Optimized Stress Level Classification: A

Fine-Tuning Approach Using TOPSIS

Lingaraj Nayak Ayush Patra Ashutosh Pattanaik
Computer Science & Engineering Computer Science & Engineering Computer Science & Engineering
C.V. Raman Global University C.V. Raman Global University C.V. Raman Global University
Bhubaneswar, Odisha Bhubaneswar, Odisha Bhubaneswar, Odisha
[email protected] [email protected] [email protected]

Abstract—This research explores an innovative approach to prediction systems. These models utilize historical data to
stress level classification using machine learning techniques establish behavioral patterns and predict stress levels based on
applied to environmental and physical activity data. The dataset real-time inputs [5]. The ability to continuously monitor stress
consists of features such as humidity, temperature, and step
count to predict an individual’s stress level. The methodology patterns can facilitate timely interventions, thereby improving
involves preprocessing, feature selection, normalization, decision- overall well-being. Additionally, integrating stress classifica-
making using the TOPSIS method, and classification with various tion systems into everyday devices such as smartphones and
machine learning models. A comparative study of classifiers smartwatches can enhance accessibility and usability, making
is conducted, and the Random Forest model emerges as the stress detection a seamless part of daily life [12].
best performer, offering high accuracy and reliability. The study
contributes to stress detection technologies by providing a robust This research applies machine learning techniques to clas-
and interpretable machine learning framework. sify stress levels based on environmental and activity-related
Index Terms—Stress Classification, Machine Learning, Feature parameters such as humidity, temperature, and step count. The
Selection, TOPSIS, Logistic Regression objective is to develop an efficient model that can predict stress
levels with high accuracy, providing a practical tool for contin-
I. I NTRODUCTION uous stress assessment and management [5]. By incorporating
Stress is a prevalent issue affecting individuals’ mental and the TOPSIS decision-making approach and evaluating multiple
physical health, often leading to conditions such as anxiety, classifiers, this study aims to establish an optimal framework
depression, and cardiovascular diseases [5]. Early identifica- for stress detection using readily available data sources. The
tion and intervention are crucial in mitigating its negative findings from this study can contribute to the development of
impact. Traditional stress detection methods rely on self- real-time stress monitoring applications, enabling individuals
reported surveys and physiological sensors [5], which can be to take proactive measures in managing their stress levels
intrusive and require specialized equipment. effectively.
Advancements in wearable technology and environmental This is how the rest of the paper is structured. The related
sensors have enabled the collection of real-time data on various work is summarized in Section II. Section III illustrates the
factors influencing stress levels [5]. These include physical stated methodology. The comparative analysis and experimen-
activity, ambient temperature, and humidity, which have been tal findings are covered in Section IV. Finally, the concluding
found to correlate with stress responses [6]. Leveraging ma- remarks were included in Section V.
chine learning techniques for stress classification based on
such non-invasive data sources presents a promising avenue II. R ELATED W ORK
for scalable and cost-effective stress monitoring solutions [7].
Furthermore, the integration of decision-making techniques Stress detection has been a widely studied field, with
such as the Technique for Order Preference by Similarity research spanning physiological, behavioral, and environmen-
to Ideal Solution (TOPSIS) enhances the interpretability of tal data sources [2]. Traditional approaches have primarily
stress classification results [8]. Machine learning models, when relied on physiological signals such as heart rate variability
combined with structured decision-making processes, provide (HRV), electrodermal activity (EDA), and electroencephalo-
a more comprehensive and accurate assessment of stress lev- gram (EEG) readings to assess stress level. These biomet-
els.The machine learning techniques have used to solve various ric signals are often captured through wearable devices and
diseases like breast cancer [8]–[11], brain tumor detection medical-grade sensors, enabling researchers to develop stress
[13] [15]. The ability to analyze real-time data from wearable classification models using machine learning techniques such
devices and environmental sensors allows for continuous stress as Support Vector Machines (SVM), Decision Trees, and
monitoring without the need for intrusive methods. Neural Networks [2]. While these methods have demonstrated
Recent advances in artificial intelligence and data-driven high accuracy, their reliance on specialized hardware limits
models have enabled the development of personalized stress their scalability and accessibility for broader applications.
 More recent research has explored behavioral data, includ- the most relevant features improve classification accuracy, as
ing physical activity, sleep patterns, and mobile phone usage, illustrated in Fig. 1 [3].
as alternative indicators of stress [2]. Studies have lever-
A. Dataset Description
aged data from smartphone sensors, accelerometers, and GPS
tracking to analyze movement patterns, step counts, and user The dataset comprises 2001 records with three input vari-
interactions [2]. Machine learning models such as Random ables—humidity, temperature, and step count—and one target
Forest, K-Nearest Neighbors (KNN), and deep learning-based variable, stress level (0: low, 1: moderate, 2: high) [2]. Hu-
approaches have been used to classify stress levels based on midity and temperature reflect environmental conditions, while
these behavioral patterns [2]. However, challenges remain in step count represents physical activity [6]. The dataset includes
ensuring generalizability across different populations, as stress numerical features (humidity, temperature, step count) and a
responses can vary significantly between individuals. categorical target variable (stress level). Correlation analysis
A growing body of work has also focused on environmental highlights the influence of environmental factors and physical
factors, such as temperature, humidity, noise levels, and air activity on stress levels, with a balanced distribution ensuring
quality, in stress classification. Researchers have found corre- unbiased model training [7].
lations between environmental conditions and stress, leading to
the development of hybrid models that integrate multiple data
sources for improved accuracy. Feature selection techniques,
including Information Gain, Correlation Analysis, and Jaccard
Similarity, have been applied to refine predictive models and
enhance classification performance [5].
Despite advancements in stress classification, limited re-
search has explored integrating Multi-Criteria Decision-
Making (MCDM) techniques like the Technique for Order
of Preference by Similarity to Ideal Solution (TOPSIS) with
machine learning models [4]. TOPSIS ranks stress levels
using weighted decision matrices, enhancing interpretability
and structured classification [2]. By combining feature selec-
tion, decision-making, and classification, this study aims to
develop a scalable and accurate stress detection framework.
Logistic Regression provides probabilistic outputs that help in
understanding feature importance.
Our approach improves upon traditional models by [2] inte-
Fig. 1. The overall work flow of the proposed model
grating Logistic Regression with the TOPSIS ranking method,
enhancing feature selection and decision-making [5] . This
structured methodology refines classification accuracy by pri- B. Dataset used
oritizing significant stress indicators while reducing noise [7]. To enhance model performance, the dataset underwent
Compared to complex tree-based models like Random Forest preprocessing to handle inconsistencies, followed by feature
and deep learning methods, our model maintains high accuracy selection to identify the most relevant attributes. The data was
with reduced computational overhead. The combination of then split into training and testing sets, ensuring a structured
feature selection, ranking, and Logistic Regression ensures a approach for evaluation, as presented in TABLE I.
scalable and accessible framework for stress detection without
relying on specialized biometric sensors [2]. TABLE I
D ISTRIBUTION OF S AMPLES IN T RAINING AND T ESTING S ETS
III. P ROPOSED M ETHODOLOGY Class Training Set Testing Set
Low Stress (0) 400 samples 101 samples
The primary objective of this study is to develop a robust Medium Stress (1) 632 samples 158 samples
and interpretable stress classification model using environ- High Stress (2) 568 samples 142 samples
mental and physical activity data. The methodology involves
data preprocessing (handling missing values, normalization),
feature selection (Information Gain, Correlation, Jaccard Sim- C. Data Preprocessing
ilarity), and stress level ranking with TOPSIS. The selected Data preprocessing is a crucial step in preparing the dataset
features are classified using models like Logistic Regression, for model training and evaluation [2]. Raw data often contain
KNN, SVM, Naı̈ve Bayes, Decision Tree, AdaBoost, XG- inconsistencies, missing values, and features with different
Boost, ANN, and Random Forest, with performance evaluated scales, which can impact model performance. Therefore,
through accuracy, precision, recall, sensitivity, and specificity. several preprocessing techniques were applied to ensure the
This approach enhances model reliability and ensures that only quality and efficiency of the stress classification model [5].
 a) Handling Missing Values: One of the primary chal- a) Constructing the Decision Matrix: The decision ma-
lenges in real-world datasets is missing data [7]. The dataset trix consists of multiple features that influence stress levels,
used in this study included various environmental and activity- such as environmental parameters (e.g., temperature, humidity)
based parameters such as temperature, humidity, step count, and activity data (e.g., step count, heart rate) [2].
and other physiological indicators [6]. Missing values in nu- b) Normalization of Decision Matrix: To ensure fair
merical attributes were handled using mean imputation, where comparisons, feature values are normalized using the following
missing entries were replaced with the mean of the respective equation (2) [5].
feature. This approach helps maintain data integrity without
significantly altering the distribution [5]. ′ Xij
Xij = qP (2)
b) Feature Normalization: Since the dataset consists of n 2
i=1 Xij
numerical values with different ranges, it was essential to
normalize the features [2]. For instance, temperature values ′
where Xij is the normalized value, and Xij is the original
range from 20°C to 40°C, while step counts can vary from a
feature value.
few hundred to thousands. If left unnormalized, features with
larger numeric values may dominate model learning, leading c) Weighting the Normalized Matrix: Feature importance
to biased predictions [7]. To address this issue, Min-Max is determined based on predefined weights using statistical
normalization was applied to scale all numerical attributes to techniques such as Information Gain, Correlation, and Jaccard
a fixed range of [0,1], ensuring that each feature contributes Similarity [5]. The weighted normalized matrix is computed
equally to the model [5]. function is described in equation (1) as:
as follows:
X − Xmin ′
X′ = (1) Vij = wj · Xij (3)
Xmax − Xmin
c) Dataset Splitting: After preprocessing, the dataset was where wj represents the weight assigned to feature j.
split into training and testing sets to ensure unbiased model d) Determination of Ideal and Negative-Ideal Solutions:
evaluation [2]. The training set was used to develop the model, The ideal and negative-ideal solutions are identified by equa-
while the testing set assessed its generalization capability [5]. tion (4).
A stratified split was performed to maintain class balance,
ensuring that each class was proportionally represented in both
subsets [7]. The dataset split details are provided in TABLE I. A+ = {max Vij | j ∈ J} , A− = {min Vij | j ∈ J} (4)
D. Feature Selection
The ideal solution represents the best-case scenario, while
To enhance model performance and reduce computational the negative-ideal solution represents the worst-case scenario
complexity, feature selection was performed. Three key tech- [8].
niques were utilized: e) Distance Calculation: The Euclidean distance of each
• Information Gain: Measures the importance of each fea- alternative from the ideal and negative-ideal solutions is cal-
ture in predicting stress levels based on entropy reduction. culated using equation (5).
• Correlation Analysis: Evaluates the relationship between
different features and their impact on stress classification. v v
Features with high correlation were either removed or um um
uX uX
combined to avoid redundancy. Di+ = t (Vij − A+ 2
j ) , Di− = t (Vij − A−
j )
2 (5)

• Jaccard Similarity: Used to assess the overlap between j=1 j=1

categorical features and determine their significance in

stress classification. f) Relative Closeness Computation: The relative close-
By selecting only the most relevant features, the model focuses ness to the ideal solution is computed equation (6).
on essential data points, leading to better generalization and
interpretability. Di−
Ci = (6)
1) Decision Making using TOPSIS: The Technique for Di+ + Di−
Order Preference by Similarity to Ideal Solution (TOPSIS) is a
multi-criteria decision-making approach used in the proposed A higher Ci value indicates a stronger association with the
model to rank stress levels effectively [4]. TOPSIS works by ideal solution, leading to better stress classification.
identifying the best (ideal) and worst (negative ideal) solutions 2) Ranking and Decision Making: The computed TOPSIS
based on selected feature values and then ranking instances scores are used to rank stress levels, where higher scores
according to their relative closeness to these solutions [2]. This indicate higher stress levels [8]. The final ranked dataset is
approach enhances the interpretability and reliability of stress passed to the machine learning classification phase for further
classification [5]. processing.
E. Machine Learning Models d) Specificity: Specificity assesses the model’s capability
to correctly classify non-stress cases, minimizing false posi-
Multiple machine learning models were trained and eval-
tives. It is calculated by equation (10).
uated to classify stress levels based on environmental and
activity-based parameters. The goal was to determine the most TN
Specif icity = (10)
effective model by comparing performance metrics such as TN + FP
accuracy, precision, recall, sensitivity, and specificity [2]. A e) F1-Score: The F1-score is the harmonic mean of
diverse set of machine learning models was chosen, ranging precision and recall, balancing both metrics to provide a
from traditional statistical models to advanced ensemble and single performance measure. It is especially useful when
deep learning approaches. dealing with imbalanced datasets. The subsequent equation
• Logistic Regression: Logistic Regression is used for (11) demonstrates it.
classifying stress levels because it is an efficient and
P recision × Recall
interpretable model for binary classification [5]. Since the F 1-Score = 2 × (11)
dataset likely has two classes (”Stress” and ”No Stress”), P recision + Recall
Logistic Regression effectively models the probability IV. EXPERIMENTAL RESULT AND ANALYSIS
of each class. It works well when the data is linearly
This section presents a comprehensive evaluation of the
separable and provides insights into how factors like
proposed stress classification model, assessing its performance
humidity and step count influence stress [7]. Additionally,
in accurately predicting stress levels based on environmental
its built-in L2 regularization prevents overfitting, ensuring
and activity-based parameters [2]. The model was trained
good generalization. While the notebook explores other
on a preprocessed dataset that underwent feature selection,
models, Logistic Regression remains a strong baseline
normalization, and data balancing to enhance classification
due to its simplicity, efficiency, and reliability.
accuracy [5]. Various evaluation metrics, including accuracy,
precision, recall, sensitivity, specificity, and F1-score, were
F. Performance Evaluation Metrics used to analyze the model’s effectiveness [7]. Additionally,
To assess the effectiveness of the proposed stress classi- a comparative analysis was conducted with multiple machine
fication model, various performance evaluation metrics were learning algorithms, such as Logistic Regression, K-Nearest
utilized [2]. These metrics provide insights into the model’s Neighbors (KNN), Support Vector Machines (SVM), Naı̈ve
ability to correctly classify stress levels while minimizing Bayes, Decision Trees, Random Forest, AdaBoost, XGBoost,
errors. The following key metrics were considered: and Artificial Neural Networks (ANN), to determine the most
optimal classifier for stress prediction [12].
a) Accuracy: Accuracy represents the overall correctness
of the model by measuring the proportion of correctly clas- A. Confusion Matrix Analysis
sified instances among the total samples [5]. The following
equation (7) is used to carry out this calculation: The confusion matrix provides a detailed breakdown of the
model’s predictions, displaying the distribution of correctly
TP + TN and incorrectly classified instances across stress levels. It helps
Accuracy = (7) analyze misclassification trends and areas where the model
TP + TN + FP + FN
needs improvement. The confusion matrix for the proposed
where TP (True Positives) and TN (True Negatives) repre- model is shown in Fig. 2.
sent correctly classified stress and non-stress instances, while
FP (False Positives) and FN (False Negatives) indicate mis-
classified cases.
b) Precision: Evaluates the proportion of correctly pre-
dicted stress cases among all instances classified as stress.
A high precision score indicates fewer false positives. The
following formula (8) represents precision:

TP
P recision = (8)
TP + FP
c) Recall (Sensitivity): Measures the model’s ability to
correctly identify stress cases. A high recall score suggests
the model effectively detects stress instances without missing
significant cases. Mathematically, recall is given by equation
(9).
TP
Recall = (9) Fig. 2. Confusion matrix obtained for the model
TP + FN
 TABLE II presents the classification performance of vari- ECG-based methods, and deep learning approaches [7]. While
ous machine learning models trained using the TOPSIS-based these models have shown promising results, their accuracy
approach [8]. The Logistic Regression model achieves the remains lower than that of our proposed method [12].
highest accuracy of 99.75%, demonstrating superior perfor- Our proposed Logistic Regression model outperforms all
mance in stress classification [7]. Other models, including previous approaches, achieving an impressive accuracy of
KNN, Decision Tree, Random Forest, AdaBoost, and XG- 99.75% [8]. This superior performance is attributed to the
Boost, exhibit comparable accuracy, precision, recall, and robust feature selection process, optimized training strategy,
sensitivity, highlighting their robustness [5]. The SVM and and the model’s ability to generalize well across different
Naı̈ve Bayes models show slightly lower accuracy but still stress levels [4].
maintain competitive results [2]. These findings reinforce the
effectiveness of traditional and ensemble-based models in TABLE IV
achieving reliable stress detection [12]. PERFORMANCE COMPARISON WITH EXISTING MODELS

Authors Methods Accuracy

TABLE II Giannakakis et al. [2] Stress detection Using Biosignals 90.10%
P ERFORMANCE M ETRICS OF M ACHINE L EARNING M ODELS Gunawardhane et al. [3] Using Key Stroke Dynamics 90.97%
Gedam et al. [5] Review on Mental Stress 91.67%
Model Accuracy (%) Karthikeyan et al. [4] Using Short term ECG and HRV 94.44%
Logistic Regression 99.75 Li et al. [1] Using Deep Neural Network 97.38%
KNN 99.75 Proposed Method Logostic Regression 99.75%
SVM 99.50
Naı̈ve Bayes 99.25
Decision Tree 99.75 The accuracy progression of the proposed model across
Random Forest 99.75 training epochs is illustrated in Fig. 3 [2]. The training and
AdaBoost 99.75
XGBoost 99.75 validation accuracy curves demonstrate a rapid increase in
ANN 99.75 accuracy during the initial epochs, followed by a gradual
stabilization [5]. Both curves closely align throughout the
TABLE III presents the classification report for the pro- training process, indicating minimal overfitting [7]. The model
posed Logistic Regression model, summarizing key perfor- reaches optimal accuracy around the later epochs, showcasing
mance metrics such as precision, recall, and F1-score [5]. its robustness and generalization capability [12]. The con-
The model demonstrates consistent accuracy across all stress sistent convergence of the training and validation accuracy
levels, effectively distinguishing between different classes [7]. highlights the model’s effectiveness in learning patterns from
The support values reflect the natural distribution of instances the data while maintaining high predictive performance on
in the test dataset, ensuring a fair evaluation of the model’s unseen samples [8].
predictive capabilities [2].
The overall model accuracy confirms its reliability in clas-
sification tasks [12]. The macro and weighted average scores
further indicate balanced performance across all classes [8].
Given these results, Logistic Regression is identified as the
optimal classifier, as it effectively maintains high classification
performance without favoring any specific class [4].

TABLE III
CLASSIFICATION REPORT FOR STRESS-LYSIS DATASET

Class Precision Recall F1-score Support

Low Stress (0) 1.00 1.00 1.00 101
Moderate Stress (1) 0.994 1.00 0.997 158
High Stress (2) 1.00 0.993 0.996 142 Fig. 3. Training & Validation accuracy plot of proposed model
accuracy 0.997 0.997 0.997 401
macro avg 0.998 0.998 0.998 401 The loss convergence trend of the proposed model is de-
weighted avg 0.998 0.998 0.998 401
picted in Fig. 4 [2]. Both training and validation loss exhibit
a consistent decline over the epochs, demonstrating effective
B. Comparative Analysis learning [5]. The initial loss starts at a relatively high value,
TABLE IV presents a comparative analysis of the proposed gradually decreasing as the model optimizes its parameters
Logistic Regression model against previously proposed models [7]. The close alignment between training and validation loss
for stress detection [2]. The comparison highlights the effec- curves indicates that the model does not suffer from significant
tiveness of different methodologies used in past research and overfitting, ensuring good generalization [12]. The smooth
demonstrates the superior performance of our approach [5]. and steady reduction in loss values suggests that the learning
Several studies have explored various techniques for stress process is stable, ultimately leading to minimal error and
detection, including biosignal analysis, keystroke dynamics, improved predictive performance [8].
 [2] G. Giannakakis, D. Grigoriadis, K. Giannakaki, O. Simantiraki, A.
Roniotis, and M. Tsiknakis, ”Review on psychological stress detection
using biosignals,” IEEE Transactions on Affective Computing, vol. 13,
no. 1, pp. 440-460, 2019.
[3] S. D. W. Gunawardhane, P. M. De Silva, D. S. B. Kulathunga, and
S. M. K. D. Arunatileka, ”Non-invasive human stress detection using
keystroke dynamics and pattern variations,” in Proc. IEEE Int. Conf. on
Advances in ICT for Emerging Regions (ICTer), 2013, pp. 240-247.
[4] P. Karthikeyan, M. Murugappan, and S. Yaacob, ”Detection of human
stress using short-term ECG and HRV signals,” Journal of Mechanics
in Medicine and Biology, vol. 13, no. 2, pp. 1350038, 2013.
[5] S. Gedam and S. Paul, ”A review on mental stress detection using
wearable sensors and machine learning techniques,” IEEE Access, vol.
9, pp. 84045-84066, 2021.
[6] G. Giannakakis, D. Grigoriadis, D. Kollias, M. Pediaditis, and M.
Fig. 4. Loss convergence plot obtained for proposed model Tsiknakis, ”Stress and anxiety detection using facial cues from videos,”
Biomedical Signal Processing and Control, vol. 31, pp. 89-101, 2017.
[7] X. Li, Q. Zhang, and J. Wang, ”Deep learning-based stress detection
from physiological signals,” IEEE Transactions on Affective Computing,
The proposed Logistic Regression model achieves a remark- vol. 10, no. 4, pp. 471-484, 2019.
able accuracy of 99.75%, significantly outperforming existing [8] Muduli, Debendra, Ratnakar Dash, and Banshidhar Majhi. ”Automated
models, as shown in TABLE IV [2]. Compared to stress diagnosis of breast cancer using multi-modal datasets: A deep convolu-
tion neural network based approach.” Biomedical Signal Processing and
detection using biosignals, keystroke dynamics, and deep Control 71 (2022): 102825.
neural networks, the proposed method provides a more reliable [9] Muduli, Debendra, Ratnakar Dash, and Banshidhar Majhi. ”Auto-
and efficient approach [5]. This highlights the effectiveness of mated breast cancer detection in digital mammograms: A moth flame
optimization-based ELM approach.” Biomedical Signal Processing and
feature selection and preprocessing in improving classification Control 59 (2020): 101912.
performance [7]. [10] Muduli, Debendra, Ratnakar Dash, and Banshidhar Majhi. ”Fast discrete
The training and validation accuracy plot shows smooth curvelet transform and modified PSO based improved evolutionary
extreme learning machine for breast cancer detection.” Biomedical
convergence, indicating effective learning without overfitting Signal Processing and Control 70 (2021): 102919.
[12]. The model maintains a consistent gap between training [11] Muduli, Debendra, et al. ”An empirical evaluation of extreme learning
and validation accuracy, ensuring strong generalization [8]. machine uncertainty quantification for automated breast cancer detec-
tion.” Neural Computing and Applications (2023): 1-16.
The loss convergence plot confirms stability, with a steady [12] Sharma, Santosh Kumar, et al. ”A Diabetes Monitoring System and
decline in training and validation loss [12]. Unlike deep Health-Medical Service Composition Model in Cloud Environment.”
learning models, it avoids large fluctuations, ensuring a well- IEEE Access 11 (2023): 32804-32819.
[13] Sharma, Santosh Kumar, et al. ”An evolutionary supply chain man-
regulated optimization process [4]. agement service model based on deep learning features for automated
Despite its simplicity, Logistic Regression outperforms deep glaucoma detection using fundus images.” Engineering Applications of
neural networks in accuracy and efficiency, making it a reliable Artificial Intelligence 128 (2024): 107449.
[14] Muduli, Debendra, et al. ”Retinal imaging based glaucoma detection
and practical choice for real-world applications [2]. using modified pelican optimization based extreme learning machine.”
Scientific Reports 14.1 (2024): 29660.
V. CONCLUSION [15] Sharma, Santosh Kumar, et al. ”Discrete ripplet-II transform feature
This study presents an optimized stress classification model extraction and metaheuristic-optimized feature selection for enhanced
glaucoma detection in fundus images using least square-support vector
using environmental and physical activity data, integrating machine.” Multimedia Tools and Applications (2024): 1-33.
data preprocessing, feature selection, TOPSIS-based decision-
making, and machine learning classification. The proposed
Logistic Regression model achieves 99.75% accuracy, out-
performing deep learning approaches while maintaining com-
putational efficiency [2]. The training and validation accu-
racy plot confirms the model’s stability, showing smooth
convergence without overfitting, while the loss convergence
plot illustrates steady learning dynamics. The integration of
TOPSIS further enhances classification reliability, making the
approach suitable for real-world stress monitoring [4]. This re-
search demonstrates that well-optimized traditional models can
achieve state-of-the-art performance in stress classification.
Future work can explore integrating additional physiological
and behavioral data to improve real-time stress detection and
generalizability.
R EFERENCES
[1] Li, Russell, and Zhandong Liu, ”Stress detection using deep neural
networks,” BMC Medical Informatics and Decision Making, vol. 20,
pp. 1-10, 2020.