Received 4 April 2024, accepted 28 May 2024, date of publication 5 June 2024, date of current version 1 July 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3407711

Developing Intelligent and Immutable Vaccine

Supply and Operation Platform Using
Blockchain and Artificial Intelligence
Technologies
SAMAH ALHAZMI 1 , MOHAMMAD KHALID IMAM RAHMANI 1 , (Senior Member, IEEE),
MOHAMMAD ARIF 2 , AND MD TABREZ NAFIS3 , (Senior Member, IEEE)
1 Collegeof Computing and Informatics, Saudi Electronic University, Riyadh 11673, Saudi Arabia
2 Schoolof Computer Science and Engineering, Vellore Institute of Technology, Vellore 632014, India
3 Department of Computer Science and Engineering, Jamia Hamdard, New Delhi 110062, India

Corresponding authors: Samah Alhazmi ([email protected]), Mohammad Arif ([email protected]), and Mohammad Khalid
Imam Rahmani ([email protected])
This work was supported by the Deanship of Scientific Research at Saudi Electronic University under Grant 9476.

ABSTRACT During times of crisis, like in healthcare emergencies or disruptions in logistics and supply
chains, decision-makers encounter significant challenges in managing resources efficiently. This study
proposes an approach combining different information types and advanced computer techniques to improve
how we handle crises. By looking at data connected with blockchain technology, we aim to understand
crises better and help people make smarter decisions. We use a special kind of computer model called
Long Short-Term Memory (LSTM) to guess what might happen during a crisis, like how many people
might get sick or if there might be problems with getting supplies. Our model is right about 87% of the
time. To understand why the computer makes these guesses, we use a method called SHapley Additive
exPlanations (SHAP), which is widely known. However, we notice that SHAP sometimes doesn’t explain
things well for the entire period we’re looking at. So, we suggest some changes to the SHAP method to
explain things better for smaller parts of time. These changes help the method explain things better for
small bits of time, but it might not be as good at explaining everything overall. To ensure our idea works,
we compare our computer model’s guesses and explanations with those from other models that look at data
differently. We show the good and bad parts of our method in different crises.

INDEX TERMS Blockchain, artificial intelligence technologies, intelligent and immutable vaccine supply,
long short-term memory, SHapley additive explanations, XAI.

I. INTRODUCTION Artificial Intelligence (AI) [4], [5], offer better crisis manage-
Pandemics, natural calamities, or disturbances in supply ment strategies. By using diverse datasets [6] and advanced
chains can create crises. Managing the available resources analytical techniques [7], researchers and decision-makers
efficiently and making effective decisions for proper aim to develop better models for crisis prediction and effec-
responses are crucial in such situations. Rapid crisis predic- tive response execution. The COVID-19 or similar pandemic
tion and better response can minimize their impact and save demands an urgent need for proactive data-driven crisis
precious lives [1]. Applications of the recent advancements response strategies [8]. In healthcare institutions, respondents
in Data Science with Machine Learning (ML) [2], [3] and in Intensive Care Units (ICUs) face challenges like allocating
resources and triaging patients [9]. This urgency synergized
The associate editor coordinating the review of this manuscript and the necessity for exploring innovative solutions in real-time
approving it for publication was Gyorgy Eigner . decision-making.

2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 88189
 S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

Previous efforts in crisis management have streamlined ML models, particularly LSTMs, have become powerful
the utilization of data and technology to address these chal- tools for predicting critical events in various fields like health-
lenges. Researchers explore ML algorithms in healthcare care and supply chain management, by studying data trends
services to forecast patient outcomes based on ICU time over time. These models support decision-makers in prepar-
series data [10]. Authors [11], [12] utilizing datasets like the ing action plans for crisis management and making better
Medical Information Mart for Intensive Care (MIMIC) show decisions to use the resources for proper response to tackle
the efficiency of ML models, particularly Long Short-Term the situation. However, due to the complexities involved,
Memory (LSTM) networks, in accurately predicting patient it may be difficult to understand the crises. In this research,
mortality. Additionally, Explainable Artificial Intelligence we develop methods like SHAP to solve this issue giving
(XAI) methods such as SHapley Additive exPlanations insights into how these models make predictions. Besides,
(SHAP) have been used to provide better model predictions, we use blockchain technology because of its popular features
assisting clinicians in understanding the factors influenc- of transparency and immutability people use it to share data
ing patient outcomes [13]. However, while these efforts are securely. Blockchain in crisis management systems brings
promising, they impose several key challenges; one challenge trust among all parties involved due to data accuracy and
lies in XAI-ML models, with complex time series data [14]. traceability of any changes made in data to the source.
While the SHAP method offers valuable global insights into This makes collaboration easy and facilitates better decision-
model behavior, it may not fully capture patient’s trajecto- making. But these advances are not enough. There is still a
ries over time. This limitation has significant implications in need to explore different types of data and methods to deal
clinical decision-making, wherein precise insights are often with complex crises. This research paper proposes a novel
necessary to support effective patient care. Moreover, existing solution to this problem by combining ML, XAI methods
research mainly focuses on healthcare, ignoring the inter- like modified SHAP, and blockchain technology. By look-
disciplinary nature of crises and meaningful insights from ing at data from different dimensions and exploiting that to
diverse datasets. Crises share characteristics such as uncer- explain how the models predict, our approach makes crisis
tainty and complexity and need quick responses in healthcare, predictions more accurate and easier to understand. Using
logistics, or supply chains. By expanding the scope of inquiry blockchain, we facilitate a safe and transparent mechanism
to cover various crisis scenarios, researchers can identify for the people involved to share data and make optimized
common patterns and develop generalized approaches that decisions during a crisis as quickly as possible.
transform specific domains.
Looking into this scenario, this research proposes an III. THE STATE-OF-THE-ART
interdisciplinary approach to crisis management. By using Over the years, artificial intelligence (AI) has become one
datasets of diverse nature covering different dimensions of of the most talked-about and influential fields. It was first
crisis data and utilizing advanced ML techniques, we aim introduced in the 1950s by John Von Neumann [15] and
to develop a comprehensive understanding of crisis man- Alan Turing [16] as a Computer Science discipline that
agement tasks and enhance real-time decision-making using could think like humans. Today it covers many disciplines
ML models like LSTM networks to predict critical out- such as Mathematics, Neurology, etc. [17]. This cognitive
comes such as mortality rates or supply chain disturbances. superiority over humans has enabled its use across various
Additionally, we explore XAI methods, including modified industries like technology, banking, and marketing. In this
versions of SHAP, to provide actionable insights into model regard, several key terms in AI research to be understood.
predictions. Through this interdisciplinary approach, we aim Authors [18] and [19] have defined AI as a collection of
to contribute to the field of crisis management by offering tools and techniques for intelligent cognition, learning, and
new insights and methodologies applicable across multiple adapting to solve specific problems. Authors [20] describe
domains. By exploiting data and technology, we seek to the role of AI in exploiting big data and ML for analysis and
support decision-makers with the knowledge and tools to prediction tasks in respective domains, like healthcare and
handle crises effectively and protect precious lives and their supply chain. Authors [21] segregate AI into distinct forms
livelihoods. such as ML, learning analytics, and data mining (KDD),
The remainder of the paper is outlined as follows; the each serving specific roles for extracting insights from data.
research contribution is shown in Section II. Section III dis- Moreover, recent studies [22] and [23] explore the evolving
cusses the state-of-the-art related works. The materials and AI nature and its impact on society and business decision
methods describe data and model analysis in Section IV. support. Authors [22] describe the importance of feasible AI
Section V narrates how the experiment is set up. The research behaviors and usage in enhancing confidence in AI-based
outcomes are covered in Section VI, and Section VII gives technologies for society. Besides, authors [23] explore the
conclusions and plans for future research. role of AI in solving challenging problems and promising
sustainability through innovative technologies and dynamic
II. RESEARCH CONTRIBUTION learning processes. Moreover, researchers [24] and [25] offer
Data-driven analysis with ML, AI, and blockchain technolo- valuable insights into the behavioral strategic advantages of
gies has introduced a new solution for crisis management. AI adoption. Authors [24] demonstrate a positive correlation

88190 VOLUME 12, 2024

S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

Algorithm 1 LSTM Model Training and Evaluation

START
1. Define the LSTM model architecture
Specify the number of LSTM layers, units per layer, and activation functions
Add a dense layer for prediction with an appropriate activation function
2. Prepare the data
Normalize or preprocess the data as necessary
Split the data into training, validation, and testing sets
3. Train the LSTM model
for each set of hyperparameters
Compile the model with appropriate loss function and optimizer
Train the model on the training data
Monitor performance on the validation data
Implement early stopping if overfitting is detected
4. Evaluate the model
Evaluate model performance on the testing set using standard metrics
Accuracy: How often the model’s predictions are correct
Precision: How often the model’s positive predictions are correct
Recall: How much of the actual positive cases the model predicts correctly
F1-score: Harmonic mean of precision and recall
5. Analyze decision-making
Create a chart showing the model’s decisions and confidence levels
Assess decision changes based on confidence levels
6. Test the model in diverse scenarios
Perform cross-validation by splitting data into multiple folds
Evaluate the model’s performance across different scenarios
7. Iterate and refine
Adjust hyperparameters based on performance
Re-train and re-evaluate the model
END

between AI and consumer behavior, supporting the poten- The COVID-19 World Vaccination Progress dataset con-
tial of AI in shaping complex market dynamics. Similarly, tains information from the global vaccination drives during
authors [25] discuss the benefits of AI in decision sup- the COVID-19 crisis. This dataset comprehensively reports
port and automation to create new business models and the cases of global vaccination initiatives covering various
management techniques. It represents a new paradigm for nations, vaccine types, daily and cumulative statistics, and
automation and enhanced operational business efficiency. vaccination rates. It also includes information about vaccine
Organizations are adopting AI- and data-driven solutions to sources that offer better insights into the various strategies
elevate performance benchmarks and adapt to ever-evolving adopted by different control regions.
market demands [26] recognizing the potential capability Using the Food Delivery Time Prediction dataset, the chal-
of AI. The allied fields of AI support diversified solution lenges of estimating delivery times in food delivery platforms
techniques for technological advancements, societal empow- Zomato and Swiggy can be explored. This dataset provides
erment, and organizational transformations. By exploring a base for developing a reliable predictive model to esti-
these techniques and their promises, researchers aim to sim- mate precise delivery times considering factors of distance
plify the complexities of AI and exploit its potential for the between pickup and delivery points and delivery history.
betterment of society and business. Therefore, these datasets provide a multi-dimensional per-
spective for extracting useful knowledge in supply chain
IV. MATERIALS AND METHODS
management, public health, and optimized logistics decision-
A. DATA ANALYSIS
making.
The datasets used for the research contain useful informa-
tion from different fields. Using the Supply Chain Analysis B. MODEL ANALYSIS
dataset complex product supply chain dynamics, covering The proposed research model combines ML, XAI, and
product types, prices, sales data, inventory records, manufac- blockchain technology for predicting crises. The proposed
turing details, transportation methods, etc. can be explored. model implements LSTM to predict crisis outcomes and
So, the dataset facilitates understanding the complexities of SHAP to comprehend model predictions apart from inte-
the supply chain management ecosystem in the fashion and grating blockchain technology to support the integrity trans-
beauty industries. parency, and security of data.

VOLUME 12, 2024 88191

 S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

Algorithm 2 Enhanced SHAP Method for Explaining Predictions Over Time

START
1. Define the basic SHAP method
Compute SHAP values for each input feature to understand their contributions to predictions
Draw summary plots to visualize the main features and their effects on predictions
Provide global explanations to understand trends in the dataset
Provide local explanations to understand individual predictions
2. Address the limitations in time series data
Modify the SHAP method to cover temporal relationships between features and predictions
Draw plots to depict the evolution of feature contributions to time
Analyze feature importance to time intervals or different phases of crises
Explain the historical data leading up to the prediction point
3. Implement changes to the SHAP method
Adapt existing SHAP algorithms or create new ones to effectively handle time series data
Create visualizations that show temporal feature contributions and their effect on predictions
Apply time-based features to explain and provide better insights into model behavior over time
4. Test and validate the enhanced SHAP method
Apply the modified SHAP method to time series datasets of different domains
Evaluate its effectiveness in capturing temporal dynamics and providing interpretable explanations
Compare results with SHAP analyses to assess improvements
5. Repeat and improve the enhanced SHAP method
Consider suggestions for further improvements and refinements
Keep updating the method based on new insights and advancements in time series analysis and
interpretability techniques
END

1) LONG SHORT-TERM MEMORY FOR PREDICTIVE 2) XAI FOR UNDERSTANDING MODEL PREDICTIONS
MODELING While LSTM networks are great at making predictions, they
The LSTM-based predictive models are effective at extracting can be hard to understand. This makes it tough for people
patterns over time. So, it has been used for predicting useful involved to figure out why the model predicts what it does.
outcomes during crises. We provide a brief explanation of To help with this, we use a method called SHAP. SHAP
the LSTM model’s working procedure on a combined dataset breaks down the model’s predictions to show how much each
that was collected from data sources of varying fields. Its input feature contributes. This helps see which features are
different layers are specialized in understanding sequences most important and how they affect the prediction. We show
in data followed by a layer for making predictions. We try these SHAP values using summary plots. These plots high-
out different settings, like how many layers to use, how many light the main features and how they influence the predictions.
units each layer has, and how quickly the model learns, We also use SHAP to explain the model’s behavior overall
to explore the best setup. We also use some mechanisms, like (global explanations) and for individual cases (local explana-
stopping the training early if the model starts to memorize the tions). Global explanations help us understand trends across
data too much, and adjusting how fast it learns to improve the whole dataset, while local explanations help us under-
its ability to handle new data. To see the working of the stand why the model made a specific prediction for a single
model, we use standard measures like how often it’s right, case. However, SHAP has some limitations in dealing with
how often it’s precise, how often it remembers everything and complex time series data. Therefore, we adopt changes to
a combination of these. We also look at a chart to help us see the SHAP method to better the prediction over time. These
how the model distinguishes between different situations, like changes aim to capture how things change over time and
whether someone will survive or not in a medical situation, provide better explanations for the model’s predictions (refer
or if there will be crises or things will run smoothly in a supply to Algorithm 2).
chain or not. We also use a special chart to show how the
model’s decisions change based on how sure it is, which helps
3) INTEGRATION OF BLOCKCHAIN TECHNOLOGY FOR DATA
us understand how well it’s making decisions. Additionally,
INTEGRITY AND TRANSPARENCY
we test the model in different ways to make sure it works
well in different situations, like splitting our data into parts We explore blockchain technology integration into our crisis
for training, checking, and testing, and doing this multiple management framework after implementing the predictive
times with different parts, so we can be confident that what model and incorporating XAI analysis to ensure data integrity
we find is true for lots of situations, not just one. Algorithm 1 transparency, and security. Blockchain offers a decentral-
illustrates the corresponding mathematics. ized and immutable ledger that records transactions in a

88192 VOLUME 12, 2024

S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

network of nodes with the help of a secure and transparent TABLE 1. Hyperparameters of the LSTM model in COVID-19 vaccination
records.
data-sharing and collaboration mechanism.
We record a tamper-proof audit trail that tracks the verac-
ity of data from its source to destination by integrating
blockchain technology. This ensures the integrity and authen-
ticity of data, and mitigation of the risk of data tampering
or its unauthorized manipulation by malicious nodes. More-
over, blockchain facilitates transparent data sharing, creating
trust and hence supporting collaboration among the agreed TABLE 2. Covid-19 vaccination records of confusion matrix.

parties in the crisis management ecosystem. We implement

self-executing smart contracts with predefined rules and con-
ditions to automate data verification and validation processes.
Through smart contracts, agreed parties can define data
access and sharing rules in compliance with regulations and
data privacy standards.
In the end, we explore Decentralized Identifiers (DIDs) and score on validation data. We employed early stopping and
verifiable credentials to create a secure and privacy-preserving halting training strategies when performance metrics stag-
identity management framework for the agreed parties to nated to ensure the model’s highest performance and prevent
securely share and verify their credentials without the cen- overfitting.
tralized authorities. Besides, we employed XAI to understand and explain the
Besides, we analyze scalability, efficiency, and sustain- model’s decision-making procedure. We gained insights into
ability in blockchain-based solutions for crisis management the factors influencing the model’s predictions by calculating
ecosystems. The mechanisms of Proof-of-Work (PoW), SHAP values for individual data points and visualizing them
Proof-of-Stake (PoS), and Delegated Proof-of-Stake (DPoS) through summary and force plots (refer to Section VI-B).
to render consensus among the parties and validate trans- Furthermore, we enhanced the SHAP library to enhance
actions efficiently [27], [28] are also explored. Moreover, explanations for time-series data, ensuring more accurate and
we assess blockchain technology’s impact on the environ- detailed insights. These improvements were achieved through
ment and propose alternative approaches: Proof-of-Authority testing with benchmark datasets to establish their effective-
(PoA) and Proof-of-Space (PoSpace) to minimize energy ness.
consumption and carbon impact. A flowchart depicting the Lastly, we integrated blockchain technology into our
blockchain technology integration is displayed in Fig. 1. framework to ensure data integrity, transparency, and secu-
rity. By deploying blockchain networks using platforms like
V. EXPERIMENTAL SETUP Ethereum, we established secure data verification processes
First, we briefly describe the dataset used in the research. using decentralized smart contracts. The implemented model
The dataset comprises COVID-19 vaccination records, food demonstrated a robust and transparent data ecosystem ready
delivery patterns, supply chain dynamics, and blockchain to support decision-makers in crisis management decision-
transactions. For the experiments, Python libraries Pan- making.
das and NumPy were used to clean and organize it to
enhance quality and compatibility with different sources. VI. RESULT ANALYSIS
This includes handling missing information, categorizing A. PREDICTIVE ANALYSIS
variables, adjusting numerical values, and relevant features The LSTM models’ performance varies across domains
selection for analysis. including COVID-19 vaccination records, food delivery pat-
Then we implemented an LSTM model for crisis pre- terns, and supply chain dynamics. This variability arises due
diction and configured it with the dataset by exploiting to the complexities and dynamics of each domain. As a
user-friendly deep learning tools like TensorFlow. We con- result, customized approaches to model design, parameter
structed the model, adjusting parameters; the number of tuning, and optimization are essential to effectively capture
layers, hidden units, dropout rates, and learning rates. and predict outcomes during crises as depicted in Table 1.
For optimization grid search was employed to fine-tune In the case of COVID-19 vaccination records, LSTM mod-
these parameters and improve the model’s performance. els show strong performance in forecasting vaccination rates
To prevent the model from becoming overly specialized dur- and detecting trends as shown in Table 2.
ing training, we implemented the regularization method L1 This domain indicates temporal relationships defined by
regularization. factors- demographics, distribution strategies, and public
Once the model was ready, we trained and evaluated it health measures. The proposed LSTM architecture has multi-
using our prepared dataset. For utilizing GPU resources on ple layers with several hidden units to capture these complex
cloud platform Google Cloud Platform, we monitored key relationships. A 0.2 dropout rate is applied to rule out over-
performance metrics; loss, accuracy, precision, recall, and F1- fitting while maintaining model flexibility. A learning rate of

VOLUME 12, 2024 88193

 S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

FIGURE 1. Flowchart for the blockchain technology integration.

TABLE 3. Food delivery pattern of LSTM model hyperparameters. TABLE 4. Food delivery pattern of confusion matrix.

TABLE 5. Supply chain dynamics of LSTM model hyperparameters.

0.001 ensures convergence stability in training, capturing the

gradual changes in vaccination trends over time. The eval-
uation based on the confusion matrix indicates the model’s
capability to accurately classify the success and failure of TABLE 6. Supply chain dynamics of confusion matrix.
vaccination cases, subject to some observed misclassifica-
tions. These findings suggest scope for further improvement
to enhance model performance and predictive accuracy.
The implemented LSTM model demonstrates effective-
ness in demand prediction and finding optimized delivery
routes from food delivery patterns. However, the complex-
ities in this domain tend to diverge its effectiveness from tified to enhance the model’s ability to detect small variations
those found in COVID-19 vaccination records. Food deliv- in delivery patterns.
ery trends are affected by short-term temporal dependency, In supply chain management operations, the implemented
demand fluctuation, route optimization, and inventory man- model can predict the disturbances in the supply chain and
agement dynamics. Therefore, adjustments are made to the optimize inventory management. Supply chain operations
LSTM model’s structure and key parameters as shown in involve complex communications and interdependencies in
Table 3. inventory management, logistics, and demand forecasting.
Selecting a simplified architecture comprising two LSTM The model’s design and settings are precisely adjusted to
layers with reduced hidden units per layer, the model aims to handle these complexities. By implementing a more detailed
maintain a balance in capturing useful temporal patterns and architecture with four LSTM layers and more hidden units
computational efficiency. To mitigate the risk of overfitting per layer, the model captures the complex patterns and inter-
due to the inherent variability in delivery patterns, a rela- connections within the supply chain network data as shown
tively higher dropout rate of 0.3 is considered, signifying in Table 5.
the importance of regularization. Moreover, a higher learning A dropout rate of 0.4 is considered to effectively manage
rate of 0.01 is chosen for faster convergence during optimiza- the model’s complexity and prevent it from overfitting, con-
tion, facilitating adaptation to changes in delivery dynamics. sidering the volatility and unpredictability in supply chain
Evaluation using a confusion matrix shows promising perfor- dynamics. Therefore, a learning rate of 0.001 is chosen to
mance, although there are some misclassifications as shown ensure stable convergence during the optimization process
in Table 4. Therefore, further areas of improvement are iden- due to the complex nature of the optimization challenges

88194 VOLUME 12, 2024

S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

TABLE 7. Results of cross-validation in LSTM model within three datasets.

FIGURE 2. Implications of Kernel SHAP on time series data.

associated with training on sequential supply chain data as ity of LSTM models across different areas, aiding in better
shown in Table 6. decision-making and risk management during crises.
The evaluation using a confusion matrix shows promising Therefore, to evaluate the performance and hence applica-
results but with some misclassification cases. This indicates bility of the implemented model with the dataset, we conduct
scope for further improvement to address the challenges in cross-validation tests. By splitting the dataset into training,
supply chain disturbances or stability factors. validation, and testing subsets with k-fold cross-validation,
The varying performance of LSTM models across domains we mitigate the risk of overfitting and hence ensure the
showcases the importance of customizing model structure, reliability of the findings. The cross-validation outcomes for
settings, and optimization methods to fit uniquely into the datasets show the resilience and adaptability of the imple-
the traits and complexities of each domain. Despite all mented model for crisis outcomes prediction.
domains involving sequences over time and predictive mod- The model demonstrates high average accuracies for all
eling, the specific characteristics of each domain necessitate datasets showing its effectiveness in capturing underlying
domain-specific strategies for optimal performance. Further patterns and dynamics in each domain. However, there
research is crucial to improve the accuracy and dependabil- are slight variations in performance across the datasets.

VOLUME 12, 2024 88195

 S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

and exploiting a background dataset, Kernel SHAP allows for

predictions over time. However, while Kernel SHAP provides
rich global explanations, the local explanations may not accu-
rately reflect the model’s behavior due to the independent
evaluation of each day’s data. This shows a limitation in
the local accuracy property of SHAP because local expla-
nations do not consistently match the model’s output over
time. Therefore, modifications to SHAP are proposed to
better handle time series data by incorporating the hidden
state of recurrent neural network models, such as LSTMs.
While this approach improves local accuracy by considering
the sequential nature of the data, it may compromise global
explanations, as demonstrated in experiments with an addi-
tional LSTM model. The trade-off between local and global
interpretability underscores the need for nuanced approaches
in SHAP implementations tailored to specific data char-
acteristics (refer to Fig. 2). Moving on to TreeExplainer,
a specialized version of SHAP designed for tree-based mod-
els like random forests, we observe its application to ICU
mortality prediction models trained on different subsets of
patient data (refer to Fig. 3). TreeExplainer provides insights
FIGURE 3. TreeExplainer on ICU mortality prediction models. into feature importance and interactions, enabling a deeper
understanding of the decision-making process of random
forest models. Notably, comparisons between TreeExplainer
For example, in COVID-19 vaccination records, the model and LSTM-based models reveal differences in feature impor-
achieves the highest average accuracy of 0.918, followed by tance rankings as shown in Fig. 4, suggesting variations in
food delivery patterns with an average accuracy of 0.885, and the predictive capabilities of different model architectures.
supply chain dynamics with an average accuracy of 0.893 as Furthermore, the analysis of global explanations from Tree-
shown in Table 7. Explainer highlights distinct trends in feature importance
These variations in average accuracies suggest that there across subsets of patient data as shown in Fig. 5. For instance,
are different levels of complexity and predictability in the age emerges as a significant predictor of mortality in mod-
datasets. The low standard deviations indicate consistent els trained on admission data, reflecting the inherent risk
model performance and hence strengthen the stability of associated with aging in healthcare outcomes. Conversely,
the LSTM model’s predictions for different data subsets. features of physiological parameters and laboratory mea-
The cross-validation results validate the effectiveness of the surements exhibit varying importance levels depending on
LSTM model in predicting critical crisis outcomes with its the time window considered, indicating the dynamic nature
potential applications in decision-making and risk manage- of patient data and its impact on predictive performance.
ment strategies. However, despite the insights gained from SHAP-based XAI
methods, certain challenges and limitations persist. The paper
B. XAI ANALYSIS discusses discrepancies between local explanations provided
We analyze the results obtained from applying SHAP, includ- by SHAP and the actual model predictions over time, under-
ing Kernel SHAP1 and TreeExplainer2 in this research to scoring the need for improved methods to ensure consistency
interpret the ML models’ predictions on healthcare data and accuracy in local interpretations. The trade-off between
focusing on predicting mortality in ICU patients. The SHAP local and global explanations raises questions about the
framework offers insights into the available features and their applicability of SHAP in real-world healthcare settings,
interactions, supporting the understanding of model predic- where both types of interpretations are crucial for clinical
tions and providing interpretability in complex healthcare decision-making.
scenarios.
Firstly, the implications of using SHAP, particularly Kernel
SHAP, on time series data are explored. The paper discusses C. BLOCKCHAIN ANALYSIS
the challenges in time series data and how Kernel SHAP is We observe a thorough examination of the integration of
adapted to handle such data. By reorganizing the input data blockchain technology into our strategy for handling crises
1 https://shap-lrjball.readthedocs.io/en/latest/generated/shap.
to provide the authenticity and openness of data. Our objec-
KernelExplainer.html
tive is to validate the role and impact of blockchain-driven
2 https://shap-lrjball.readthedocs.io/en/latest/generated/shap. solutions across three separate datasets, focusing on different
TreeExplainer.html factors such as the ability to predict accurately, maintain

88196 VOLUME 12, 2024

S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

FIGURE 4. Comparison of feature importance rankings between TreeExplainer and LSTM-based models.

FIGURE 5. Analysis of global explanations from TreeExplainer across different subsets of patient data.

data authenticity, ensure transparency, enhance efficiency and Blockchain implementation ensures data integrity and
promote sustainability. transparency by establishing an immutable and transparent
ledger system. This enables stakeholders to trace the origin
1) COVID-19 VACCINATIONS ANALYSIS and flow of data securely, thereby mitigating risks associated
The integration of blockchain technology into predictive with tampering or manipulation by malicious nodes. How-
models has shown promising results as shown in Fig. 6, ever, the efficiency of blockchain-based solutions remains
indicating enhanced predictive accuracy with minimal com- moderate in this context. This moderate efficiency is pri-
promise due to blockchain. marily attributed to the overhead incurred during transaction

VOLUME 12, 2024 88197

 S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

FIGURE 6. Date vs total vaccination predictive accuracy.

FIGURE 8. Delivery-person-age vs time-taken predictive performance.

FIGURE 7. PoW graph on total vaccinations vs people vaccinated.

FIGURE 9. Time-taken vs type-of-vehicle predictive performance.

validation and the consensus-building process needed in

blockchain networks. While blockchain offers enhanced
security features, the consensus mechanism may introduce Thus, while blockchain technology enhances data integrity
latency and overhead, affecting overall system performance and transparency, addressing energy consumption concerns is
and underlying costs. Sustainability considerations are rated imperative to ensure long-term sustainability.
as moderate, with energy consumption being a significant
factor. The sustainability aspect chiefly revolves around the 2) FOOD DELIVERY PATTERNS ANALYSIS
energy-intensive nature of consensus mechanisms, notably The models combined with blockchain technology demon-
PoW algorithms commonly employed in blockchain net- strate moderate accuracy in making predictions as shown in
works such as Bitcoin. Although PoW [28] ensures enhanced Fig. 8 and Fig. 9. This indicates some trade-offs between
security through complex cryptographic puzzles, it also how well the models work and using blockchain solutions.
yields substantial energy consumption as shown in Fig. 7. However, including blockchain technology ensures that data

88198 VOLUME 12, 2024

S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

collaborative data sharing among agreed parties. The efficacy

of blockchain-based solutions within supply chain dynamics
is quite beneficial even if there is a decline in predictive
accuracy. This worthy efficiency is attributed to the efficient
execution of streamlined consensus mechanisms offered by
smart contracts.
Alternative methods, such as PoA or PoSpace are deployed
for less energy consumption and minimize environmental
side effects hence augmenting the overall system efficiency.
Moreover, the sustainability aspect within this dataset is an
added advantage. The adoption of energy conservation con-
sensus and eco-friendly practices promotes better livelihood
and environmental friendliness of blockchain-based solu-
tions. By prioritizing sustainability, stakeholders can effec-
tively mitigate the ecological risks associated with blockchain
deployments while enjoying the benefits of enhanced data
reliability and transparency.

VII. CONCLUSION AND FUTURE WORKS

The proposed approach in our research paper presents an
FIGURE 10. Price vs product-type predictive performance.
advancement in crisis management strategies in healthcare,
logistics, and supply chain disruptions. By using diverse
datasets and employing advanced ML techniques and LSTM
remains trustworthy and transparent, similar to what we see networks, the proposed work aimed to enhance crisis man-
in COVID-19 vaccinations. The tamper-proof trail of actions agement techniques by providing accurate predictions and
and clear ways of sharing data are important features here, responsive insights into critical crisis outcomes. The results
preventing anyone from messing with the data and help- from this research validate using AI, data science, and
ing everyone feel confident in the information. The way allied technologies to solve the challenges faced during
blockchain works in Food Delivery Patterns is considered crises. Exploitation of LSTM models demonstrated promis-
very efficient. This is because the methods used to reach ing results in crisis prediction, such as mortality rates in
agreements of transactions are set up to work quickly and healthcare settings and supply chain disruptions. The high
automatically through ‘‘smart contracts’’. These agreements, accuracy rates achieved by the LSTM model suggest its
like PoS or DPoS, make it faster to agree on transactions potential to support decision-makers in gauging the impact
without making things less secure. Plus, using smart contracts of crises and effectively responding to them. Furthermore,
helps that data gets checked and verified automatically, which the applications of XAI methods like SHAP gave valuable
helps everything run smoother. But, when thinking long-term support to the decision-making process by the LSTM model.
and taking care of the environment, sustainability is a problem While SHAP provided both local and global explanations for
in this dataset. Even though blockchain is efficient, some model predictions, challenges were faced in accurately cap-
ways it works use a lot of energy, especially PoW methods. turing the nuances of complex time series data. The proposed
Therefore, we need to think about lesser energy consumption modifications to the SHAP library addressed these challenges
and hence lower the impact on the environment. We need to by improving local interpretability but compromising global
explore other ways of agreeing on transactions and conserv- interpretability. Integrating blockchain technology with the
ing energy in smarter ways to justify the use of blockchain crisis management framework introduced data integrity,
without affecting the planet. transparency, and security. However, establishing a decen-
tralized and immutable ledger with blockchain ensured the
3) SUPPLY CHAIN DYNAMICS ANALYSIS authenticity and traceability of data, and risk mitigation
Blockchain technology integration in crisis management of tampering or manipulation of data. Smart contracts and
frameworks adds the advantages of data reliability and open- DIDs further enhanced privacy and data security facilitating
ness on multiple datasets. While supply chain dynamics carry secure data sharing and collaboration among parties. The
challenges in maintaining predictive accuracy as shown in experimental analysis conducted in this study validated the
Fig. 10, it introduces constraints in model effectiveness when strengths and limitations of the approach. While the LSTM
integrating blockchain technology. However, the fusion of model demonstrated better performance in predicting critical
blockchain technology promises data reliability and openness outcomes, variations in predictive accuracy were observed
to the selected datasets. The incorruptible audit trail and across domains such as healthcare, logistics, and supply
transparent data exchange mechanisms serve as effective bar- chain management. The trade-off between local and global
riers to unauthorized data manipulations and hence support interpretability in SHAP methods indicated the need for

VOLUME 12, 2024 88199

 S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

better approaches to resolve the unique characteristics in time [11] Y. Lu, H. Wu, S. Qi, and K. Cheng, ‘‘Artificial intelligence in intensive care
series data. medicine: Toward a ChatGPT/GPT-4 way?’’ Ann. Biomed. Eng., vol. 51,
no. 9, pp. 1898–1903, Sep. 01, 2023, doi: 10.1007/s10439-023-03234-w.
The outcomes of our research offer practical insights [12] J. H. Yoon, M. R. Pinsky, and G. Clermont, ‘‘Artificial intelligence in
in real-time settings for decision-makers in the healthcare critical care medicine,’’ in Annual Update in Intensive Care and Emer-
profession and supply chain management. Our approach pro- gency Medicine. Cham, Switzerland: Springer, 2022, pp. 353–367, doi:
10.1007/978-3-030-93433-0_27.
vides a blueprint for handling crises effectively, exploiting [13] C.-C. Lee, T. Comes, M. Finn, and A. Mostafavi, ‘‘Roadmap towards
the power of data science and other allied technologies to responsible AI in crisis resilience management,’’ 2022, arXiv:2207.09648.
safeguard precious lives and livelihoods. Any future research [14] S. Ghaffarian, F. R. Taghikhah, and H. R. Maier, ‘‘Explainable artificial
intelligence in disaster risk management: Achievements and prospec-
efforts can be toward XAI techniques to manage the time tive futures,’’ Int. J. Disaster Risk Reduction, vol. 98, Nov. 2023,
series data and augment localized disturbances effectively. Art. no. 104123, doi: 10.1016/j.ijdrr.2023.104123.
Moreover, investigating alternative mechanisms and imple- [15] A. Bhattacharya. (2022). The Man From the Future: The Visionary Life of
John Von Neumann. Allen Lane. Accessed: Mar. 24, 2024. [Online]. Avail-
menting energy-efficient strategies in blockchain technology able: https://play.google.com/store/books/details?id=flh0zgEACAAJ
will enhance scalability and sustainability. [16] T. Guo, ‘‘Alan turing: Artificial intelligence as human self-knowledge,’’
Anthropology Today, vol. 31, no. 6, pp. 3–7, Dec. 2015, doi: 10.1111/1467-
CONFLICTS OF INTEREST 8322.12209.
[17] P. E. Ekmekci and B. Arda, ‘‘History of artificial intelligence,’’ in Artificial
The authors declare no conflict of interest.
Intelligence and Bioethics (SpringerBriefs in Ethics), Cham, Switzerland:
Springer, 2020, pp. 1–15, doi: 10.1007/978-3-030-52448-7_1.
ACKNOWLEDGMENT [18] H. Lu, Y. Li, M. Chen, H. Kim, and S. Serikawa, ‘‘Brain intelligence:
The authors extend their appreciation to the Deanship of Sci- Go beyond artificial intelligence,’’ Mobile Netw. Appl., vol. 23, no. 2,
entific Research at Saudi Electronic University for funding pp. 368–375, Apr. 2018, doi: 10.1007/s11036-017-0932-8.
[19] M. Haenlein and A. Kaplan, ‘‘A brief history of artificial intelligence: On
this research (9476). the past, present, and future of artificial intelligence,’’ California Manage.
Rev., vol. 61, no. 4, pp. 5–14, Jul. 2019, doi: 10.1177/0008125619864925.
REFERENCES [20] R. Vaishya, M. Javaid, I. H. Khan, and A. Haleem, ‘‘Artificial intelli-
[1] C. Sevim, A. Oztekin, O. Bali, S. Gumus, and E. Guresen, ‘‘Develop- gence (AI) applications for COVID-19 pandemic,’’ Diabetes Metabolic
ing an early warning system to predict currency crises,’’ Eur. J. Oper. Syndrome, Clin. Res. Rev., vol. 14, no. 4, pp. 337–339, Jul. 2020, doi:
Res., vol. 237, no. 3, pp. 1095–1104, Sep. 2014, doi: 10.1016/j.ejor. 10.1016/j.dsx.2020.04.012.
2014.02.047. [21] L. Chen, P. Chen, and Z. Lin, ‘‘Artificial intelligence in educa-
[2] Z. Zhao, D. Li, and W. Dai, ‘‘Machine-learning-enabled intelligence tion: A review,’’ IEEE Access, vol. 8, pp. 75264–75278, 2020, doi:
computing for crisis management in small and medium-sized enter- 10.1109/ACCESS.2020.2988510.
prises (SMEs),’’ Technol. Forecasting Social Change, vol. 191, Jun. 2023, [22] E. Glikson and A. W. Woolley, ‘‘Human trust in artificial intelligence:
Art. no. 122492, doi: 10.1016/j.techfore.2023.122492. Review of empirical research,’’ Acad. Manage. Ann., vol. 14, no. 2,
[3] T. Saeed, C. Kiong Loo, and M. Shahreeza Safiruz Kassim, ‘‘Artificial pp. 627–660, Jul. 2020, doi: 10.5465/annals.2018.0057.
intelligence based sentiment analysis for health crisis management in smart [23] R. Nishant, M. Kennedy, and J. Corbett, ‘‘Artificial intelligence for
cities,’’ Comput., Mater. Continua, vol. 71, no. 1, pp. 143–157, 2022, doi: sustainability: Challenges, opportunities, and a research agenda,’’
10.32604/cmc.2022.021502. Int. J. Inf. Manage., vol. 53, Aug. 2020, Art. no. 102104, doi:
[4] K. Jitkajornwanich, N. Vijaranakul, S. Jaiyen, P. Srestasathiern, and 10.1016/j.ijinfomgt.2020.102104.
S. Lawawirojwong, ‘‘Enhancing risk communication and environmen- [24] M. H. Mussa, ‘‘The impact of artificial intelligence on consumer
tal crisis management through satellite imagery and AI for air quality behaviors an applied study on the online retailing sector in Egypt,’’
index estimation,’’ MethodsX, vol. 12, Jun. 2024, Art. no. 102611, doi: Sci. J. Econ. Trade, vol. 50, no. 4, pp. 293–318, Dec. 2020, doi:
10.1016/j.mex.2024.102611. 10.21608/jsec.2020.128722.
[5] S. Andrews, T. Day, K. Domdouzis, L. Hirsch, R. Lefticaru, and [25] A. F. S. Borges, F. J. B. Laurindo, M. M. Spínola, R. F. Gonçalves,
C. Orphanides, ‘‘Analyzing crowd-sourced information and social media and C. A. Mattos, ‘‘The strategic use of artificial intelligence in the
for crisis management,’’ in Application of Social Media in Crisis Manage- digital era: Systematic literature review and future research direc-
ment. Cham, Switzerland: Springer, 2017, pp. 77–96, doi: 10.1007/978-3- tions,’’ Int. J. Inf. Manage., vol. 57, Apr. 2021, Art. no. 102225, doi:
319-52419-1_6. 10.1016/j.ijinfomgt.2020.102225.
[6] X. Song, H. Zhang, R. Akerkar, H. Huang, S. Guo, L. Zhong, Y. Ji, [26] S. Raisch and S. Krakowski, ‘‘Artificial intelligence and management: The
A. L. Opdahl, H. Purohit, A. Skupin, A. Pottathil, and A. Culotta, ‘‘Big automation–augmentation paradox,’’ Acad. Manage. Rev., vol. 46, no. 1,
data and emergency management: Concepts, methodologies, and applica- pp. 192–210, Jan. 2021, doi: 10.5465/amr.2018.0072.
tions,’’ IEEE Trans. Big Data, vol. 8, no. 2, pp. 397–419, Apr. 2022, doi: [27] M. K. I. Rahmani, M. Shuaib, S. Alam, S. T. Siddiqui, S. Ahmad, S. Bhatia,
10.1109/TBDATA.2020.2972871. and A. Mashat, ‘‘Blockchain-based trust management framework for
[7] F. Tena-Chollet, J. Tixier, A. Dandrieux, and P. Slangen, ‘‘Training cloud computing-based Internet of Medical Things (IoMT): A systematic
decision-makers: Existing strategies for natural and technological cri- review,’’ Comput. Intell. Neurosci., vol. 2022, pp. 1–14, May 2022, doi:
sis management and specifications of an improved simulation-based 10.1155/2022/9766844.
tool,’’ Saf. Sci., vol. 97, pp. 144–153, Aug. 2017, doi: 10.1016/j.ssci. [28] M. K. I. Rahmani, ‘‘Blockchain technology: Principles and algorithms,’’
2016.03.025. in Blockchain Technology and Computational Excellence for Society 5.0.
[8] M. Hassankhani, M. Alidadi, A. Sharifi, and A. Azhdari, ‘‘Smart city and Hershey, PA, USA: IGI Global, 2022, pp. 16–27.
crisis management: Lessons for the COVID-19 pandemic,’’ Int. J. Env-
iron. Res. Public Health, vol. 18, no. 15, p. 7736, Jul. 21, 2021, doi:
10.3390/ijerph18157736.
[9] J. F. Jensen, J. Ramos, M. Ørom, K. B. Naver, L. Shiv, G. Bunkenborg,
A. M. Kodal, and U. Skram, ‘‘Improving patient’s intensive care admission
through multidisciplinary simulation-based crisis resource management: A
qualitative study,’’ J. Clin. Nursing, vol. 32, nos. 19–20, pp. 7530–7542, SAMAH ALHAZMI received the Ph.D. degree from The University of
Oct. 2023, doi: 10.1111/jocn.16821. Manchester, U.K. She is currently an Associate Professor with the Com-
[10] A. Mahmoudian-Dehkordi and S. Sadat, ‘‘Sustaining critical care: Using puter Science Department, College of Computing and Informatics, Saudi
evidence-based simulation to evaluate ICU management policies,’’ Health Electronic University, Riyadh, Saudi Arabia. She is the Vice-Dean of the
Care Manage. Sci., vol. 20, no. 4, pp. 532–547, Dec. 2017, doi: Students Academic Affairs, College of Computing and Informatics. Her
10.1007/s10729-016-9369-z. research interests include artificial intelligence, the IoT, and Blockchain.

88200 VOLUME 12, 2024

S. Alhazmi et al.: Developing Intelligent and Immutable Vaccine Supply and Operation Platform

MOHAMMAD KHALID IMAM RAHMANI MD TABREZ NAFIS (Senior Member, IEEE) is

(Senior Member, IEEE) was born in Patharghatti, passionate about harnessing the power of technol-
Kishanganj, Bihar, India, in 1975. He received ogy for real-world impact, particularly in health-
the B.Sc. (Engg.) degree in computer engi- care informatics and big data. With 15 years of
neering from Aligarh Muslim University, India, experience in academia and research, he brings
in 1998, the M.Tech. degree in computer engineer- a comprehensive understanding of computer sci-
ing from Maharshi Dayanand University, Rohtak, ence and engineering to his role as an Associate
in 2010, and the Ph.D. degree in computer sci- Professor with Jamia Hamdard. His expertise lies
ence engineering from Mewar University, India, in the intersection of machine learning, health-
in 2015. From 1999 to 2006, he was a Lec- care, big data, and the IoT. Three Ph.D. scholars
turer with the Maulana Azad College of Engineering and Technology, have been successfully produced under his guidance and nine are pursuing
Patna. From 2006 to 2008, he was a Lecturer and a Senior Lecturer under his supervision/co-supervision. He has successfully guided more than
with the Galgotias College of Engineering and Technology, Greater Noida. 75 master’s students. He has extensive experience in academic adminis-
From 2010 to 2011, he was an Assistant Professor with MVN, Palwal. He is tration, including an Assistant Proctor, the Deputy Superintendent (Exam),
currently an Associate Professor with the Department of Computer Science, the In-Charge-Examination, Admission, Time Table, M.Tech. CSE Part-
College of Computing and Informatics, Saudi Electronic University, Riyadh, time Coordinator. He has published several research and conference papers
Saudi Arabia. He has published more than 75 research papers in journals indexed in SCI/Scopus. His publications include edited books and authored
and conferences of international repute, three book chapters, and holds books. He has four international patents and three national patents to his
one USA patent and another Australian patent of innovation. His research credit.
interests include algorithms, the IoT, cryptography, image retrieval, pattern He has nominated as the Nodal Officer for 5G Use Case Laboratory
recognition, machine learning, and deep learning. (Amount Rs. one crore) awarded by the Ministry of Telecommunications,
Government of India.

MOHAMMAD ARIF received the B.Tech. degree

in computer science and engineering from CCS
University, Meerut, India, in 2001, the M.Tech.
degree in computer science and engineering from
MNNIT, Allahabad, India, in 2008, and the Ph.D.
degree in computer science and engineering from
Integral University, Lucknow, India, in 2018.
He has nearly 22 years of teaching experience.
He is currently with the School of Computer
Science and Engineering, Vellore Institute of Tech-
nology, Vellore, Tamil Nadu. He has published nearly 40 peer-reviewed
papers in Scopus, other international journals, and IEEE and other inter-
national and national conferences. He authored one book and two book
chapters and holds two Indian patents. His research interests include Ad-hoc
networks, vehicular networks, cloud computing, mobile computing, security,
machine learning, and deep learning.

VOLUME 12, 2024 88201