Coronavirus disease (COVID-19) vaccination was adopted worldwide due to the advent of the COVID-19 pandemic in 2019. However, many post-vaccination adverse events, such as death and severe illness, were reported. So far, the specific case of post-vaccination adverse events pertaining to persons with cardiovascular risk factors and comorbidities has not been explored empirically, which limits the understanding of the underlying causes of adverse reactions to vaccination by this category of persons. This paper explored Explainable AI (XAI) to identify the critical determinants of post-vaccination mortality in persons with cardiovascular risk factors. To do this, we extracted 16657 records of persons with cardiovascular risk factors from the Vaccine Adverse Event Reporting System (VAERS) open dataset (from 2020 to May 2024). We then employed predictive modelling using a process that involved four stages. The first stage involved extracting relevant data from VAERS, data preprocessing, and handling class imbalance. In the second stage, we conducted a comparative performance evaluation of seven machine learning (ML) algorithms (Logistic Regression - LR, K-Nearest Neighbour - KNN, Deep Multilayer Perceptron - Deep MLP, Support Vector Machines - SVM, Random Forest - RF, Extreme Gradient Boosting XGBoost, and Categorical Boost - CatBoost). In the third stage, we compared the performance of two stacked ensemble models composed of six base models, using Catboost and XGBoost as the meta-learners in each case. The fourth stage involved using SHAPley Additive Explanations (SHAP) to interpret the predictions of the best-performing model. The result showed that CatBoost has the best performance among the base ML models (Acc = 0.96, F1=0.96, AUC = 0.96), while Stacked ensemble - XGBoost had the best overall performance (Acc = 0.96, F1=0.96, AUC = 0.99). Also, we found the important predictors of post-vaccination mortality in persons with cardiovascular comorbidity. Generally, older age, a higher number of days spent in the hospital increases the risk of mortality, while the absence of current illness, life-threatening condition, hospitalization, prolonged hospitalization, disability, birth defect, doctor visit, and emergency care; and vaccination dose completion will enhance the probability of survival. However, the presence of diabetes, high cholesterol, high blood pressure, and other illnesses increases the risk of mortality. This study's findings contribute to a better understanding of critical factors that could enable better handling of adverse events related to post-vaccination in persons with cardiovascular disease comorbidity.
Cardiovascular disease poses a serious risk to the quality of life of vulnerable populations, particularly
middle-aged and older individuals. There is a long history of significant morbidity, death, and financial
losses associated with cardiovascular illnesses and their consequences [
In December 2019, China saw the start of the global pandemic known as Coronavirus disease
(COVID19), which was brought on by SARS-CoV-2. Over 776 million confirmed cases and 7 million deaths were
reported worldwide by September 2024 [
Among the many instruments at our disposal for lowering morbidity and death, vaccinations are by
far the most successful [
Because Artificial Intelligence (AI) is being utilized to help healthcare professionals forecast patient outcomes, it is continuously transforming biomedical research and healthcare management [18]. It is dificult for the general public and medical professionals to understand the outcomes of AI systems, which makes their acceptability problematic. Explainable AI (XAI) is the product of research attempts to make the decisions or outcomes of AI systems clearly comprehensible. XAI aids in comprehending the reasoning behind the results of machine learning algorithms [19][20]. Finding possible weaknesses and locating the sources of errors more quickly can be facilitated by explaining how AI systems operate internally [21][22].
On the other hand, ethics and transparency are necessary for people to have faith in medical devices that use AI [23]. Furthermore, by highlighting the combination of rigorous validation of AI decisionsupport systems in real-life clinical settings and valuable explanations, explainability promotes ethically sound medical decision-making [24]. It follows that the healthcare sector is one of the sectors where user acceptability of AI algorithms depends on both explainability and accuracy [25]. Explainability can also be used to evaluate whether the system makes fair decisions and allows users to confirm that it does not rely on noise or artefacts in the training data, especially in situations where the training data may present a partial or biased image of the population [26]. Moreover, explanations can help us better grasp what the algorithm is optimized for and the associated trade-ofs, as well as provide fresh perspectives on what the AI system has learnt from the data [27]. This paper uses explainable AI to understand the post-vaccination mortality in persons with cardiovascular risk factors.
The rest of the paper is structured as follows: Section 2 discusses the related work. Section 3 presents the methodology. Section 4 presents the results of the research. The discussion was presented in Section 5. Section 6 discusses the limitations of the study while Section 7 presents the conclusion and further work.
Authors in [28] examined vaccination adverse events utilizing the Vaccine Adverse Event Reporting System (VAERS) database through the application of ontology and machine learning. In order to provide easy access to side efect information for patients, healthcare practitioners, and oficials, a relational/graph database was established. Machine learning algorithms also predict important symptoms that lead to hospitalization and treatment. Using the VAERS dataset, [29] developed a prognostic tool to determine the risks connected to COVID-19 vaccinations. Hospitalization, mortality, and COVID-19 outcomes were predicted using machine learning models such as Multilayer Feedforward Perceptron, Random Forest, Naive Bayes, Light Gradient Boosting Algorithm, and Linear Regression. Males between the ages of 50 and 70 and those with significant diseases were found to be the most vulnerable. Authors in [30] used patient data to identify common factors in adverse reactions and develop strategies to reduce their incidence. They found that factors such as prior illnesses, hospital admission, and SARS-CoV-2 reinfection were significantly associated with poor patient reactions. Preexisting conditions like age, gender, and medication use were also significant predictors. Machine learning classifiers trained with medical history were successful in predicting complication-free vaccinations with an accuracy score above 90%.
Authors in [31] employed machine learning to classify COVID-19 vaccination adverse efects in people with sensitivities to foods, animals, and weather. Machine learning classifiers have been utilized to predict bad responses to COVID-19 vaccinations in patients with allergies. This has been helpful in identifying those who are more likely to have side efects. Authors in [ 32] studied COVID-19 vaccine adverse events by gender, age, manufacturer, and dose using the Vaccine Adverse Event Reporting System datasets. They found higher frequency in women, but diferent characteristics of vaccine adverse events, including gender, manufacturer, age, and underlying diseases, were associated with fatal cases.
Authors in [33] propose a multi-label classification method for Vaccine Adverse Event (VAE) detection,
utilizing term- and topic-based label selection strategies. They use one-vs-rest, problem transformation,
algorithm adaptation, and deep learning methods. Experimental results show that topic-based PT
methods improve accuracy by up to 33.69%, OvsR methods achieve optimal accuracy of 98.88%, and AA
methods increase accuracy by 87.36%. The method enhances model accuracy and VAE interpretability.
Authors in [
In order to identify ongoing research eforts using machine learning techniques to comprehend comorbidity processes and make clinical predictions taking these intricate patterns into account, [18] conducted a review. Four predictive analytics tasks were identified: risk prediction, network analysis, clustering, and illness comorbidity data extraction. The results show that certain machine learningdriven applications interpret the model and identify important risk factors for the development of comorbidity while also addressing intrinsic data inadequacies in healthcare datasets.
Authors in [34] investigated the impact of COVID-19 immunization in intubated patients with acute respiratory distress syndrome associated with COVID-19 in Taiwan. Data on patients who were intubated between May 1, 2022, and October 31, 2022, owing to COVID-19 pneumonia were retrospectively evaluated by the authors. A person was deemed completely immunized if they received two or more doses of the vaccine. There were 84 patients in all (40 completely immunized and 44 controls). The two groups had comparable baseline characteristics on the day of intubation, such as age, comorbidities, and Sequential Organ Failure Assessment (SOFA) score. There was no statistically significant diference in the Intensive Care Unit (ICU) death rate between the completely vaccinated and control groups. Body mass index (BMI) and SOFA score had a strong correlation with ICU mortality.
Numerous adverse occurrences following vaccination, including fatalities and serious illnesses, have been documented in the literature [35][36][37]. This paper examines a particular instance that has not been experimentally studied: post-vaccination adverse events involving individuals who had comorbidities and cardiovascular risk factors. The findings of this paper will foster an understanding of the underlying factors that could lead to mortality when adverse events due to post-vaccination occur for this category of patients.
This section presents the details of methods/techniques used in predicting post-vaccination mortality in persons with cardiovascular risk factors using XAI. This was done in the following stages. Stage 1 contains data collection, preprocessing, and feature engineering. In stage 2, we have models’ selection, hyperparameter tuning, model training and model evaluation. Architecture design, hyperparameter tuning, training, and evaluation of the Stacked ensemble were done in stage 3. In stage 4, we provided model interpretation and explainability. The process workflow is presented in Figure 1.
The reported incidences of COVID-19 post-vaccination adverse events were extracted from the Vaccine Adverse Event Reporting System (VAERS), which can be accessed through the following link: https://vaers.hhs.gov/data/datasets.html was used in this study. The dataset over five (5) years from 2020 to May 2024 was downloaded in CSV format. Each year dataset has three (3) CSV files tagged VAERSDATA (history/profile), VAERSYMPTOMS (symptoms), and VAERSVAX (vaccine). All the three datasets for each year were merged and named as 2020COVIDCOMBINED, 2021COVIDCOMBINED, 2022COVIDCOMBINED, 2023COVIDCOMBINED, and 2024COVIDCOMBINED.
Based on our interest in examining patients with cardiovascular disease risk factors, we extracted data on instances where any two of high blood pressure, diabetes, and high cholesterol had been acknowledged as a preexisting condition by a patient who reported post-vaccination adverse events. We then selected the records from each year and then merged them for all the five (5) years under consideration. To get relevant patient records, each of the symptoms was expanded in terms of their medical terminologies as follows: High blood pressure/Hypertension; Diabetes/Type 1 Diabetes/Type 2 (includes instances where similar/closely related words like diabetic, diabetes mellitus, diabetes/Type 1 diabetic/Type 2 were used); High cholesterol/Hyperlipidemia/Hypercholesterolemia. We had a total of 16657 records with 52 features. Out of the 52 features, there were 34 categorical features and 18 numerical features.
The records of patients less than 18 years of age (11 records) were removed because the focus was on adult patients. Patient records with unknown vaccine (50 records) and patients with unknown vaccine dosage (1552 records) were also removed. Redundant features irrelevant to our analysis or empty fields were also dropped. We introduced additional fields, as shown in Table 1. For DOSE_COMPLETE, we input Y for a complete dose and N for an incomplete dose. For Moderna, Pfizer-Biontech, and Novavax vaccines, the patient must take 2 primary doses for a complete dosage. For Janssen, Pfizer-Biontech Bivalent, and Moderna Bivalent, only 1 primary dose is required.
For patients who took more than the required dose, we removed their records (3,843 records) because they did not follow the prescription.
We input Y for at least one symptom shown by the patient. Otherwise, N was inputted. Our focus was to look at the three cardiovascular disease risk factors (high blood pressure, diabetes, and high cholesterol). We created four additional fields for high blood pressure (HBP), diabetes (DIABETES), high cholesterol (H_CHOL), and other illnesses (OTHER_ILL) Since our focus was to look at the three cardiovascular disease risk factors (high blood pressure, diabetes, and high cholesterol, we extracted the value automatically from the patient's HISTORY field of the dataset Due to the nature of the dataset, which is usergenerated, there was no uniformity in the reporting by the users. Extracting whether a patient has other illnesses automatically becomes dificult since there are many other illnesses reported.
We manually checked the entries one after the other to determine whether a patient had other illnesses apart from at least two combinations of the three diseases of interest (high blood pressure, diabetes, and high cholesterol.
We represented each type of vaccine with A, B, C, D, E, and F for easy representation and analysis. A: JANSSEN; B: MODERNA; C: MODERNA BIVALENT; D: NOVAVAX; E: PFIZER-BIONTECH;
F: PFIZER-BIONTECH BIVALENT
To handle the missing values in the dataset, we used univariate imputation or data removal techniques as follows: 1. AGE_YRS: There are 173 cases of missing age values. We calculated the mean (67.24) and median (68) of the patients’ age from the existing observations. Since the mean and median are close, we picked the median value and inserted it in place of the missing observations.
There are 22 attributes consisting of 19 categorical features and 3 numerical features (See Table 2).
The categorical features were transformed using one-hot encoding, and Min-Max scaling was applied to the numerical features. LASSO Regression and Random Forests algorithms were used for feature selection. We split the data into train and test data in the ratio 70:30, respectively. The GridSearchCV was used to get the best hyperparameter for alpha needed for LASSO Regression. The result shows 1e-05 for alpha as the best parameter for the GridSearchCV. The two feature selection algorithms used were compared using log-likelihood and AUC-ROC scores. Lasso Log-Likelihood (DIED_N): -0.17485207557765708, (DIED_Y): -0.1748520755782875; Random Forest Log-Likelihood (DIED_1): 0.0006035625338038024 (DIED_2): -0.0006035625338038024 as shown in Figure 2. The AUC-ROC scores for LASSO Regression and Random Forests are presented in Figure 3.
The RF selected the top 19 features and LASSO also has 19 distinct selected features. The results of Log-likelihood (which shows the measure of goodness of fit), and AUC-ROC Scores showed that RF performed better than Lasso. Therefore, we chose features selected by RF for further analysis.
Based on the features selected, the original and distinct features from Excel in CSV format were then loaded. At the end of the features selection, we had a total of 20 distinct features: 18 categorical variables (including the target variable) and 2 numerical variables, as shown in Table 3. Table 3 provides the selected features and their descriptions.
Frequencies 4835, 6242 326, 10751 4160, 6917 11, 11066 313, 10764 3913, 4493, 2671 1752, 9325 11040, 37 6958, 4119 6947, 4130 9142, 1935 516, 10561 6, 11071 2870, 8207 2492, 8585 5317, 5760 768, 5089, 35, 2, 5141, 42 1035, 10042 Data value Male/Female Yes/No Yes/No Yes/No Yes/No Yes/No/Unknown Yes/No Yes/No A/B/C/D/E/F DIED
Vaccine recipient died or not Numerical variable
Feature name/description AGE_YRS HOSPDAYS
The age of the vaccine recipient in years Number of days Hospitalised
Yes/No Data range 18-103 0-91
Mean 66.62 2.00
Standard deviation 13.29 5.02
Since our data is largely skewed based on our target variable, we have a class imbalance as presented in Figure 4 with Alive (N) = 10042 and Died (Y) = 1035 records.
We used Random Oversampling and SMOTE techniques to balance the datasets. We compared the performance of Random Oversampling with SMOTE using AUC_ROC and Average Precision metrics. SMOTE performed better in terms of AUC-ROC and Average Precision. Hence, SMOTE was selected as the resampling technique used to balance the dataset (see Table 4).
We trained seven (7) ML algorithms on the dataset. The algorithms are Logistic Regression (LR), K-Nearest Neighbour (KNN), CatBoost, XGBoost, Support Vector Machine (SVM), Deep Multilayer Perception (DMLP), and Random Forests. We also trained two stacked ensembles, each of these had 6 base models and a meta-learner. The stacked ensembles are Stack Ensemble-CatBoost (CatBoost as the meta-learner) and Stack Ensemble-XGBoost (XGBoost as the meta-learner). We had a total of 20084 records after resampling with SMOTE. We split the data into training and testing sets in the ratio of 70:30, that is (14058: 6026), respectively.
To select the best hyperparameters for the seven models and the stacked ensembles, we performed hyperparameter tuning for the seven models and the two stack ensembles using GridSearchCV and Bayesian Optimization, respectively. For the stack ensembles, the best hyperparameters were selected out of fifty (50) trials. The best hyperparameters for each model are presented in Table 5.
In the hyperparameter optimization, we employed k-fold cross validation, which averages the scores across all iterations to get the final assessment of the models. One of the advantages of the ensemble techniques which was used in this study is to address the overfitting issue. In addition, the evaluation of the models was performed on the test set of the data used in this study.
In this study, we have explored the post-vaccination adverse event for patients with cardiovascular comorbidity indicators by targeting those who are alive or dead. We used the following metrics to assess the performance of the seven algorithms and the two stacked ensembles: Accuracy, Precision, Recall, F1, AUC, and MCC. We also plot the confusion matrix and ROC curve for the algorithms. SHAP was also used to determine the importance of the features and their impact on the best-performing model. The results of the seven models and two stack ensemble models are presented in Table 6.
From Table 6, CatBoost had the best performance among all the individual models that were trained and tested. As a result, CatBoost was used as a meta-model for the stacked ensemble, while the remaining six models were used as base models. However, we noticed a slight drop in the performance of Stack ensemble-CatBoost compared to the performance of CatBoost as an individual model. We went on to pick the individual model that performed next to CatBoost, which was XGBoost, as the meta-learner of the stacked ensemble, and the result showed a superior performance.
From Figure 5, the receiver operating characteristic (ROC) curves for all the models are shown. The ROC is used to evaluate the performance of a binary diagnostic classification method. The ROC tells how much the model is capable of distinguishing between classes. All the models show outstanding performance of more than 0.9, except in the logistic regression model, which has 0.89. The ROC results show that the models can clearly distinguish between the two classes, Alive or Died, in the case of this study.
SHAPley Additive ExPlanations (SHAP) is used to explain the output of a machine learning model. It uses game theory to assign credit for a model’s prediction to each feature or feature value [38][39]. Out of all the seven (7) base models, the performance of the CatBoost model was the best. The CatBoost model’s predictions on the test data were used as input to the SHAP for the model’s predictions’ explanations. The SHAP summary plot is presented in Figure 6.
From Figure 6, the Y-axis indicates the feature names in order of importance from top to bottom. The X-axis represents the SHAP value, which indicates the degree of change in log odds. The colour of each point on the graph represents the value of the corresponding feature, with red indicating high values and blue indicating low values. Each point represents a row of data from the original dataset. From Figure 6, it is clear that the feature ‘AGE_YRS’ has the highest impact on the model while the feature ‘HBP’ has the least impact.
We identified the significant determinants of post-vaccination mortality in individuals with cardiovascular comorbidities from the SHAP summary plot. The risk of death generally rises with age and the number of days spent in the hospital; on the other hand, the likelihood of survival increases with the absence of current illness, life-threatening conditions, hospitalization, prolonged hospitalization, disability, birth defects, doctor visits, and emergency care. On the other hand, the risk of death is raised by diabetes, high blood pressure, high cholesterol, and other conditions. The results of this study add to our knowledge of important variables to consider when managing post-vaccination adverse events for patients who also have cardiovascular disease comorbidity indicators.
Authors in [40] identified the top ten factors that increase the risk of cardiovascular disease (CVD) as unhealthful nutrition, physical inactivity, dyslipidemia (high cholesterol), hyperglycemia (high blood sugar), high blood pressure (hypertension), obesity, considerations of select populations (older age, race/ethnicity, and sex diferences), thrombosis/smoking, kidney dysfunction and genetics/familial hypercholesterolemia. This study investigates how COVID-19 post-vaccination adverse events afect persons with CVD risk factors, particularly focusing on the co-occurrences of dyslipidemia (high cholesterol), hyperglycemia (high blood sugar), diabetes, and high blood pressure (hypertension). Our results show that the risk of death from COVID-19 post-vaccination adverse events increases for persons with diabetes, high blood pressure, high cholesterol, and other illness conditions. This observation agrees with the results of previously reported clinical studies.
According to [41], based on a study on 187 COVID-19 patients in which 144 patients survived and 43 patients died, found that the probability of death was higher (69.44%) for those with underlying CVD and elevated TnTs (troponin T) levels. Also, patients with underlying CVD were more likely to exhibit elevation of TnT levels (54.5%) compared with the patients without CVD (13.2%). In addition, patients with elevated TnT levels had evidence of more severe respiratory dysfunction and developed more frequent complications. Also, [32], a systematic review study that captured findings from 3912 participants, found that patients with preexisting CVD had worse outcomes and increased risk of death from COVID-19. Also, CVD risk factors such as hypertension, diabetes mellitus, and obesity were associated with the severity of COVID-19 infection, intensive care unit admission and poor prognosis. Thus, the results of this study add to our knowledge of important variables to consider when managing post-vaccination adverse events for patients who also have cardiovascular disease comorbidity indicators.
The VAERS dataset used in this research work is user generated information and there was no validation done on the reports. The type of features in the dataset also constitutes a limitation as most of the features extracted from the dataset are categorical variables; the dataset did not have the actual values of cholesterol, blood pressure, glycemia values as numerical variables. The absence of these numerical variables limit the depth of computational analysis that is possible.
This paper explores the critical determinants of mortality in post-vaccination adverse events for persons with cardiovascular disease risk factors using XAI. We did this by extracting 16657 records (from 2020 to May 2024) of COVID-19 vaccinated persons with preexisting cardiovascular disease risk factors (any two of high blood pressure, diabetes, and high cholesterol) from the VAERS public dataset. Next, we applied a predictive modelling process that has four stages, which are 1) data preprocessing; 2) model training and performance evaluation of seven ML algorithms: Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Categorical Boosting (CatBoost), Random Forest (LR), and Extreme Gradient Boosting (XGBoost); 3) modelling using two stacked ensembles consisting of six base models — Using Catboost and XGBoost as the meta-learners respectively; 4) model explainability using SHAP.
The results show that Stacked ensemble - XGBoost had the best overall performance (Acc = 0.96, F1 = 0.96, AUC = 0.99), whereas CatBoost has the best performance among the individual machine learning models (Acc = 0.96, F1 = 0.96, AUC = 0.96). We also identified the critical determinants of mortality in persons with cardiovascular disease risk comorbidity when a post-vaccination adverse event occurs. The risk of death generally rises with age and the number of days spent in the hospital; on the other hand, the likelihood of survival increases with the absence of current illness, life-threatening conditions, hospitalization, prolonged hospitalization, disability, birth defects, doctor visits, and emergency care. Also, the risk of death is raised by diabetes, high blood pressure, high cholesterol, and other conditions. The results of this study foster an understanding of critical factors that could enable better handling of adverse events related to post-vaccination in patients with cardiovascular disease comorbidities.
In further research, we will explore the severity of COVID-19 post-vaccination adverse events in patients with cardiovascular disease risk factors. We will also try to explore real-life datasets on the same scenario proposed in this study.
The research was supported by the National Research Foundation (NRF) of South Africa, and the University of Pretoria, South Africa. [13] W. W. Leitner, H. Ying, N. P. Restifo, DNA and RNA-based vaccines: principles, progress and prospects, Vaccine 18(9-10) (1999) 765-777. doi: 10.1016/s0264-410x(99)00271-6. [14] N. P. Restifo, H. Ying, L. Hwang, W. W. Leitner, The promise of nucleic acid vaccines, Gene Ther.
7(2) (2000) 89-92. doi: 10.1038/sj.gt.3301117. [15] Centers for Disease Control and Prevention, Vaccines and immunizations: viral vector COVID-19 vaccines. Atlanta, GA: Centers for Disease Control and Prevention (2021). www.cdc.gov/vaccines/ covid-19/hcp/viral-vector-vaccine-basics.html. [16] M. H. Paknahad, F. B. Yancheshmeh, A. Soleimani, Cardiovascular complications of COVID-19 vaccines: a review of case-report and case-series studies, Heart Lung 59 (2023) 173-180. doi: 10.1016/j.hrtlng.2023.02.003. [17] Z. Akhtar, M. Trent, A. Moa, T. C. Tan, O. Fröbert, C. R. MacIntyre, The impact of COVID19 and COVID vaccination on cardiovascular outcomes, European Heart Journal Supplements 25(Supplement_A) (2023) A42–A49. https://doi.org/10.1093/eurheartjsupp/suac123. [18] D. Xu, Z. Xu, Machine learning applications in preventive healthcare: A systematic literature review on predictive analytics of disease comorbidity from multiple perspectives, Artificial Intelligence in Medicine 156 (2024) 102950. https://doi.org/10.1016/j.artmed.2024.102950. [19] S. Ali, T. Abuhmed, S. El-Sappagh, K. Muhammad, J. M. Alonso-Moral, R. Confalonieri, R. Guidotti, J. Del Ser, N Díaz-Rodríguez, F. Herrera, Explainable artificial intelligence (XAI): What we know and what is left to attain trustworthy artificial intelligence. Information Fusion 99 (2023) 101805. https://doi.org/10.1016/j.infus.2023.101805 [20] T. Kolajo, O. Daramola, Human-centric and semantics-based explainable event detection: a survey,
Artificial Intelligence Review 56 (2023) 119-158. https://doi.org/10.1007/s10462-023-10525-0 [21] J. Amann, D. Vetter, S. N. Blomberg, H. C. Christensen, M. Cofee, S. Gerke S, To explain or not to explain?—Artificial intelligence explainability in clinical decision support systems. PLOS Digit Health 1(2) (2022) e0000016. https://doi.org/10.1371/journal.pdig.0000016 [22] U. Bhatt, A. Xiang, S. Sharma, A. Weller, A. Taly, Y. Jia, J. Ghosh, R. Puri, J. M. F. Moura, P. Eckersley, Explainable machine learning in deployment. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency pp.648 - 657 [23] L. Farah, J. M. Murris, I. Borget, A. Guilloux, N. M. Martelli, S. I. Katsahian, Assessment of performance, interpretability, and explainability in artificial intelligence–based health technologies: what healthcare stakeholders need to know. Mayo Clinic Proceedings: Digital Health 1(2) (2023) 120-138. https://doi.org/10.1016/j.mcpdig.2023.02.004 [24] Gerdes, A. The role of explainability in AI-supported medical decision-making. Discov Artif Intell 4 (2024) 29. https://doi.org/10.1007/s44163-024-00119-2. [25] V. Hassija, V. Chamola, A. Mahapatra, A. Singal, D. Goel, K. Huang, S. Scardapane, I. Spinelli, M.
Mahmud, A. Hussain, Interpreting black-box models: a review on explainable artificial intelligence, Cogn Comput 16 (2024) 45-74. https://doi.org/10.1007/s12559-023-10179-8. [26] T. Kirat, O. Tambou, V. Do, A. Tsoukiàs, Fairness and explainability in automatic decision-making systems. a challenge for computer science and law, EURO Journal on Decision Processes 11 (2022) 100036. https://doi.org/10.1016/j.ejdp.2023.100036. [27] M. M. Soliman, E. Ahmed, A. Darwish, A. E. Hassanien, Artificial intelligence powered metaverse: analysis, challenges and future perspectives. Artif Intell Rev 57 (2024) 36. https://doi.org/10.1007/ s10462-023-10641-x. [28] Z. Liu, X. Gao, C. Li, Modeling COVID-19 vaccine adverse efects with a visualized knowledge graph database, Healthcare 10 (2022) 1419. https://doi.org/10.3390/healthcare10081419. [29] H. Gupta, O. M. Verma, Vaccine hesitancy in the post-vaccination COVID-19 era: a machine learning and statistical analysis driven study, Evolutionary Intelligence 16 (2023) 739–757 https: //doi.org/10.1007/s12065-022-00704-3. [30] M. Ahamad, S. Aktar, J. Uddin, R. Al-Mahfuz, A. K. M. Azad, S. Uddin, S. A. Alyami, I. H. Sarker, A.
Khan, P. Liò, J. M. W. Quinn, M. A. Moni, Adverse efects of COVID-19 vaccination: machine learning and statistical approach to identify and classify incidences of morbidity and postvaccination reactogenicity, Healthcare 11 (2023) 31. https://doi.org/10.3390/healthcare11010031.