er to send data to the access point, also known as ustolisi, G., Grasso, A. D., Imbruglia, A., Mauro, G.
the base station (AP or BS), and several hops are S., Nastasi, G. A. M., Palumbo, G., Pennisi, S., Sor-
used in the transmission process. bello, G., and Scuderi, A. (2019). Autonomous energy-
• Sensors are grouped into clusters in hierarchi- efficient wireless sensor network platform for home/
cal networks, and cluster heads are in charge office automation. IEEE Sens. J., 19(9), 3501–3512.
of gathering and aggregating data from sensors https://doi.org/10.1109/JSEN.2019.2892604.
and reporting to the access point. In hierarchi- Ahmad, Arshad, Ayaz Ullah, Chong Feng, Muzammil Khan,
cal networks, single-hop transmission to the BS Shahzad Ashraf, Muhammad Adnan, Shah Nazir, and
and multihop transmission between clusters are Habib Ullah Khan. (2020). Towards an improved
used. energy efficient and end-to-end secure protocol for
iot healthcare applications. Security and Commu-
• In a sensor network with mobile access, mobile
nication Networks. vol. 2020: 1–10. https://doi.
BS roaming the sensor field directly communi- org/10.1155/2020/8867792
cates with the sensors, and transmission from the Ahmed, G., Zhao, X., Fareed, M. M. S., Asif, M. R.,
sensors to BS is one-hop. and Raza, S. A. (2019). Data redundancy-control
energy-efficient multi-hop framework for wireless
B. Wireless sensor nodes sensor networks. Wire. Person. Comm., 108(4),
Motes are another name for sensor nodes. The mar- 2559–2583. https://doi.org/10.1007/s11277-019-
ket offers a wide variety of sensors. The following are 06538-0.
some of the sensor nodes in use right now: Ahmed, M., Salleh, M., and Channa, M. I. (2018). Rout-
ing protocols based on protocol operations for un-
University of California Los Angeles (UCLA) Wireless derwater wireless sensor network: A survey. Egyptian
Integrated Network Sensors (WINS) UCLA created Informat. J., 19(1), 57–62. https://doi.org/10.1016/j.
some of the low power wireless integrated micro sen- eij.2017.07.002.
Al Qundus, J., Dabbour, K., Gupta, S., Meissonier, R.,
sors that are utilized as sensor nodes. Using a 1 mW
and Paschke, A. (2022). Wireless sensor network for
transmitter, it provides wireless communication at a AI-based flood disaster detection. Ann. Oper. Res,
speed of 100 Kbps over a distance of 10 m. 319(1), 697–719. https://doi.org/10.1007/s10479-
UC Berkeley quotes: Crossbow’s Mica family con- 020-03754-x.
tains the members Mica, Mica2, Mica2Dot, and Alghamdi, T. A. (2020). Energy efficient protocol in wire-
MicaZ. With a 16 MHz CPU, 4 KB of RAM, 2 KB less sensor network: Optimized cluster head selection
of flash memory, and a data rate of up to 250 Kbps, model. Telecomm. Sys., 74(3), 331–345. https://doi.
MicaZ supports the IEEE 802.15.4 standard and org/10.1007/s11235-020-00659-9.
Ali, Ahmad, Yu Ming, Sagnik Chakraborty, and Saima
ZigBee protocols.
Iram. (2017). A comprehensive survey on real-time
AMPS from MIT: It is a low-end standalone guard- applications of WSN. Future internet. 9(4): 77 pp.1-
ing node that can also act as a fully complete node 22. https://doi.org/10.3390/fi9040077
for middle-end sensor networks or as a supporting Al-Turjman, F. (2019). Cognitive-node architecture and a de-
element in a higher-end sensor system. The institution ployment strategy for the future WSNs. Mobile Netw.
uses these motes for a variety of purposes. Appl., 24(5), 1663–1681. https://doi.org/10.1007/
s11036-017-0891-0.
Tiny node: 512 KB of external flash memory, 8 KB Behera, T. M., Mohapatra, S. K., Samal, U. C., Khan, M.
of RAM for programmes and data, and the TinyOS S., Daneshmand, M., and Gandomi, A. H. (2020).
operating system. Moreover, there are BTNode, I-SEP: An improved routing protocol for heteroge-
Imote, Iris Mote, TelosB, Wasp Mote, and others. neous WSN for IoT-based environmental monitoring.
 Applied Data Science and Smart Systems 107
IEEE Internet of Things J., 7(1), 710–717. https://doi. for precision agriculture: A review. Sensors, 17(8), 1–45:
org/10.1109/JIOT.2019.2940988. 1781. doi: https://doi.org/10.3390/s17081781
Benayache, A., Bilami, A., Barkat, S., Lorenz, P., and Ta- Bajaj, Karan, Bhisham Sharma, and Raman Singh. (2020).
leb, H. (2019). MsM: A microservice middleware for Integration of WSN with IoT applications: a vision,
smart WSN-based IoT application. J. Netw. Comp. architecture, and future challenges. Integration of
Appl., 144, 138–154. https://doi.org/10.1016/j. WSN and IoT for Smart Cities. 79–102. doi: https://
jnca.2019.06.015. doi.org/10.1007/978-3-030-38516-3_5
Chaitra, M., and B. Sivakumar. (2017). Disaster debris de- Preeth, SK Sathya Lakshmi, R. Dhanalakshmi, R. Kumar,
tection and management system using wsn & iot.. and P. Mohamed Shakeel. (2018). An adaptive fuzzy
International Journal of Advanced Networking and rule based energy efficient clustering and immune-in-
Applications. 9(1): 3306–3310. spired routing protocol for WSN-assisted IoT system.
Díaz, S. E., Pérez, J. C., Mateos, A. C., Marinescu, M. C., Journal of Ambient Intelligence and Humanized Com-
and Guerra, B. B. (2011). A novel methodology for puting. vol. (2018): 1–13. https://doi.org/10.1007/
the monitoring of the agricultural production process s12652-018-1154-z
based on wireless sensor networks. Comp. Elec. Ag- Shahraki, A., Taherkordi, A., Haugen, O., and Eliassen, F.
ricul., 76(2), 252–265. https://doi.org/10.1016/j.com- (2021). A survey and future directions on clustering:
pag.2011.02.004. From WSNs to IoT and modern networking para-
Elappila, M., Chinara, S., and Parhi, D. R. (2018). Surviv- digms. IEEE Trans. Netw. Ser. Manag., 18(2), 2242–
able path routing in WSN for IoT applications. Pervas. 2274. https://doi.org/10.1109/TNSM.2020.3035315.
Mobile Comput., 43, 49–63. https://doi.org/10.1016/j. Shukla, A. and Tripathi, S. (2020). A multi-tier based cluster-
pmcj.2017.11.004. ing framework for scalable and energy efficient WSN-
Elappila, M., Chinara, S., and Parhi, D. R. (2020). Surviv- assisted IoT network. Wireless Netw., 26(5), 3471–
ability aware channel allocation in WSN for IoT ap- 3493. https://doi.org/10.1007/s11276-020-02277-4.
plications. Pervas. Mobile Comput., 61. https://doi. Khan, JavedAkhtar. (2019). —Multiple Cluster-Android
org/10.1016/j.pmcj.2019.101107. lock Patterns (MALPs) for Smart Phone Authentica-
Farhan, L., Shukur, S. T., Alissa, A. E., Alrweg, M., Raza, U., tion. In 2019 3rd International Conference on Com-
and Kharel, R. (2017). A survey on the challenges and puting Methodologies and Communication (ICCMC).
opportunities of the Internet of Things (IoT). Proc. 619–623. IEEE. doi: 10.1109/ICCMC.2019.8819635.
Int. Conf. Sens. Technol., ICST, 2017-December, 1–5. Sutagundar, A. V. and Manvi, S. S. (2013). Wheel based
https://doi.org/10.1109/ICSensT.2017.8304465. event triggered data aggregation and routing in wire-
Gao, Q., Zuo, Y., Zhang, J., and Peng, X. H. (2010). Improv- less sensor networks: Agent based approach. Wire.
ing energy efficiency in a wireless sensor network by Per. Comm., 71(1), 491–517. https://doi.org/10.1007/
combining cooperative MIMO with data aggregation. s11277-012-0825-x.
IEEE Trans. Vehicul. Technol., 59(8), 3956–3965. Tian, H., Shen, H., and Roughan, M. (2008). Maximizing
https://doi.org/10.1109/TVT.2010.2063719. networking lifetime in wireless sensor networks with
Gardašević, G., Katzis, K., Bajić, D., and Berbakov, L. regular topologies. Par. Distribut. Comput. Appl. Tech-
(2020). Emerging wireless sensor networks and in- nol., PDCAT Proc., 211–217. https://doi.org/10.1109/
ternet of things technologies—foundations of smart PDCAT.2008.29.
healthcare. Sensors (Switzerland), 20(13), 1–30. Tyagi, S. and Kumar, N. (2013). A systematic review on
https://doi.org/10.3390/s20133619. clustering and routing techniques based upon LEACH
Haseeb, Khalid, Ikram Ud Din, Ahmad Almogren, and Nav- protocol for wireless sensor networks. J. Netw. Comp.
eed Islam. (2020). An energy efficient and secure IoT- Appl. 36(2), 623–645. https://doi.org/10.1016/j.
based WSN framework: An application to smart ag- jnca.2012.12.001.
riculture. Sensors, 20(7): 2081 pp. 1–14. doi: https:// Weng, C. E. and Lai, T. W. (2013). An energy-efficient rout-
doi.org/10.3390/s20072081 ing algorithm based on relative identification and
Azzam, Riad, and Nabil Aouf. (2014). Embeded fusion of direction for wireless sensor networks. Wire. Per.
visual and acoustic for active acoustic source detec- Comm., 69(1), 253–268. https://doi.org/10.1007/
tion with SGGMM. In Proceedings ELMAR-2014. s11277-012-0571-0.
1–4. IEEE. doi: 10.1109/ELMAR.2014.6923352 Yu, X., Wu, P., Han, W., and Zhang, Z. (2013). A survey
Jaiswal, K. and Anand, V. (2020). EOMR: An energy-effi- on wireless sensor network infrastructure for agricul-
cient optimal multi-path routing protocol to improve ture. Comp. Stan. Interf., 35(1), 59–64. https://doi.
QoS in wireless sensor network for IoT applications. org/10.1016/j.csi.2012.05.001.
Wire. Pers. Comm., 111(4), 2493–2515. https://doi. Zhou, Z., Zhou, S., Cui, S., and Cui, J. H. (2008). Energy-
org/10.1007/s11277-019-07000-x. efficient cooperative communication in a clustered
Singh, J., Singh, S., Singh, S., and Singh, H. (2019). Evalu- wireless sensor network. IEEE Transac. Vehicular
ating the performance of map matching algorithms Technol., 57(6), 3618–3628. https://doi.org/10.1109/
for navigation systems: an empirical study. Spatial In- TVT.2008.918730.
form. Res., 27, 63–74. Zhu, C., Zheng, C., Shu, L., and Han, G. (2012). A survey
Jawad, Haider Mahmood, Rosdiadee Nordin, Sadik Kamel on coverage and connectivity issues in wireless sen-
Gharghan, Aqeel Mahmood Jawad, and Mahamod Is- sor networks. J. Netw. Comp. Appl., 35(2), 619–632.
mail. (2017). Energy-efficient wireless sensor networks https://doi.org/10.1016/j.jnca.2011.11.016.
15 Internet of health things-enabled monitoring of vital signs
in hospitals of the future
Amit Sundas1,a, Sumit Badotra2, Gurpreet Singh3 and Amit Verma4
1,3
Department of Computer Science and Engineering, Lovely Professional University, Phagwara, Punjab, India
2
School of Computer Engineering and Technology, Bennett University, Greater Noida, India
Department of Computer Science and Engineering, University Center for Research and Development, Chandigarh
4

University, Gharuan Mohali, India

Abstract
Vital signs and other extensive patient data are among the many types of information typically obtained by hand in hospitals
utilizing discrete medical equipment. It might be challenging for careers to integrate and analyze this information since it is
often kept in separate spreadsheets and not part of patients’ electronic health records. Connecting medical equipment via a
decentralized network such as the Internet is one way to get around these restrictions. By combining data from many sources,
we can more accurately assess a patient’s health and plan for preventative measures. In this study, we present the notion of
the internet of health things (IoHT) and conduct a broad landscape analysis of the methods that may be used to collect and
integrate data on vital signs in healthcare facilities. The potential use of intelligent algorithms is investigated, and common
heuristic techniques such weighted early warning score systems are addressed. In order to maximize efficiency, make the most
of available resources, and prevent unnecessary patient health decline, this article suggests potential avenues for merging pa-
tient data on hospital wards. It is stated that the IoHT paradigm will continue to provide better options for patient treatment
on hospital wards, and that a patient-centered approach is crucial.

Keywords: Machine learning, sepsis, vital sign, prediction, electronic health records

I. Introduction patient health decline and optimize hospital resources

by anticipating future patient needs.
Hospitalized patients undergo regular vital sign This paper explores the internet of health things
monitoring, a crucial practice that can prevent (IoHT), an emerging technology that enables inter-
health deterioration, reduce morbidity and mortal- connected devices to monitor patients’ health and
ity, shorten hospital stays, and alleviate financial share data. We propose the integration of ML with
burdens (Gultepe et al., 2013; Sundas et al., 2022). this architecture to correlate data and forecast future
However, the techniques employed for vital sign col- health trends and requirements. When information
lection in hospital wards lack standardization on a and communication technology (ICT) is applied in the
global scale. In some cases, manual data collection is healthcare context, it is known as eHealth or mHealth
still utilized, with patient-specific spreadsheets often (Vistisen et al., 2019). These terms are all centered on
discarded upon discharge. Alternatively, vital signs enhancing patient outcomes, with a particular focus on
can be recorded on devices such as tablets, personal mobile health services and ubiquitous health (uHealth).
digital assistants (PDAs), or other electronic tools and uHealth leverages pervasive and mobile computing to
stored in the patient’s electronic health record (EHR) continuously monitor an individual’s health, empha-
(Sundas et al., 2022). These recorded vital signs can sizing preventive and personalized care, departing
be leveraged to assess a patient’s health status through from the current paradigm (Chen et al., 2016).
heuristic methods like early warning or modified early In this article, we delve into the potential of IoHT
warning scoring (EWS/MEWS) (Sundas et al., 2023), for monitoring vital signs in hospital wards and
particularly in the United Kingdom. explore automated and intelligent approaches for
The internet of things (IoT) facilitates the interac- predicting patient health deterioration. The article is
tion and data analysis of devices (Sundas et al., 2021). structured into six sections. Section 2 discusses the
As a result, nurses can potentially automate the vital latest advancements in hospital patient treatment.
sign recording process. IoT employs cloud comput- Section 3 introduces the IoHT concept and techniques
ing in its distributed platform for data processing and for collecting vital signs in hospital wards. Section 4
storage (Sundas et al., 2022). This platform enables discusses the application of ML for data interpreta-
the application of machine learning (ML) algorithms tion. Section 5 addresses the major challenges and
(Gultepe et al., 2013; Sundas et al., 2022) to predict provides solutions. Section 6 concludes the study.

[email protected]
a
 Applied Data Science and Smart Systems 109

II. Hospital treatment focused on the individual rules determine which vital signs are measured, which
patient is contentious.
Elliott and Coventry recommended eight vital signs
Patient-centered care (PCC) stands as a pivotal
(Rana et al., 2018), adding pain, consciousness, and
hospital quality indicator (Gultepe et al., 2013). The
urine output (Iyer et al., 2022). Patients lived longer
assessment of PCC hinges on factors such as patient
Table 15.1 defines, normalizes, and impacts eight vital
needs, effective provider communication, and the
indicators hospitals may monitor for patient health.
availability of services. From an information tech-
Monitoring vital signs raises challenges about how
nology (IT) perspective, PCC can be equated to the
frequently and what to report. Unlike Table 15.1, not
patient’s EHR. This differs significantly from vari-
all vitals may be obtained instantly.
ous enterprise resource planning (ERP) systems of
Pain assessment is subjective. WILDA verifies pain
the past, which primarily aimed to optimize work-
terms, intensity on a 0–10 scale, location, duration,
flow and procedural aspects (Bloch et al., 2019). In
aggravating factors, and pain-relieving variables
the context of PCC, hospital ward vital signs play a
(Sundas et al., 2021). The patient-caregiver WILDA
crucial role, serving as essential markers for identify-
method may use computerized data recording. Tablets
ing patient health concerns and their correlation with
and smartphones can capture patient data for EHRs.
other pertinent data.
Assessing consciousness requires patient-provider
communication. The Glasgow Coma scale measures
A. Vital signs monitoring
eye opening, verbal, and motor responses (Khan et
Patient-centered care improves hospital treatment
al., 2021). These assessments’ numerical outcomes
(Tang et al., 2020). Patient-centered care is measured
depend on patient reactions to stimuli.
by how well it meets patient needs, how fast clinicians
Electronically capturing these numbers may assist
communicate health data, and how readily patients
evaluate the patient’s neurological condition. Catheters
may get treatments. The EHRs are PCC medical
may automatically record urine output. Manual uri-
information systems. Thus, ERPs enhance workflow
nometers are still used (Sundas et al., 2021).
(Khan et al., 2014).
Hospital ward vital signs, vital signs used to high- B. Patient risk assessment
light patient health issues, and vital signs connected Hospital PCC examines vital signs regularly and more
to other data may describe PCC. Hospital nurses have often if concerns arise. Data and graded response tech-
measured the same vitals since 1900 (Sundas et al., niques lower risk. Monitoring and triggering define
2021). how frequently, what, and when to check in. Table
Blood pressure, temperature, heart rate, and respi- 15.2 outlines healthcare institution risk assessment
ration have oxygen saturation. To effectively assess a approaches. Check metric first (Liu et al., 2014). This
patient’s state, National Institute for Health and Care group uses MET (Medical Emergency Team) calling
Excellence (NICE) suggests monitoring oxygen satu- criteria. MET is determined by airway threats, respi-
ration in addition to the five vitals. Consider urine ratory or cardiac arrest, state alterations, and convul-
output, discomfort, and biochemical testing. Hospital sions (Singh et al., 2019; Sundas et al., 2021). Group

Table 15.1 Typical hospital vitals used for patient monitoring

Definition Vital sign Some influencing factors Normal range

A pain scale is used to measure Pain of level The view from the patient The patient reports no pain
the patients’ pain intensity on the 0–10 pain scale (1–3
mild, 4–6 moderate, 7–10
severe)
The force multiplied by the Blood pressure Variables such as age, posture, 90/60–120/80 mmHg
period between heartbeats effort, sleep, slant, and
(systole) that blood exerts on confounding variables (such as
arteries White-Coat-Syndrome or anxiety)
How many times in 60 seconds Breathing rate Variables such as age, oxygen Breathing rate: 12–1/8/min
the chest moves up and down levels in the environment, pain and
anxiety levels, and physical effort
Estimates the quantity of oxygen SPO2 Workload, oxygen levels, and From 95% to 100%
in the blood by measuring the other confounding variables (such
saturation level of its peripheral as activity and pain intensity)
capillaries (SpO2)
110 Internet of health things-enabled monitoring of vital signs in hospitals of the future
Table 15.2 Methods, frameworks, and systems for assessing risk

Common practices Type Characteristics

EWS + MET, MEWS + PART intensity Combination Monitoring development, graded response, varying
sensitivity and specificity
Acceptable calling standards Single parameter Easy to use, but no improvement tracking
PART Multiple parameter high sensitivity and low specificity, yet it allows
progress tracking and progressive reaction
The worthing physiological scoring Aggregate scoring Allows development monitoring, progressive response,
system (EWS, MEWS, ViEWS) and high sensitivity and specificity, based on the score

2 needs one abnormal vital sign. PART (Prehospital Heuristics underpin all these methods. These meth-
Acuity Rating for Triage) calling requirements are an ods compare physiological measures to preset criteria,
example. The PART method measures respiration, resulting in many false positives. These examinations
heart rate, systolic blood pressure, consciousness, may employ artificial intelligence (AI) to better diag-
oxygen saturation, and urine output. nose the patient.
The third category of risk assessment focuses on
evaluating vital signs to detect early signs of health C. Health information recording
deterioration. The early warning score (EWS) was ini- The PCC programmers include physiological obser-
tially introduced as the scoring system. Notably, EWS vations at admission and throughout hospitaliza-
has gained widespread adoption in most UK hospitals tion. For this reason, healthcare workers employ
due to its endorsement by the NICE, and its proven EHR systems. Healthcare practitioners manage most
effectiveness (Khan et al., 2021). The EWS relies on EHRs. These systems solely monitor the patient’s cur-
calculated data, assessing a patient’s health by ana- rent healthcare provider. Combining data from sev-
lyzing multiple parameters at varying intervals. The eral EHR systems doesn’t provide a comprehensive
vital sign data needed for EWS can be entered into the patient EHR (Chen et al., 2016).
patient’s EHR either manually or automatically, and Personal health records (PHR) are one option for
this allows for continuous calculation and visualiza- achieving a holistic and unified picture of a patient’s
tion of the EWS score over time. The process of man- health. A PHR is an individual’s own representation
aging this data can be carried out using mechanical or of their health records, which may consist of separate
manual EWS devices, including paper-based systems. pieces of data or include data from several other sources.
The initial EWS score is computed based on vital Patients have full authority over their PHRs, including
signs such as systolic blood pressure, temperature, the ability to appoint a proxy or set access privileges.
heart rate, breathing rate, and level of awareness (Liu Involving patients in the management of their health
et al., 2014). Each vital sign is compared to estab- information improves collaboration and participation
lished norms to determine an individual score, with a in therapy (Sundas et al., 2021). Furthermore, the intro-
range of 0–6 for systolic blood pressure and 0–3 for duction of mobile devices and wearables drastically
the remaining parameters. The overall EWS score is alters the patient’s role in engaging with their PHR by
derived by summing up the scores for each vital sign enabling real-time monitoring of vital signs, supple-
and adding them to the respective norms. It’s worth menting health information, and allowing for more
noting that there exist several versions of the EWS. proactive intervention. The current tendency is for
One such variant is the modified early warning score patients to supplement their healthcare providers’ EHR
(MEWS), which incorporates urine output as the sixth data with data collected from their own wearables and
vital indicator. Additionally, the vitalpac early warning mobile devices (Sundas et al., 2022) (Figure 15.1).
score (ViEWS) offers a solution for bedside vital sign
monitoring using a PDA. Another system, the worth- III. Internet of health things
ing physiological scoring system, takes into account a
broader range of vital signs, including respiration rate, Since 1999, the IoT has grown into a worldwide
heart rate, arterial pressure, body temperature, oxy- sensor, wireless communication, and information
gen saturation, and level of awareness, in estimating a processing network. The IoT relies on smart items,
patient’s risk of adverse outcomes. These risk assess- which can transmit and analyze information to
ment methods in the third category combine the ease interact autonomously. Recent efforts to define, IoT
of use from the previous categories with the enhanced have included sensing environmental data, providing
sensitivity provided by the EWS and its variants. communication services, analytics, applications, and
 Applied Data Science and Smart Systems 111

et al., we propose using cloud computing because of

its inherent qualities, such as on-demand self-service,
widespread network access, resource pooling to meet
scaled demand, quick flexibility, and metering capa-
bilities. Recently, a kind of cloud computing has been
promoted for IoHT as a solution to the excessive
latency of health monitoring systems. Fog computing
is a kind of cloud computing that integrates locally
based devices (also known as edge computing) with
cloud-based resources in a decentralized method.
Patient data is registered in a PHR that is semanti-
Figure 15.1 Provider-controlled electronic health re- cally interoperable, which is another significant feature.
cords (EHR) and patient-controlled, mobile, wearable The processing of patient information involves
sensor-based personal health records (PHR) are two analyzing such information. Here, we suggest that
types of electronic health records (EHR) smart algorithms grounded on ML should be used
in place of more conventional heuristic methods.
Improved resource allocation is anticipated as a result
of improved patient health deterioration inference
made possible by cutting-edge data fusion and predic-
tive analytics.
Results are presented as a synthesis of the preced-
ing levels, or “presentation,” at the final stage. These
may appear as notifications, recommended next steps,
charts, or graphs. De-identified data from many PHRs
within a given geographical area, metropolitan area,
or healthcare facility may be combined to provide
epidemiological perspectives. IoT healthcare appli-
cations are just getting off the ground. Cost savings,
improved quality of life, and enhanced user experi-
ence are all conceivable outcomes (Tang et al., 2020).
Figure 15.2 Acquisition, storage, processing, and dis- From the standpoint of healthcare providers, IoHT
play are shown at the bottom of this schematic of the can minimize disruptions in service, pinpoint when it
Internet of Health Things (IoHT) is most convenient to restock supplies, and make the
most effective use of scarce assets.
information exchange. The IoT can be things-centric
(from sensors), internet-centric (from middleware and A. Connectivity tools for IoT
architecture), semantic-centric (from knowledge), or Various wireless protocols are currently employed to
user-centric (enabling innovative applications focused coordinate wearable health monitors within IoHT.
on people) (Vistisen et al., 2019). We utilize the IoT Additionally, communication technologies based on
to monitor hospital patients’ vital signs in this article. electromagnetic fields, such as radio-frequency identi-
Connected medical devices that can share and fication (RFID) and near-field communication (NFC),
interpret data in order to better care for patients make have been explored as options for IoHT applications.
up what is known as the “IoHT”. This focus on the Initially, RFID embraced in the logistics sector,
patient entails four levels of analysis (Figure 15.2). involves the use of readers and tags. The RFID tech-
Smart health objects (SHO) include things like nology gained early exposure in the context of IoT
wearables and medical equipment and are acquired within logistics (Tang et al., 2020). The RFID systems
in the first phase. An SHO’s primary responsibility is can utilize either passive or active tags. Active RFID
to record information about a patient’s vital signs and tags initiate communication via an internal battery
other physiological statuses. and can operate at greater ranges, while passive RFID
Standard protocols (e.g., those under the wing of tags rely on the reader’s signal for communication
ISO, HL7, DICOM, and others) and other technolo- and do not require a battery. The development of
gies (e.g., Bluetooth, Wi-Fi) are often used for their ultra-high frequency (UHF) tags, with advanced sens-
communicative capacities. ing and computing capabilities, has paved the way for
For example, “storage” is in charge of representing battery-free, cost-effective monitoring and transmis-
the gathered data in a manner that is both scalable sion of patients’ vital signs, contributing to RFID’’s
and interoperable. In line with the research of Rana promising role in healthcare (Li-wei et al., 2016).
112 Internet of health things-enabled monitoring of vital signs in hospitals of the future

NFC, on the other hand, is a contactless proximity possible remedies. The increased complexity and vari-
communication technology that operates at close dis- ety of data has led to a rise in research on big data
tances, typically within approximately 4 cm in prac- analysis in healthcare. We focus on ML approaches
tice (though theoretically, it can work at distances of for data modeling. These approaches typically include
less than 10 cm). The proximity nature of NFC, cou- three steps: data collection, feature selection and
pled with its user-friendliness, makes it an excellent extraction, and learning (Li-wei et al., 2014).
choice for enabling communication between patients IoHT devices and other healthcare equipment with
and their medical data. registration and synchronization capabilities gather
Finally, low-power area networks like 6LoWPAN data. After pre-processing (filtering, standardizing, and
enable the delivery of IPv6 packets in wireless sensor aggregating), features (signal descriptive statistics, tem-
networks (WSNs), extending the reach of the IoT to poral and frequency domain characteristics) that dis-
the level of individual sensor nodes. IPv6’s scalability, criminate the patient’s health condition are identified
improved mobility features, and support for multiple and chosen. A classifier or regressor algorithm is taught
stakeholders’ management have enhanced the admin- to relate the data to health deterioration. Deep learn-
istration of smart objects. Consequently, 6LoWPAN ing uses algorithms to extract information from raw
is widely recognized as the foundational technol- data instead of hand-crafted qualities. After training,
ogy for the IoHT in a substantial body of literature the model may be used as a decision support system
(Baidillah et al., 2023). to evaluate a patient’s health or offer relevant actions.

B. IoHT with real-time health status tracking IV. Algorithms with intelligence for monitoring
The IoHT may help hospitals manage PCC and patient vital signs
data. Nurses manually take vital signs. Manual sphyg-
momanometers, stethoscopes, pain and consciousness Many ML algorithms incorporate critical factors to
questionnaires are employed. In these cases, a smart- enhance their predictive capabilities. Support vector
phone or tablet might help the caretaker by providing machines (SVMs), for instance, are capable of assess-
additional information or collecting data. Electronic ing patient risk by considering various indicators,
vital sign registration saves time and labor. including patient demographics, laboratory findings,
A WSN of wireless personal devices may collect and vital signs. SVMs can predict daily risk ratings for
vital indicators. IPv6 over 6LoWPAN is replacing patients and then aggregate these ratings to stratify
manufacturer specifications for linking smart health overall risk.
devices. Sundas et al., presented a multisensor pain In the healthcare domain, decision trees are com-
assessment approach. These sensors include acceler- monly used for disease classification, enabling the
ometers and GPS trackers for activity levels, micro- identification and categorization of different medi-
phones, and a computer’s capacity to interpret speech cal conditions. Medical research has increasingly
and facial expressions. The authors noted the abun- explored the use of individual neural networks (NN)
dance of smartphone and tablet pain measuring appli- and their integration with other approaches. Notably,
cations. Similar to Aung and colleagues’ technique, one of the early experiments applied long short-term
one may employ image processing to analyze eye memory (LSTM) recurrent neural networks (RNN)
movement, speech analysis to evaluate verbal replies, to detect patterns in EHR data. These NN are well-
and an accelerometer or gyroscope to evaluate motor suited for handling time series data, irregular sam-
responses to determine the patient’s awareness. BP, pling, and gaps in medical records.
temperature, HR, RR, and SpO2 may be monitored To address issues related to missing data, research-
using IoHT. Otero and colleagues recommended auto- ers have developed deep models incorporating gated
matic urine monitoring for severely unwell patients. recurrent units (GRU), which provide effective rep-
RFID, NFC, and Bluetooth can communicate sensor resentations for incomplete or intermittent data. The
data to smartphones. Smartphones convey this data application of RNN to medical data represents a bur-
to a fog or cloud middleware. Intelligent algorithms geoning area of research that continues to evolve and
for processing huge patient data requires further development (Li-wei et al., 2014).
The IoHT, high-throughput sequencing platforms
(genomics, proteomics, and metabolomics), real-time A. Deep learning techniques [11–15] are another op-
imaging, and point-of-care diagnostic devices have tion to consider
made health informatics a data-rich field. Environment NN with several nested layers and neurons lies at the
and social media may provide health information (Liu heart of these approaches (Iyer et al., 2022). Using a
et al., 2014). “Big data analysis” is the processing of large number of neurons enables the coverage of a great
massive amounts of vital sign data using sophisticated deal of raw data, and the option of cascading many
algorithms to identify health decline risks and predict layers enables the automated abstraction of a higher
 Applied Data Science and Smart Systems 113

level, eliminating the need for human involvement patients with closed-loop alarm. IoT-enabled Smart
When applied to vital signs data, this function may help Healthcare Sys. Serv. Appl., 143–176.
extract potentially nuanced and obscure insights from Vistisen, S. T., Johnson, A. E. W., and Scheeren, T. W. L.
simple observation. Ravi and coworkers point to con- (2019). Predicting vital sign deterioration with artifi-
cial intelligence or machine learning. J. Clin. Monit.
volutional neural networks (CNNs) as the kind of deep
Comput., 33(6), 949–951.
learning having the most influence on health informatics
Singh, Jaiteg, and Nandini Modi. (2019). Use of information
at the moment (Bloch et al., 2019). CNN has been stud- modelling techniques to understand research trends in
ied extensively, but most of the time it’s used to analyze eye gaze estimation methods: An automated review. He-
photos of the human body for diagnosis. liyon, 5(12), 1–12.
Lujie, C., Dubrawski, A., Wang, D., Fiterau, M., Guillame-
V. Conclusion Bert, M., Bose, E., Kaynar, A. M. et al. (2016). Using
supervised machine learning to classify real alerts and
This review included vital sign monitoring and analy- artifact in online multi-signal vital sign monitoring
sis to the IoHT to predict patient health risks. The data. Crit. Care Med., 44(7), e456.
first portion of the review included the eight main Bloch, Eli, Tammy Rotem, Jonathan Cohen, Pierre Sing-
physiological observations: blood pressure, body tem- er, and Yehudit Aperstein. (2019). Machine learn-
perature, heart rate, respiration rate, oxygen satura- ing models for analysis of vital signs dynamics: a
case for sepsis onset prediction. Journal of health-
tion, pain, degree of consciousness, and urine output.
care engineering. vol. 2019. 1–12. https://doi.
The article highlighted the first five vital indicators
org/10.1155/2019/5930379
as most important. We then examined how hospitals Baidillah, Marlin Ramadhan, Pratondo Busono, and Riyanto
assess patients’ health risks using tracking and trigger- Riyanto. (2023). Mechanical ventilation intervention
ing systems. Most current approaches (typically EWS based on machine learning from vital signs monitor-
or versions thereof) are heuristics with hard-and-fast ing: A scoping review. Measurement Science and Tech-
thresholds, and just a fraction apply AI. The move nology. 34, 062001. Doi: 10.1088/1361-6501/acc11e
from EHRs to PHRs highlights the need of semantic Shengpu, T., Chappell, G. T., Mazzoli, A., Tewari, M., Choi,
interoperability in integrating and exchanging health- S. W., and Wiens, J. (2020). Predicting acute graft-ver-
care data. Today, vital signs may be collected using sus-host disease using machine learning and longitudi-
wearable devices with Bluetooth, NFC, RFID, or nal vital sign data from electronic health records. JCO
Clin. Cancer Inform., 4, 128–135.
UWB connections and gateways to connect hospital
Khan, M. I., Jan, M. A., Muhammad, Y., Do, D.-T., Ur
ward medical equipment. Next, ML processed cru-
Rehman, A., Mavromoustakis, C. X., and Pallis, E.
cial indicators. The IoHT notion introduces various (2021). Tracking vital signs of a patient using channel
issues, allowing for additional research and develop- state information and machine learning for a smart
ment. Prevention and individualization replace symp- healthcare system. Neural Comput. Appl., 1–15.
tom- and disease-focused therapy in the IoHT. This Liu, N. T., Holcomb, J. B., Wade, C. E., Darrah, M. I., and
vital sign monitoring system may help doctors pre- Salinas, J. (2014). Utility of vital signs, heart rate vari-
dict future treatments and interventions. Thus, IoHT ability and complexity, and machine learning for iden-
will improve ward-based patient care. This requires a tifying the need for lifesaving interventions in trauma
patient-centered approach. patients. Shock, 42(2), 108–114.
Li-wei, H. L., Mark, R. G., and Nemati, S. (2016). A model-
based machine learning approach to probing autonom-
References ic regulation from nonstationary vital-sign time series.
Gultepe, Eren, Jeffrey P. Green, and Hien Nguyen. (2013). IEEE J. Biomed. Health Informat., 22(1), 56–66.
From vital signs to clinical outcomes for. 1–11. doi: Li-wei, H. L., Nemati, S., Moody, G. B., Heldt, T., and
10.1136/amiajnl-2013-001815 Mark, R. G. (2014). Uncovering clinical significance
Amit, S., Badotra, S., Bharany, S., Almogren, A., Tag-ElDin, of vital sign dynamics in critical care. Comput. Car-
E. M., and Rehman, A. U. (2022). HealthGuard: An in- diol., 1141–1144.
telligent healthcare system security framework based Srikrishna, I., Zhao, L., Mohan, M. P., Jimeno, J., Siyal,
on machine learning. Sustainability, 14(19), 11934. M. Y., Alphones, A., and Karim, M. F. (2022). mm-
Amit, S., Badotra, S., Rani, S., and Gyaang, R. (2023). Wave radar-based vital signs monitoring and arrhyth-
Evaluation of autism spectrum disorder based on the mia detection using machine learning. Sensors, 22(9),
healthcare by using artificial intelligence strategies. J. 3106.
Sensors, 2023, 1–12. Rana, Soumya Prakash, Maitreyee Dey, Robert Brown,
Amit, S. and Panda, S. N. (2021). Real-time data communi- Hafeez Ur Siddiqui, and Sandra Dudley. (2018). Re-
cation with IoT sensors and things speak cloud. Wire. mote vital sign recognition through machine learning
Sen. Netw. Internet Things, 157–173. augmented UWB. 12th European Conference on An-
Amit, S., Badotra, S., Rani, S., and Gajare, C. M. (2022). tennas and Propagation (EuCAP 2018), London, UK,
WSN-and IoT-based smart surveillance systems for 1–5, doi: 10.1049/cp.2018.0978.
16 Artificial intelligence-based learning techniques for
accurate prediction and classification of colorectal cancer
Yogesh Kumar1,a, Shapali Bansal2, Ankush Jariyal3 and Apeksha Koul4
1
Department of CSE, School of Technology, Pandit Deendayal Energy University, Gandhi Nagar, Gujarat, India
2,3
Department of Computer Applications, USMS, Rayat Bahra University, Mohali, India
4
Department of Computer Science and Engineering, Punjbai University, Patiala, Punjab, India

Abstract
Colorectal cancer (CRC) is a prominent source of illness and death worldwide. Detection and precise diagnosis of CRC at an
early stage can significantly enhance patient outcomes. Artificial intelligence (AI) has yielded promising results in the detec-
tion and classifications of CRC. The application of machine learning (ML) algorithms, deep learning (DL), and computer-
assisted diagnosis systems are only a few of the most current advances in the use of AI techniques for CRC detection and
diagnosis that we discuss in this study. In the article, we also compared and evaluated the CRC detection work of various
researchers using various performance parameters such as accuracy and loss. We also examine the types and epidemiology of
CRC, which aids in the diagnosis of the numerous CRC cancer types. AI has the possible to substantially enhance the detec-
tion and diagnosis of cancer, leading to improved patient health and lower healthcare costs.

Key words: Colorectal cancer, artificial intelligence, epidemiology, deep learning, machine learning, computer-assisted diagnosis

I. Introduction The colon or large bowel is an important part of the

gastrointestinal tract which starts from the esophagus
Colorectal cancer (CRC) is a form of cancer that
to the anus (Chaplot et al., 2023). The large intestine
mostly affects the rectum or colon part of the body. It
of the human body is mainly made up of the colon,
occurs when cells in the colon or rectum lining pro-
which is around 1.5 m long and is divided into vari-
liferate and divide uncontrollably, forming a tumor.
ous sections such as (Kumar et al., 2021):
Certain risk factors for CRC have been identified,
including age (the risk increases with age), family his- Ascending colon (15–20 cm): The first section starts
tory of the disease, a diet high in red and low in fruits with the pouch called the caecum. Its role is to receive
and vegetables, smoking, and a sedentary lifestyle the undigested food from the small intestine.
(Chaplot et al., 2023).
Transverse colon (50 cm): The second section goes
Changes in gastrointestinal habits, abdomi-
across the body to the left from the right side. The
nal pain or discomfort, blood in the stool, abrupt
transverse and ascending colon are collectively called
weight loss, and fatigue may be symptoms of CRC.
as proximal colon.
However, some individuals with CRC may exhibit
no symptoms. The screening procedures for CRC Descending colon (25 cm): It is the third section that
such as colonoscopy and colon occult blood tests, descends on the left side.
can detect the disease at an early, more treatable The sigmoid colon (7.5–12 cm): It is the last and the
stage. Treatment for CRC be subject to on the stage fourth section, which is S-shaped. This part of the
and location of the cancer, but may include surgery, colon joins the rectum, which later connects to the
radiation therapy, and chemotherapy. It is crucial to anus. The sigmoid and descending colon are collec-
prioritize a healthy lifestyle, which involves incorpo- tively called the distal colon.
rating regular physical activity and a well-balanced
diet, in order to minimize the chances of developing When some abnormal cells start growing from the
CRC. Additionally, early detection through screening inner lining of the colon or rectum, it is called colorec-
can significantly recover the probabilities of effec- tal cancer, and such uncontrollable growth is called
tive treatment and recovery. This section covers the polyps (Hamabe et al., 2022). The CRC is invasive neo-
reason for applying artificial intelligence (AI) tech- plasia that occurs as intestinal epithelium tumor in situ
niques to detect and diagnose CRC, its brief study, (TIS) and grows in different morphological ways. It has
types, epidemiology, and finally, traditional and AI been demonstrated that adenomatous polyps can be a
methods to analyze it. precursor of invasive cancer, although only 5–10% of
them turn into malignant tumors (Fearon, 2011).

[email protected]
a
 Applied Data Science and Smart Systems 115

The paper is ordered in the following method: Table 16.1 Cases and deaths in the US 2020 due to CRC
Section 2 presents the types and epidemiology and
Age Cases Deaths
various types of CRC. Section 3 presents the con-
(years)
ventional and AI-based diagnosis method. Section 4 CRC Colon Rectum Colorectum
describes the current state-of-the-art techniques for
detecting CRC and highlights any gaps or limitations 0–49 17,930 11,540 6390 3640
in the existing methods. Whereas Section 5 defines the 50–64 50,010 32,290 17720 13,380
methodology and steps to follow the CRC detection 65+ 80.010 60,780 19,230 36,180
using deep learning (DL)-based approaches. Section
All ages 14,7950 10,4610 43,340 53,200
6 concludes the study and presents the significance of
learning models for CRC detection.

II. Types and epidemiology of CRC malignancies are combined and shown in the Table
(American Cancer Society 2020).
There are various types of CRC which are shown in
Table 16.1, along with its brief description and symp-
toms. The CRC can be broadly classified into two III. Diagnosis of CRC
main types: Conventional techniques: People who do physical
activities have been linked to a higher incidence of
Adenocarcinoma: It represents 96% of cases, this is rectal cancer but not colon cancer. According to the
the most prevalent kind of CRC. The cells that lining research, those who are physically active are at the
the inside of the colon and rectum are where adeno- risk of 25% of having distal and proximal colon
carcinoma develops. cancers compared to those who are not. Consuming
Carcinoid tumors: An uncommon form of colon can- aspirin and other non-steroid anti-inflammatory
cer that develop in the intestine’s hormone-producing medicines have also been shown to decrease the risk
cells. Less than 1% of CRCs are caused by them. There of CRC (Howard et al., 2008). Furthermore, other
are also several subtypes of adenocarcinoma of the medications, such as oral bisphosphonates, are used
colon and rectum, which are classified based on their for treating and preventing osteoporosis, which may
microscopic appearance and genetic characteristics. lessen the risk of CRC.
AI techniques: The increasing workload of the pathol-
CRC was rarely identified at least 10 years ago.
ogist in terms of more time and labor consumption
Having 9,00,000 deaths annually is considered the
has tried to incorporate the introduction of computa-
fourth most fatal malignancy globally. It is the most
tional-based pathology for CRC diagnosis. We know
prevalent cancer in men, accounting for 10% of all
that AI has changed the pathology sector. It has been
cases globally, followed by lung cancer (17.2%) and
used to inspect Whole-slide imaging (WSI) data which
prostate cancer (20.3%), and it is the second most
may provide a computer-aided diagnosis of tumors
common cancer in women, accounting for 9.4%
using medical image analysis and various learning
of all cases worldwide, trailing only breast cancer
models such as machine learning (ML) and DL (Cui
(30.9%) (Kanna et al., 2023). In the United States,
et al., 2021).
CRC is the third leading cause of cancer-related mor-
tality among men and women and the second leading Present investigation has demonstrated that AI
cause of cancer deaths among men and women com- plays a vital role to diagnose and treat CRC patients.
bined. It is estimated that 52,580 persons will perish It is a responsible for improving early screening effi-
by 2022 (Sisodia et al., 2023). For several decades, ciency and dramatically improving CRC patients’
the death rate from CRC (per 100,000 persons per 5-year survival rate after treatment. Since 2010, there
year) has decreased in both men and women. There has been a substantial increase in the study and appli-
are several possible explanations for this. One rea- cation of AI in medically assisted gastrointestinal dis-
son is that colorectal polyps are now being discov- ease diagnosis and therapy. AI can help doctors with
ered and removed more frequently through screening the qualitative diagnosis and stage of colon cancer,
before they can develop into malignancies, or cancers which is now reliant on colonoscopy and pathologi-
are being discovered sooner when they are simpler to cal biopsies (Wang et al., 2020).
cure. Furthermore, CRC treatments have improved The researchers, such as Takemura et al. (2012),
during the last few decades (Wolf et al., 2018). utilized narrow-band imaging (NBI) along with a sup-
Table 16.1 projects the number of cases and deaths port vector machine algorithm, a supervised machine
in the United States for 2020. Due to the misclassifica- learning algorithm to calculate extreme points at
tion of rectal cancer deaths as colon, deaths for both the margin. These extreme points were employed to
116 Artificial intelligence-based learning techniques for accurate prediction and classification

identify exceptional parameters on the boundary, histopathology images to review existing research
enabling the differentiation between neoplasia polyps on AI in CRC. According to the authors, DL algo-
and nonneoplasia polyps. The approach achieved a rithms in histopathology are capable of diagnosing,
detection accuracy of 97.8%. identifying the features of histological images related
This shows that an AI can reliably evaluate colo- to prognosis, predicting clinical-based molecular phe-
noscopy biopsies at a rate that is on par with a prac- notypes, and evaluating the specific components of
ticing pathologist. The progress of AI applications the tumor.
in the medical arena suggests that AI will eventually Similarly, (Mitsala et al., 2021) investigated the
be employed for the diagnostic of CRC despite the usefulness of AI systems in medical therapy and diag-
dearth of systematic research. nosis by yielding numerous outstanding outcomes.
They stated that AI-assisted procedures in routine
screening are a critical step in lowering CRC inci-
IV. Related work
dence rates. In this approach, many researchers have
Significant advances have been made by AI techniques used AI algorithms to identify and diagnose CRC, but
in the health arena to demonstrate clinical applica- they also confront significant challenges, as shown in
tion potential. As a result, (Davri et al., 2022) used Table 16.2. This section covers the work done by the

Table 16.2 Comparative analysis of CRC

Author’s name Year of Dataset Techniques Outcome Limitation

publication

Zhang et al. 2019 1104 endoscopic CNN Accuracy: 86% Class imbalance
non-polyp images, AUC: 1
826 polyp images
Yamada et al. 2019 ImageNet dataset CNN Sensitivity: 97% The system performed
weak in order to detect
AUC: 0.98% lesions in the different
areas of the medical
image
Misawa et al. 2016 1079 narrow band CNN Specificity: 63.3% Unable to classify
imaging images Accuracy: 76.5% correctly because of the
limited dataset
Geetha et al. 2016 703 images Hand Sensitivity: 95% Model trained with
crafted LBP Specificity: 97% limited dataset
Ito et al. 2018 41 cases of colon CNN using Accuracy: 81.2% High cost, low efficiency
endoscopies yielded machine
190 pictures of colon learning
lesions algorithms
Yu et al. 2016 18 colonoscopy CNN Sensitivity: 71% Limited GPU memory,
videos specific length of video
clips were used
Figueiredo 2019 1680 cases of polyps SVM Sensitivity: 99% The model failed to
et al. and 1360 frames of Specificity: 85% evaluate the dimension of
healthy mucosa Accuracy: 91% colorectal polyp
Billah et al. 2017 14,000 still images CNN Prediction rate: Consumes more
98.6% processing time
Ozawa et al. 2020 16,418 images CNN Sensitivity: 92% Less number of training
Accuracy: 83% images were used
Urban et al. 2018 8,641 hand-labeled CNN Sensitivity: 90% The model failed to
images indicate the histology of
polyps
Tsai et al. 2009 CRC-VAL-HE-7K ResNet101 Accuracy: 98.81% Class imbalance issue

Ho et al. 2022 66,191 images AI learning Sensitivity: 97.4% Small dataset

models Specificity: 60.3%
Accuracy: 79.3%
AUC: 91.7%
 Applied Data Science and Smart Systems 117

researchers to detect CRC using various ML and DL The research methodology for the proposal is men-
techniques along with the research gaps. tioned as under:

V. Research methodology • The research primarily focuses on a literature

review, in which the datasets, approaches, and
The study of AI is becoming more interested in areas outcomes of various researchers working on pre-
such as algorithms and gadgets that enable people dicting CRC using multiple AI techniques are pre-
to tackle technically challenging issues. Researchers sented.
also use AI to identify epidemics’ environmental and • Various current techniques to reduce the use of
epidemiological patterns to anticipate outbreaks. the limited dataset, modeling errors, dereliction
This is being done to prevent epidemics from occur- of models, and class imbalance will be examined
ring. Mathematical models and DL can analyze vast to find new possible outcomes.
amounts of data to offer insight into the next likely • As illustrated in Figure 16.1, an open-source da-
source of illness. Ecologists can protect and moni- taset of different types of CRC such as tumors,
tor prospective host species more effectively with the stroma, complex, lymph, debris, mucosa, and adi-
help of these projections, which ultimately helps them pose can be used for implementation.
prevent future outbreaks. In epidemiology, AI is cur- • Initially, the data can pre-process to eliminate
rently being utilized to help with disease prevention noisy signals, missing values, NAN values, etc.,
and management and tracking and forecasting (Koul thereby enhancing the data quality.
et al., 2023). New situations, such as the current • Later, exploratory data analysis can be performed
coronavirus pandemic, provide opportunities for AI to classify the types of CRC to aid us in a better
to have the most impact. We should have high hopes understanding of the data.
that international cooperation will improve due to • Cancerous features can be extracted from the
the increased use of AI in medical systems. This will CRC dataset using various feature extraction and
allow us to battle epidemics better. Medical practitio- scaling strategies.
ners can employ AI techniques to aid them in mak- • To identify and classify different types of CRC,
ing more accurate and simpler judgments based on multiple learning models can be used, and their
patients’ experiences and historical facts (Koul et al., performance will be assessed.
2022; Kumar et al., 2023).

Figure 16.1 System design for CRC detection

118 Artificial intelligence-based learning techniques for accurate prediction and classification

• Later, propose a novel hybrid deep learning model rectal cancer using high-resolution MRI. PLoS One,
for the early prediction of different types of CRC. 17(6), e0269931.
• In the end, accuracy, loss, recall, precision, F- Ho, C., Zitong, Z., Xiu, F. C., Jan, S., Sahil, A. S., Rajasa,
score, performance testing, etc., can be used to J., Kaveh, T. et al. (2022). A promising deep learning-
assistive algorithm for histopathological screening of
validate the proposed model’s implemented re-
colorectal cancer. Scientif. Reports, 12(1), 2222.
sults during both the training and testing phase.
Howard, R. A., Michal Freedman, D., Yikyung, P., Albert,
H., Arthur, S., and Michael, F. L. (2008). Physical ac-
VI. Conclusion tivity, sedentary behavior, and the risk of colon and
rectal cancer in the NIH-AARP diet and health study.
The study assists readers (physicians, analysts, doc- Can. causes Con., 19, 939–953.
tors, and so on) in identifying previously utilized Ito, N., Hiroshi, K., Hirotaka, N., Masaya, U., Hideaki,
strategies or algorithms used by the researchers to M., and Hisahiro, M. (2018). Endoscopic diagnostic
detect CRC. In this research, we first focus on the lim- support system for cT1b colorectal cancer using deep
itations of the researchers’ work, such as optimizing learning. Oncol., 96(1), 44–50.
the model and loss function and examine its ability Kanna, G. P., Jagadeesh Kumar, S. J. K., Parthasarathi, P.,
on the significant histopathological dataset. Later, an and Yogesh, K. (2023). A review on prediction and
AI-based model can be designed to assist end users prognosis of the prostate cancer and gleason grading
of prostatic carcinoma using deep transfer learning
in detecting anomalies such as polyps to enhance the
based approaches. Arch. Computat. Methods Engg.,
diagnosis of CRC in a short period. The suggested
1–20.
models can be verified further for real-time images to Koul, A., Rajesh, K. B., and Yogesh, K. (2023). Artificial in-
test its efficacy. telligence techniques to predict the airway disorders
illness: a systematic review. Arch. Computat. Methods
References Engg., 30(2), 831–864.
Koul, A., Yogesh, K., and Anish, G. (2022). A study on
American Cancer Society. (2020). Colorectal cancer facts & bladder cancer detection using AI-based learning tech-
figures 2020–2022. Published Online, 48. niques. 2022 2nd Int. Conf. Technol. Adv. Computat.
Billah, Mustain, Sajjad Waheed, and Mohammad Motiur Sci. (ICTACS), 600–604.
Rahman. (2017). An automatic gastrointestinal pol- Kumar, Y., Inderpreet, K., and Shakti, M. (2023). Food-
yp detection system in video endoscopy using fusion borne disease symptoms, diagnostics, and predictions
of color wavelet and convolutional neural network using artificial intelligence-based learning approaches:
features. International journal of biomedical imag- A systematic review. Arch. Computat. Methods Engg.,
ing. vol. 2017. 1–9. https://doi.org/10.1155/2017/ 1–26.
9545920 Kumar, Y., Surbhi, G., Ruchi, S., and Yu-Chen, H. (2021). A
Chaplot, N., Dhiraj, P., Yogesh, K., and Pushpendra Singh, systematic review of artificial intelligence techniques
S. (2023). A comprehensive analysis of artificial intel- in cancer prediction and diagnosis. Arch. Computat.
ligence techniques for the prediction and prognosis of Methods Engg., 1–28.
genetic disorders using various gene disorders. Arch. Misawa, M., Shin-ei, K., Yuichi, M., Hiroki, N., Shinichi,
Computat. Method Engg., 30(5), 3301–3323. K., Yasuharu, M., Toyoki, K. et al. (2016). Character-
Cui, M. and David, Y. Z. (2021). Artificial intelligence and ization of colorectal lesions using a computer-aided
computational pathology. Lab. Invest., 101(4), 412– diagnostic system for narrow-band imaging endocy-
422. toscopy. Gastroenterol., 150(7), 1531–1532.
Davri, A., Effrosyni, B., Theofilos, K., Georgios, N., Niko- Mitsala, A., Christos, T., Michail, P., Constantinos, S.,
laos, G., Alexandros, T. T., and Anna, B. (2022). Deep and Alexandra, K. T. (2021). Artificial intelligence in
learning on histopathological images for colorectal colorectal cancer screening, diagnosis and treatment.
cancer diagnosis: A systematic review. Diagnostics, A new era. Curr. Oncol., 28(3), 1581–1607.
12(4), 837. Ozawa, T., Soichiro, I., Mitsuhiro, F., Youichi, K., Satoki,
Fearon, E. R. (2011). Molecular genetics of colorectal can- S., and Tomohiro, T. (2020). Automated endoscopic
cer. Ann. Rev. Pathol. Mec. Dis., 6, 479–507. detection and classification of colorectal polyps using
Figueiredo, P. N., Isabel, N. F., Luís, P., Sunil, K., Yen-his, convolutional neural networks. Ther. Adv. Gastroen-
R. T., and Alexander, V. M. (2019). Polyp detection terol., 13, 1756284820910659.
with computer-aided diagnosis in white light colonos- Sisodia, P. S., Gaurav, K. A., Yogesh, K., and Neelam, C.
copy: comparison of three different methods. Endo. (2023). A review of deep transfer learning approaches
Int. Open, 7(02), E209–E215. for class-wise prediction of Alzheimer’s disease using
Geetha, K. and Rajan, C. (2016). Automatic colorectal pol- MRI images. Arch. Computat. Methods Engg., 30(4),
yp detection in colonoscopy video frames. Asian Pac. 2409–2429.
J. Can. Preven. APJCP, 17(11), 4869. Takemura, Y., Shigeto, Y., Shinji, T., Rie, K., Keiichi, O.,
Hamabe, A., Masayuki, I., Rena, K., Saeko, S., Koichi, O., Shiro, O., Toru, T. et al. (2012). Computer-aided sys-
Kenji, O., Emi, A. et al. (2022). Artificial intelligence– tem for predicting the histology of colorectal tumors
based technology for semi-automated segmentation of by using narrow-band imaging magnifying colonos-
 Applied Data Science and Smart Systems 119
copy (with video). Gastrointes. Endoscop., 75(1), Colorectal cancer screening for average-risk adults:
179–185. 2018 guideline update from the American Cancer So-
Tsai, H.-L., Koung-Shing, C., Yu-Ho, H., Yu-Chung, S., ciety. CA Can. J. Clin., 68(4), 250–281.
Jeng-Yih, W., Chao-Hung, K., Chao-Wen, C., and Jaw- Yamada, M., Yutaka, S., Hitoshi, I., Masahiro, S., Shigemi,
Yuan, W. (2009). Predictive factors of early relapse in Y., Hiroko, K., Hiroyuki, T. et al. (2019). Development
UICC stage I–III colorectal cancer patients after cura- of a real-time endoscopic image diagnosis support sys-
tive resection. J. Surg. Oncol., 100(8), 736–743. tem using deep learning technology in colonoscopy.
Urban, G., Priyam, T., Talal, A., Mohit, M., Farid, J., Wil- Scientif. Reports, 9(1), 14465.
liam, K., and Pierre, B. (2018). Deep learning local- Yu, L., Hao, C., Qi, D., Jing, Q., and Pheng, A. H. (2016).
izes and identifies polyps in real time with 96% accu- Integrating online and offline three-dimensional deep
racy in screening colonoscopy. Gastroenterol. 155(4), learning for automated polyp detection in colonosco-
1069–1078. py videos. IEEE J. Biomed. Health Informat., 21(1),
Wang, Y., Xiaoyun, H., Hui, N., Jianhua, Z., Pengfei, C., 65–75.
and Chunlin, O. (2020). Application of artificial in- Zhang, X., Yang, Y., Yalan, W., and Qi, F. (2019). Detection
telligence to the diagnosis and therapy of colorectal of the BRAF V600E mutation in colorectal cancer by
cancer. Am. J. Can. Res., 10(11), 3575. NIR spectroscopy in conjunction with counter propa-
Wolf, A., Elizabeth, T. H. F., Timothy, R. C., Christopher, R. gation artificial neural network. Molecules, 24(12),
F., Carmen, E. G., Samuel, J. L., Ruth, E. et al. (2018). 2238.
17 SLODS: Real-time smart lane detection and object
detection system
Tanuja Satish Dhope1,a, Pranav Chippalkatti2, Sulakshana Patil3,
Vijaya Gopalrao Rajeshwarkar3 and Jyoti Ramesh Gangane4
1
Department of Electronics and Communication, Bharati Vidyapeeth (Deemed to be University) College of Engineer-
ing, Pune, Maharashtra, India
2
Department of Computer Science and Engineering, School of Computing, MIT Art, Design and Technology Univer-
sity, Pune, Maharashtra, India
3
Department of Electronics and Telecommunication, Sinhgad Institute of Technology, Lonavala, Pune, Maharashtra,
India
4
Department of Electronics and Telecommunication, Vishwaniketan’s Institute of Management Entrepreneurship and
Engineering Technology, India

Abstract
With the advances in technologies, autonomous cars/self-driving cars are now-a-days gaining more demand due to the in-
crement in mortality rate by road accidents caused due to human errors. Detecting obstacles on a road is one of the biggest
challenges in autonomous vehicle/self-driving navigation system. In this paper, we have proposed the real-time smart lane
detection and object detection system (SLODS) which captures the real-time road traffic using two cameras, one in the front
and the other one at the back of the car. The front one detects the lane while the other one detects if any other vehicle is ap-
proaching while changing the lanes, ensuring safe lane change. Region of interest (ROI) determines object and lane detection.
The performance of the edge detection algorithms like Roberts, Sobel, Prewitt’s, and Canny edge detectors, are evaluated
based on precision, recall, F1 score, and peak signal to noise ratio (PSNR) values. For PSNR, Canny is outperforming oth-
ers by the difference of -39dB with Sobel, -14 dB with Prewitt, and -48 dB. Further the proposed system also calculates the
speed of the approaching vehicle.

Keywords: Lane detection, edge detection, object detection, machine learning, Hough transform

I. Introduction when required. The organization of this paper is as

follows: Section II – Related work is discussed. Section
Road accidents are responsible for several deaths, hos-
III – Deals with our proposed systems. Section IV –
pitalization and disability amongst individuals world-
Methodology. Section V elaborates with results and
wide. One out of 10 people killed on roads across the
analysis. Section VI deals with conclusion followed by
globe is from India (Annual report, 2020). As shown
future scope in Section VII.
in Figure 17.1, during the year 2020, road accidents
decreased due to the imposition of lockdown world-
wide due to covid. Unfortunately, the people within II. Related work
the age bracket most affected in road traffic accidents The authors focused on different kernels of support
are 18–45-years-old, accounting for about 70% of all vector machine (SVM) to analyze performance of
fatalities. Some causes that result in accidents due to object classification for traffic objects. The experimen-
human errors are – Over speeding, drunken driving, tation results have helped to calculate recall, precision,
distractions to drivers, red light jumping, and avoid- F1 score and accuracy during classification (Madhura
ing safety gear like seat belts and helmets (Annual Bhosale et al., 2022). Various issues related to lane
report, 2020). To minimize accidents, the idea of detection and departure warning has been discussed
autonomous vehicles comes forward, which uses arti- by Sandipann Narote et al. (2018). Anuj Mohan et
ficial intelligence (AI) and machine learning (ML) for al. (2001) in his paper, the objects detected in the
traffic observation and analysis. Figure 1 shows the images are localized and then classifiers are used.
statistics related to road accidents that occurred in This method mainly focuses on localizing objects in
India. a nexus of other objects. It assisted in detecting many
A ML algorithm can collect data from its surround- items in a single frame. For efficient lane detection, the
ing using cameras and sensors and then starts inter- frame must be converted into a plot bird’s eye view of
preting it so that it is able to judge and take actions the highway, giving a good vision as the lines of the

[email protected]
a
 Applied Data Science and Smart Systems 121

Figure 17.2 SLODS

localizing curve and straight pathways. This approach

has proven to be highly resilient and stable for the
Figure 17.1 Statistics for road accidents (Annual vast majority of roads and walkways. The focus of
­report, 2020) VanQuang Nguyena et al. (2018) work is on a strat-
egy based on real-time data that allows the driver to
change lanes as efficiently as feasible. For accurate
detected lane appear vertical. This has aided authors results, the information about the vehicles and lanes
in reviewing the positioning of the cameras on the to be identified is considered, and a combination of
vehicle (Bertozzi et al., 1998). Mukesh Tiwari et al. a driver aid system and a lane change system is used.
(2017) proposed that for object tracking, it is neces- To detect numerous lanes, it must focus on the lanes
sary to first select an item and then track it using its in front of the vehicle and detect vehicles behind the
features. Some prior knowledge related to the shape primary vehicle using many cameras.
of the object being detected, size, color, etc. has been
used. William Ng et al. (2005) deals with tracking
III. Proposed system: smart lane detection and
objects by first localizing the objects and then perform-
object detection system (SLODS)
ing classification techniques on the localized objects.
It focuses on detecting multiple objects present inside SLODS overview
a frame by using SMC (Monte Carlo). This enabled The SLODS system uses two cameras to capture
to localize the region of interest (ROI) for the pro- lanes and oncoming traffic. When changing lanes, a
posed system. Sunil Kumar Vishwakarma et al. 2015 camera in the back detects any oncoming vehicles,
has offered a lane detecting approach using OpenCV and another in the front identifies the lanes. In vehi-
for roads and highways with prominent lanes. But at cle tracking, the speed of the vehicle is calculated
the same time, it faces problems with the structure of by the distance it travels per unit time. Using this
the roads, texture, hindrance, road type and visibility technique and machine learning (ML) enabled dash-
issues. Least median square (LMed Square) approach board, the speed of the impending car is displayed.
is utilized for obtaining an optimal subset by combin- The camera installed on the vehicle continuously
ing it with the least squares method for automatically records photos in order to detect and track vehicles
finding lanes (Xu Yang et al., 2015). It is applicable (Figure 17.2).
to both curved and straight pathways. The resulting
product is based on real-time presentation and has an Methodology
exact value as well as resilience. Zhong-Xun Wang
et al. (2018) explained lane detection by localizing A. Lane detection
them. It entails obtaining the photos and applying Lane detection is the technique of determining the
them to pre-processing. This is followed by frame seg- lanes on highways and expressways, thereby assisting
mentation algorithm and edge detection. The image the driver in safely maneuvering his car. We apply the
features must be extracted before segmentation can following equation for a line depiction to identify the
be performed. Finally, feature point identification is lines in an input image.
completed, followed by lane-line recognition. Singh et
al. (2019), Xining Yang et al. (2011) includes a road (1)
model as well as a lane recognition tool. The image
goes through a categorization process based on the where,
radiance of the frames. The image is then subjected to ρ = distance between center and line along the vec-
the Hough transform which aids in recognizing and tor perpendicular to the line
122 SLODS: Real-time smart lane detection and object detection system
Table 17.1 Area of interest vertices

S. No. Vertex X Y

1 Lower left 0 539

2 Lower right 959 539
3 Upper left 450 330
4 Upper right 490 330

Table 17.2 Hough transform parameters

S. No. Parameters Value

1 Rho 1
2 Beta π / 180
3 Minimum votes 15
threshold
4 Minimum line 7
Figure 17.3 Lane detection flowchart length
5 Maximum line 3
gap
β = angle between of x-axis along with vector. 6 Line thickness 1
A flowchart showing the sequence of events that
take place during lane detection is shown in Figure
17.3.
Steps for lane detection according to the Figure 17.1: 6. Applying Hough transform: Hough transform is
a technique to extract features from the image
1. Reading video and dividing into frames: The in- to analyze it. It is extensively used for image an-
put footage is recorded by a camera mounted alyzation based on shapes like rectangles, circles,
on the vehicle. The video input is then divided etc. It assumes all the white pixels of the image to
into frames (images) which are used to determine be the points and converts them into ρ-β plane.
lanes and boundaries on a laned road or high- ρ line connects polar coordinates to the origin
way. where the x-axis intersects the y-axis (Peerawat
2. Converting image into gray scale form: This is Mongkonyong et al., 2018) (Table 17. 2).
done to avoid recording unnecessary pixels. As a
result, far less information is collected and evalu- (2)
ated as compared to a colorful image.
3. Noise reduction using filter: The image produced Figures 17.4–17.8 gives a detailed idea of the step
after gray scale conversion is of poor quality. To 1–5 performed by the algorithms.
improve the accuracy of the recognized items,
noise filtration from the gray scaled image is B. Edge detection techniques
done before applying ML algorithm for object An edge is defined as an area of significant change
detection. in image intensity/contrast. Locating the areas with
4. Detecting edges: One of the topic’s cornerstones great intensity contrasts is called edge detection. Now
is edge detection. To detect a picture, the input it’s also possible that a certain pixel can accommo-
gray scaled image is exposed to the various edge date any variation and we can mistake it for an edge.
detectors (discussed in next section) after being Different situations can lead to this, for example, in
filtered (Zakir Hussain et al., 2015; Zhi Zhang et low light conditions or there can be noise that can
al., 2016; VanQuang Nguyena et al., 2018). This show all the characteristics of edge color segmentation.
gives us the power to modify the intensity of the
frames. B.1 Robert’s operator
5. Choosing region of interest (ROI): The area Robert’s operator (Zakir Hussain et al., 2015; Zhi
of the image in which the lane is detected and Zhang et al., 2016; VanQuang Nguyena et al., 2018)
placed in an area referred to as ROI. In our sys- is a type of an operator that works using cross prod-
tem ROI is taken as follows (Table 17.1). ucts to determine the grade of the detected image
 Applied Data Science and Smart Systems 123

Figure 17.4 Original image

Figure 17.8 Hough transform

using differential operations. The equation for the

gradient is given by Equation 3.

(3)

Using convolution masks, this becomes as given in

Equation 4.

(4)
Figure 17.5 Frame gray scaling

Where A = source image

B.2 Prewitt’s operator

Prewitt’s operator convolutes the frames, although we
use two masks, in case of common mask, it is repre-
sented by Hx and Hy (see Equation 5.)

(5)

Figure 17.6 Denoising Now we can also calculate the gradient direction
given by Equation 6

(6)

B.3 Sobel’s operator

It is used for processing of blurred images. Here
frames to be processed are divided into two distinct
Figure 17.7 Edge detection directions – x and y. In order to get the elements of the
124 SLODS: Real-time smart lane detection and object detection system

gradient along the directions, we apply a mask over

the frames (Zakir Hussain et al., 2015; Zhi Zhang
et al., 2016; VanQuang Nguyena et al., 2018). Below
is the mask for the Sobel’s operator i.e., Hx and Hy
(VanQuang Nguyena et al., 2018).

(7)

Where the partial derivatives are computed by,

(8)

(9)

B.4 Canny operator

It takes a grayscale image as input and then pro-
cesses it as an output using algorithms scattered
over numerous tiers (John Canny , 1986; Assidiq et
al., 2008; Ziqiang Sun , 2020). This method for edge
detection entails eliminating noise from frames and
then extracting data from frames while ensuring that
the functionality of the frame stays unchanged. The
gradient for a subtle edge is given by (Figures 17.9,
17.10, and Table 17.3):

(10)

C. Object detection and object tracking

Object detection is the phenomena of detecting seman-
tic objects of a specific kind in the form of images and
videos. It is a vision-based technique that can even
detect faces of pedestrians via face detection (Bertozzi
et al., 1998). The object detected by the algorithm
needs to be tracked down. Object tracking stores the Figure 17.9 Edge detection for each edge detector
initial set of coordinates of the detected object and
assigns a unique identity to each set of detections.
get are converted into grey format from RGB or
C.1 Object detection and object tracking steps multicolored images. The two colors present in
the gray scaled images are black and white.
1. Collecting video input: The video recorded by 3. Selecting the ROI: Region of interest is the area
the camera consists of thousands of frames that which is defined to detect object in the partic-
are in repetition. The recorded video is firstly ular area. Every other object is ignored in this
converted into frames. The images we get are area and rest all the operations that is, masking,
worked upon by applying gray scaling and filtra- thresholding, contouring is done within this re-
tion for noise reduction. gion only.
2. Gray scaling of frame: The next step involves 4. Masking the ROI: Masking of frames is done to
gray scaling of the images. The image that we differentiate the moving objects wrt the back-
 Applied Data Science and Smart Systems 125

Figure 17.11 Frame capture

Figure 17.10 Flowchart for object detection

Table 17.3 Canny algorithm parameters

S. No. Parameters Value

1 Low threshold 50
2 High threshold 150
Figure 17.12 Frame gray scaling

ground. It highlights the objects to be moving in

white, and rest of the background is put in white.
5. Bounding the detected object within the box:
Bounding the object as soon as the masked ob-
ject enters the ROI.

C.2 Speed of oncoming vehicle

To detect the speed of oncoming vehicle, ROI is
used. The ROI can be modified as per requirements.
Moreover, the speed of the oncoming vehicle is esti-
mated by the concept of distance and time, that is, the
distance the vehicle travels in a second.
Figure 17.13 Extracting ROI

(11)

Results and analysis

The real time traffic video of DMART road, Katraj
(18.4518331°, 73.8439111°) in Pune, Maharashtra,
India has been taken into considerations for lane
detection, object detection and its tracking. In this
section we are going to discuss the real time results of
lane and object detection using the above said algo-
rithms and proposed model that we have developed
using machine learning. Figures 17.11–17.15 indicate
the output of steps which is discussed in Section III.C.1
under object detection and tracking (Table 17.4). Figure 17.14 ROI masking
126 SLODS: Real-time smart lane detection and object detection system
Table 17.6 Parameters for Sobel operator

PSNR
Image Pixel Precision Recall F1 score (dB)

Img 1 Image Pixel Precision Recall F1

score
Img 2 300*168 0.96 0.81 0.44 45
Img 3 303*166 0.43 0.39 0.38 44
Img 4 300*168 0.76 0.79 0.40 51
Img 5 259*194 0.92 0.86 0.37 59

Figure 17.15 Final output

Table 17.7 Parameters for Canny operator

F1 PSNR
Table 17.4 Parameters for Robert operator Image Pixel Precision Recall score (dB)

F1 PSNR Img 1 300*168 0.95 0.97 0.96 82

Images Pixel Precision Recall score (dB)
Img 2 300*168 0.88 0.89 0.91 79
Img 1 300*168 0.53 0.49 0.72 66 Img 3 303*166 0.89 0.90 0.93 80
Img 2 300*168 0.51 0.45 0.77 70 Img 4 300*168 0.74 0.77 0.90 83
Img 3 303*166 0.59 0.51 0.64 68 Img 5 259*194 0.82 0.86 0.90 85
Img 4 300*168 0.57 0.50 0.71 77
Img 5 259*194 0.61 0.53 0.57 79

Table 17.5 Parameters for Prewitt operator

Image Pixel Precision Recall F1 PSNR
score (dB)

Img 1 300*168 0.83 0.77 0.90 70

Img 2 300*168 0.79 0.68 0.84 56
Img 3 303*166 0.39 0.42 0.88 59
Img 4 300*168 0.47 0.38 0.81 68
Img 5 259*194 0.93 0.86 0.80 62
Figure 17.16 Real time lane detection

Below is the comparison table for all detector’s that Canny outperforms for the various images in
algorithms for five images extracted from real time terms of other parameters also.
video. For lane detection, on an empty road, with a
The different edge detecting operators were tested straight lane, we apply Hough transform on the
based on various parameters like Precision, Recall, F1 detected edges using Canny edge detector. The lanes
score and PSNR values. Both Sobel and Prewitt edge on the road are detected and then highlighted using
detector were able to detect edges successfully, but the orange color markings (see Figure 17.16). Thus, even-
number of edges detected were far lower than Canny tually making it easier for the driver to navigate on
edge detection method (see Tables 17.5–17.7). For the road.
example, PSNR provided by Robert, Sobel, Prewitt, As soon as the incoming vehicle enters the specified
Canny is 66 dB, 40 dB, 70 dB and 82 dB for img ROI, the object detection algorithm starts detecting
1, respectively. Apart from low processing time, the the vehicle and it is finally bounded in a bounding
canny operator also displays a higher precision rate box (see Figure 17.17). Thus, the driver is alerted for
and better PSNR as compared to other operators. safe lane change.
Also, F1 score provided by Canny is 0.91 compared To detect the speed of oncoming vehicle ROI is
to others for img 1. Further Tables 17.5–17.7 indicate used. ROI has been set up as a distance up to 80 m
 Applied Data Science and Smart Systems 127

VII. Future scope

The same object detection algorithm can also be used
to recognize stop signs or pedestrians in a self-driving
vehicle. This helps the vehicle to stop or maneuver at
a safe distance from the pedestrian.

References
Annual report on Road Accidents in India. (2020). Re-
trieved from https:// morth.nic.in/sites /default/files/
RA_2020.pdf.
Madhura, B., Tanuja, D., Akshay, V., and Dina, S. (2022).
Figure 17.17 Real time object detection Performance analysis of object classification system
for traffic objects using various SVM Kkernels. Adv.
Data Comput. Comm. Sec., 423–432, doi:https://doi.
org/10.1007/978-981-16-8403-6_39.
Table 17.8 Speed analysis
Sandipann, N., Pradnya, N. B., and Dhiraj, M. D. (2018). A
Object Distance (m) Speed (m/s) review of recent advances in lane detection and depar-
ture warning system. Pattern Recogn., 73, 216–234.
Object 1 70 35 doi:https://doi.org/ 10.1016/ j.patcog.2017.08.014.
Anuj, M., Constantine, P., and Tomaso, P. (2001). Exam-
Object 2 75 27.5
ple-based object detection in images by components.
Object 3 60 30 IEEE Trans. Pattern Anal. Mac. Intel., 23(4), 349–
Object 4 65 37.5 361. doi:https: //doi.org/ 10.1109/ 34.917571.
Bertozzi , M. and Broggi, A. (1998). GOLD: A parallel real-
Object 5 72 36
time stereo vision system for generic obstacle and lane
detection. IEEE Trans. Image Proc., 7(1), 62–81. doi:
https://doi.org/10.1109/ 83.650851.
Mukesh, T. and Rakesh, S. (2017). A review of detec-
which can be varied according to user requirement. tion and tracking of object from image and video
The speed at which the test vehicle is moving is 25 sequences. Int. J. Computat. Intel. Res., 13(5),
m/s. A time interval of 2 s has been chosen for a vehi- 745–765. https://www.ripublication.com/ijcir17/
cle to cover its distance (see Table 17.8). ijcirv13n5_07 .pdf.
William, N., Jack, L., Simon, G., and Jaco, V. (2005). A re-
view of recent results in multiple target tracking. Proc.
VI. Conclusion 4th Int. Sym. Image Sig. Proc. Anal., 3807–3812.
Four distinct edge detection algorithms are men- doi:https://doi.org/ 10.1109/ ISPA.2005. 195381.
tioned in the study for our proposed SLODS. PSNR Sunil, K., Vishwakarma, A., and Divakar, S. Y. (2015).
Values, Precision, Recall, and F1 Score were the Analysis of lane detection techniques using OpenCV.
2015 Ann. IEEE India Conf., 1–4. doi:https://doi.
metrics utilized to assess these approaches. These
org// 10.1109/ INDICON. 2015.7443166.
parameters were utilized in this study to assess the
Xu, Y. and Zhang, L. (2015). Research on lane detec-
performance of the edge detection approaches devel- tion technology based on Opencv. 3rd Int. Conf.
oped by Canny, Sobel, Prewitt, and Robert. After Mech. Engg. Intel. Sys., 994–997. doi: https://doi.
careful examination, we determined that the edge org//10.2991/icmeis-15.2015.187.
detection approach successfully detected the great- Zhong-Xun, W. and Wenqi, W. (2018). The research on edge
est number of edges for both vertical and horizontal detection algorithm of lane. EURASIP J. Image Video
edges. If we visualize images in Figure 17.9, we can Proc., 98, 1–9. doi: https://doi.org/10.1186/s13640-
clearly see that Robert, Prewitt and Sobel give a low- 018-0326-2.
quality image as output when compared to Canny. Xining, Y., Dezhi, G., Jianmin, D., and Lei, Y. (2011). Re-
The Canny method on the other hand can detect search on lane detection based on machine vision. Proc.
2011 Int. Conf. Informat. Cybernet. Comp. Engg.,
both weak and strong edges. The paper also men-
110, 539–547. doi: https://doi.org/10.1007/978-3-
tions a method to detect and track objects in an effi-
642-25185-6_69.
cient manner. It suggests selecting and then masking VanQuang, N., Heungsuk, K., SeoChang, J., and Kwang-
the ROI after gray scale conversion of the image. The suck, B. (2018). A study on real-time detection method
algorithm used in this paper was able to detect 96% of lane and vehicle for lane change assistant system
of all the vehicles in the image. The SLODS provides using vision system on highway. Engg. Sci. Technol.
accurate speed estimation of the oncoming vehicle as Int. J., 21(5), 822–833. doi: https://doi.org/ 10.1016/
well. j.jestch.2018.06.006.
128 SLODS: Real-time smart lane detection and object detection system
Singh, J., Singh, S., Singh, S., and Singh, H. (2019). Evaluat- transform. IOP Conf. Ser. Mat. Sci. Engg., 297(1),
ing the performance of map matching algorithms for 1–11. doi: https:///iopscience.iop.org/ article/ 10.1088
navigation systems: an empirical study. Spat. Inform. /1757-899X/297/1/012050.
Res., 27, 63–74. Assidiq, A. A., Khalifa, O. O., Islam, M. R., and Khan, S.
Hussain, Zakir, and Diwakar Agarwal. (2015). A com- (2008). Real time lane detection for autonomous ve-
parative analysis of edge detection techniques used in hicles. Int. Conf. Comp. Comm. Engg., 82–88. doi:
flame image processing. International Journal of Ad- https://doi.org/10.1109/ICCCE.2008.4580573.
vance Research In Science And Engineering IJARSE, Ziqiang, S. (2020). Vision based lane detection for self-
4, 3703–3711. driving car. 2020 IEEE Int Conf Adv. Elec. Engg.
Zhi, Z., Zhihai, H., Guitao, C., and Wenming, C. (2016). Comp. Appl., 635–638. doi: https://doi.org/10.1109/
Animal detection from highly cluttered natural AEECA49918.2020.9213624.
scenes using spatiotemporal object region pro- John, C. (1986). A computational approach to edge de-
posal sand patch verification. IEEE Trans. Mul- tection. IEEE Trans. Pat. Anal. Mac. Intel., 8(6),
timed., 18(10), 1–14. doi: https://doi.org/10.1109/ 679–698. doi: https:// doi.org /10.1109/TPAMI.1986.
TMM.2016.2594138. 4767851.
Peerawat, M., Chaiwat, N., Supakorn, S., and Masaki, Y.
(2018). Lane detection using randomized Hough
18 Computational task off-loading using deep Q-learning in
mobile edge computing
Tanuja Satish Dhopea, Tanmay Dikshit, Unnati Gupta and Kumar Kartik
Department of Electronics and Communication, Bharati Vidyapeeth (Deemed to be University) College of Engineering,
Pune, India

Abstract
Because of the growing proliferation of networked Inter of Things (IoT) devices and the demanding requirements of IoT
applications, existing cloud computing (CC) architectures have encountered significant challenges. A novel mobile edge com-
puting (MEC) can bring cloud computing capabilities to the edge network and support computationally expensive applica-
tions. By shifting local workloads to edge servers, it enhances the functionality of mobile devices and the user experience.
Computation off-loading (CO) is a crucial mobile edge computing technology to enhance the performance and minimize
the delay. In this paper, the deep Q-learning method has been utilized to make off-loading decisions whenever numerous
workloads are running concurrently on one user equipment (UE) or on a cellular network, for better resource management in
MEC. The suggested technique determines which tasks should be assigned to the edge server by examining the CPU utiliza-
tion needs for each task. This reduces the amount of power and execution time needed.

Keywords: Computation off-loading, edge server, mobile edge server, deep Q-learning

I. Introduction Analyzing the load on mobile edge servers is

required prior to jobs being offloaded to edge serv-
Existing cloud computing (CC) architectures have
ers, including whether to do so and, if so, which edge
faced considerable hurdles as a result of the ongoing server. Analyzing the load on mobile edge servers is
proliferation of networked Internet of Things (IoT) necessary to respond to the question above. The task/
devices and the demanding needs of IoT applica- data off-loading decision is important because it is
tions, notably in terms of network congestion and predicted to have a straightforward impact on the QoS
data privacy. Relocating computing resources closer of the user application, including the resulting latency
to end users can help overcome these problems and caused by the off-loading mechanism (Yeongjin Kim
improve cloud efficiency by boosting its processing et al., 2018). When there is a lot of stress at the edge
power (Elhadj Benkhelifa et al., 2015). This strategy node as a result of a staggeringly high number of user
has developed with the introduction of many para- devices using the same edge network for every task
digms; fog computing, edge computing, all of which as it could result in considerable processing delays
have the same objective of increasing the deployment and the cessation of some processes (Fengxian Guo
of resources at the network edge. The significant et al., 2018). The reasoning architecture of the MEC
issues that traditional cloud computing (as central- notion is utilized to obtain cloud computing applica-
ized) is experiencing include increased latency in real- tions. By placing several information centers at the
time applications, low spectral efficiency. As a result network’s node, users of smartphones will be more
of new technologies, distributed computing capabili- accessible. The network terminal can refer to a multi-
ties are increasingly being used by organization’s or tude of places, including indoor areas like Wi-Fi and
network’s edge devices in an effort to explain all these 3G/4G. In today’s world, computing off-loading (CO)
challenges. Mobile edge computing (MEC) enables discusses both boosting smartphone performance as
certain apps to be off-loaded from resource-con- well as attempting to guarantee energy savings simul-
strained devices like smartphones, saving resources. taneously (Abbas Kiani et al., 2018). Although meet-
MEC’s characteristics set it apart from typical cloud ing the delay pre-requisites, MEC allows the edge to
computing because, unlike remote cloud servers, the perform computation-intensive applications rather
network can aggregate tasks in areas near the user than user equipment. Additionally, IoT users will
and device. By moving cloud processing to local serv- participate in later detecting and processing duties in
ers, MEC improves user quality of experience (QoE), user-centric 5G networks (Liang Huang et al., 2019;
in addition to reducing congestion in cellular infra- Thakur et al., 2021). In reality, using MEC to off-load
structure and cutting delay (Khadija Akherfi et al., computation processes results in wireless networks
2016). being used to transmit data. It is feasible for wireless

a
[email protected]
130 Computational task off-loading using deep Q-learning in mobile edge computing

connections to become severely congested if many best application approach while adhering to work-
application stations forcefully dump their processing flow applications deadline constraints. Numerous
resources to the edge node, which would dramatically experimental evaluations have been carried out to
slow down MEC (Gagandeep Kaur et al., 2021). A demonstrate the usefulness and efficiency of suggested
unified management system for CO and the accom- strategy.
panying wireless resource distribution in order to In MEC wireless networks, an Software Defined
benefit from compute off-loading is required (Khadija Networking (SDN) -based solution for off-loading
Akherfi et al., 2016). In section 2, this paper describes compute. Based on reinforcement learning, a solu-
the job off-loading research in MEC. The task off- tion to the energy conservation problem that consid-
loading system model is described in section 3 as local ers both incentives and penalties have been assessed
computing, edge computing, and the deep Q-learning (Nahida Kiran et al., 2020).
method. Section 4 elaborates on the results and charts Distributed off-loading method with deep rein-
for various task off-loading techniques. Finally con- forcement learning that allows mobile devices to
clusion is presented in section 5. make their off-loading decisions in a decentralized
way has been proposed. Simulation findings dem-
II. Related work onstrated that suggested technique may decrease the
ratio of dropped jobs and average latency when com-
In (Khadija Akherfi et al., 2016) many edge com- pared to numerous benchmark methods (Ming Tang
puting paradigms and their various applications, as et al., 2020). A multi-layer CO optimization frame-
well as the difficulties that academics and industry work appropriate for multi-user, multi-channel, and
professionals encounter in this fast-paced area has multi-server situations in MEC has been suggested.
been examined. Author suggested options, including Energy consumption and latency parameters are used
establishing a middleware-based design employing an for CO decision from the perspective of edge users.
optimizing off-loading mechanism, which might help Multi-objective decision-making technique has been
to improve the current frameworks and provide the proposed to decrease energy consumption and delay
mobile cloud computing (MCC) users more effective of the edge client (Nanliang Shan et al., 2020).
and adaptable solutions by conserving energy, speed-
ing up reaction times, and lowering execution costs. III. Methodology
Ke Zhang et al. (2016) has given an energy efficient
computation off-loading (EECO) method, which We took into account energy-sensitive UEs in this
combinedly optimizes the decisions of CO and alloca- paper, such as IoT devices and sensor nodes, which
tion of radio resources thereby minimizing the cost have low power requirements but are not delay-sen-
of system energy within the delay constraints in 5G sitive. We take into account N energy-sensitive UEs
heterogeneous networks. that are running concurrently on a server, and the
An energy-efficient caching (EEC) techniques for a server must choose which task from the task queue
backhaul capacity-limited cellular network to reduce needs to be done first in order to reduce the power
power consumption while meeting a cost limitation and execution time for each work. When a user device
for computation latency has been proposed (Zhaohui lacks the energy resources to complete the computa-
Luo et al., 2019; Gera et al., 2021). The numerical tion-intensive task locally, an edge server can step in.
findings demonstrate that 20% increase in delay effi- To make decisions on off-loading, we employ deep
ciency. The proposed method may be very close to the Q-learning algorithm. Based on the state and reward
ideal answer and far superior to the most likely out- of the Q function at state t, the Q-learning algorithm
come, i.e., the approximation bound. acts.
With the use of two time-optimized sequential
decision-making models and the optimal stopping A. Local computing
theory, author (Ibrahim Alghamdi et al., 2019) Let’s assume that E represents the energy needed for
address the issue of where to off-load from and when each User Equipment (UE) to operate locally n= num-
to do so. Real-world data sets are used to offer a ber of UE, pn =the power coefficient of energy used
performance evaluation, which is then contrasted for local computing per CPU cycle, cn= the CPU cycles
with baseline deterministic and stochastic models. desired for each bit in numbers, βn = the percentage of
The outcomes demonstrate that, in cases involving tasks computed locally, and Sn = the size of the com-
a single user and rival users, our technique optimizes putation task are all represented by the numbers n.
such decisions. Therefore, the amount of energy needed for UE to
Kai Peng et al. (2019) examine the multi-objective operate locally can be determined by discretion.
computation off-loading approach for workflow
applications (MCOWA) in MEC which discovers the (1)
 Applied Data Science and Smart Systems 131

B. Edge computing model (2)

Let’s assume that N numbers of UEs are anticipat-
ing tasks to be off-loaded and executed on edge where, “P(s, a)” is the Q-value for state “s” and action “a”
server since local server does not have enough power “α” is the learning rate.
resources. The task queue contains every single “R(s, a)” is the immediate reward for taking action
task. When the Q value function is modified based “a” in state “s”.
on reward and state, the task queue is supposed to “γ” is the discount factor.
update each time. The processes that are being used “max(Q(s’, a’))” represents the maximum Q-value
grow if the number of tasks (component list/UEs) in for the next state “s” and all possible actions “a”.
the task queue rises.
The system model consists of workload off-loading Algorithm
in MEC (see Figure 18.1). The task has been uploaded Input: Pt, Pt0, Pre_node, Comp_list, Trans_amount
to the any of three servers based on tasks that have Output: Trans_energy
come from UE. Selection of the any one of the server Initialization: Trans_energy ® 0;
is based on deep Q-learning algorithm. We have used If Pt≠ 0 and Pre_node (Pt(0)(0)) ≤ Comp_list then
TCP/IP, User Datagram Protocol (UDP) for transmit- Trans_energy+= ε ptr
ting task from UE to edge server depending on type of If Trans_amount≥ Pt(0)(2) then Pt0.append (Pt(0))
application viz., image processing, AR/VR, healthcare sort tasks on Pt0
applications, agriculture applications. The server will else Pt(0)(2) -= Trans_amount
track the Q value using a Q-learning algorithm and it
will update the entire task queue if one task consumes Results
less power than others.
Q-learning is a reinforcement learning algorithm We have considered total three edge servers and tasks
used in MEC that trains itself based on parameters sup- which are requesting for the edge server services. The
plied during environment building and server allocation following parameters have been taken into account
algorithms. Later, an optimum job distribution on the for deep Q-learning algorithm (see Table 18.1).
servers can be accomplished using the learned model. The other parameters like transmit power, band-
The tasks that off-load delay will be more difficult with width and noise PSD has been taken into consider-
local computing. In MEC, the state space could stand ation. We analyzed the number of UEs that are now in
in for the present network conditions, device status, the task queue. If there is only one UE, we can decide
and resource availability (such as CPU, memory, and whether to off-load the job and compute the trans-
bandwidth). Making judgments about how to allo- mission energy using the Epsilon-Greedy model. We
cate resources requires access to this state information. checked if there are multiple UEs in the task queue and
Q-learning assesses the effectiveness of activities con- the amount of transmission needed for the task before
ducted in a specific condition using a reward mecha- it in the queue is greater than the amount needed for
nism. Rewards in MEC can be determined based on the task after it. If so, we simply sort the task queue
a variety of performance indicators, including latency, based on the amount of transmission needed for pro-
energy use, throughput, or user happiness. Higher cessing, off-load the task in question, and execute it
rewards are associated with better decisions. on the edge server from the queue after sorting, using
The learning method entails exploring the state- less power in the process. According to the prior state
action space iteratively and updating the Q-table
based on the rewards gained. Q-learning uses the
Bellman equation to update Q-values iteratively: Table 18.1 Parameters for deep Q-learning algorithm

Parameters Value

Learning rate for the neural network’s optimizer 0.1

Reward decay factor in the Q-learning update 0.001
Initial Epsilon-Greedy exploration probability 0.99
Frequency of updating target network 200
parameters
Size of replay memory 10KB
Batch size for training 32
Exploration probability 0.9
Maximum number of episodes for training 3000
Figure 18.1 Block diagram of edge computing model
132 Computational task off-loading using deep Q-learning in mobile edge computing

Figure 18.2 Number of tasks in queue with respect to

time (s)

Figure 18.4 CPU utilization of different CPU’s and

number of tasks in queue during a single window ex-
ecution, time (s) vs. number of tasks

Figure 18.5 Episodes vs. number of tasks for local com-

putation and edge server off-loading using Q-learning
algorithm

and maximum, the Q value function produced this

transmission quantity.
Figure 18.2 shows the number of tasks in queue
with respect to time on the basis of provided num-
ber of nodes, environment variables, CPU requested
and processing time. The deep Q-learning algorithm
assigns the requested task to any three of the serv-
ers based on the reward. Figure 18.3 analyses the
three server utilization taken into consideration with
respect to time.
CPU utilization of different CPU’s and number of
tasks in queue during a single window execution has
been shown in Figure 18.4.
Figure 18.5 reflects the tasks in number which can
be off-loaded with respect to the edge server and
tasks that can be computed locally based on deep
Figure 18.3 Server utilization with respect to time (ms) Q-learning algorithm.
 Applied Data Science and Smart Systems 133
Guo, F., Zhang, H., Ji, H., Li, X., and Victor, C. M. L.
(2018). Energy efficient computation offloading for
multi-access MEC enabled small cell networks. IEEE
Int. Conf. Comm. Workshops, 1–6. doi: https://doi.
org 10.1109/ICCW.2018.8403701.
Gera, T., Singh, J., Mehbodniya, A., Webber, J. L., Shabaz,
M., and Thakur, D. (2021). Dominant feature selec-
tion and machine learning-based hybrid approach
to analyze android ransomware. Sec. Comm. Netw.,
1–22. https://doi.org/10.1155/2021/7035233.
Kiani, A. and Ansari, N. (2018). Edge computing aware
NOMA for 5G networks. IEEE IoT J., 5(2), 1299–
1306. doi: 10.1109/JIOT.2018.2796542.
Huang, L., Feng, X., Zhang, C., Qian, L., Wu, Y. (2019).
Figure 18.6 Number of tasks vs. average delay (s) for Deep reinforcement learning based joint task offload-
local and off-loading using Q-learning algorithm ing and bandwidth allocation for multi-user MEC.
Dig. Comm. Netw., 5, 10–17. doi: https://doi.
org/10.1016/j.dcan.2018.10.003.
As the number of tasks rises, the average time taken Kaur, G., Batth, R. S. (2021). Edge computing: classifica-
to complete each job also rises (see Figure 18.6). tion, applications, and challenges. 2nd Int. Conf. In-
When there are more tasks running simultaneously, tel. Engg. Manag., 1–6. doi: https://doi.org /10.1109/
the Q-learning method requires less time for task exe- iciem51511.2021.94453.
Zhang, K., Mao, Y., Leng, S., Zhao, Q., Li, L., Peng, X.,
cution than local computing.
Pan, L., Maharjan, S., and Zhang, Y. (2016). Energy-
efficient offloading for mobile edge computing in 5G
V. Conclusion heterogeneous networks. IEEE Acc., 4, 5896–5907.
doi: https://doi.org /10.1109/access.2016.259716.
In MEC, deep Q-learning algorithms play a cru- Thakur, D., Singh, J., Dhiman, G., Shabaz, M., and Gera,
cial role in the off-loading of computational tasks. T. (2021). Identifying major research areas and minor
We make the assumption in this study that numer- research themes of android malware analysis and de-
ous tasks are running concurrently on various user tection field using LSA. Complexity, 1–28. https://doi.
devices, and that the jobs are both delay and power org/10.1155/2021/4551067.
insensitive. We use the TCP/IP or UDP based on appli- Luo, Z., LiWang, M., Lin, Z., Huang, L., Du, X., and Gui-
cation to link user equipment to the server. The work zani, M. (2017). Energy-efficient caching for mobile
queue on the server adjusts based on the amount of edge computing in 5G networks. Appl. Sci., 7(6),
transmission needed to complete jobs, and the reward 1–13. doi: https://doi.org /10.3390/app7060557.
Alghamdi, I., Anagnostopoulos, C., and Pezaros, D. P.
value is updated in the Q-learning process. The
(2019). On the optimality of task offloading in mobile
Q-learning algorithm, which is based on reinforce- edge computing environments. IEEE Global Comm.
ment learning, considers both rewards and penalties Conf., 1–6. doi:https://doi.org/10.1109/GLOBE-
to minimize power usage. Off-loading workload to a COM38437.2019.9014081.
node server instead of remote server increases power Peng, K., Zhu, M., Zhang, Y., Liu, L., Zhang, J., Leung, V. C.
efficiency, lowers processing delay, and lowers total M., and Zheng, L. (2019). An energy- and cost-aware
infrastructure costs. computation offloading method for workflow appli-
cations in mobile edge computing. EURASIP J. Wire.
Comm. Netw., 1, 1–15. doi:https://doi.org/10.1109/
References globecom38437.2019.9014081.
Benkhelifa, E., Welsh, T., Tawalbeh, L., Jararweh, Y., and Ba- Kiran, N., Pan, C., and Changchuan, Y. (2020). Reinforce-
salamah, A. (2015). User profiling for energy optimisa- ment learning for task offloading in mobile edge com-
tion in mobile cloud computing. Proc. Comp. Sci., 52, puting for SDN based wireless networks. Seventh Int.
1275–1278, doi: https://doi.org/10.1016/j.procs.2015. Conf. Softw. Defined Sys. (SDS), 1–6. doi : https://doi.
05.151. org /10.1109/sds49854.2020.9143888.
Akherfi, K., Gerndt, M., and Harroud, H. (2016). Mobile Tang, M. and Wong, V. W. S. (2020). Deep reinforcement
cloud computing for computation offloading: Issues learning for task offloading in mobile edge comput-
and challenges. Appl. Comput. Informat., 14(1), 1–16. ing systems. IEEE Trans. Mob. Comput., 1, 1–12. doi:
doi: https://doi.org/10.1016/j.aci.2016.11.002. https://doi.org /10.1109/TMC.2020.3036871.
Kim, Y., Lee, H.-W., and Chong, S. (2018). Mobile compu- Shan, N., Li, Y., and Cu, X. (2020). A multilevel optimiza-
tation offloading for application throughput fairness tion framework for computation offloading in mobile
and energy efficiency. IEEE Trans Wire. Comm., 1–16. edge computing. Math. Prob. Engg., 1–17. doi: https://
doi:https://doi.org/10.1109/TWC.2018.2868679. doi.org 10.1155/2020/4124791.
19 A comprehensive analysis of driver drowsiness detection
techniques
Aaditya Chopra1, Naveen Kumar2,a and Rajesh Kumar Kaushal3
1
Thapar Institute of Engineering and Technology, Patiala, Punjab, India
2,3
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India

Abstract
Drowsiness or tiredness is a leading cause of accidents on the road, posing a serious threat to safety. Many accidents can
be prevented if drowsy drivers can be alerted within time. Several drowsiness detection methods are available to observe
drivers’ alertness during their journey and alert them if they are distracted. These methods gauges drowsiness by looking for
signs like yawning, closed eyes, or unusual head movements. They also consider the driver’s physical condition and vehicle
behavior. This paper offers a comprehensive analysis of the existing drowsiness detection methods and a detailed review of
the common classification techniques. It first categorizes the current methods into those based on subjective, behavioral, ve-
hicular, and physiological parameters based. Finally, it examines the strengths and weaknesses of these various methods and
compares them. In conclusion, the paper summarizes the research findings from this comprehensive survey to guide other
researchers toward potential future work in this field.

Keywords: Driver drowsiness detection technique, analysis, road safety

I. Introduction been studying how to detect and predict driver fatigue

for over ten years, it is still a problem that needs to
As we all know, automobiles have become integral
be solved. This paper aims to review the progress that
to our lives. In 2022, 81.6 million vehicles were
has been made so far and identify the main challenges
sold globally (Sales Statistics | Www.Oica.Net, n.d.).
that are preventing driver drowsiness prediction tech-
Despite the undeniable benefits of transportation
nologies from being used.
advancements, such as increased traveling speed,
comfort, and convenience, it has several adverse
effects. Injuries resulting from traffic accidents are II. Related work
the leading cause of death for children and young The existing review documents focused on the per-
people aged 5–29. Approximately 1.3 million people formance of four main aspects, such as signal acqui-
die in car crashes each year, and approximately 20–50 sition, extract features, and detect driver drowsiness
million suffer non-traumatic injuries that often lead itself. These technical details are typically categorized
to disability. Road traffic crashes have a significant in four primary categories: subjective, behavioral,
economic impact on most countries, accounting for vehicle-based, and physiological. Mckernon (2009)
approximately 3% of their gross domestic product emphasized ongoing initiatives to manage general
(World Health Organization: WHO, Road Traffic fatigue and its harmful effects on driving perfor-
Injuries, 2022). A report by the National Highway mance. Charles et al. (2009) and Brown et al. (2009)
Traffic Safety Administration (NHTSA) found that reviewed vehicle-based drowsiness detection research.
there were 7.3 million car accidents in the United Sigari et al. (2014) and Mittal et al. (2016) focused
States in 2016, which resulted in 37,461 deaths and their investigations on driver-behavioral measures.
3.1 million injuries. Fatigue driving was the cause of Sanjaya et al. (2016) presented research advancements
approximately 20−30% of traffic accidents. in physiological signal measurement. Sahayadhas et
Tiredness is considered as one of the top “fatal five” al. (2013) and Kang et al. (2013) discussed various
risks to driving safety, along with driver distraction, approaches and provided a comprehensive overview
alcohol or drug influence, speeding, and not wearing a of driver drowsiness detection solutions.
seat belt. While legal measures have been put in place
to address the other four risks, such as tracking and 2.1 Subjective measures
enforcing and speed limits, alcohol limits, mandatory Subjective indicators of drowsiness encompass signs
seat belt regulations, and restrictions on phone usage and sensations that individuals personally discern and
while driving. Driving fatigue remains a challenge that report when experiencing sleepiness or fatigue. We
needs to be tackled. Even though researchers have can collect subjective measures of driver drowsiness

[email protected]
a
 Applied Data Science and Smart Systems 135

in a variety of ways, such as by conducting surveys, 2.2 Behavioral measures

questionnaires, or interviews. These measures are rel- Unlike subjective assessments, behavioral measures
atively easy and inexpensive to collect, and they can are objective measures used to assess a person’s
provide valuable insights into how drivers experience behavior, such as personality traits, cognitive abilities,
fatigue and drowsiness. Nonetheless, they have limi- and emotional states.
tations such as drivers might be influenced by social These non-invasive measures monitor behavioral
expectations, resulting in responses that are perceived patterns to check the driver’s fatigue by focusing
as acceptable or desirable, rather than being fully on three main features: eye movement, head posi-
accurate. Additionally, subjective measures might tion, and facial expressions. Drowsy displays sev-
not always be dependable, as drivers may struggle to eral characteristic facial signs, including slow eye
accurately assess their own level of alertness. Notably, movements, eye closure, pupil dilation, head nod-
several established scales are employed to quantify ding, swinging or drooping, and frequent yawn-
sleepiness and drowsiness, aiming to capture the ing. This video-based approach extracts behavioral
subjective perceptions of affected individuals. These features from the camera and computer vision
indicators are valuable tools in various contexts, techniques.
including research investigations, clinical assess-
ments, and drowsiness detection systems. The most 2.2.1 Eye movement
commonly used scales are discussed in the following This measure focuses on eye monitoring through the
sections. slow eye movements (SEM), blinking rate, and eye
closure activities, including the PERCLOS metric and
2.1.1 Epworth sleepiness scale (ESS) the average eye closure speed (AECS) that character-
Dr. Murray Johns developed ESS. It is commonly izes eye movement. Unusual blinking and eye closure
used in clinical and research settings to assess daytime can be a sign of drowsiness.
sleepiness and other sleep-related issues, including Rahman et al. (2015) proposed a method for
narcolepsy or sleep apnea. It is a self-report question- detecting driver drowsiness based on eye blinking.
naire asking individuals to rate their likelihood of Firstly, video is captured from the camera and con-
dozing off in eight situations commonly encountered verted to frames. Viola-Jones algorithm is applied
in daily life. to detect the driver’s face, subsequently defining
region of interest (ROI) around the facial region.
2.1.2 Stanford sleepiness scale (SSS) The Viola-Jones cascade classifier technique is used
William C. Dement and Nathanial Kleitman devel- on the ROI to detect eyes using Haar-like features.
oped SSS at Stanford University and is widely used Both eyes are then extracted for further processing.
in sleep research, clinical sleep medicine, and various The colored image is converted to gray scale using
other fields to determine the alertness level of an indi- the Luminosity algorithm. Harris corner detection
vidual at a certain time. This scale asks individuals to method detects two upper eye corners and one lower
rate their current level of sleepiness on a scale from 1 eyelid. The midpoint value between the upper two
(feeling active) to 7 (feeling sleepy, almost in a trance). corner points is calculated (d1). Then, the midpoint
from the lower eyelid is calculated using Pythagoras
2.1.3 Karolinska sleepiness scale (KSS) theorem (d2). Finally, the d2 value is used to make the
Sleep medicine center at the Karolinska Institute, decision. If d2 is zero or d2 approaches zero, the eye-
Sweden developed KSS. It assesses a person’s sleepi- lid is considered closed; otherwise, it deems it open.
ness level on a 9-point scale, ranging from 1 (feeling The duration of a standard blink typically ranges
awake and alert) to 7 (feeling sleepy and barely able between 0.1 and 0.4 s. An increase in blink rate is
to stay awake). indicative of driver drowsiness. To detect drowsi-
Other scales include the Pittsburgh Sleep Quality ness, a threshold of 2 s is established. The alarm
Index (PSQI), The Groningen Sleep Quality Scale is triggered to alert the driver if this threshold is
(GSQS), the Visual Analogue Scale (VAS), and the breached. The proposed algorithm was tested under
Daytime Sleepiness Scale (DSS). ESS and SSS are the various lighting conditions and performed poorly in
most commonly used subjective scales to measure poor lighting conditions. The proposed method has
sleepiness. ESS is quick, easy to administer, reliable, been compared with other methods, including face
and valid. SSS is not as sensitive to changes in sleepi- and eye tracking using neural networks and visual
ness as the ESS. KSS is a newer scale still being studied, data, computer vision and machine learning (ML)
but it has proved to be a reliable and valid measure algorithms, and tools that measure EOG. The solu-
of sleepiness. The PSQI is a more comprehensive mea- tion demonstrated a 94% accuracy rate while main-
sure of sleep quality than the other scales but is more taining a relatively simplified structure compared to
time-consuming. alternative methods.
136 A comprehensive analysis of driver drowsiness detection techniques

2.2.2 Mouth and Yawning analysis compared with those obtained from testing images to
Yawning, often a result of tiredness or boredom, determine the appropriate classification. The system
can signal a potential risk for drivers, suggesting computes the duration of closed eyes and identifies
they might doze off while driving. Techniques exist drowsiness if this duration surpasses a predefined
to gauge the extent of mouth widening, serving as a time. Furthermore, it evaluates various head move-
means to detect signs of yawning in drivers. ments such as left, right, forward, backward, and
Yan et al. (2016) proposed an effective method for tilting motions in both directions. To conduct this
monitoring driver fatigue using Yawning extraction. analysis, the video footage is divided into frames,
To begin, the method uses the support vector machine with the system examining head images and com-
(SVM) technique to extract the face region from paring their positions to determine head postures.
images, reducing associated costs. The method pro- Subsequently, it merges the duration of closed eyes
ceeds to locate the mouth: facial edges are detected with the assessment of head positions to ascertain
through an edge detection technique, followed by drowsiness. The methodology was tested using six
a vertical projection to determine the right and left videos simulating genuine driving conditions and the
boundaries in the lower face area. Then, a horizon- findings are displayed through a confusion matrix. It
tal projection helps identify the upper and lower achieved a 98% accuracy rate, proving more effective
mouth limits, defining the mouth’s localized region. than alternative detection methods.
For yawning detection, the system employs circular The main problem with vision-based approach
Hough transform (CHT) on mouth region images is lighting. Regular cameras struggle at night.
to spot wide-open mouths. An alert is generated if Furthermore, many methods have been tested using
a notable number of consecutive frames capture a data from drivers imitating drowsiness instead of
widely open mouth. The method’s effectiveness is com- using real videos capturing a driver naturally becom-
pared with various edge detectors like Sobel, Roberts, ing sleepy.
Prewitt, and Canny. The experiment utilizes six videos
simulating real driving conditions, and the results are 2.3 Vehicle-based measures
depicted in a confusion matrix. The proposed method Vehicle-based measures detect driver fatigue using
attains a 98% accuracy rate, surpassing the perfor- vehicular features, including steering wheel angle,
mance of all other edge detection techniques. steering wheel grip force, lane changing patterns, and
vehicle speed variability. These measures necessitate
2.2.3 Head position the installation of sensors on various vehicle compo-
The head’s position is another sign of tiredness and nents, such as the steering wheel, accelerator, or brake
drowsiness in drivers. When feeling drowsy, drivers pedal, among others. The signals produced by these
often tend to tilt, lower, or nod their heads, particu- sensors serve as the basis for evaluating the drowsi-
larly in the later stages of sleepiness. Several factors, ness levels of drivers.
including decreased muscle tone, reduced vigilance,
and brief periods of sleep, can cause these changes in 2.3.1 Lane detection
the head position. Monitoring head position is a nota- This approach checks the vehicle’s position with
bly effective method for identifying driver drowsiness, respect to the middle of the lane. It is also known as
given that it is relatively easy and inexpensive to mon- the standard deviation of lane position (SDLP). Katyal
itor. Moreover, the head position remains unaffected et al. (2014) proposed a driver’s drowsiness detection
by environmental elements like lighting or noise. system using lane and driver’s fatigue level. Hough
Teyeb et al. (2014) proposed a method for drowsy transform is used to detect lanes and canny edge
driver detection using eye closure and head postures. detection is applied over viola-jones to detect eyes and
The system begins by capturing video through a web- driver’s fatigue level. This information is then used to
cam and conducting following operations on each detect improper driving. Ingre et al. (2006) conducted
frame of the video. It employs the Viola-Jones method multiple experiments and concluded that KSS ratings
to identify the ROI, encompassing the face and eyes. are directly proportional to SDLP metrics.
Subsequently, the facial area is sub-divided into sec-
tions, and the Haar classifier is utilized to focus on 2.3.2 Steering wheel analysis
the upper segment, specifically targeting the region Steering wheel analysis (SWA) is a widely used
corresponding to the eyes for analysis. Following this, vehicle-based measure to detect driver drowsiness
identifying the eye state entails the utilization of a (Fairclough et al., 1999; Thiffault, 2003). An angle
Wavelet network, a neural network-based approach, sensor is attached on the steering wheel axis to collect
which is trained using image data. The learning pro- the data. Abnormal steering wheel reversals, steering
cess involves ascertaining coefficients from training correction periods, and a vehicle’s jerky motion indi-
images. These learned coefficients are subsequently cate fatigue and a drowsy driver. Li et al. (2017) uses
 Applied Data Science and Smart Systems 137

SWA and proposed an online drowsiness detection apt for detecting drowsiness. Leveraging physiologi-
system to monitor the fatigue level of drivers under cal signals for drowsiness detection holds the poten-
natural conditions by extracting approximate entropy tial to mitigate the issue of false positives, which
features and using a decision classifier for detection. is a common challenge with existing approaches.
Zhenhai et al. (2017) proposed a solution by analyz- Furthermore, it enables timely alerts, thereby averting
ing the time series of the angular velocity of the steer- road accidents.
ing wheel. Fairclough and Graham (1999) proposed a
solution by checking the steering wheel’s reversals and 2.4.1 Electroencephalography (EEG)
small SWMs. They found that drowsy drivers make EEG measures the brain’s electrical activity by plac-
fewer steering wheel reversals than typical drivers. ing some electrodes on the head and forehead. The
Many studies have shown that vehicle-based mea- frequency of signals ranges from 1 to 50 Hz and
sures are not the best way to judge a driver’s drowsiness amplitude from 20 to 200 μV. Some frequency bands
and often lead to inaccurate results. Assessing driver are defined as alpha waves (8–12 Hz, 25–100 μV),
fatigue solely based on vehicle movement has limita- which measure relaxation; beta waves (faster than 13
tions, as the measurement metrics can be susceptible Hz and below 40 μV), which measure alertness; theta
to external influences like the road’s geometric attri- waves (4–7 Hz, 20–120 μV), which measure drowsi-
butes and prevailing weather conditions. Other factors ness; and delta waves (0.5–3.5 Hz, 75–200 μV) helps
can also affect these measures, such as road, traffic, to check if the subject is asleep.
lighting conditions, and driving under the influence of Several studies support the connection between
alcohol or other drugs. Steering wheel grip force on a EEG signals and driver behavior (Campagne et al.,
curvy mountain road differs significantly from that of 2004; Akin et al., 2008; Liu et al., 2010; Lin et al.,
a straight highway. Furthermore, the driver’s grip can 2012; Lin et al., 2013). Changes in the alpha frequency
vary with road conditions. Driver may not grip the band, where the power decreases, and an increase in
steering with that pressure on a busy road with which the theta frequency band are indicative of drowsiness.
he grips the steering on an empty expressway. Akin et al. (2008) observed that combining EEG and
EMG signals is more successful in detecting drowsi-
2.4 Physiological measures ness compared to using either signal alone.
As drivers experience fatigue, they may observe a
subtle swaying of their heads, and there is an elevated 2.4.2 Electrocardiography (ECG)
risk of the vehicle deviating from the center of the The ECG method records the heart’s electrical activity
lane. Previously discussed methods for detecting this by positioning electrodes on the chest, arms, and legs,
behavior, such as behavior-based and vehicle-based, capturing the small electrical signals generated with
possess limitations, primarily because they can only each heartbeat.
detect fatigue after the driver has already entered a Tsuchida et al. (2009) research claims that heart
drowsy state. rate variability (HRV) can be used to detect driver
However, it is worth noting that physiological sig- fatigue and drowsiness. As drivers get tired, their
nals undergo discernible changes early in the onset parasympathetic activity decreases, and their sympa-
of drowsiness. Hence, physiological signals are more thetic activity increases. This causes a notable shift in
138 A comprehensive analysis of driver drowsiness detection techniques

cardiac rhythm from a high-frequency range of 0.15– Table 19.1 List of various work done on driver drowsiness
0.4 Hz to a lower frequency range of 0.04–0.15 Hz. detection.
Several studies have explored driver fatigue and
S. Measure Method Algorithm Accuracy
drowsiness detection using photo plethysmogram No. used (%)
(PPG) and electrocardiogram (ECG) wavelet spec-
trum analysis. Tsuchida et al. (2009), Arun et al. 1 Behavioral Eye-blink Viola Jones 94
(2012), Lee et al. (2014), reporting an average predic- rate
tion accuracy of 96% in their experimental findings. 2 Behavioral Yawning SVM 98
analysis
2.4.3 Electromyography (EMG) 3 Behavioral Head Viola Jones 98
EMG measures the electrical activity of muscles and position with Haar
is commonly obtained from the chin (Hostens, 2005). classifier
When a muscle contracts, it sends electrical signals to 4 Physiological PPG and 96
the brain. ECG
Katsis et al. (2004) observed up to 20% frequency 5 Physiological, EOG Haar 80
decrease and up to 50% amplitude increase after behavioral with eye classifier
movement
monotonous driving tasks and used them to indicate
fatigue and drowsiness. Balasubramanian et al. (2007) 6 Physiological, Heart PERCLOS 96
behavioral rate with
also had similar observations in EMG from shoulder eyelid
and neck muscles during 15 min of simulated driving. closure
ratio
2.4.4 Electrooculography (EOG)
EOG measures the electrical potential difference
between the human eye’s front (cornea) and back (ret-
elucidated, and their advantages and drawbacks are
ina). It’s one of the primary functions is to gauge the
considered. Nonetheless, certain gaps have been pin-
amplitude and the direction of eye movements, which
pointed in the existing literature, such as the need to
is applicable in detecting driver drowsiness (Shuvan et
evaluate current techniques in real time. This is par-
al., 2009). The electric potential difference between the
ticularly crucial in dynamic driving conditions and
retina and cornea generates an electrical field which is
diverse environmental factors, presenting opportuni-
measured using EOG sensors and determines eyes ori-
ties for refinement and enhancement. A comparative
entation. By employing a disposable Ag–Cl electrode
analysis reveals that no single method achieves abso-
on each eye’s outer corner and a third electrode at
lute accuracy, although techniques relying on physi-
the forehead’s center, the system observes horizontal
ological parameters tend to yield more precise results
eye movements (Shuvan et al., 2009). These electrodes
than others. A combination of these methods, includ-
assist in identifying behavioral patterns such as rapid
ing physiological, vehicular, or behavioral measures,
eye movements (REM) and slow eye movements
can address the limitations present in each technique
(SEM), contributing to drowsiness detection in driv-
when used individually (Table 19.1).
ers (Lal et al., 2001; Sharma et al., 2020).
Khushaba et al. (2010) and Kukreja et al. (2022)
discovered that EOG alone could not produce accu- References
rate results compared to EEG alone for detecting Sales Statistics | Www.Oica.Net. (n.d.). Accessed September
drowsiness. Chieh et al. (2005) monitored eye move- 17, 2023. https://www.oica.net/category/sales-statistics/.
ment using EOG rather than a video-based eye moni- World Health Organization: WHO. (2022). Road Traffic
tor and achieved 80% accuracy. Injuries. 2022. https://www.who.int/news-room/fact-
sheets/detail/road%09traffic-injuries.
National Highway Traffic Safety Administration. (n.d.). Ac-
III. Result and discussion cessed September 17, 2023. https://crashstats.nhtsa.
The issue of driver drowsiness represents a significant dot.gov/#!/.
threat to road safety. Detecting driver drowsiness and McKernon, S. (2009). A literature review on driver fatigue
promptly issuing alerts is essential to avert a substan- among drivers in the general public. Land Transport,
New Zealand. 1–62.
tial number of road accidents. The primary objective
Liu, C. C., Simon, G. H., and Michael, G. L. (2009). Predict-
of this systematic review is to explore the most cur-
ing driver drowsiness using vehicle measures: Recent
rent advancements in drowsiness detection systems. insights and future challenges. J. Safety Res., 40(4),
This review examines drowsiness detection methods 239–245.
based on subjective, behavioral, vehicular, and physi- Brown, T., John, L., Chris, S., Dary, F., and Anthony, M.
ological parameters. These methods are thoroughly (2014). Assessing the feasibility of vehicle-based
 Applied Data Science and Smart Systems 139
sensors to detect drowsy driving. No. DOT HS 811 ness prediction system by using a self-organizing neu-
886. ral fuzzy system. IEEE Trans. Circuit. Sys. I Reg. Pa-
Sigari, M.-H., Muhammad-Reza, P., Mohsen, S., and Mah- pers, 59(9), 2044–2055.
mood, F. (2014). A review on driver face monitoring Liu, J., Chong, Z., and Chongxun, Z. (2010). EEG-based
systems for fatigue and distraction detection. Int. J. estimation of mental fatigue by using KPCA–HMM
Adv. Sci. Technol., 64, 73–100. and complexity parameters. Biomed. Sig. Proc. Con.,
Mittal, A., Kanika, K., Sarina, D., and Manvjeet, K. (2016). 5(2), 124–130.
Head movement-based driver drowsiness detection: Campagne, A., Thierry, P., and Alain, M. (2004). Correlation
A review of state-of-art techniques. 2016 IEEE Int. between driving errors and vigilance level: influence of
Conf. Engg. Technol. (ICETECH), 903–908. the driver’s age. Physiol. Behav., 80(4), 515–524.
Sanjaya, K. H., Soomin, L., and Tetsuo, K. (2016). Review Lin, C.-T., Kuan-Chih, H., Chun-Hsiang, C., Li-Wei, K., and
on the application of physiological and biomechanical Tzyy-Ping. J. (2013). Can arousing feedback rectify
measurement methods in driving fatigue detection. J. lapses in driving? Prediction from EEG power spectra.
Mechat. Elec. Power Vehicul. Technol., 7(1), 35–48. J. Neural Engg., 10(5), 056024.
Sahayadhas, A., Kenneth, S., and Murugappan, M. (2013). Tsuchida, A., Md Shoaib, B., and Koji, O. (2009). Estima-
Drowsiness detection during different times of day us- tion of drowsiness level based on eyelid closure and
ing multiple features. Aus. Phy. Engg Sci. Med., 36, heart rate variability. 2009 Ann. Int. Conf. IEEE
243–250. Engg. Med. Biol. Soc., 2543–2546.
Kang, H.-B. (2013). Various approaches for driver and driv- Arun, S., Kenneth, S., and Murugappan, M. (2012). Hypo-
ing behavior monitoring: A review. Proc. IEEE Int. vigilance detection using energy of electrocardiogram
Conf. Comp. Vis. Workshops, 616–623. signals. Journal of scientific & Industrial Research.
Rahman, A., Mehreen, S., and Aliya, K. (2015). Real time 71(12), 794–799.
drowsiness detection using eye blink monitoring. 2015 Lee, B.-G., Jae-Hee, P., Chuan-Chin, P., and Wan-Young, C.
Nat. Softw. Engg. Conf. (NSEC), 1–7. (2014). Mobile-based kernel-fuzzy-c-means-wavelet
Yan, C., Frans, C., Yong, Y., Xiaosong, Y., and Bailing, Z. for driver fatigue prediction with cloud computing.
(2016). Video-based classification of driving behavior Sensors 2014 IEEE, 1236–1239.
using a hierarchical classification system with mul- Hostens, I. and Herman, R. (2005). Assessment of muscle
tiple features. Int. J. Pat. Recogn. Artif. Intel., 30(05), fatigue in low level monotonous task performance
1650010. during car driving. J. Electromyograp. Kinesiol., 15(3),
Teyeb, I., Olfa, J., Mourad, Z., and Chokri, B. A. (2014). 266–274.
A novel approach for drowsy driver detection using Katsis, C. D., Ntouvas, N. E., Bafas, C. G., and Fotiadis, D.
head posture estimation and eyes recognition system I. (2004). Assessment of muscle fatigue during driving
based on wavelet network. IISA 2014 5th Int. Conf. using surface EMG. In Proceedings of the IASTED in-
Inform. Intel. Sys. Appl., 379–384. ternational conference on biomedical engineering, vol.
Katyal, Y., Suhas, A., and Shipra, D. (2014). Safe driving by 262. doi: 10.2316/Journal.216.2004.2.417-112
detecting lane discipline and driver drowsiness. 2014 Balasubramanian, V. and Adalarasu, K. (2007). EMG-based
IEEE Int. Conf. Adv. Comm. Con. Comput. Technol., analysis of change in muscle activity during simulated
1008–1012. driving. J. Bodywork Mov. Ther., 11(2), 151–158.
Ingre, M., Torbjörn, Å., Björn, P., Anna, A., and Göran, K. Hu, S. and Gangtie, Z. (2009). Driver drowsiness detection
(2006). Subjective sleepiness, simulated driving per- with eyelid related parameters by support vector ma-
formance and blink duration: examining individual chine. Exp. Sys. Appl., 36(4), 7651–7658.
differences. J. Sleep Res., 15(1), 47–53. Lal, S. K. L. and Ashley, C. (2001). A critical review of the
Fairclough, S. H. and Graham, R. (1999). Impairment of driv- psychophysiology of driver fatigue. Biol. Psychol.,
ing performance caused by sleep deprivation or alcohol: 55(3), 173–194.
a comparative study. Human Factors, 41(1), 118–128. Khushaba, R. N., Sarath, K., Sara, L., and Gamini, D.
Thiffault, P. and Jacques, B. (2003). Monotony of road en- (2010). Driver drowsiness classification using fuzzy
vironment and driver fatigue: a simulator study. Acc. wavelet-packet-based feature-extraction algorithm.
Anal. Preven., 35(3), 381–391. IEEE Trans. Biomed. Engg., 58(1), 121–131.
Li, Z., Shengbo, E. L., Renjie, L., Bo, C., and Jinliang, S. Kukreja, V. and Sakshi. (2022). Machine learning models
(2017). Online detection of driver fatigue using steer- for mathematical symbol recognition: A stem to stern
ing wheel angles for real driving conditions. Sensors, literature analysis. Multimedia Tools Appl., 81(20),
17(3), 495. 28651–28687.
Zhenhai, G., Le, D., Hu, H., Yu, Z., and Wu, X. (2017). Driv- Chieh, T. C., Mohd, M. M., Aini, H., Seyed, F. H., and
er drowsiness detection based on time series analysis Burhanuddin, Y. M. (2005). Development of vehicle
of steering wheel angular velocity. 2017 9th Int. Conf. driver drowsiness detection system using electroocu-
Meas. Technol. Mech. Autom. (ICMTMA), 99–101. logram (EOG). 2005 1st Int. Conf. Comp. Comm. Sig.
Akin, M., Muhammed, B. K., Necmettin, S., and Muhittin, Proc. Special Track Biomed. Engg., 165–168.
B. (2008). Estimating vigilance level by using EEG and Sharma, R. and Vinay, K. (2022). Segmentation and multi-
EMG signals. Neural Comput. Appl., 17, 227–236. layer perceptron: An intelligent multi-classification
​Lin, F.-C., Li-Wei, K., Chun-Hsiang, C., Tung-Ping, S., and model for sugarcane disease detection. 2022 Int. Conf.
Chin-Teng, L. (2012). Generalized EEG-based drowsi- Dec. Aid Sci. Appl. (DASA), 1265–1269.
20 Issues with existing solutions for grievance redressal
systems and mitigation approach using blockchain
network
Harish Kumar, Rajesh Kumar Kaushala and Naveen Kumar
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India

Abstract
Grievance redressal has always been vital for any organization to maintain a good work environment for its stakeholders.
Some organizations follow online portals, websites, or mobile applications to register grievances to provide more privacy to
the complainant’s identity. However, online platforms provide better solutions to the existing manual methods for grievance
redressal. Still, there are a lot of issues and challenges associated with them. This research has comprehensively analyzed the
existing grievance redressal systems to identify and discuss all the challenges. After comprehensive analysis, it is found that
presently there are several issues such as delayed response, opaque processes, biases, complexity and accessibility issues, lack
of personalization, and other privacy and security concerns associated with existing grievance redressal methods. To address
all these issues this study is proposing a blockchain-based solution for grievance redressal systems. The proposed solution
will be a blockchain-based web and mobile application that consists of multiple entities such as complainants, redressal
committee, and higher authorities. This system will provide the necessary privacy and confidentiality to the complainants
through the immutable distributed ledger technology and auditability of the entire process with complete transparency.

Keywords: Blockchain, grievance redressal system, immutability, privacy, smart contracts

I. Introduction affected by societal developments, advancements in

technology, and the desire for justice and fairness
In today’s interconnected and information-driven (Aggarwal, Dhaliwal, and Nobi, 2018). Grievance
culture, the importance of effective procedures for redressal procedures are widely observed in several
addressing grievances has become increasingly sig- areas of contemporary society, including governmen-
nificant. Instances of grievances, disagreements, and tal agencies, corporate organizations, online plat-
conflicts are commonly encountered by individuals in forms, and social communities. However, despite their
many areas such as the public sector, business orga- prevalence, certain persisting issues hinder the effec-
nizations, and online communities. These situations tiveness and dependability of these systems (Prajapat,
often require prompt and fair settlement. Throughout Sabharwal, and Wadhwani, 2018).
history, conventional methods employed to address
these issues have often been linked to inefficiency, a
dearth of openness, and a deficit of confidence (Denny II. Blockchain technology
et al., 2021). The advent of blockchain technology has Blockchain is a peer-to-peer network that lets peo-
presented novel prospects for augmenting grievance ple all over the world do business with each other.
redressal systems through the facilitation of transpar- The immutable ledger stores all events as a chain of
ency, resistance to tampering, and enhanced efficiency. blocks, and each node keeps an offline copy of the
This study addresses the issues with the existing whole ledger. There is no central or middle authority
procedures for resolving grievances and explores the that stores and verifies transactions; instead, all nodes
potential of blockchain technology as a viable solu- in the network are responsible for verifying new
tion for enhancement. The phenomenon of addressing transactions. Each blockchain framework uses a dif-
grievances is profoundly embedded throughout the ferent set of techniques called consensus algorithms
social framework of human civilization. Throughout to do this. Proof of work (PoW), proof of stake (PoS),
history, the necessity to confront and resolve conflicts and others are some of the most common consensus
and disagreements has consistently held considerable algorithms. When it comes to completing deals, each
significance, encompassing various regulatory frame- consensus algorithm takes a different approach. Some
works from ancient times to present-day institutions. blockchain frameworks use “smart contracts” to
The approaches employed to address grievances have complete the deal. Smart contracts are computer pro-
experienced significant changes throughout history, grams that are stored in the blockchain network and

[email protected]
a
 Applied Data Science and Smart Systems 141

run automatically when certain conditions are met

(Kumar et al., 2021). Blockchain stores a record of all
transactions and keeps it safe in an unalterable led-
ger which provides auditability and complete trans-
parency. Any authorized network node can check the
complete history of a transaction.

Research objectives
The primary objective of this research article is to
investigate the challenges faced by existing grievance
redressal systems and explore how blockchain tech-
nology can mitigate these issues. To achieve this, the
following research goals will be pursued:

• Identify the key limitations of traditional griev-

ance redressal systems across various sectors.
• Analyze the core features of blockchain technol-
ogy and how they can address the identified limi-
tations.
• Present case studies and examples of blockchain-
Figure 20.1 Prisma flow diagram for literature study
based grievance redressal systems (GRS) to show-
case their practical applications and benefits to
mitigate the issues with the existing solutions.

IV. Methodology
The research work is conducted using a prisma
approach in order to ensure its conclusion. A stan-
dardized methodology encompassing the stages of
planning, execution, and reporting was implemented.
The sections that follow outline the procedural phases
of the approach employed in this study. The initial
stage involves the formulation of search keywords.
The subsequent stage is conducting a search for sys-
tems designed to address grievances, (Figure 20.1)
online portals dedicated to grievance redressal, and
research papers pertaining to blockchain technology,
utilizing certain keywords. The purpose of utilizing
certain keywords is to effectively distinguish between Figure 20.2 Structure of grievances redressal system
research articles that are relevant and those that are
irrelevant. In the third phase, an analysis is conducted
on the operational characteristics of online portals. of a proficient and impactful method for addressing
This examination involves the identification of short- grievances is an essential requirement for any orga-
comings in current solutions and the subsequent nization or institution to demonstrate responsibility
proposal of blockchain technology as a fundamental and accountability (Tripathi, Srivastava, and Singh,
remedy for the aforementioned concerns within this 2021). The workflow of the traditional grievances
industry (Figure 20.2). redressal system is shown in Figure 20.2.
Grievances may emerge at several levels within an
organization, including educational institutions such
V. Related work
as universities and schools. When any individuals per-
A complaint is typically characterized as a form of ceive that their rights, needs, or expectations have not
communication, whether spoken or written that been adequately fulfilled. This issue becomes highly
articulates dissatisfaction with a particular course delicate when it pertains to the students of an aca-
of conduct or neglect, or with the quality of service demic institution, given that students are the most
provided by an organization. The implementation vulnerable individuals in this context. Frequently,
142 Issues with existing solutions for grievance redressal systems

individuals encounter difficulties in effectively com- have encompassed the crucial features which are
municating their concerns and encounter challenges required for an effective and efficient system, such as
in receiving adequate help from relevant authorities at immutability, transparency, auditability, and distrib-
different stages of their academic progression within uted storage to mitigate the risk of a single point of
the institution (Prajapat, Sabharwal, and Wadhwani, failure within the system.
2018). In a research article, the authors (Magner, Table 20.1 exhibits a comprehensive literature
1995) have examined a particular case wherein a sub- assessment of the existing solution for grievance
stantial group of students collectively endorsed and redressal on the basis of the key features of an effec-
submitted a petition alleging substandard teaching by tive and efficient system.
their teacher, citing an inability to effectively deliver The authors Prajapat, Sabharwal, and Wadhwani
the curriculum in accordance with the updated edu- (2018) in their study have proposed an automated
cational framework. The authors (Miklas and Kleiner, system for grievance registration and redressal, but it
2003) discussed the case of a foreign university where follows a very basic architecture and it is still human-
a group of female students registered a complaint dependent to forward the complaint at almost every
against their professor for harassment. stage, and due to that resolution to the student griev-
Many researchers proposed a variety of theoreti- ance may be delayed, the privacy of the user’s identity
cal frameworks, prototypes, online solutions, mobile is not preserved, security of sensitive data is not men-
applications, and web portals utilizing diverse tech- tioned and covered, due to centralized system archi-
nologies such as artificial intelligence and machine tecture, tampering with the data may be possible,
learning (ML) techniques to manage grievance and due to weak architecture it cannot be implemented
redressal processes. However, none of these proposals at larger scale. The authors Kandhari and Mohinani

Table 20.1 Literature review for grievances redressal system.

Year Technology used 1 2 3 4 5 6 7 8 9 Ref.

2014 PhoneGap, GPS, Google   x X  x x x x (Kandhari and

Maps & MySQL Mohinani, 2014)

2017 Not mentioned x  x X x x x x x (Prajapat, Sabharwal,

and Wadhwani, 2018)
2018 Android, AI, NLP, machine   x X  x x x x (Kormpho et al., 2018)
learning techniques
2019 Not mentioned   x X x x x x  (Palanissamy and
Kesavamoorthy, 2019)
2020 PHP & MySQL   x x x x x x  (Aravindhan et al.,
2020)
2020 Android, Google Maps,   x x  x x x  (Laxmaiah and
cloud vision, geo-coding and Mahesh, 2020)
firebase and machine learning
2020 Ethereum, Android,   x       (Hingorani et al.,
Encryption 2020)
2020 Hyperledger Fabric x   x     x (Shettigar et al., 2021)
2020 Ethereum, Node.JS, Web3.JS, x  x  x     (Jattan et al., 2020)
MetaMask
2020 Not mentioned   x x x x x x (Shahnawaz et al.,
2020)
2021 Node.JS, MongoDB   x x  x x x x (Oguntosin Victoria et
al., 2021)
2021 Not mentioned  x x  x x x x (Bhadouria and others,
2021)
2022 Django, HTML, CSS, SQL,   x x  x x x  (Jha, Sonawane, and
artificial intelligence and others, 2022)
machine learning
1. Implemented 2. Data security 3. Privacy 4. Decentralized 5. Automated 6. Immutable 7. Distributed storage
8. Auditable 9. Transparent
 Applied Data Science and Smart Systems 143

(2014) have designed a mobile application for the citi- In the event of prolonged inactivity, the system auto-
zens to register their municipal services-related griev- matically elevates the complaint’s status to the supe-
ances. The application enables the users to register rior officer and District Magistrate through an email
their complaints along with the image of problems notification, providing an update on the registered
and location coordinates. The author Kormpho et al. complaint. The authors Shettigar et al. (2021) in a
(2018) proposed a solution that involves the devel- separate study proposed a blockchain-based solution
opment of a mobile application and a chatbot that for a grievances management system for college stu-
allows end-users to effectively register their concerns. dents, being a blockchain-based solution, it provides
Additionally, an innovative web-based solution is all the inherent features of blockchain but the most
provided for the organization to address these issues, important phase, which is implementation, is miss-
with the added benefit of preventing duplicate com- ing. The proposed solution uses the permissioned
plaints. The authors Palanissamy and Kesavamoorthy blockchain hyperledger fabric framework. In another
(2019) in their proposed solution comprise a widely study, the researchers Jha, Sonawane, and others
used online application for addressing grievances. (2022) presented a proposal for the development
This approach employs a multi-step negotiation pro- of a web portal designed specifically for students to
cess to identify and resolve issues. Which depends on register their complaints across different categories.
human involvement at each level. However, the pro- This portal offers transparency and monitoring capa-
posed method provides an easy user interface but is bilities, allowing the tracking of complaint statuses
deficient in key attributes such as tamper-proofing, at any given stage. Additionally, the authors suggest
immutability, privacy, and transparency, all of which the incorporation of ML and artificial intelligence
are of utmost significance when dealing with sensitive (AI) techniques to identify and address instances of
information. offensive language and complaints that propagate
In another study, the authors Aravindhan et al. misinformation on sensitive subjects such as racism,
(2020) developed a web-based solution that relies gender, and religion.
on human intervention at various stages to address In another study, authors Musa et al. (2021) have
complaints. However, this dependence on human proposed a centralized web portal to handle the stu-
involvement can potentially lead to delays in resolv- dents’ grievances at the university level, where stu-
ing student grievances. Furthermore, the preservation dents can register their academic and non-academic
of user identification and the protection of sensitive related grievances. Another govt. of India, initiative
data are not well-addressed or discussed. The poten- Rana et al. (2016) has introduced an online portal
tial for data tampering exists due to the central- for Indian citizens, where they can register their com-
ized system architecture, and the proposed solution plaints against any central or state govt. departments,
is not feasible for larger-scale implementation due sub-departments, or any public service-providing
to its inherent weaknesses in architecture. The sole agencies.
advantageous aspect of the suggested solution is to The Director of Public Grievances, The Department
the incorporation of a web interface into the preex- of Administrative Reforms and Public Grievances has
isting manual system. The researchers Laxmaiah and implemented a web-based portal to address and mon-
Mahesh (2020) of the study developed an automated itor public grievances. This portal is interconnected
and intelligent mobile application for the citizens to with all ministries and departments of the Indian gov-
register their grievances, the mobile application uses ernment as well as state governments. It allows citi-
ML techniques to segregate the types of complaints zens to access the portal through a mobile application
and automatically forward them to the concerned and register their grievances about any service pro-
department of an official without any delay, most of vided by the Indian government or state government.
the phases of this system is totally automated without Additionally, the portal offers transparency to users
any human dependency or intervention, researchers by enabling them to track the progress of their griev-
use cloud vision server and geo-coding and reverse ances using a unique registration number (DARPG,
geo-coding to label and identify the problem location 2023). The success of online portals for grievance
without any human involvement. It also provides the redressal systems has been comprehensively assessed
user’s accounts and their registered grievances and and analyzed by the authors Rana et al. (2015) in a
also provides the tracking information about the reg- separate study. This evaluation was conducted using
istered complaints. an E-government-based IS success model, which
In another study, the authors Hingorani et al. was constructed utilizing existing IS success models.
(2020) proposed a police complaint system by utiliz- Multiple hypotheses were examined and supported
ing blockchain technology in the development of web by empirical evidence, indicating that the implemen-
and mobile applications. It enables complainants to tation of the online public grievances redressal system
conveniently register and monitor their complaints. is likely to be highly effective. However, it is important
144 Issues with existing solutions for grievance redressal systems

mistrust, and skepticism, ultimately undermining the

credibility of the system.

Delayed resolutions
The effectiveness of conventional methods for
addressing grievances is often hampered by lengthy
and extended procedures for resolving disputes. It can
result in extended suffering for the aggrieved parties,
especially in cases where time-sensitive issues are at
stake. Delays can also increase tensions, and conflicts
which aggravate disputes, hence emphasizing the sig-
nificance of quick resolution.
Figure 20.3 Publication trends in existing solutions on
grievances redressal system
Susceptibility to manipulation
Several grievance redressal systems exhibit vulnerabil-
to note that despite the convenience and accessibility ity to manipulation, stemming from either unethical
offered by the online portal for registering grievances, practices or organizational inefficiency. This suscep-
the privacy and security of user data remain signifi- tibility undermines the justice of the system and may
cant concerns. discourage individuals from seeking resolution for
The researchers Alawneh, Al-Refai, and Batiha their issues early.
(2013) in their study investigated the factors that
influence user satisfaction with Jordan’s e-govern- Lack of accountability
ment services portal. The research paper outlines five Accountability counted as a key component of an effi-
primary criteria that have the potential to influence cient grievance redressal system. However, in several
the level of satisfaction among Jordanian individu- cases, it proves to be quite a challenge to ensure that
als with the portal. These factors encompass security the individuals or entities involved are held liable for
and privacy, trust, accessibility, awareness of public their actions. The absence of accountability can give
services, and the quality of public services. The study rise to a culture of freedom, wherein instances of mis-
presents significant findings derived from the analysis conduct remain unaddressed.
of survey data, emphasizing the importance of com-
prehending these factors to enhance the design and Inadequate data security
functionality of e-government portals. It also provides The rising dependence on online platforms for the
recommendations for practitioners and policy-mak- resolution of grievances has led to increased attention
ers to effectively improve user experience and cater on data security. The occurrence of breaches and data
to the needs of citizens. The outcomes of this study leaks can result in significant implications, such as the
underscore the shortcomings of the current system for disclosure of confidential data and a decline of confi-
addressing issues. dence in the system.
The literature review indicates that most of the
studies have implemented a centralized solution, few VII. Proposed system
articles only discuss the theoretical models, and few
have done the analysis of the effectiveness of the exist- The main purpose of this study is to investigate the
ing centralized solution, and the blockchain-based existing literature and identify the issues with the
studies are merely proposing a theoretical model. As existing solutions and propose an effective and effi-
per the reviewed literature, Figure 20.3 illustrates the cient system for grievance redressal which will cover
publication trends observed in published articles in all the issues identified during the literature review
this domain. of the existing solutions. The proposed system will
be an online web/mobile application that will use a
blockchain framework to provide distributed stor-
VI. Issues identified with existing grs age and provide immutability and full auditability.
Lack of transparency The structure of the proposed system is shown in
One of the significant challenges in existing grievance Figure 20.4.
redressal procedures is the absence of openness. In
several instances, people are often uninformed of the VIII. Result and discussion
status of their grievances, the procedures involved in
decision-making, and the eventual outcome. The lack The conventional methods of addressing grievances
of transparency can result in feelings of dissatisfaction, are considered to be inadequate due to their limited
 Applied Data Science and Smart Systems 145

Figure 20.4 Structure of the blockchain-based proposed system

effectiveness in delivering crucial elements necessary IX. Conclusion and future work
for an efficient solution, such as transparency, immu-
Blockchain can provide transparency, immutability,
tability, privacy, quick redressal, security, and audit-
and a distributed storage facility and its integration
ability. The prevailing approach in online grievance
with various sectors can improve the existing services.
management solutions relies on a centrally managed
This article explores the limitations and issues of the
server system, rendering them more vulnerable to
present grievance redressal procedures. Although some
potential removal or tampering of data. Conversely,
of the online portals provide some sort of privacy and
the implementation of a decentralized grievance
transparency but they fail to provide immutability and
redressal system may hinder these efforts due to the
solution to a single point of failure. While identifying
widespread availability of all grievances across every
the limitations of the existing grievances redressal sys-
peer within the network.
tem this study proposes blockchain technology as a
During the literature review on the traditional and
mitigation approach to all the identified issues with
other online solutions for the grievances redressal sys-
the existing solutions. Blockchain-powered systems
tem, some important issues are identified. This exhib-
offer benefits like increased transparency, record pres-
its that there is a strong requirement for an efficient
ervation, and decentralized trust mechanisms.
grievance redressal system that should be both trans-
The findings of our investigation indicate that the
parent and tamper-proof and operate on a distributed
adoption of a blockchain-powered grievance redressal
peer-to-peer network which eliminates any potential
system presents numerous benefits, such as increased
instances of ignorance and abuse of power by higher-
transparency, the preservation of unalterable records,
level officials.
and the utilization of decentralized trust mechanisms.
Our proposed system extends the security, pri-
Through the utilization of smart contracts and cryp-
vacy, and other key features of the online portal by
tographic methodologies, blockchain technology has
adding blockchain technology, which provides all
the potential to enable a secure and effective mecha-
the inherent features of blockchain such as immu-
nism for addressing grievances. However, scalability,
tability, auditability, transparency, and distributed
privacy, and regulatory constraints are crucial and
ledger storage which mitigate the single point of
that can be managed with the selection of an appro-
failure of a centralized system and auditability fea-
priate blockchain framework.
ture enable the authorized user to check the com-
plete transaction history, and transparency feature
of blockchain allows the users to get updates about References
every change/transaction made to the registered Aggarwal, A., Ran Singh, D., and Kamrunnisha, N. (2018).
complaint. Impact of structural empowerment on organizational
146 Issues with existing solutions for grievance redressal systems
commitment: The mediating role of women’s psycho- Laxmaiah, M. and Mahesh, K. (2020). An intelligent public
logical empowerment. Vision, 22(3), 284–294. grievance reporting system-IReport. ICCCE 2020 Proc.
Alawneh, A., Al-Refai, H., and Batiha, K. (2013). Mea- 3rd Int. Conf. Comm. Cyber Phy. Engg., 197–207.
suring user satisfaction from e-government services: Magner, D. K. (1995). Mid-semester removal of Professor
Lessons from Jordan. Gov. Inform. Quart., 30(3), Roils University of Montana. Chron. Higher Edu., A25.
277–288. Miklas, E. J. and Brian, H. K. (2003). New developments
Aravindhan, K., Periyakaruppan, K., Aswini, K., Vaishnavi, concerning academic grievances. Manag. Res. News,
S., and Yamini, L. (2020). Web portal for effective stu- 26(2/3/4), 141–147.
dent grievance support system. 2020 6th Int. Conf. Musa, W. M. W., Azahari, A. A., Noorimah, M., and Suhaily,
Adv. Comput. Comm. Sys. (ICACCS), 1463–1465. M. A. M. (2021). E-justice: Students’ complaints made
Bhadouria, L. S. and others. (2021). Online complaint man- easy. SEARCH J. Media Comm. Res. (SEARCH), 1.
agement system. Turkish J. Comp. Math. Edu. (TUR- Oguntosin, V., Oluwadurotimi, M., Adoghe, A., Abdulka-
COMAT), 12(6), 5144–5150. reem, A., and Adeyemi, G. (2021). Development of a
DARPG. (2023). Centralized public grievance redress and web-based complaint management platform for a Uni-
monitoring system. https://pgportal.gov.in/Home/ versity community. J. Engg. Sci. Technol. Rev., 14(1),
AboutUs. 150–159.
Denny, J., Ramya, C., Sweta, R. L., Srija Reddy, A., and Sa- Palanissamy, A. and Kesavamoorthy, R. (2019). Automated
hithya, V. (2021). A Web Portal for Student Grievance dispute resolution system (ADRS) – A proposed ini-
Support System. International Research Journal of tial framework for digital justice in online consumer
Engineering and Technology. 8(5), 1261–1263. transactions in India. Proc. Comp. Sci., 165, 224–231.
Hingorani, I., Rushabh, K., Deepika, P., and Nataasha, R. Prajapat, S., Vaibhav, S., and Varun, W. (2018). A prototype
(2020). Police complaint management system using for grievance redressal system. Proc. Int. Conf. Recent
blockchain technology. 2020 3rd Int. Conf. Intel. Sus- Adv. Comp. Comm. ICRAC 2017, 41–49.
tain. Sys. (ICISS), 1214–1219. Rana, N. P., Yogesh, K. D., Michael, D. W., and Banita, L.
Jattan, S., Vineeth, K., Akhilesh, R., Rachith, R. N., and Sne- (2015). Examining the success of the online public
ha, N. S. (2020). Smart complaint redressal system us- grievance redressal systems: An extension of the IS
ing Ethereum blockchain. 2020 IEEE Int. Conf. Dis- success model. Inform. Sys. Manag., 32(1), 39–59.
tribut. Comput. VLSI Elec. Cir. Robot. (DISCOVER), Rana, N. P., Yogesh, K. D., Michael, D. W., and Vishanth, W.
224–229. (2016). Adoption of online public grievance redressal
Jha, S., Pankaj, S., and others. (2022). Smart student system in India: Toward developing a unified view.
grievance redressal system with foul language detec- Comp. Human Behav., 59, 265–282.
tion. 2022 8th Int. Conf. Adv. Comp. Comm. Sys. Shahnawaz, M., Prashant, S., Prabhat, K., and Anuradha, K.
(ICACCS), 1, 187–192. (2020). Grievance redressal system. Int. J. Data Min.
Kandhari, V. K. and Keertika, D. M. (2014). GPS based com- Big Data, 1(1), 1–4.
plaint redressal system. 2014 IEEE Global Human. Shettigar, R., Nishant, D., Ketan, I., Farhan, A., and Ram-
Technol. Conf. South Asia Satel. (GHTC-SAS), 51–56. krushna, C. M. (2021). Blockchain-based grievance
Kormpho, P., Panida, L., Narut, P., and Siripen, P. (2018). management system. Evol. Computat. Intel. Front.
Smart complaint management system. 2018 Seventh Intel. Comput. Theory Appl. (FICTA 2020), 1,
ICT Int. Student Project Conf. (ICT-ISPC), 1–6. 211–222.
Kumar, A., Sharad, S., Nitin, G., Aman, S., Xiaochun, C., Tripathi, U. N., Amit Kumar, S., and Bhanu P. S. (2021). Ef-
and Parminder, S. (2021). Secure and energy-efficient fectiveness of online grievance redressal and manage-
smart building architecture with emerging technology ment system: A case study of IGNOU learners. Ind. J.
IoT. Comp. Comm., 176, 207–217. Edu. Technol., 3(2), 92.
21 A systematic approach to implement hyperledger fabric
for remote patient monitoring
Shilpi Garg, Rajesh Kumar Kaushala and Naveen Kumar
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India

Abstract
The integration of blockchain technology, particularly hyperledger fabric, into the domain of remote patient monitor-
ing, presents a new era that has the potential to greatly improve healthcare systems. This research paper introduces a
methodical strategy for integrating hyperledger fabric into remote patient monitoring systems. It provides a structure that
efficiently addresses key challenges pertaining to data security, privacy, and interoperability. This study aims to establish a
methodology for remote patient monitoring environments by carefully analyzing the distinctive requirements and constraints
associated with these types of environments. The methodology encompasses various key phases, including network configu-
ration, smart contract design and sensitive data management specifically tailored to healthcare contexts. In addition, the
study explores practical methods of implementation and conducts performance evaluation of the suggested strategy using
minifab and hyperledger explorer, respectively. This analysis provides valuable insights into the effectiveness and efficiency
of the approach in safeguarding the privacy and security of patient information. Through an examination of the mutually
beneficial capabilities of hyperledger fabric and remote patient monitoring, this study makes a valuable contribution to the
advancement of healthcare systems that are both secure and efficient.

Keywords: Hyperledger fabric, remote patient monitoring, blockchain, hyperledger explorer, minifab

I. Introduction This research endeavors to explore the systematic

implementation of hyperledger fabric in the context
In recent years, the convergence of advanced technolo-
of remote patient monitoring. By leveraging the capa-
gies has assist in a new era of healthcare delivery, char-
bilities of blockchain technology, the aim is to estab-
acterized by improved patient outcomes, enhanced
lish a robust and secure framework that not only
data management, and increased accessibility. One of
safeguards patient data but also enhances interop-
the most significant advancements in this domain is
erability and trust among stakeholders, including
the integration of blockchain technology, particularly
patients, healthcare providers, and researchers.
hyperledger fabric, with remote patient monitoring
Through a systematic approach, this study discloses
(RPM) system. Remote patient monitoring, enabled
the intricate steps necessary to effectively integrate
by the proliferation of internet of things (IoT) devices
hyperledger fabric into RPM systems. From the initial
and wearable sensors, allows for continuous and real-
setup of the blockchain network to the development
time tracking of patients’ vital signs and health-related
of smart contracts tailored to healthcare scenarios,
data from the comfort of their homes. This paradigm
the research work provides a comprehensive guide for
shift has the potential to revolutionize healthcare by
implementation. Furthermore, this work also assesses
facilitating early diagnosis, personalized treatment,
the performance and effectiveness of the proposed
and reduced hospitalizations (Kumar et al., 2021;
systematic approach through testing and simulation.
Zhang et al., 2021; Garg, Kaushal, and Kumar, 2022).
However, the integration of RPM and blockchain
technology poses unique challenges and opportuni- II. Hyperledger fabric
ties. RPM systems deal with sensitive patient infor- Hyperledger fabric is an open-source blockchain
mation, making data security, privacy, and integrity platform specifically designed for enterprise applica-
paramount concerns. Traditional centralized data tions. The major objective of this platform is to enable
storage models often fall short in ensuring the con- customers to develop robust and scalable blockchain
fidentiality and authenticity of patient data (Pap et solutions. The platform under consideration exhibits
al., 2018; McGee et al., 2022; Kantorowska et al., a novel structure that coordinates the processing of
2023). Here in lies the potential of blockchain tech- transactions by means of executing smart contracts,
nology, particularly hyperledger fabric, to provide a referred to as chaincode, which can be written in pro-
decentralized and tamper-resistant framework that gramming languages like as Go, Java, or JavaScript
addresses these challenges. (Ichikawa, Kashiyama, and Ueno, 2017; Dabbagh

a
[email protected]
148 A systematic approach to implement hyperledger fabric for remote patient monitoring

Figure 21.1 Transaction flow for hyperledger fabric

et al., 2020; Tanwar, Parekh, and Evans, 2020). The sibility of transmitting information among net-
technology was developed within the framework of work participants while upholding data integrity.
the “Hyperledger Foundation”, an organization led Additionally, it enables the establishment of spe-
by IBM. It possesses several notable features, such as cific criteria or permissions to encapsulate the
the ability to create private data collections, strong transmitted data. In situations where maintain-
security measures for Docker containers, a flexible ing the confidentiality of particular information
programming framework, and a consensus model that is crucial, the option exists to create a separate
can be adjusted based on the host nodes. Hyperledger channel distinct from the rest, accessible only
fabric consists of various major components includ- to select organizations. This feature underscores
ing peers, orderer, chaincode, membership service the potential for multiple blockchains to coexist
provider (MSP), channels, and fabric certificate within the same network, as a channel essentially
authority (CA). Figure 21.1 illustrates the transaction operates as an independent blockchain.
flow diagram of the hyperledger fabric (Pongnumkul, • Certification authorities (CAs) are a fundamental
Siripanpornchana, and Thajchayapong, 2017; component of public key infrastructures (PKIs)
Performance, Group, and others, 2018; Jennath, and have been assigned with ensuring the distri-
Anoop, and Asharaf, 2020; Woznica and Kedziora, bution of digital certificates. The primary func-
2022). tion of this layer is to verify the identities of the
Peer refers to the individual nodes that comprise parties or actors involved in the communication,
the network organizations. The aforementioned ensuring that they are indeed who they claim to
pieces are responsible for providing information to be. Websites commonly possess a digital certifi-
the ordering nodes within the network, enabling them cate that is issued by a reputable CA in order to
to configure the blocks that are being transacted. authenticate the trustworthiness of the visited
Orderer: One of the pivotal components within the website.
network, the orderer assumes a critical role in config- • The membership service provider (MSP) is re-
uring blocks according to specified criteria and dis- sponsible for gathering all cryptographic tech-
tributing them to their respective peers. These peers niques employed for network interaction. It is
can be affiliated with one or multiple organizations, imperative for every organization to own a man-
necessitating the attainment of a consensus agreement aged security provider (MSP) that encompasses
as per the network’s requirements. All transactions its cryptographic data, including keys and the CA
related to network configuration flow through the responsible for issuing its certificates. The cre-
orderer. Additionally, these computing entities enforce dentials are utilized by clients for the purpose of
fundamental access control for channels, determining authenticating their transactions, while peers em-
who has read and write privileges and the authority ploy them to authenticate the outcomes of trans-
to configure them. action processing, specifically endorsements.
• Chaincode, often referred to as smart contracts
• Channel functions as a communication medium within the context of hyperledger fabric, serves
among network participants. In this context, it as the mechanism through which contractual
serves as a mechanism for conducting private agreements are implemented. A smart contract
communications, ensuring data isolation and refers to a block of code that is triggered by an
confidentiality. This layer takes on the respon- external client application, operating outside
 Applied Data Science and Smart Systems 149

the blockchain network. Its purpose is to over- comprising a network of peers. Language java is uti-
see the manipulation and control of a collection lized for the purpose of writing chaincode. Figure 21.3
of key-value pairs inside the present state of the is a screenshot of spec.yaml file that is a configuration
network, accomplished through the execution of file about the network used by minifab. Table 21.1
transactions. Smart contracts are encapsulated
and distributed as chaincode. Subsequently, the
chaincode is deployed onto the peers and subse-
quently defined and utilized within one or many
channels.

III. Implementation
In order to effectively handle the patient data it is
necessary to establish a correlation between the vari-
ous components of the fabric and the demands of the
RPM-based EHR systems. All medical centers func-
tion as entities inside a fabric network. The patient
data has been regarded as valuable resources stored
within the ledger. Currently, patient records consist
of a limited number of categories, encompassing
personal and medical information such as age, resi- Figure 21.3 Spec.yaml file for network
dence, allergies, symptoms, therapy, follow-up, and
so on. When a physician administers medication to a
Table 21.1 Process to build up the minifab network.
patient, they will have access to the patient’s medical
history data, which assists them in determining the Steps Command Description
most suitable type of medical care. Figure 21.2 illus-
1 minifab netup -s Start the network
trate the system architecture. couchdb -e true -i by adding hospital1.
The medical database is utilized to establish a 2.4.8 -o hospital1. health.com as a
repository of transactions within the proposed sys- health.com current organization
tem. The orderer and peer nodes are executed within 2 minifab create -c Create the
the Docker container. Hyperledger fabric framework healthchannel channel named as
is designed to be configured with a minimum of two healthchannel
organizations-hospital1 and hospital2. Every organi- 3 minifab join -c Network will join the
zation will be assigned to a single peer node, a chan- healthchannel healthchannel
nel, and an orderer node within the ordering service. 4 minifab Update the anchor
Each peer node within the network possesses a dupli- anchorupdate peer node
cate of the ledger. A chaincode is developed with the 5 minifab profilegen -c Generate the profiles
purpose of facilitating access to the two organizations healthchannel for healthchannel

Figure 21.2 System architecture for hyperledger fabric

150 A systematic approach to implement hyperledger fabric for remote patient monitoring

Figure 21.7 Result for query a patient

Figure 21.4 Smart contracts for entities

Figure 21.5 Screenshot for creating a patient

Figure 21.6 Screenshot for query a patient

depicts the steps for creating the network using

minifab.
After successfully build up the network, chaincode
is deployed on the health channel. One chaincode has
been created for admin, patient and doctor entities.
Basically chaincode is a collection of smart contracts.
For creating a block invoke command is used and for Figure 21.8 (a) Blocks per min, (b) Transactions per
evaluation query command is used. Functionality of organization
admin, patient and doctor smart contract has shown
into the Figure 21.4.
Health information is extremely sensitive and
Results
must be protected. Only the healthcare providers
and institutions to which the patient has granted Hyperledger explorer is a tool that designed to pro-
access should have access to their medical records. vide a user-friendly interface for viewing, analyzing,
Private data collections are an option for storing and interacting with blockchain data in hyperledger
sensitive information in hyperledger fabric. Certain fabric networks. Minifab network is evaluated using
patient information must be shielded from investiga- the hyperledger explorer tool which is open source
tors at other medical facilities. Figure 21.4 illustrate tool. Figure 21.8(a) and (b) depicts the total number
the invoke command for creating a new patient by of blocks created per min and blocks created by orga-
admin smart contract. The patient data added here nizations. It gives the complete information about the
is kept private. Figures 21.5–21.7 depicts the query block like data hash, previous hash, block hash, num-
command for reading a patient private data and out- ber of tractions, channel name. A total of 9 blocks are
put for a read query. created for 9 transactions with 2 nodes.
 Applied Data Science and Smart Systems 151

Conclusion abling trusted artificial intelligence. 15–23. https://doi.

org/10.9781/ijimai.2020.07.002.
This systematic approach to implement hyperledger Kantorowska, Agata, Koral Cohen, Maxwell Oberlander,
fabric for remote patient monitoring, addressing criti- Anna R. Jaysing, Meredith B. Akerman, Anne-Marie
cal challenges in healthcare viz., data management Wise, Devin M. Mann et al. (2023). Remote patient
and security. The deployment of blockchain tech- monitoring for management of diabetes mellitus in
nology in healthcare, offers promising prospects for pregnancy is associated with improved maternal and
enhancing data integrity, privacy, and interoperability. neonatal outcomes. American Journal of Obstetrics
Throughout this study, outlined a comprehensive and Gynecology. 228, 6 pp. 1-11.
Kumar, A., Sharad, S., Nitin, G., Aman, S., Xiaochun, C.,
implementation framework, including smart con-
and Parminder, S. (2021). Secure and energy-efficient
tracts and data management strategies, to facilitate
smart building architecture with emerging technology
secure and efficient remote patient monitoring. As the IoT. Comp. Comm., 176, 207–217.
healthcare industry continues to evolve and embrace McGee, M. J., Max, R., Stepehn, C. B., Shanathan Srith-
digital transformation, the integration of hyperledger aran, Andrew J. Boyle, Nicholas, J., James, W. L., and
fabric for remote patient monitoring has the poten- Aaron, L. S. (2022). Remote monitoring in patients
tial to revolutionize patient care by enabling real-time with heart failure with cardiac implantable electron-
data sharing among healthcare providers, ensur- ic devices: A systematic review and meta-analysis.
ing data accuracy, and safeguarding patient privacy. Open Heart, 9(2), 6–11. https://doi.org/10.1136/
Nevertheless, challenges and barriers remain, includ- openhrt-2022-002096.
ing scalability concerns and the need for widespread Pap, I. A., Stefan, O., Ioan, O., and Alexandru, A. (2018).
IoT-based ehealth data acquisition system. 2018 IEEE
adoption.
Int. Conf. Automat. Qual. Test. Robot. AQTR 2018 -
In the future, further research and practical imple-
THETA 21st Ed., Proc., 1–5. https://doi.org/10.1109/
mentations should explore scalability solutions and AQTR.2018.8402711.
continue to engage stakeholders across the healthcare Performance, Hyperledger, Scale Working Group, and oth-
ecosystem to overcome barriers to adoption. ers. (2018). Hyperledger blockchain performance
metrics white paper. https://www. Hyperledger. Org/
References resources/publications/blockchain-Performance-Met-
rics. Accessed on 31(1), 2020.
Dabbagh, M., Mohsen, K., Mohammad, T., and Angela, A. Pongnumkul, S., Chaiyaphum, S., and Suttipong, T. (2017).
(2020). Performance analysis of blockchain platforms: Performance analysis of private blockchain platforms
Empirical evaluation of hyperledger fabric and ethe- in varying workloads. 2017 26th Int. Conf. Comp.
reum. 2020 IEEE 2nd Int. Conf. Artif. Intel. Engg. Comm. Netw. (ICCCN), IEEE, 1–6.
Technol. (IICAIET), 1–6. Tanwar, S., Karan, P., and Richard, E. (2020). Blockchain-
Garg, S., Rajesh Kumar, K., and Naveen, K. (2022). based electronic healthcare record system for healthcare
Blockchain-based electronic health record and open 4.0 applications. J. Inform. Sec. Appl., 50, 102407.
research challenges. 2022 10th Int. Conf. Reliabil. Woznica, A. and Michal, K. (2022). Performance and
Infocom Technol. Optimiz. (Trends and Future Direc- scalability evaluation of a permissioned blockchain
tions)(ICRITO), 1–5. based on the hyperledger fabric, Sawtooth and Iroha.
Ichikawa, D., Kashiyama, M., and Ueno, T. (2017). Tamper- Comp. Sci. Inform. Sys., 19(2), 659–678. https://doi.
resistant mobile health using blockchain technology. org/10.2298/CSIS210507002W.
JMIR mHealth and uHealth, 5(7), 1–11. https://doi. Zhang, X., Kantilal, P. R., Ismail, K., and Mohammad, S.
org/10.2196/mhealth.7938. (2021). Research on vibration monitoring and fault
Jennath, H. S., Anoop, V. S., and Asharaf, S. (2020). Block- diagnosis of rotating machinery based on internet of
chain for healthcare: Securing patient data and en- things technology. Nonlin. Engg., 10(1), 245–254.
22 Developing spell check and transliteration tools for Indian
regional language – Kannada
Chandrika Prasada, Jagadish S. Kallimani, Geetha Reddy and
Dhanashekar K.
Department of Computer Science and Engineering, M S Ramaiah Institute of Technology, (Affiliated to Visvesvaraya
Technical University Belagavi, Karnataka), Bangalore, Karnataka, India

Abstract
Kannada is one of the major regional languages of Karnataka, a prominent state of India. The text processing tasks are very
important and highly required for the development of the language in this digital world. Spell checking is one of the needs
in creating an effective document. Even though one can find several tools on the internet, it allows you to type or paste the
Kannada text on the text editor and submit the text then the result will appear on the another editor. The proposed work de-
lineates on developing an efficient interactive spell checking and transliteration tools for the Kannada language based on the
Blooms filter algorithm, Symspell technique and International Phonetic Alphabet (IPA) representation. This work carried out
with an intention to provide handy text processing tools to the public. The proposed work has been tested on several datasets
and found to be useful with more than 85% and 87.29% accuracy for both spell check and transliteration tools, respectively.

Keywords: Blooms filter, Symspell algorithm, Levenshtein distance, transliteration, International Phonetic Alphabet, candi-
date words

Introduction it in English. This paper provides solution for these

identified challenges.
One of the key uses of natural language processing
Since its impossible to create a corpus for each eng-
is the spell checker. It aids the user in producing a
kan translation, hence we go for the transliteration
document free of errors. In several languages, the
approach where the pronunciation is preserved.
work of spell checking has already been extensively
developed. MS-word is a widely used editor that gives
Dictionary based tool
users all the tools they need to create effective docu-
ments, especially in English. India contains more than Set of Kannada words from different internet sources
100 local languages, although only 22 are regarded as are collected and stored as a database. A list of mis-
regional. Text processing is still in its infancy in many spelled words and its equivalent correct words are
languages, notably Kannada. While there are a few mapped and stored as dictionary. The module invokes
transliteration tools available, there are no editors a routine which performs look up operation in the
where users can freely enter and create documents dictionary and predicts the possible word for the mis-
that are error-free. spelled Kannada word using efficient techniques like
According to the literary survey, more than 1.35 Bloom filter and Symspell algorithms.
billion people on the world speak and understand
English, hence it is safe to assume that English is Kannada transliteration
widely accepted all around the globe. The popula- Transliteration is the process of converting text from
tion of India is around 1.4 billion surpassing China one writing system to another. One widely used
and hence becoming the most populated country in method of transliteration is to use the International
the world. Around 57 million Kannada speakers exist Phonetic Alphabet (IPA), which provides a standard-
in India making it one of the most used languages ized set of symbols for representing the sounds of
for communication in India. The government has human language. IPA-based transliteration allows for
come up with various solutions to enable even the a more accurate representation of the sounds of one
most remote parts of the country to access informa- language in another, which can be useful for language
tion. Hence the need to convert language into native learners, linguists, and others who need to work
tongue arises. with multiple writing systems. The IPA is based on
Proper names such as name, place, object and, etc., the principle of one sound, one symbol, which means
are the fields in all government forms and offices, it’s that each symbol represents a single sound. It allows
practically impossible to ask the public to only fill for a standardized representation of the sounds of a

[email protected]
a
 Applied Data Science and Smart Systems 153

language, regardless of the writing system used. The more effort into creating a real-word spell checker
IPA includes symbols for consonants, vowels, and that incorporates additional language principles.
other sounds, as well as diacritic marks that indicate A spell check tool using Levenshtein’s edit distance
variations in pronunciation, such as stress and tone. algorithm, rule-based algorithm, Soundex algorithm,
While IPA-based transliteration can be more precise and LSTM (Long- Short-Term Memory) model is
than other methods, it can also be more complex developed for Tamil language (Sampath et al., 2022).
and time-consuming, especially for those who are The model handles three categories of errors with a
not familiar with the IPA. Additionally, not all lan- good performance of 95.67%.
guages have a one-to-one correspondence between A Telugu spell-checker’s innovative concept and
their sounds and IPA symbols, which can lead to some implementation are presented in (Parameshwari et
ambiguity in transliteration. Despite these challenges, al., 2012). The core of Telugu spell-checking is mor-
IPA-based transliteration remains a valuable tool for phological validation using a morphological analyzer.
those who need to work with multiple languages and Along with issues affecting orthography and morphol-
writing systems. ogy, difficulties associated with Telugu document spell
In the proposed work, both dictionary and translit- checking are examined. On these lines, a spell-checker
eration tools are developed, the implementation part has been created. The spelling checker’s architecture
will focus more on these two models. and algorithm, which is based on Sandhi splitter and
morphological analysis principles, are described.
Related work Additionally, it contains tables of spelling variations
gleaned from Telugu’s spatiotemporal dialects.
A comprehensive survey is done on various languages A common approach used to develop a spell check
to understand the methodology/ technique used by tool is minimum edit distance algorithm (Patil et al.,
researchers. 2021). By carrying out numerous operations including
The researchers have explored methodologies character replacement, insertion, and deletion, it fixes
(Randhawa et al., 2014) used for developing spell spelling mistakes. The proposed work focuses on cor-
check tool for various Indian regional languages recting the errors for Marathi text and it works better
including the performance analysis. This helps a for short words with a good accuracy of 85.5%.
researcher to understand the pros and cons of avail- The challenges associated with multi-lingual
able techniques. A spell checker tool on Bangla is speech recognition and propose solutions to address
explored in (Chaudhuri et al., 2002). Researchers these challenges in the Indian context are explored in
have handled the errors based on the phonetic in two (Manjunath et al., 2019; Khattar et al., 2020). The
stages using phonetically similar character error cor- proposed model contributes to the advancement of
rection and reversed word dictionary and error cor- speech technology in the context of Indian languages,
rection. Experiment is conducted on three million which is crucial for enabling effective communication
words which are arranged in Trie data structure and and technology access for the diverse linguistic popu-
obtained satisfied results. lation in India.
Speech recognition is explored in (Priya et al., Both forward and backward transliteration of
2022), authors have used novel Automatic Speech Punjabi names was performed between Gurmukhi
Recognition system for seven low-resource languages and English Roman scripts using an n-gram language
based on deep sequence modeling with an enhanced model (Goyal et al., 2022). Over one million paral-
spell checker. The researchers have obtained word lel entities of person names in both scripts were used
error rate (WER) of 0.62 using recurrent neural as the training corpus. The study created extensive
network-gated recurrent unit (RNN-GRU) and the English-to-Punjabi and Punjabi-to-English n-gram
transformer-based INDIC Bidirectional Encoder pre- databases, comprising more than 10 million n-grams
sentations significantly enhance performance by 10% with multiple script mappings. Categorizing n-grams
and lower the average WER to 0.52. into starting, middle, and ending n-grams was essen-
A spell check tool is explored on Dawurootsuwa tial due to variations in pronunciation based on let-
which is one of the Ethiopian languages (Arya et al., ter placement in words. The transliteration process
2021; Gamu et al., 2023), it has poor dataset. The involved searching for the longest matching n-gram
root words in this study were built using the Hunspell in the database, recursively splitting the string until a
dictionary format and consisted of 5,000 total root match was found, and then merging the transliterated
words, more than 2,500 morphological rules, and strings to produce the final output.
3,156 unique terms for testing. total spell error detec- The challenges of speech recognition and spell cor-
tion performance was 90.4%, and total spell error rection in low-resource Indian language are discussed
repair performance was 79.31%, according to the in the work did by Priya et al. (2022). The authors
experimental results. Additionally, we are putting propose a solution using Indic BERT. A multi-lingual
154 Developing spell check and transliteration tools for Indian regional language – Kannada

transformer-based language model, the model per- regular keyboard (English language keyboard)
forms speech recognition and spell correction for the and can check the correct Kannada words on the
text written in Tamil, Telugu, and Kannada languages. editor.
By leveraging the power of transfer learning, the
authors demonstrate the effectiveness of Indic BERT Methodology
in improving the accuracy of speech recognition and
spell correction tasks in these languages. The general steps to develop a spell-checking tool in
Kannada speech corpus for automatic speech rec- Kannada language is as follows:
ognition system based on phoneme (Praveen et al.,
2022) is developed for Kannada corpus. The authors • Corpus collection: Gather a large collection of
describe the methodology employed in creating the correctly spelled Kannada text. This can include
corpus, which includes collecting speech samples books, articles, websites, and other reliable sourc-
from native Kannada speakers and annotating them es written in Kannada.
with phoneme-level transcriptions. The resulting • Corpus pre-processing: Clean and preprocess the
corpus serves as a valuable resource for researchers collected corpus data by removing any unwanted
and practitioners working on Kannada speech recog- characters, punctuation marks, and special sym-
nition, enabling the development and evaluation of bols. Normalize the text to ensure consistent rep-
accurate and efficient speech recognition models for resentation.
this language. • Tokenization: Split the pre-processed text into in-
A convolutional neural network-based speech rec- dividual words or tokens. This step helps in ana-
ognition model for Kannada Language is demon- lyzing and processing each word separately.
strated in work by Rudregowda et al. (2020). The • Build a dictionary: Create a dictionary of correct-
authors propose a methodology for visual speech ly spelled Kannada words based on the tokenized
recognition in Kannada. The findings of this study corpus. This dictionary will serve as the reference
contribute to the advancement of speech recognition for spell-checking.
technology for Kannada, which could have signifi- • Error generation: Generate a set of common
cant implications for speech-based applications in the spelling errors that occur in Kannada. This can
Kannada-speaking community. include typos, phonetic errors, and other com-
mon mistakes made by Kannada speakers.
• Spell-checking algorithm: Implement a spell-
Scope of the work
checking algorithm that compares each word in
From the survey, it is found extensive research work the input text with the words in the dictionary.
has not carried out in this domain. There is a lot of The algorithm should identify potential spelling
scope in this area. Summary of the survey is as fol- errors and suggest corrections based on the clos-
lows: After analyzing the survey, it is found that est matching words in the dictionary.
• User interface: Develop a user-friendly inter-
• In Kannada languages, minimum work has been face where users can input their text for spell-
carried out in spell check and transliteration do- checking and receive suggestions for correc-
main. tions. This can be a web-based interface or an
• Getting the proper Kannada datasets for training application.
and testing is not an easy task. • Testing and refinement: Test the spell-checking
• There is no open-source optical recognition tool tool with a variety of Kannada texts, including
available to convert pdf to word which is required different genres and writing styles. Collect user
for the text processing. feedback and refine the algorithm and user inter-
face based on the feedback received.
Objectives • Continuous improvement: Maintain and update
the spell-checking tool by periodically updating
From the survey, it is noted that, in Kannada language the dictionary with new words and refining the
there is a lot of scope with respect to transliteration error generation algorithms to improve accuracy
and not many research articles are published. We have and coverage.
contributed in this domain by
It is worth noting that building a robust and
• Developing a spell check tool with the possible accurate spell-checking tool requires a considerable
features amount of linguistic expertise and computational
• Designing a transliteration tool for Kannada lan- resources. Collaborating with Kannada language
guage, where the user can type Kannada using experts or researchers in natural language processing
 Applied Data Science and Smart Systems 155

(NLP) would be beneficial in ensuring the effective- • Initialize a bit array of the specified size and set
ness of the tool. all bits to 0.
• Calculate the optimal number of hash functions
Dictionary-based spell checking tool based on the desired false positive probability
and the size of the dataset.
A huge dataset of 7 lakh is collected and in that • Create a list of hash functions using different seed
125,000 words are identified as unique words. These values.
words can have spelled in many ways all those mis-
spelled forms of these unique words are tabulated in a The following Figure 22.1 shows the architecture
dictionary which is used for error correction of the proposed model:
In this proposed model a user interface is devel-
oped such that it accepts the Kannada document or • Insert elements into the Bloom filter.
an editor is provided for the user to start typing the • For each element in the Kannada dataset, apply
Kannada articles. each hash function to generate hash values.
After the user uploads the document, two possible • Set the corresponding bits in the Bloom filter’s bit
scenarios can unfold. Firstly, a routine can be imple- array to 1 for each generated hash value.
mented to exhibit the precise contents of the docu- • Search for an element in the Bloom filter
ment within the designated text area. Secondly, all • Given a query element, apply each hash function
the words present in the document are divided into to generate hash values.
tokens, and the unique tokens are subsequently sub- • Check if the corresponding bits in the Bloom filter’s
jected to processing by Blooms filter. bit array are set to 1 for each generated hash value.
Here are the steps for implementing a Bloom filter • If any of the bits are not set to 1, the element is
searching algorithm for a Kannada dataset: definitely not present in the dataset.
• If all bits are set to 1, the element is probably pres-
• Create a Bloom filter: ent in the dataset (there is a false positive prob-
• Specify the desired size of the Bloom filter and the ability). Figures 22.2 and 22.3 shows the result of
number of hash functions to use. the search operation using Bloom filter.

Figure 22.1 Chronological order of face shield development

156 Developing spell check and transliteration tools for Indian regional language – Kannada

print(f”The word ‘{query_word}’ is probably

present in the dataset.”)
else:
print(f”The word ‘{query_word}’ is definitely not
present in the dataset.”)

The input undergoes scanning by the Bloom filter,

leading to the display of words on the designated edi-
tor within the user interface. In this context, two pos-
sibilities arise once again:
Figure 22.2 Input word not present in Bloom filter
1. Incorrect words are highlighted with a red un-
derline.
2. Additionally, correct words that are not found
in the dictionary are also recognized as incorrect
and flagged with a red underline.

The user has given three options with the identified

wrong words

1. User can ignore this by clicking on the ignore op-

tion present in the toolkit.
Figure 22.3 Input word present in Bloom filter 2. User can add the words to dictionary, if he feels
that it is correct. The admin later will add this
to dictionary so that next time, these words are
Here’s the pseudo code for the Bloom filter search- treated as correct words.
ing algorithm: 3. User can find the possible correct words for the
identified wrong words.
import bitarray
from hashlib import sha256 Error correction
class BloomFilter: To predict the possible correct word for the given
def __init__(self, size, num_hash): wrong word, the model uses SymSpell algorithm with
self.size = size the Levenshtein distance metric. The steps are :
self.num_hash = num_hash
self.bit_array = bitarray.bitarray(size) 1. Prepare the dictionary.
self.bit_array.setall(0) 2. Create or obtain a Kannada dictionary file in the
def add(self, item): format of “term frequency” per line.
for seed in range(self.num_hash): 3. Initialize SymSpell
index = int(sha256(item.encode(‘utf-8’) + 4. Perform spell checking with post-process sugges-
str(seed).encode(‘utf-8’)).hexdigest(), 16) % self.size tions using the Levenshtein distance, the steps to
self.bit_array[index] = 1 calculate the similarity between two words are
def search(self, item): given below:
for seed in range(self.num_hash): 5. Input: Two Kannada words, word1 and word2,
index = int(sha256(item.encode(‘utf-8’) + for which we want to calculate the Levenshtein
str(seed).encode(‘utf-8’)).hexdigest(), 16) % self.size distance.
if self.bit_array[index] == 0: 6. Initialize the matrix
return False • Create a matrix, dp, of size (m+1) × (n+1),
return True where m is the length of word1 and n is the
# Example usage: length of word2.
bloom_filter = BloomFilter(size=1000, num_hash=3) • The literature was reviewed first to find the
kannada_dataset = [“ಪ್ರಿಯಾ”, “ಸುರೇಶ”, specifications of initialize the first row of the
“ಕೃಷ್ಣ”, “ಮಂಜು”] matrix with values 0 to n, representing the
for word in kannada_dataset: number of insertions required to transform
bloom_filter.add(word) an empty string into word2.
query_word = “ಸುರೇಶ” • Initialize the first column of the matrix with
if bloom_filter.search(query_word): values 0 to m, representing the number of
 Applied Data Science and Smart Systems 157

deletions required to transform word1 into the given dataset. Similarly, the term fail in the graph
an empty string. refers to the percentage of failure in predicting the
7. Calculate the Levenshtein distance wrong words.
• Iterate through the characters of word1 From the Table 22.2 and the graph in Figure 22.5,
(from i=1 to m) and word2 (from j=1 to n). it is clear that for a small dataset like 10 words it
• If word1[i-1] is equal to word2[j-1] (i.e., the works pretty well with 90% accuracy. As we increase
characters are the same), the cost of the cur- the dataset it performs better, for 1 lakh of words
rent operation is 0. Set dp[i][j] to the value the accuracy is still better with 87%. Frequency of
of dp[i-1][j-1]. the words in the dictionary and different forms of
• If word1[i-1] is different from word2[j-1], grammatical words for a given word has an impact
we have three possible operations: on the performance of the model. If the data-
• Insertion: Calculate the cost of inserting set has more wrong words, then it will learn and
word2[j-1] into word1 at position i. Set dp[i] perform the prediction better. Figures 22.6–22.9
[j] to dp[i][j-1] + 1.
• Deletion: Calculate the cost of deleting
word1[i-1] from word1. Set dp[i][j] to dp[i- Table 22.1 Dataset details.
1][j] + 1.
• Substitution: Calculate the cost of substitut- Dataset Files
ing word1[i-1] with word2[j-1]. Set dp[i][j]
Articles 1026
to dp[i-1][j-1] + 1.
• Choose the minimum cost among the three Stories 51
possible operations and assign it to dp[i][j]. Wikipedia dataset 201
• Output: Grammar data 3
• The final Levenshtein distance is stored in
Dataset Content Size
dp[m][n], representing the minimum num-
ber of edits required to transform word1 Words 726,654
into word2.
Unique Words 179,863
The following Figure 22.4 demonstrates the results of
searching a word in a dictionary using Bloom filter.
Table 22.2 Performance analysis.
Results S.No. Number of words Accuracy in %
The proposed work with complete user interface is
1 10 90
uploaded on a website and released to the public. The
website is designed by taking requirements from the 2 100 91
users working from various domains. Initially the 3 1000 89
model is tested with 7 lakh words and later with dif- 4 10,000 85
ferent set of words. Tables 22.1 and 22.2 shows the 5 100,000 87
dataset type, size and the accuracy of the model. The
graph in Figure 22.5 shows the performance of the
model. In the graph, the term pass refers to the accu-
racy of the model in predicting the wrong words for

Figure 22.4 Input misspelled word to Symspell give

suggestion word Figure 22.5 Performance analysis
158 Developing spell check and transliteration tools for Indian regional language – Kannada

demonstrates the user interface, underlining of

wrong words, selecting a wrong word and its cor-
rect word, respectively.

Snapshots

Figure 22.9 Suggesting the correct word for the mis-

spelled word

Figure 22.6 User interface for spell check

Figure 22.10 Pipeline for the process

Table 22.3 Dataset split up.

Dataset split Number of words

Train 75,557
Validation 25,185
Test 25,185
Figure 22.7 Identifying wrong words (underlined in
red)
English to IPA translation
The model is built using IPA to transliterate words
written in English to Kannada language. The dataset
used has 125,927 unique words. It is represented as
each English word and all its IPA translations. The
Figure 22.10 demonstrates the abstract view of this
work.
The following are the steps involved in the
translation:

Data pre-processing
The input data is initially in the raw state, converting
the dataset into a pair of English words and their cor-
responding IPA translations is done in the preprocess-
ing stage by removing unwanted text like numbers
Figure 22.8 Selecting the wrong words with options
and the special symbols since these do not require any
translations.
The English words and IPA translations are
Kannada transliteration tool tokenized, and the unique characters in both sets are
This section describes the implementation details of extracted. The input sequences are padded to a fixed
transliteration which translates text from English to length to ensure uniformity. The dataset details are
Kannada. shown in the Table 22.3.
 Applied Data Science and Smart Systems 159

Model architecture
Character BERT is a specialized variant of the BERT
model that operates at the character level, making
it suitable for tasks such as phonetic transcription.
When translating English words to IPA transcrip-
tions, character BERT learns the relationship between
input characters and their corresponding IPA sym-
bols. The process begins by encoding each English
word into individual characters and converting them
into numerical representations using a character
vocabulary.
The model architecture of character BERT con-
sists of a multi-layer bidirectional transformer that
captures contextual information from both the left
and right contexts of each character. Prior to fine-
tuning, Character BERT undergoes pre-training on
large-scale unlabeled data, where it learns to predict
masked characters based on the context provided by
surrounding characters. Figure 22.11 Character BERT embedding
During fine-tuning, the model is trained on a par-
allel dataset of English words and their IPA tran-
scriptions, enabling it to encode the input characters
and predict the correct IPA transcriptions using the
learned contextual information. In inference, given an
English word, the characters are tokenized, encoded,
and passed through the fine-tuned character BERT
model. The model generates a sequence of numeri-
cal representations that can be decoded using the IPA
vocabulary, yielding the corresponding IPA transcrip-
tion. Character BERT’s strength lies in its ability to
capture fine-grained information from individual
characters, enabling accurate and context-aware IPA
transcriptions for English words (Figure 22.11).

• The target IPA sequence is shifted by one time

step to form the decoder input, and one-hot en- Figure 22.12(a) Mapping vowels to its IPA representa-
coding is used for training labels. tion
• For inference, the trained model is used to gener-
ate IPA translations for new English words.
• The encoder model is used to encode the input
English word and retrieve the final hidden state.
• The decoder model takes the encoded state and
generates the IPA translation sequence character
by character.

Model training
• The model is trained using the compiled model
with the RMSprop optimizer and categorical Figure 22.12(b) Mapping consonants to its IPA rep-
cross-entropy loss function. resentation
• The training is performed by providing the en-
coder input (English word sequence) and decoder
input (IPA translation sequence) to predict the de- • The generated IPA translations are outputted for
coder output (next IPA character) as shown in the evaluation or further processing.
Figures 22.12a and b.
• The model is trained on a training set and vali- After the IPA translation is generated, using the
dated on a separate validation set. IPA-Kannada mapping we map each IPA symbol
160 Developing spell check and transliteration tools for Indian regional language – Kannada

to a corresponding Kannada alphabet as shown in Table 22.4 Sample transliterations of Kannada words
the Figure 22.13. The pseudo code is given by the
following: English word IPA translation Transliteration

Annabella ˌænəˈbɛlə ಅನ್ನಾಬೆಲ್ಲಾ

Algorithm: ipa to kannada Time taɪm’ ಟೈಮ್
School skul ಸ್ಕೂಲ್
Input: an ipa string
Output: transliterated kannada string Transformer trænsˈfɔrmər ಟ್ರಾನ್ಸ್ಫಒರ್ಮೆರ್
Amazing əˈmeɪzɪŋ ಅಮೇಜಿಂಗ್
1. Initialize consonant = 0, kan = “ “
Public ˈpəblɪk ಪಬ್ಲಿಕ್
2. While there are unprocessed characters in the Beautiful ˈbjutəfəl ಬ್ಯೂಟಿಫುಲ್
generated IPA word:
if consonant == 0 and IPA word is a vowel:
3. Append the corresponding vowel from the
given dataset. The model may perform well with
Kannada mapping to ‘kan’
increase in cleaned dataset.
else if consonant! = 1 and IPA word is a
vowel:
Discussion
4. Combine the consonant and vowel, and ap-
pend the result to ‘kan’ The proposed work incorporates both dictionary
5. Decrement ‘consonant’ based spell checking tool and transliteration. Spell
else if consonant != 1 and IPA word is a checking tool is based on dictionary and the per-
subscript: formance of the model mainly based on the volume
of the dictionary. As long as the dictionary is grow-
6. Append the subscript to the consonant ing the performance starts improving. This is not an
else if IPA word is a consonant: effective nature instead if the model understands the
7. Increment ‘consonant’ and append it to ‘kan’ rules of the grammar, the model doesn’t depend on
else: print(“error”) the words in the dictionary. This is the planned work
in the future.
Table 22.4 shows the sample examples of the Kannada transliteration tool is a productive one
transliteration. since it is based on the rules of the Kannada gram-
The proposed model performs well on the valida- mar still more feature can be added to it in future and
tion set by achieving an accuracy of 86.9% for the released for public use.

Conclusion
The proposed work incorporates both dictionary
based spell checking tool and transliteration. Spell
checking tool is based on dictionary and the per-
formance of the model mainly based on the volume
of the dictionary. As long as the dictionary is grow-
ing the performance starts improving. This is not an
effective nature instead if the model understands the
rules of the grammar, the model doesn’t depend on
the words in the dictionary. This is the planned work
in the future.
Kannada transliteration tool is a productive one
since it is based on the rules of the Kannada gram-
mar still more feature can be added to it in future and
released for public use.

References
Randhawa, Er, Sumreet, K., and Saroa, Er C. S. (2014).
Study of spell checking techniques and available spell
checkers in regional languages: a survey. Int. J. Tech-
Figure 22.13 Mapping consonants to its IPA nol. Res. Engg., 2(3), 148–151.
 Applied Data Science and Smart Systems 161
Chaudhuri, B. B. (2002). Towards Indian language spell- Goyal, K. D., Muhammad, R. A., Vishal, G., and Yasir, S.
checker design. Lang. Engg. Conf., 2002 Proc., 139– (2022). Forward-backward transliteration of Punjabi
146. Gurmukhi script using n-gram language model. ACM
Priya, M. C., Shunmuga, D., Karthika, R., Ashok Kumar, Trans. Asian Low-Res. Lang. Inform. Proc., 22(2),
L., and Lovelyn Rose, S. (2022). Multilingual low re- 1–24.
source Indian language speech recognition and spell Manjunath, K. E., Dinesh Babu, J., Sreenivasa Rao, K., and
correction using Indic BERT. Sa-dhana-, 47(4), 227. Ramasubramanian, V. (2019). Development and anal-
Gamu, D. T. and Michael, M. W. (2023). Morphology-based ysis of multilingual phone recognition systems using
spell checker for Dawurootsuwa language. Scientif. Indian languages. Int. J. Speech Technol., 22, 157–168.
Prog., 2023. Arya, R., Singh, J., and Kumar, A. (2021). A survey of
Sampath, A. and Varadhaganapathy, S. (2023). Hybrid multidisciplinary domains contributing to affective
Tamil spell checker with combined character splitting. computing. Comp. Sci. Rev., 40, 100399, https://doi.
Concur. Computat. Prac. Exp., 35(1), e7440. org/10.1016/j.cosrev.2021.100399.
Patil, K. T., Bhavsar, R. P., and Pawar, B. V. (2021). Spell- Prakash, A. and Hema, A. M. (2022). Exploring the role of
ing checking and error corrector system for Marathi language families for building Indic speech synthesis-
language text using minimum edit distance algorithm. ers. IEEE/ACM Trans. Audio Speech Lang. Proc., 31,
Adv. Comput. Data Sci. 5th Int. Conf. ICACDS 2021, 734–747.
Nashik, India, April 23–24, 2021, Revised Selected Priya, M. C., Shunmuga, D., Karthika, R., Ashok Kumar,
Papers, Part I 5, 102–111. Springer International Pub- L., and Lovelyn Rose, S. (2022). Multilingual low re-
lishing, 2021. source Indian language speech recognition and spell
Priyadarshani, H. S., Rajapaksha, M. D. W., Ranasing- correction using Indic BERT. Sa-dhana-, 47(4), 227.
he, M. M. S. P., Kengatharaiyer, S., and Dias, G. V. Praveen, N. and Shashidhar, K. (2022). Phoneme based
(2019). Statistical machine learning for translitera- Kannada speech corpus for automatic speech recogni-
tion: Transliterating names between Sinhala, Tamil tion system. 2022 IEEE Int. Conf. Distribut. Comput.
and English. 2019 Int Conf. Asian Lang. Proc. Elec. Cir. Elec. (ICDCECE), 1–5.
(IALP), 244–249. Rudregowda, S., Sudarshan Patil, K., Gururaj, H. L., Vi-
Khattar, N., Singh, J., and Sidhu, J. (2020). An energy effi- nayakumar, R., and Moez, K. (2023). Visual speech
cient and adaptive threshold VM consolidation frame- recognition for Kannada language using VGG16 con-
work for cloud environment. Wire. Per. Comm., 113, volutional neural network. Acoustics, 5(1), 343–353.
349–367.
23 Real-time identification of traffic actors using YOLOv7
Pavan Kumar Polagania, Lakshmi Priyanka Siddib and Vani Pujitha M.c
Velagapudi Ramakrishna Siddhartha Engineering College, Vijayawada, India

Abstract
Real-time traffic object detection is a key topic in computer vision, especially for improving traffic safety and management. This
research describes a novel strategy for detecting traffic actors in real-time using YOLOv7, a cutting-edge deep learning system.
Traditional computer vision algorithms, such as Single Shot Detector, R-CNN, and older versions of You Only Look Once
(YOLO), frequently exhibit slow response times and poor accuracy in high-traffic areas.YOLOv7, an advanced object detec-
tion method based on convolutional neural networks (CNNs), is used in the proposed approach to address these difficulties
straight on. YOLOv7 not only achieves real-time object detection, but also greatly increases accuracy by removing superfluous
candidate boxes and employing a non-maximum suppression module to choose the best bounding boxes from overlapping
ones. Furthermore, the spatial pyramid pooling block improves accuracy by enhancing the network’s receptive field without in-
troducing additional parameters. In this study, we demonstrate the performance of our model under various driving scenarios,
including clear and cloudy skies, varying lighting, occlusions, and noisy input data. This model detects traffic participants such
as automobiles, pedestrians, cyclists, and traffic signs, which contributes to improved traffic safety and management.

Keywords: YOLOv7, traffic detection, convolutional neural networks

Introduction The manuscript follows a structured format com-

prising seven sections and they are as follows: An
Real-time object detection in traffic poses a formi-
extensive literature review on image classification and
dable challenge within the realm of computer vision.
object detection. The dataset used in this study. An
This task involves processing live video streams cap-
overview of the architectural framework utilized. The
tured by cameras or sensors deployed in traffic envi-
methodology employed in our research. Experimental
ronments. The objective is to detect and precisely
setup. The experimental data are presented together
localize a wide variety of items, such as vehicles,
with a thorough evaluation of the model’s perfor-
people on foot, bicycles, and traffic signals. The pri-
mance and a discussion of the results. Finally, the
mary goal of real-time object detection in traffic is to
paper is concluded by summarizing key insights and
enhance traffic safety by providing real-time informa-
suggesting potential avenues for future research.
tion to drivers regarding the condition of traffic in
nearby locations.
Main contribution of the work
This task is complex due to factors such as the
1. The primary objective in this project is to cre-
diverse appearance and motion of objects, chang-
ate a computer vision system that can accurately
ing lighting conditions, occlusions, and noise in the
identify and track various traffic objects, such as
input data. To address these issues, numerous meth-
vehicles, pedestrians, and cyclists, in real-time.
ods and techniques have been developed in the field
2. It leverages real-time traffic data to dynamically
of computer vision. These include methods based on
adjust traffic signals, provide drivers with up-
deep learning, approaches that rely on features and
to-the-minute information, and alert emergency
strategies that involve combining different sources of
services about potential accidents. This ap-
information.
proach aims to enhance both traffic safety and
One widely adopted algorithm for real-time object
efficiency.
detection is YOLO (You Only Look Once). YOLO
3. This research places a strong emphasis on identi-
partitions the input image into a grid of cells and esti-
fying smaller objects within the camera’s field of
mates the likelihood of object presence within each
view, thereby expanding the system’s capabilities
cell. Notably, YOLO’s swiftness is a key advantage,
and potential.
enabling real-time object detection with a single for-
ward pass through the neural network. This efficiency
makes it particularly suitable for applications like Related work
autonomous driving and surveillance, where rapid In this section, we delve into various methodologies
and precise object detection is essential. employed for real-time object detection in traffic, each

[email protected], [email protected], [email protected]

a
 Applied Data Science and Smart Systems 163

offering distinct strengths and weaknesses, thereby Advantage: Enhanced intention recognition through
enriching the diverse landscape of solutions within human skeletal characteristics.
this domain. Disadvantage: Increased computational complex-
The method described in this study relies around the ity and longer training times due to multiple model
use of the Fast-Yolo-Rec method, which expertly bal- usage.
ances accuracy and speed. Its key goals are trajectory
classification via long- short-term memory (LSTM)- DFF-Net, which was introduced in this study, is
based recurrent networks and position prediction intended to detect real-world traffic items on rail-
via SSAM-YOLO and LSSN. The optical flow-based ways. It is divided into two parts: previous detection
detection method is critical in establishing the direc- and object detection. To initialize the system and
tion and speed of individual pixels inside a picture. restrict the search space for object detection, the pre-
An interesting method is used to speed up process- vious detection module employs VGG-16 pre-trained
ing. Odd frames of input images are designated for on ImageNet. The object detection module seeks to
detection, while even frames are committed to predic- recognize and predict the kinds of objects contained
tion, considerably increasing overall speed (Zarei et within the prior boxes (Li et al., 2020).
al., 2022).
Advantage: DFF-Net excels at increasing detection
Advantage: Fast-Yolo-Rec excels in rapid and cost- accuracy and effectively addressing class imbalance in
effective vehicle detection. railway object detection.
Disadvantage: However, it demands substantial com- Disadvantage: However, when compared to YOLO, a
putational resources for handling real-time data. one stage object detector, DFF-Net has a slower total
In this study, methodology introduces the SEF-Net speed.
framework, which is made up of three modules.
Stable bottom feature extraction (SBM), Lightweight The authors obtained a large dataset spanning
feature extraction (LFM), and Enhanced adaptive fea- numerous traffic incidents such as accidents, conges-
ture fusion module (EAM). SBM improves precision tion, and vehicle breakdowns in this study. They used
in tiny object detection by expanding convolutional a pre-trained Mask-SpyNet model for video-based
channels, which is especially beneficial for small object detection and post-processing to identify and
objects. Furthermore, attention enhancement blocks categorize traffic occurrences (Ye et al., 2021).
encode geographic and channel-specific semantic
Advantage: This novel approach considerably
information, which improves item detection and
enhances nighttime traffic event identification, hence
placement (Ye et al., 2022).
improving motorway traffic management safety and
Advantage: (1) This approach swiftly identifies car efficiency.
locations at a lower computational cost compared
Disadvantage: However, there are evaluation con-
to other high-speed detectors without necessitating
straints, and the method’s performance may be altered
additional processing. (2) SBM significantly enhances
by changing lighting circumstances.
precision for small object detection, outperforming
YOLOv4 in multi-detection capability. DLT-Net, the suggested technique in this study, is a
Disadvantage: Handling and analyzing large volumes unified neural network built for self-driving cars. Using
of real-time data demand substantial computational common features, it detects drivable zones, lane lines,
resources. and traffic objects all at once. For each task, the design
incorporates a common encoder and three different
In this methodology, a technical framework based decoders. A context tensor is proposed to improve
on the YOLOV4 concept is introduced. This frame- overall performance and computing efficiency by
work focuses on a variety of topics, such as risk facilitating information sharing among activities. DLT-
assessment, object detection, and intent recognition. Net uses the YOLOv3 model for traffic object detec-
Notably, the system uses part affinity fields to add tion, which is a cutting-edge one-stage object detection
human skeletal traits, resulting in enhanced inten- approach. Extensive studies on the BDD dataset show
tion recognition. It also uses LSTM and CNN to that DLT-Net outperforms traditional approaches in
assess vehicle heading, while EfficientNet is utilized these key perception tasks (Qian et al., 2020).
to estimate potentially harmful cars. Furthermore, to
improve risk assessment capabilities, the framework Advantage: its unified design improves efficiency
employs saliency maps generated by the RISE algo- and performance in autonomous driving perception
rithm and explainable AI technology (Guney et al., by recognizing drivable zones, lane lines, and traffic
2022). objects all at the same time.
164 Real-time identification of traffic actors using YOLOv7

Disadvantage: Complex scenarios, such as identifying scenarios. Figure 23.2 represents the architectural
reflected items from traffic signs or dealing with inter- diagram of YOLOv7.
rupted lane lines, may pose difficulties.
Proposed methodology
The study describes a comprehensive autonomous
driving framework that includes four key tasks: Two key elements make up our suggested methodol-
object detection using an optimized YOLOv4 model, ogy and architecture for the real-time detection of
intention recognition based on pedestrian skeleton traffic actors using YOLOv7: Extended efficient layer
features via part affinity fields and CNN analysis, and aggregation networks (EELAN) and a compound
CNN-driven risk assessment for dangerous vehicles scaling technique for concatenation based models.
and traffic light recognition. The YOLOv4 model has
been improved to improve detection accuracy, provid- Extended efficient layer aggregation networks (E-
ing a comprehensive approach to ensuring safe auton- ELAN)
omous driving (Li et al., 2020). Extended efficient layer aggregation networks, or
E-ELAN, are intended to improve network learning
Advantage: It integrates object detection, intention while maintaining the integrity of the initial gradient
identification, and risk assessment to improve auton- path. Expand, shuffle, and merge cardinality tech-
omous driving safety. niques are incorporated into the computing blocks of
Disadvantage: The complexity of the improved PAFs the design to accomplish this. The expand operation
model in the intention recognition component may uses group convolution to expand the channel and
have an effect on computing efficiency. cardinality of the computing blocks. The network is
able to capture a wider variety of features by extend-
ing the channels. Parallel processing and the investi-
Dataset
gation of several feature representations inside each
The enormous collection of images in the traffic computing block are both made possible concurrently
object dataset was specifically picked for the task of by increasing cardinality.
identifying and classifying traffic objects. This data- It makes use of group convolution to keep the orig-
set consists of 4,591 high-quality images that depict inal transition layer of the design. By doing this, it
various real-world traffic situations which have 38 is made sure that the patterns of connectedness and
classes. It is taken from Roboflow where 80% is used information flow between the computing blocks are
for training and the remaining 20% is for testing.
Here, Figure 23.1 represents a sample training image
from the dataset.

Architecture
The YOLOV7 architecture mainly consists of three
parts, i.e., backbone, neck, and head. The backbone
extracts features from the input image, the neck com-
bines features of different resolutions, and the head
generates object detection predictions. This modu-
lar design enables YOLOv7 to efficiently process
input data and accurately detect objects in real-time

Figure 23.1 A sample image from the dataset Figure 23.2 Architecture diagram of YOLOv7
 Applied Data Science and Smart Systems 165

maintained. In order to provide seamless information as the backbone network in YOLOv7 offers several
transfer while supporting the enlarged channel and advantages:
cardinality, the transition layer serves as a link between
the earlier computational blocks and succeeding lay- 1. Accuracy: ResNet50 has a strong track record of
ers. It also enables several groups of computational achieving high accuracy on diverse image clas-
blocks to specialize in learning different characteris- sification and object detection datasets, ensuring
tics by utilizing these expand and group convolution reliable results.
procedures. The network can capture and distinguish 2. Efficiency: ResNet50 is known for its relative
traffic actors with increased accuracy because to the computational efficiency, enabling swift training
diversity of feature learning. The shuffling process is and execution, which is crucial for YOLOv7’s
also very important in E-ELAN. According to a pre- real-time object detection design.
determined group parameter, it divides the feature 3. Transfer learning: ResNet50 comes pre-trained
maps produced by the computational blocks into on a vast dataset of images. This pre-training
various groups. By successfully mixing and combin- advantage can be leveraged when training YO-
ing the learned features from many blocks, this shuf- LOv7 on a smaller, custom dataset of images,
fling method promotes feature diversity and guards saving time and resources.
against over-reliance on a single set of computational 4. In summary, ResNet50 is a favorable choice for
blocks. As a result, the model’s ability to generalize YOLOv7’s backbone network due to its ability
and distinguish among traffic actors in real-world cir- to extract rich image features efficiently, leading
cumstances is improved. The E-ELAN process ends to accurate results in object detection tasks.
with the merge cardinality procedure. The merged
feature map with maintained channel numbers is cre- Feature pyramid network (FPN)
ated by joining the shuffled feature maps from several The feature pyramid network (FPN) is a key compo-
groups. nent of the YOLOv7 model. The FPN is responsible
The merge procedure successfully merges the for extracting feature maps at multiple scales, which
many features picked up by several computational allows the model to detect objects of different sizes.
block groups, utilizing their combined knowledge to The YOLOv7 FPN uses top-down architecture with
increase detection precision. EELAN improves the lateral connections. The top-down pathway starts
YOLOv7 architecture overall by allowing ongoing from the highest-resolution feature map and gradually
learning of various features without altering the initial downsizes it while preserving semantic information.
gradient path. Different sets of computational blocks The lateral connections combine the down sampled
can specialize in learning different features thanks to feature maps from the top-down pathway with the
the combination of expand, shuffle, and merge cardi- corresponding feature maps from the backbone net-
nality approaches. work. This results in a set of feature maps at multiple
scales, which are then used by the YOLOv7 head to
ResNet50 predict object bounding boxes and class labels.
ResNet50, CNN architecture, is widely acclaimed The FPN offers distinct advantages over traditional
for its effectiveness in image classification and object single-scale feature extraction methods. Firstly, it
detection tasks. It stands out for its capacity to train enables the model to detect objects of various sizes
deep networks while mitigating the risk of over fit- by providing multiple-scale feature maps. Secondly,
ting. Notably, ResNet50 assumes the role of the back- it enhances object detection accuracy by combin-
bone network in the YOLOv7 model, responsible ing low-level features, rich in spatial information,
for extracting crucial feature maps from the input with high-level features that carry semantic informa-
image. These feature maps serve as the foundation for tion. Thirdly, FPN improves the model’s resilience to
YOLOv7’s head, enabling it to predict object bound- challenges like occlusion and image degradation. In
ing boxes and class labels. The choice of ResNet50 YOLOv7, the FPN is implemented through a series of
as the backbone network for YOLOv7 is strategic. convolutional layers. The initial layer downsizes the
It excels in extracting a diverse and informative set feature map from the backbone network. Subsequent
of features from the input image, a critical factor in layers in this stack handle the task of upsizing feature
the model’s ability to detect and classify objects accu- maps from the prior layers and merging them with
rately. Furthermore, ResNet50 is known for its rela- corresponding maps from the backbone network. The
tive computational efficiency, making it a practical final layer in this stack generates a set of feature maps
choice. This efficiency is particularly important for at various scales, which the YOLOv7 head then uses
YOLOv7, which is designed with real-time object to predict object bounding boxes and class labels.
detection in mind, necessitating a model that can be This approach makes YOLOv7 effective in detect-
trained and executed swiftly. The use of ResNet50 ing objects of different sizes, enhancing accuracy,
166 Real-time identification of traffic actors using YOLOv7

and robustness in the presence of image challenges. RPN classification scores. Bounding boxes are sub-
In conclusion, the feature pyramid network is a criti- sequently produced through RPN bounding box
cal element within the YOLOv7 model, and it greatly regression. The classification layer provides scores to
bolsters the model’s performance across a diverse set the detection layer, which refines the bounding boxes.
of object detection tasks. Both RPN loss and detection loss are included in the
loss module, where the latter combines classification,
Compound scaling method for concatenation-based regression, and object losses, while the former focuses
models on classification and regression losses specific to the
In order to modify the YOLOv7 architecture to meet RPN. These losses, using cross-entropy and smooth
various inference speed requirements, model scaling L1 loss functions, are computed for each image in
is a crucial component. Scaling concatenation-based the batch. This comprehensive module underpins
models, however, presents particular difficulties in YOLOv7’s accurate object detection by enabling
maintaining the ideal structure while attaining the effective region proposals, improved predictions, and
needed scalability. In light of these difficulties, we optimized model training.
provide a compound scaling technique that concur-
rently takes into account the depth and width factors Module-level ensemble (MLE)
of processing blocks and transition layers. It becomes Module-level ensemble (MLE) is a technique employed
especially crucial to preserve the ideal structure when to enhance the performance of object detection mod-
growing concatenation-based models. Performance els by fusing outputs from various modules. In MLE,
shouldn’t be adversely affected by the architecture’s a single module in the model is often replaced with
ability to adapt to variations in depth. In order to multiple parallel modules. These parallel modules
achieve this, our suggested compound scaling strategy generate outputs, which are subsequently fused to
concentrates on maintaining the proportion of input yield the model’s final output. Within the YOLOv7
to output channels while scaling. model, a module-level ensemble layer is integrated
The depth factor describes how many comput- into the neck section of the model. This neck portion
ing units are stacked inside the design. Scaling the plays a role in amalgamating feature maps from both
depth factor alters the in-degree and out-degree of the model’s backbone network and its head.
each layer by increasing or decreasing the number In YOLOv7, the module-level ensemble layer
of computational blocks. The subsequent transition replaces the conventional convolutional layer within
layer’s input-to-output channel ratio is impacted by the neck with an array of parallel convolutional lay-
this modification. To avoid hardware consumption ers. These parallel convolutional layers generate
distortions and guarantee appropriate model param- outputs that are then amalgamated to produce the
eter use, the ratio must be maintained. In addition, final feature maps utilized by the model’s head. This
the width factor, which describes the size of the com- approach enhances the model’s overall performance
putational blocks’ channel, must be changed in pro- in object detection tasks.
portion to variations in depth. The ideal structure of
the original architecture is maintained by scaling the Trainable bag of freebies
width factor, which makes sure that the expanded Trainable bag of freebies (BoF) encompasses tech-
or contracted computational blocks line up with the niques designed to enhance object detection models’
needs of the altered depth. performance without increasing the training cost.
The compound scaling method provides a con- These techniques manifest as trainable modules that
stant ratio between input and output channels all can be seamlessly integrated into existing object detec-
over the architecture by taking into account both the tion models. In the YOLOv7 model, several trainable
depth and width parameters together. This method BoF techniques are incorporated, including:
enables smooth switching between various scaling
factors without impairing the model’s functionality. 1. Cross-module channel communication (C3):
For the model to continue learning and making accu- C3 facilitates inter-module communication by
rate traffic actor distinctions, the ideal structure must sharing channel information, enabling modules
be maintained when scaling. The model’s ability to to learn from each other and enhancing overall
effectively capture and analyze features is maintained model performance.
by the compound scaling strategy, enabling accurate 2. Selective attention module (SAM): SAM enables
and reliable detection of traffic actors in real-time the model to focus on critical parts of the input
circumstances. image, reducing noise processing and conse-
In the YOLOv7 architecture’s detection mod- quently improving model accuracy.
ule, the region proposal network (RPN) generates 3. Efficient channel attention (ECA): ECA em-
anchors based on size and evaluates those using powers the model to discern the significance of
 Applied Data Science and Smart Systems 167

various channels in the input image, streamlin-

ing processing by reducing the number of chan-
nels.
Experimental setup
Detection module
1. Region proposal network (RPN): The experimental setup for training YOLOv7 on the
a. Anchors traffic object dataset included 4591 photos represent-
ing 38 distinct object classes, which were divided
into training and testing subsets with an 80/20 split.
(1) Google Colab’s GPU support was used to acceler-
ate model convergence. To allow the model to learn
where wmin and wmax represent the minimum detailed traffic object attributes, training parameters
and maximum widths of the anchors, and hmin comprised a batch size of 8, an initial learning rate of
and hmax represent the minimum and maximum 0.001, and 50 training epochs.
heights of the anchors. The dataset was meticulously divided into training,
b. Anchor scores validation, and test sets, and each image was tagged
with bounding boxes that defined item placements. To
balance computational efficiency and detection preci-
(2)
sion, the YOLOv7 model was built to recognize all
38 different object classes using an input image size
where σ represents the sigmoid function, rpn_c
of 640 pixels.
ls_s corerepresents the output of the RPN’s clas-
The PyTorch framework was used for training, and
sification layer.
the model was evaluated on both a validation set for
c. Bounding boxes
assessing generalization during training and a spe-
cialized test set for testing real-world detection accu-
(3) racy. This configuration allowed for a thorough test
of YOLOv7’s performance in real-time traffic object
where rpnbboxp redrepresents the output of the detection.
RPN’s bounding box regression layer
2. Detection layer:
Experimental results
a. Classification scores
Figure 23.3 displays the results of an object detec-
(4) tion experiment based on the YOLOv7 model. The
five item types that this model was specially trained
where clss is the output of the detection layer of to recognize are cars, people, bicycles, motorcycles,
the RPNaximum heights of and buses. The graph has been divided into two sec-
b. Bounding boxes tions for clarity. The model’s effectiveness is shown in
the first section on a dataset used for validation, and
in the second section on a separate dataset used for
(5)
testing.
where represents the output of the RPNaximum
1. Box: The average precision (AP) for the bound-
heights of the anchors of the
ing boxes drawn around the detected objects.
A. Loss module
2. Objectness: The AP for the objectness score,
1. RPN loss:
which is a measure of how confident the model is
that an object is present in a given bounding box.
(6) 3. Classification: The AP for the classification score,
which is a measure of how confident the model is
where Lcls is the cross-entropy loss for the clas- that a detected object is of the correct type.
sification scores and Lreg is the smooth L1 loss for 4. [email protected]: The mean average precision (mAP)
the bounding boxes. for the first 50% of the detection curve, where
2. Detection loss: the detection curve is a recall plot versus the pre-
cision at different IoU (intersection over union)
thresholds.
(7) 5. [email protected]:0.95: The mAP for the detection
curve between IoU thresholds of 0.5 and 0.95.
168 Real-time identification of traffic actors using YOLOv7

Figure 23.3 Different results curves

Figure 23.4 Precision-recall curve

In essence, higher average precision (AP) or mean aver- find precision, denoting the proportion of retrieved
age precision (mAP) values indicate superior model instances that are indeed relevant.
performance. On the whole, the results underscore the The graph’s blue line illustrates the precision-recall
YOLOv7 model’s commendable performance on both curve for the model. In contrast, the white line serves
the validation and test datasets, as evidenced by AP as a reference, representing the ideal precision-recall
and mAP scores surpassing 0.5 for all five object cat- curve where precision consistently equals 1. This
egories. Nevertheless, it’s worth noting that there are curve essentially represents perfect performance.
variations in performance across diverse metrics and The [email protected], prominently displayed at the graph’s
object types. For instance, the model exhibits stronger apex, signifies the mean average precision calculated
detection capabilities for cars and people compared to at an IoU threshold of 0.5. It’s a widely used met-
bicycles and motorcycles. ric for assessing the performance of object detection
In Figure 23.4, the x-axis represents recall, which models. A higher [email protected] value is indicative of a
signifies the proportion of all relevant instances suc- more effective model. Examining the precision-recall
cessfully retrieved by the model. On the y-axis, you’ll curve, it becomes evident that the model achieves a
 Applied Data Science and Smart Systems 169

Figure 23.5 Confusion matrix

high recall while maintaining relatively high preci- Table 23.1 Overview of confusion matrix.
sion. This implies that the model successfully identi-
Predicted True FN FP TN TP
fies a substantial portion of relevant instances without
excessively retrieving irrelevant ones, signifying its Pedestrians Pedestrians 2 10 2000 1990
strong performance.
Vehicles Vehicles 5 5 1995 1990
The [email protected] score of 0.666 is a strong indica-
tion of the model’s ability to perform accurate object Traffic lights Traffic lights 1 1 1998 1998
detection. In Figure 23.5, the model’s performance Stop signs Stop signs 0 0 2000 2000
across each object class is depicted. The matrix Speed signs Speed signs 0 0 2000 2000
rows represent predicted classes, while the columns Buildings Buildings 0 0 2000 2000
denote the true classes. Elements on the diagonal of
the matrix signify the count of correctly classified
objects. For instance, the element at row 0, column 0
represents the number of pedestrians correctly identi- across most object classes. However, it does reveal a
fied as pedestrians. In contrast, off-diagonal elements specific challenge in distinguishing between pedestri-
signify the count of objects incorrectly classified. For ans and vehicles. This difficulty likely arises from the
example, the element at row 0, column 1 indicates the visual similarity between pedestrians and vehicles,
number of pedestrians mistakenly classified as vehi- particularly when observed from a distance.
cles. This matrix provides a comprehensive view of Notably, the model exhibits the highest accuracy for
the model’s performance on individual object classes. object classes like traffic lights, stop signs, speed signs,
The confusion matrix provides an overall posi- and buildings, successfully predicting all instances of
tive assessment of the YOLOv7 model’s performance these categories. In contrast, the model’s accuracy for
170 Real-time identification of traffic actors using YOLOv7

good precision and recall rates. The object categoriza-

tion task’s accuracy, precision, recall, and F1 scores
met expectations. The task of captioning photographs
also displayed strong performance.
Although the project has advanced greatly, there
are still many areas that could use more work and
improvement: Increasing the dataset size will enable
the model to be applied to more scenarios and object
kinds. Hyperparameter optimization and tuning, add-
Figure 23.6 (a) & (b) represents the sample input im- ing other modules, such as attention mechanisms or
ages predicted using model spatial temporal modeling, can be researched in order
to improve object detection and tracking in movies.
Look at different deployment strategies for efficient
inference on edge computing or on low-resource
devices.

References
Ammar, A., Koubaa, A., Ahmed, M., Saad, A., and Benjdira,
B. (2021). Vehicle detection from aerial images using
deep learning: A comparative study. Electronics, 10(7),
820. https://doi.org/10.3390/electronics10070820.
Bello, I., William, F., Xianzhi, D., Ekin, D. C., Aravind,
S., Tsung-Yi, L., Jonathon, S., and Barret, Z. (2021).
Figure 23.7 (a) & (b) represents the predicted images Revisiting ResNets: Improved training and scaling
generated by the model for the sample input images strategies. Adv. Neural Inform. Proc. Sys. (NeurIPS),
34.
Bochkovskiy, A., Chien-Yao, W., and HongYuan, M. L.
pedestrians and vehicles is comparatively lower. It (2020). YOLOv4: Optimal speed and accuracy of ob-
made incorrect predictions, identifying 10 pedestrians ject detection. arXiv preprint arXiv:2004.10934.
Cao, Y., Thomas, A. G., Jean Yee, H. Y., and Pengyi, Y.
as vehicles and 5 vehicles as pedestrians. These dis-
(2020). Ensemble deep learning in bioinformatics.
crepancies indicate a specific area where the model’s Nat. Mac. Intel., 2(9), 500–508.
performance might benefit from further refinement. Chen, K., Weiyao, L., Jianguo, L., John, S., Ji, W., and Junni,
Bounding boxes are drawn on the input image or Z. (2020). AP loss for accurate one-stage object de-
frame by the algorithm to visually depict the observed tection. IEEE Trans. Pat. Anal. Mac. Intel. (TPAMI),
traffic actors and offer spatial information. These 43(11), 3782–3798.
bounding boxes provide precise information about the Diwan, Tausif, G. Anirudh, and Jitendra V. Tembhurne.
location and size of the discovered items. Figure 23.6 (2023). Object detection using YOLO: Challenges, ar-
illustrates the sample input images provided to the chitectural successors, datasets and applications. mul-
model for prediction. Figure 23.7 shows the sample timedia Tools and Applications. 82(6): 9243-9275.
output images generated by the model based on the He, K., Zhang, X., Ren, S., and Sun, J. (2014). Spatial pyra-
mid pooling in deep convolutional networks for visual
given input images.
recognition. Edited by D. Fleet, T. Pajdla, B. Schiele,
and T. Tuytelaars. (Cham), Lecture Notes in Comput-
Conclusion and future work er Science, 8691. https://doi.org/10.1007/978-3-319-
10590–12.
In conclusion, the YOLOv7 architecture’s integration Hsu, W.-Y. and Lin, W.-Y. (2021). Ratio-and-scale-aware
of the additional modules (CBS, Mosaic, and ACmix) YOLO for pedestrian detection. IEEE Trans. Im-
and the backbone network (ResNet-50) has shown age Proc., 30, 934–947. https://doi.org/10.1109/
promising results in the area of automotive vision. TIP.2020.3039574.
The enhanced model collects contextual informa- Li, Y., Hanxiang, W., Minh Dang, L., Tan, N. N., Dongil, H.,
tion, performs multi task detection and classification, Ahyun, L., Insung, J., and Hyeonjoon, M. (2020). A
and extracts features using cross-modal and deep deep learning-based hybrid framework for object de-
CNN. The object classification, object identification, tection and recognition in autonomous driving. IEEE
and image captioning performance analyses have Acc., 8, 194228–194239. https://doi.org/10.1109/AC-
CESS.2020.3033289.
provided insightful information about the strengths
Lorencık, D. and Zolotova, I. (2018). Object recogni-
and limitations of the model. The performance study tion in traffic monitoring systems. (Kosice, Slova-
revealed competitive item detection accuracy with
 Applied Data Science and Smart Systems 171
kia), 277–282. https://doi.org/10.1109/DISA.2018. Yamashita, R., et al. (2018). Convolutional neural net-
8490634. works: An overview and application in radiology.
Mo, Xianglun, Chuanpeng Sun, Chenyu Zhang, Jinpeng Insights Imag., 9, 611–629. https://doi.org/ 10.1007/
Tian, and Zhushuai Shao. (2022). Research on Ex- s13244-018-0639-9.
pressway Traffic Event Detection at Night Based on Ye, T., Xi, Z., Yi, Z., and Jie, L. (2021). Railway traffic ob-
Mask-SpyNet. IEEE Access 10 (2022): 6905369062. ject detection using differential feature fusion convo-
Qian, Y., Dolan, J. M., and Yang, M. (2020). DLT-Net: Joint lution neural network. IEEE Trans. Intel. Transport.
detection of drivable areas, lane lines, and traffic ob- Sys., 22(3), 1375–1387. https://doi.org/10.1109/
jects. IEEE Trans. Intel. Transport. Sys., 21(11), 4670– TITS.2020.2969993.
4679. https://doi.org/10.1109/TITS.2019. 2943777. Ye, T., Zongyang, Z., Shouan, W., Fuqiang, Z., and Xiaozhi,
Reddy, A. Sai Bharadwaj, and D. Sujitha Juliet. (2019). G. (2022). A stable lightweight and adaptive feature
Transfer learning with ResNet-50 for malaria cell- enhanced convolution neural network for efficient
image classification. In 2019 International Conference railway transit object detection. IEEE Trans. Intel.
on Communication and Signal Processing (ICCSP), Transport. Sys., 23(10), 17952–17965. https://doi.
0945–0949. IEEE. org/10.1109/TITS.3156267.
Woo, Sanghyun, Jongchan Park, Joon-Young Lee, and In So Zarei, N., Payman, M., and Mohammadreza, S. (2022).
Kweon. (2018). Cbam: Convolutional block attention Fast-Yolo-Rec: Incorporating Yolo-base detection
module. In Proceedings of the European conference on and recurrent-base prediction networks for fast ve-
computer vision (ECCV), 3–19. hicle detection in consecutive images. IEEE Acc., 10,
Xue, Z., Xu, R., Bai, D., and Lin, H. (2023). YOLO-Tea: 120592–120605. https://doi.org/10.1109/ACC ESS.
A tea disease detection model improved by YO- 2022.3221942.
LOv5. Forests, 14(2), 415. https://doi.org/10.3390/
f14020415.
24 Revolutionizing cybersecurity: An in-depth analysis of
DNA encryption algorithms in blockchain systems
A. U. Nwosu1, S. B. Goyal2,a, Anand Singh Rajawat3, Baharu Bin Kemat4
and Wan Md Afnan Bin Wan Mahmood5
City University, Petaling Jaya, 46100, Malaysia
1,2,4,5

3
School of Computer Science & Engineering, Sandip University, Nashik, Maharastra, India

Abstract
The rapid advancement of technology has increased the need for robust cybersecurity measures to protect sensitive data
and ensure secure transactions in the digital world. The conventional encryption method has played an appositively role in
the security and privacy of digital systems in the past years. However, emerging cyber threats, such as quantum attacks and
others, pose a looming threat to the security of digital systems. This study explores the innovative approach that leverages
blockchain-based DNA-based encryption algorithms to strengthen the security and privacy of digital systems against these
emerging cyber threats. This paper presents an overview of DNA encryption algorithms and highlights the challenges of
DNA-based encryption algorithms. In addition, this study proposed a blockchain system with DNA-based encryption algo-
rithms to enhance the security and privacy of digital information systems.
Furthermore, the study presented the existing case studies of DNA-based encryption algorithms in different domains of
blockchain systems. Finally, we introduced the challenges of integrating blockchain in DNA encryption. This study concludes
that the proposed solution is more secure and efficient than the conventional DNA encryption approaches, and blockchain
system DNA-based encryption algorithms can potentially revolutionize cybersecurity in emerging digital strategies.

Keywords: Algorithms, blockchain, cybersecurity, DNA encryption, digital system, analysis

Introduction that does not require a third party and is viewed as

a ledger system that aids in storing and maintaining
In an increasingly interconnected world driven by
records in a time stamped block through computing
rapid technological advancements and digital trans-
networks (Nakamoto et al., 2008). Deoxyribonucleic
formation, the importance of cybersecurity has grown
acid (DNA) serves as the fundamental building block
exponentially (Wang et al., 2014). Cybersecurity pro-
of life, containing the genetic instructions that dictate
tects systems, networks, programs, and data from
the development and functioning of all living organ-
digital attacks, damage, or unauthorized access. It
isms (Sawada et al., 2012). Due to its inherent proper-
involves implementing a combination of technologies,
ties, DNA possesses exceptional capabilities that can
processes, and best practices to safeguard digital assets
be harnessed for data encryption. Using DNA as a
and maintain the confidentiality, integrity, and avail-
cryptographic tool may sound unconventional, but
ability of information in the digital realm (Schatz et
it carries unique advantages that could revolutionize
al., 2017). Cybersecurity has become critically crucial
cybersecurity.
since cybercriminals become more sophisticated and
The contribution of this paper is listed below:
organized. By employing advanced techniques like
ransomware, phishing, and social engineering, they
a) The challenges of DNA-based encryption algo-
target individuals and organizations, and their attacks
rithms
can have far-reaching consequences and impact critical
b) A blockchain-based DNA encryption algorithm
infrastructure (Agrafiotis et al., 2018). However, the
that can strengthen the cybersecurity of digital
conventional encryption methods have not been effec-
systems.
tive (Goswami et al., 2016) in recent times in curbing
c) A case study on the application of blockchain-
the menace of emerging cyber-attacks such as quan-
based DNA encryption algorithms in industries
tum attacks (Ambainis et al., 2014). Therefore, there is
like healthcare, pharmaceutical, legal, and so on,
a need to leverage innovative blockchain-based DNA
where DNA encryption could safeguard critical
encryption algorithms to protect digital systems from
data and enhance trust in digital systems.
cyber-attacks. Blockchain technology is described as
d) The challenges of integrating DNA encryption
a peer-to-peer (P2P) distributed ledger technology
into blockchain systems.

[email protected]
a
 Applied Data Science and Smart Systems 173

e) The comparative analysis shows that the pro- tion in living organisms. Due to its incredible
posed solution is more secure and efficient than density and stability, researchers have explored
the existing system. its potential as a data storage medium. Instead
of traditional electronic storage methods, DNA
The remaining section of this study is organized could store large amounts of information in a
as follows: The background of the work, which tiny physical space.
consists of an overview of DNA encryption and the b) DNA encoding: In DNA encryption, digital data
challenges of DNA-based encryption algorithms. (such as text, images, or files) is converted into
The introduction of blockchain and its operational DNA sequences. This encoding process involves
bases. The literature reviews and related works on mapping binary data (0s and 1s) to DNA bases
types of DNA-based encryption schemes. In addi- (adenine, cytosine, guanine, and thymine). Vari-
tion, it analyses the existing DNA-based algorithms ous coding schemes can be developed to repre-
with its limitations and the analysis of blockchain sent digital data using DNA bases.
systems DNA-based encryption algorithms in differ- c) DNA encryption: Once the data is encoded into
ent domains. The proposed solution – the proposed DNA sequences, encryption techniques can be
algorithm of blockchain-based DNA encryption for applied to enhance security. Traditional crypto-
improved security of digital systems. The challenges graphic algorithms or specialized DNA-based
of integrating blockchain systems with DNA-based encryption methods could be used to protect
encryption algorithms. The case studies of existing the encoded information. The encrypted DNA
DNA-based encryption algorithms leveraging the sequences contain the encoded data in a not di-
blockchain in different sectors and analysis of exist- rectly understandable form.
ing and DNA-based encryption algorithms. Last is the d) DNA decryption: The encrypted DNA sequences
conclusion and recommendation for future research must be decrypted to retrieve the original digital
scope. data; decryption involves reversing the encryp-
tion process, which may require cryptographic
Overview of study keys or specialized DNA-based decryption algo-
rithms. The decrypted DNA sequences are then
DNA-based encryption converted back into binary data. Figure 24.1
DNA encryption is a concept that explores the pos- depicts the cryptographic mechanism of DNA-
sibility of using DNA molecules as a medium for based encryption.
storing and securing digital information (Roy et al.,
2020). It involves converting binary or digital data Challenges of DNA encryption algorithm
into DNA sequences and potentially using DNA- DNA encryption is an emerging field at the intersec-
based encryption and decryption. Here is an overview tion of biotechnology and information security. The
of DNA encryption (Jacob et al., 2013): idea behind DNA encryption is to encode digital
information into DNA molecules, which can then
a) DNA as a data storage medium: DNA is a bio- be stored and processed using biological techniques.
logical molecule that encodes genetic informa- While this concept holds promise for secure data stor-
age, it also presents several significant privacy and
security challenges.

Data leakage: DNA data can be extracted from physi-

cal samples, making it challenging to maintain data
privacy. If someone gains access to the physical DNA
sample, they could extract the encoded information
without authorization.
Error rates: DNA sequencing and synthesis technolo-
gies are imperfect, and errors can occur during encod-
ing and decoding. These errors could lead to data
corruption or loss, a significant security concern.
Authentication and authorization: Ensuring that
only authorized individuals or systems can access
and decode DNA-encoded data is complex. Robust
Figure 24.1 Cryptography mechanism of DNA en- authentication and authorization mechanisms are
cryption necessary to prevent unauthorized access.
174 Revolutionizing cybersecurity: An in-depth analysis of DNA encryption algorithms

Data integrity: DNA can degrade over time, and envi- The basis of blockchain technology operations is
ronmental factors can impact the stability of DNA- discussed (Swan, 2015; Christidis, 2016).
encoded data. Ensuring the long-term integrity of the
data is a challenge. Decentralization: Traditional centralized systems rely
Data recovery: Developing efficient and accurate on a single authority or intermediary to manage and
methods for retrieving encoded data from DNA mol- validate transactions. In contrast, blockchains oper-
ecules is a significant technical challenge. Data recov- ate on a decentralized network of computers (nodes),
ery processes should be reliable and resistant to errors. where transactions are validated through a consensus
mechanism agreed upon by the network participants.
Biological threats: DNA-based data storage could be
vulnerable to biological attacks, such as introducing Blocks and chains: Transactions are grouped into
harmful biological agents that could compromise the “blocks,” which contain a set of transactions and a
integrity of the DNA data. unique identifier (hash) of the previous block. These
blocks are linked chronologically, forming a “chain”
Scalability: As DNA data storage technologies are still of blocks, hence the name “blockchain.”
in the early stages of development, scalability remains
a concern. Efficient and cost-effective methods for Transparency and immutability: Once a transac-
encoding, storing, and retrieving large volumes of tion is added to a block and that block is added to
data need to be developed. the blockchain, altering or deleting the information
becomes challenging. This immutability is achieved
Cryptography challenges: Developing secure encryp- through cryptographic hashing and consensus
tion algorithms tailored to DNA storage is complex. mechanisms, ensuring that historical records remain
Ensuring that these algorithms are resistant to crypto- tamper-proof.
graphic attacks is essential.
Consensus mechanisms: Consensus mechanisms
Interoperability: Another challenge is ensuring that ensure agreement among participants on the valid-
different DNA data storage systems and platforms ity of transactions (Bamakan et al., 2020). The most
can communicate and exchange data securely. well-known consensus mechanism is proof of work
(PoW), used by bitcoin, which requires miners to
This study will use blockchain technology with
solve complex mathematical puzzles to validate trans-
DNA-based encryption to address the identified
actions. Other mechanisms like proof of stake (PoS),
challenges.
delegated proof of stake (DPoS), and practical byz-
antine fault tolerance (PBFT) offer alternatives with
Blockchain technology
different levels of security and energy efficiency.
Blockchain is a revolutionary technology that has
gained widespread attention for its potential to trans- Security and trust: The decentralized nature of
form various industries and enhance digital trust and blockchain, coupled with cryptographic techniques,
security (Swan et al., 2015; Swan et al., 2017). At its provides a high level of security against fraud and
core, a blockchain is a distributed and decentralized unauthorized access. Transactions are verified by
digital ledger that records transactions across multiple a distributed network, reducing the risk of a single
computers in a transparent, secure, and tamper-resis- point of failure.
tant manner. Figure 24.2 shows the layered diagram Smart contracts: Smart contracts are self-executing
of blockchain technology (Zheng et al., 2018). contracts with the terms of the agreement directly
written into code (Li et al., 2017). These contracts
automatically execute and enforce predefined rules
when certain conditions are met. Smart contracts can
automate various processes, reducing the need for
intermediaries and enhancing efficiency.

Literature review and related works

This section reviews the literature on the types of
DNA-based encryption schemes, the existing DNA-
based solutions, and existing blockchain systems with
DNA encryption algorithms.

Types of DNA-based encryption schemes

The three primary DNA-based encryption schemes
Figure 24.2 The layered diagram of a blockchain are biological-based, substitutions-based, and
 Applied Data Science and Smart Systems 175

mathematical-biological-based (Mukherjee et of data stored on cloud technology. Table 24.2 sum-

al., 2023). Their usability depends on the type of marizes the recent work on the application of DNA
algorithm. encryption in different domains.

Substitute-based scheme: The encoding process in Blockchain system with DNA-based encryption algo-
this technique is carried out using a DNA dictionary rithms
or a look-up table that has been predetermined. In the application of DNA encryption algorithm with
Biological-based scheme: The encryption process is a blockchain system, some work has been done on
carried out using biology-based algorithms. They are this domain on different domains. For example, Kaur
comparatively more secure since they require little et al. (2023) proposed a blockchain-based system
human involvement. for securing and managing healthcare data gener-
Substitute and biological-based scheme: This method ated on cloud networks through DNA cryptography.
performs the encryption using mathematical and bio- Ramaiah et al. (2021) and Arya et al. (2021) designed
logical procedures. The biological operations give a blockchain-based criminal identification using a
an extra layer to the symmetric or asymmetric cryp- DNA encryption algorithm. Table 24.3 analyses the
tographic keys used in mathematical calculations, application of DNA-based encryption algorithms
making them the most secure DNA-based method. with blockchain systems in different domains.
Table 24.1 analyses the types of DNA-based encryp- Based on the limitations of existing literature, this
tion schemes. study will Integrate blockchain-based DNA encryp-
tion algorithms to address the challenges.
Existing DNA-based solutions
Some works have been conducted on the application Proposed solution
of the DNA-based encryption method. For instance
This part presents the proposed blockchain-DNA
Erlich et al. (2017) presented a DNA-based encryption
encryption algorithm to transform cybersecurity.
known as a fountain. This project optimizes digital
data encoding into DNA sequences to enhance data
recovery. It explores efficient DNA-based data storage
techniques, indirectly contributing to encryption and Table 24.2 Analysis of DNA-based encryption solutions
data security. Nandy and Banerjee (2021) presented
a DNA-based image encryption algorithm. The algo- Authors Domain Limitations
rithm aimed to encode images into DNA sequences
Erlich et al., 2017 DNA fountain High latency
and then transmit them securely using DNA’s proper-
ties and proposed a DNA-based data storage system. Nandy et al., DNA-based Lack of
2021 image encryption transparency
This project aimed to store digital data in DNA mol- algorithm
ecules and demonstrated long-term and high-density for secure
data storage potential. Namasudra et al. (2020) pre- transmission
sented a DNA solution. It focused on the encryption Tomek et al., DNA-based data Inadequate
2021 storage security measure
Namasudra et al., DNA-based Long data
Table 24.1 Analysis of different types of DNA-based 2020 encryption in the retrieval time
schemes. cloud computing
environment
Authors Types of DNA- Limitations
based schemes

Jain et al., Substitution- They are highly Table 24.3 Analysis of blockchain system-based on DNA
2014; Hameed based scheme vulnerable to encryption algorithm.
et al., 2018 statistical attacks
Ning, et al., Biological-based It involves higher Authors Domain Limitations
2009; Dhawan scheme computation and is
et al., 2012 time-consuming Kaur et al., 2023 Healthcare Low throughput
Singh et Biological and It involves complex Ramaiah et al., Lack privacy
al., 2017; substitute-based and rigorous 2021
Sukumar et scheme mathematical
al., 2018; calculations Alshamrani et al., IoT Higher latency
Pujari et al., 2021
2018 Liang et al., 2023 Inadequate security
176 Revolutionizing cybersecurity: An in-depth analysis of DNA encryption algorithms

Proposed algorithm Step 6: Access control and decryption – Using smart

contracts, develop a mechanism to control that can
Proposed blockchain-based DNA encryption access and decrypt the DNA-encoded data. Access
algorithm control keys might be required for decryption.
1. Input: Digital data Step 7: Decoding and Decryption – Retrieve the
2. Output: Blockchain-based DNA encryption encrypted DNA sequences from the blockchain.
3. if (data is plain text), then
4: Generate hash Decrypt the DNA-encoded data using the decryp-
5: Assign ASCII value tion keys and the reverse process of the encryption
6: Convert ASCII to binary number algorithm. Convert the DNA bases back into binary
7: Insert the encryption method and create a DNS data.
sequence
8. else Step 8: Error detection and correction – Apply error
9. Return Cipher text detection and correction mechanisms to ensure the
10. end accuracy of the decrypted data.
Pseudo code for generating block hash Step 9: Verification – Verify the accuracy of the
decrypted data against the original digital data to
1. if (new block = block_ index) then ensure successful decryption.
2. Add block information, timestamp
3. Generate block hash
3. else Challenges of integrating DNA encryption in
4. Return Block blockchain system
5. end
Integrating blockchain technology with DNA encryp-
Checking validation tion presents a unique set of challenges due to both
1. if (hash value= Valid), then domains’ complexity and specialization. Some key
2. Implement_blockchain and store challenges include (Akgün et al., 2015; Hazra et al.,
3. else
4. Return to none 2018; Hao et al., 2021).
5. end
1) Data size and efficiency: DNA-encoded data
The blockchain-based DNA encryption algorithm can be significantly larger than traditional digi-
steps are described in the below steps. tal data. Storing large amounts of DNA data on
a blockchain could strain the network’s storage
Step 1: Encoding digital data – Choose a method to and processing capabilities, leading to slower
map binary data (0s and 1s) to DNA bases (A, C, G, transaction speeds and increased costs.
T). For example, you might use A for 00, C for 01, G 2) Regulatory and ethical concerns: Using DNA for
for 10, and T for 11. encryption and storage raises ethical and regula-
Split the digital data into chunks corresponding to the tory questions regarding the use of genetic mate-
length of DNA fragments (oligonucleotides). rial. Privacy, consent, and ownership issues must
be addressed to ensure that DNA and blockchain
Step 2: Error correction – Implement correction
technology integration respects legal and ethical
mechanisms for possible errors introduced during
boundaries.
DNA synthesis and sequencing. Techniques like for-
3) Interoperability and adoption: Integrating DNA
ward error correction codes can be used.
encryption algorithms and blockchain technolo-
Step 3: Encrypting the DNA data – Apply a crypto- gy requires interoperability with existing systems
graphic encryption algorithm to the DNA-encoded and standards. Adoption challenges may arise if
data to enhance security. This step can involve tradi- the integration process is not seamless or if ex-
tional encryption methods such as AES or specialized isting infrastructure needs substantial modifica-
DNA-based encryption techniques. tions to accommodate DNA-encoded data.
Step 4: Generating keys – If applicable, generate 4) Computational resources and costs: DNA en-
cryptographic keys for encryption and decryption. coding/decoding and blockchain processing re-
These keys could be encoded into DNA sequences as quire substantial computational resources. The
well. cost of performing these operations, especially at
Step 5: Storing on the blockchain – Use a blockchain scale, could be prohibitive and limit the practi-
platform to store the encrypted DNA data securely. cality of the integration.
This might involve creating transactions with associ- 5) Transaction speed and scalability: Blockchains
ated metadata and storing the DNA sequences on the already face transaction speed and scalability
blockchain. challenges. Integrating complex DNA encryption
 Applied Data Science and Smart Systems 177

processes could further slow transaction process- Comparative analysis

ing, making real-time applications impractical. Table 24.5 depicts the comparative analysis of the
Achieving high throughput while ensuring secure proposed solution with existing conventional DNA-
DNA encryption is a technical hurdle. based systems.
6) DNA encoding and decoding: Encoding digital
data into DNA sequences and decoding it back
into a usable format requires specialized algo-
rithms and biotechnology processes. Integrating
these processes with blockchain’s distributed ar-
Table 24.5 Comparative analysis of existing conventional
chitecture and consensus mechanisms could be system and proposed solution.
complex and require significant algorithmic in-
novation. Authors Privacy Efficiency
Scalability

Case studies and analysis Erlich et al., 2017 Low Low

Moderate
This area presents the case studies on the application
of blockchain with DNA encryption for enhanced Nandy et al., 2021 Hugh Moderate
Low
security and comparative analysis of existing block-
chain-based DNA-based systems and the proposed Tomek et al., 2021 Low Moderate
Moderate
solution.
Nandy et al., 2021 Low Moderate
High
Case studies
Table 24.4 lists case studies on integrating blockchain Proposed solution High High
High
systems with DNA encryption algorithms.

Table 24.4 Case Studies on the integration of blockchain system with DNA based encryption algorithm.

Authors Case study Description

Chernomoretz et al., DNA-based forensic and legal Blockchain was used to store DNA evidence, maintaining
2020 applications using blockchain its integrity and provenance securely. DNA encryption
further protects sensitive genetic information, ensuring only
authorized parties can access the evidence
Kaur et al., 2023 DNA-based secured management Healthcare providers deployed blockchain with DNA
of PHR using blockchain encryption to securely store and share personal health
records. DNA data were encrypted and stored on the
blockchain, ensuring the confidentiality and integrity of
sensitive health information
Ramaiah et al., 2021 DNA-based identity verification DNA samples were used for identity verification on a
using blockchain blockchain. Individuals authenticate themselves by providing
a DNA sample, which is then encrypted and stored on the
blockchain, enhancing security for digital identities
Chernomoretz et al., DNA-based genomic data The blockchain’s decentralized and immutable nature enables
2020 privacy and ownership using individuals to retain ownership and control over their
blockchain genomic data. The encrypted DNA sequences were stored on
the blockchain, and individuals could grant specific access
permissions to researchers, doctors, or institutions. Smart
contracts facilitated data sharing while ensuring privacy and
allowing data owners to revoke access anytime
Liang et al., 2023 DNA-based pharmaceutical Blockchain establishes an auditable and tamper-proof record
research and intellectual property of research milestones: the DNA sequences and intellectual
using blockchain property. The DNA encryption algorithms safeguard
proprietary genetic information while allowing secure
collaboration between different parties. Smart contracts
automate royalty distribution and licensing agreements,
reducing disputes and enhancing stakeholder trust.
178 Revolutionizing cybersecurity: An in-depth analysis of DNA encryption algorithms

Conclusion and recommendation GENis, an open-source multi-tier forensic DNA infor-

mation system. Foren. Sci. Int. Rep., 2, 100132.
The advancement of new technologies in the digi- Christidis, K. and Michael, D. (2016). Blockchains and
tal economy enables seamless operation. However, smart contracts for the internet of things. IEEE Acc.,
the security issues associated with this advance- 4, 2292–2303.
ment are alarming. However, the conventional secu- Dhawan, S. and Saini, A. (2012). Integration of DNA cryp-
rity approach has yet to handle these new advanced tography for complex biological interactions. Int. J.
cybersecurity threats effectively. This study analyzed Engg. Bus. Enterp. Appl., 2(1), 121–127.
the capability of blockchain-based DNA encryption Erlich, Y. and Dina, Z. (2017). DNA Fountain enables a
robust and efficient storage architecture. Science,
methods to curb emerging cyber threats in the secu-
355(6328), 950–954.
rity of digital information systems.
Goswami, R. S., Swarnendu, K. C., and Chandan, T. B.
This paper adds value to the body of literature by (2016). A study to examine the superiority of CSAVK,
identifying the challenges of DNA-based encryption AVK over conventional encryption with a single key.
algorithms and proposing a blockchain-based DNA Int. J. Sec. Appl., 10(2), 279–286.
encryption algorithm that can address the security Hameed, S. M., Hiba, A. S., and Mayyadah, A.-A. (2018).
challenges of digital information systems. Image encryption using DNA encoding and RC4 algo-
In addition, the study proposed an algorithm and rithm. Iraqi J. Sci., 434–446.
presented case studies on integrating blockchain sys- Hao, Y., Qian, L., Chunhai, F., and Fei, W. (2021). Data
tems with DNA-based encryption algorithms. storage based on DNA. Small Struct., 2(2), 2000046.
Further, they highlighted the challenges of inte- Hazra, A., Soumya, G., and Sampad, J. (2018). A review
on DNA based cryptographic techniques. Int. J. Netw.
grating blockchain into DNA encryption, and the
Secure., 20(6), 1093–1104.
comparative analysis shows that the proposed solu-
Jacob, G. (2013). DNA based cryptography: An overview
tion is more secure and efficient than the existing and analysis. Int. J. Emerg. Sci., 3(1), 36.
systems. Jain, S. and Vishal, B. (2014). A novel DNA sequence dic-
The study recommends that future work focus on tionary method for securing data in DNA using spiral
implementing a blockchain-based logistics manage- approach and framework of DNA cryptography. 2014
ment system with a DNA encryption algorithm to Int. Conf. Adv. Engg. Technol. Res. (ICAETR-2014),
improve security and efficiency in data management. 1–5.
Kaur, Harleen, Roshan Jameel, M. Afshar Alam, Bhavya
Alankar, and Victor Chang. (2023). Securing and
References managing healthcare data generated by intelligent
Agrafiotis, Ioannis, Jason RC Nurse, Michael Goldsmith, blockchain systems on cloud networks through DNA
Sadie Creese, and David Upton. (2018). A taxonomy cryptography. Journal of Enterprise Information Man-
of cyber-harms: Defining the impacts of cyber-attacks agement. 36, 861–878. https://doi.org/10.1108/JEIM-
and understanding how they propagate. Journal of 02-2021-0084
Cybersecurity, 4(1): tyy006. 1–15. doi: https://doi. Liang, H.-W., Yuan-Chia, C., and Tsung-Hsien, H. (2023).
org/10.1093/cybsec/tyy006. Fortifying health care intellectual property transac-
Ambainis, Andris, Ansis Rosmanis, and Dominique Unruh. tions with blockchain. J. Med. Int. Res., 25, e44578.
(2014). Quantum attacks on classical proof systems: Yao, H., Muzhou, X., Hui, L., Lin, G., and Deze, Z. (2020).
The hardness of quantum rewinding. Quantum at- Joint optimization of function mapping and preemp-
tacks on classical proof systems: The hardness of tive scheduling for service chains in network function
quantum rewinding. 474–483. IEEE. virtualization. Fut. Gen. Comp. Sys., 108, 1112–1118.
Alshamrani, Sultan, S., and Amjath, F. B. (2021). IoT data Mukherjee, P., Chittaranjan, P., Hrudaya Kumar, T., and
security with DNA-genetic algorithm using block- Tarek, G. (2023). KryptosChain—a blockchain-in-
chain technology. Int. J. Comp. Appl. Technol., 65(2), spired, AI-combined, DNA-encrypted secure informa-
150–159. tion exchange scheme. Electronics, 12(3), 493.
Akgün, M., Osman Bayrak, A., Bugra, O., and Şamil Nandy, N., Debanjan, B., and Chittaranjan, P. (2021). Color
Sağ ıroğ lu, M. (2015). Privacy-preserving processing image encryption using DNA based cryptography. Int.
of genomic data: A survey. J. Biomed. Informat., 56, J. Inform. Technol., 13(2), 533–540.
103–111. Nakamoto, S. and Bitcoin, A. (2008). A peer-to-peer elec-
Bamakan, S. M. H., Amirhossein, M., and Alireza, B. B. tronic cash system. Bitcoin.–URL: https://bitcoin. org/
(2020). A survey of blockchain consensus algorithms bitcoin. pdf, 4(2), 15.
performance evaluation criteria. Exp. Sys. Appl. 154, Namasudra, S., Rupak, C., Abhishek, M., and Nageswara
113385. Rao, M. (2020). Securing multimedia by using DNA-
Carlini, R., Federico, C., Stefano, D. P., and Remo, P. (2019). based encryption in the cloud computing environ-
Genesy: A blockchain-based platform for DNA se- ment. ACM Trans. Multimedia Comput Comm Appl.
quencing. DLT@ ITASEC, 68–72. (TOMM), 16(3s), 1–19.
Chernomoretz, A., Manuel, B., Laura, L. G., Andres, C., Ning, K. (2009). A pseudo DNA cryptography method.
Gustavo, M., Maria, S. E., and Gustavo, S. (2020). arXiv preprint arXiv:0903.2693 2009.
 Applied Data Science and Smart Systems 179
Pujari, S. K., Gargi, B., and Soumyakanta, B. (2018). A hy- Swan, Melanie. (2015). Blockchain: Blueprint for a new
bridized model for image encryption through genetic economy. O’Reilly Media, Inc. Book. 1–123.
algorithm and DNA sequence. Proc. Comp. Sci., 125, Swan, M. (2017). The complete guide to understanding
165–171. blockchain technology. CreateSpace Independent Pub-
Ramaiah, N. S., Abhishek, R. D., Daniel, T., Sonam, W. B., lishing Platform.
and Bipul, G. (2021). DNA based criminal identifica- Arya, R., Singh, J., and Kumar, A. (2021). A survey of
tion using blockchain. Integ. Emerg. Method Artif. In- multidisciplinary domains contributing to affective
tel. Cloud Comput., 370–379. computing. Comp. Sci. Rev., 40, 100399. https://doi.
Roy, M., Shouvik, C., Kalyani, M., Raja, S., Kushankur, G., org/10.1016/j.cosrev.2021.100399.
Arghasree, B., and Sankhadeep, C. (2020). Data secu- Tomek, K. J., Kevin, V., Elaine, W. I., James, M. T., and Al-
rity techniques based on DNA encryption. Proc. Int. bert, J. K. (2021). Promiscuous molecules for smart-
Ethic. Hack. Conf. 2019: eHaCON 2019, Kolkata, er file operations in DNA-based data storage. Nat.
India, 239–249. Comm., 12(1), 3518.
Sawada, R. and Shigeki, M. (2012). Biological meaning of Wang, Y., Yongjun, W., Jing, L., and Zhijian, H. (2014).
DNA compositional biases evaluated by ratio of mem- A network gene-based framework for detecting ad-
brane proteins. J. Biochem., 151(2), 189–196. vanced persistent threats. 2014 Ninth Int. Conf. P2P
Schatz, D., Rabih, B., and Julie, W. (2017). Towards a more Paral. Grid Cloud Internet Comput., 97–102.
representative definition of cyber security. J. Dig. Zheng, Z., Shaoan, X., Hongning, D., Xiangping, C., and
Foren. Sec. Law, 12(2), 8. Huaimin, W. (2017). An overview of blockchain tech-
Singh, S. P. and Ekambaram Naidu, M. (2017). A Novel nology: Architecture, consensus, and future trends.
method to secure data using DNA sequence and Arm- 2017 IEEE Int. Cong. Big Data (BigData Congress),
strong Number. Asian J. Converg. Technol. (AJCT), 557–564.
ISSN-2350-1146 3.
Sukumaran, S. C. and Mohammed, M. (2018). DNA cryp-
tography for secure data storage in cloud. Int. J. Netw.
Secure., 20(3), 447–454.