Heart Disease Prediction Using Statistical Feature Selection and Interpretable Machine Learning

Summary

This protocol describes a machine learning framework for heart disease prediction that combines data augmentation using generative adversarial networks, statistical and metaheuristic-based feature selection, and explainable artificial intelligence.

Abstract

Heart disease is a major cause of death worldwide, making its early prediction an important clinical and computational issue. Several studies have addressed challenges such as data scarcity, feature selection, and model interpretability individually, but fewer studies have proposed an integrated framework that addresses these challenges in a synergistic manner. This paper presents a comprehensive predictive framework that uses: (1) a generative adversarial network (GAN) to address class imbalance and data scarcity; (2) a hybrid feature selection approach that combines statistical pre-filtering via Welch’s t.-test and Cohen’s d effect size, along with metaheuristic optimization via Harris Hawk Optimization; and (3) various explainable artificial intelligence methods, including SHAP, partial dependence plots, and odds ratios. This framework was assessed on the Cleveland and Statlog datasets, yielding strong accuracy, F1-scores, and ROC-AUC values compared with selected baselines and existing methods. The model provides a robust, interpretable computational framework for heart disease prediction, linking machine learning performance with clinical interpretability.

Introduction

Cardiovascular diseases are one of the leading causes of morbidity and mortality worldwide, accounting for an estimated 17.9 million deaths per year¹. Early and accurate prediction of heart disease is important for timely intervention and improved patient outcomes. In this context, heart disease prediction using machine learning (ML) algorithms faces three main challenges: the limited availability of high-quality medical data, high-dimensional feature spaces containing redundant or irrelevant variables, and the black-box nature of complex models, which can hinder clinical trust and adoption². Recent work has combined ML with explainable artificial intelligence (XAI) methods for heart disease prediction³. Many researchers have also used ML for heart disease prediction and detection⁴. Recent developments in generative artificial intelligence, especially generative adversarial networks (GANs), hold promise for data augmentation in healthcare⁵. Simultaneously, metaheuristic algorithms such as Harris Hawk Optimization (HHO) and Particle Swarm Optimization (PSO) have proven effective in feature selection and model optimization⁶. XAI techniques, such as SHAP and partial dependence plots (PDPs), have also emerged as important tools for interpreting complex model predictions⁷. Many studies of ML models for cardiovascular risk prediction have been conducted⁸.

However, the available literature often discusses these issues in isolation. Some studies focus on data augmentation using GANs⁹, while others focus specifically on feature selection using metaheuristic algorithms¹⁰ or on model interpretability based on XAI methods¹¹. SMOTE-based augmentation has been explored for heart failure survival prediction¹². KNN-based cardiac disease diagnosis has also been reported¹³. These separate approaches do not fully harness the combined benefits that can be achieved through an integrated framework that addresses data scarcity, feature selection, model training, and interpretability together.

Recent studies have explored related approaches. A 2026 study in Frontiers in Medicine proposed PSO-optimized heterogeneous classifiers with padding interpolation and median imputation for heart disease diagnosis, achieving 91.3% accuracy on a merged dataset¹⁴. Other recent work has applied metaheuristic optimization for medical image segmentation¹⁵, XAI for stroke prediction¹⁶, hybrid optimization for cardiac arrhythmia classification¹⁷, and SHAP-enhanced clinical decision support systems¹⁸. However, few of these studies combine generative augmentation, dual-criteria statistical feature selection with HHO optimization, and multi-method XAI in a single integrated framework.

This paper aims to address this gap by proposing a heart disease prediction framework that systematically integrates data augmentation, hybrid feature selection, model optimization and training, and explainability analysis. In the data augmentation stage, GANs are used to synthesize tabular clinical data based on patient features such as age, blood pressure, cholesterol levels, and electrocardiogram measurements. While GANs are widely used for medical image generation, this study applies them to the Cleveland Heart Disease dataset, which contains 13 numerical and categorical features, to address the limited sample size (n = 303) and class imbalance. In the hybrid feature selection stage, Welch’s t.-test and Cohen’s d effect size are combined with HHO to identify statistically robust and clinically relevant feature subsets. During the model optimization and training stage, PSO is used to optimize the weights of the artificial neural network, while Logistic Regression and Random Forest models are trained due to their balance between performance and explainability. In the explainability stage, complementary XAI techniques, including SHAP, PDPs, and odds ratios, are used to provide global and local model interpretations.

The overall workflow of the proposed framework is illustrated in Figure 1. Table 1 summarizes the key differences between the proposed approach and existing feature selection methods [Table 1 here].

Figure 1: Overview of the proposed heart disease prediction framework. The workflow comprises four main phases: (1) data preprocessing and augmentation using GANs to address data scarcity; (2) hybrid feature selection combining statistical filtering (Welch’s t.-test with Cohen’s d) and Harris Hawk Optimization; (3) model training with interpretable classifiers, including Logistic Regression and Random Forest, and a PSO-optimized ANN; and (4) explainability analysis using SHAP, partial dependence plots, and odds ratios. Abbreviations: GANs = generative adversarial networks; PSO = Particle Swarm Optimization; ANN = Artificial Neural Network. Please click here to view a larger version of this figure.

Approach Category	Statistical Testing (e.g., t-test)	Effect Size (e.g., Cohen's d)	Metaheuristic Optimization (e.g., HHO/PSO)	Interpretability Focus
Traditional Statistical	Yes	Rarely	No	Moderate
Pure Optimization	No	No	Yes	Low
Existing Hybrid Methods	Sometimes	Rarely	Yes	Variable
Proposed Framework	Yes (Welch’s t-test)	Yes (Cohen’s d ≥ 0.5)	Yes (HHO)	High (XAI-integrated)

Table 1: Comparison of feature selection approaches in heart disease prediction. Approaches compared include Traditional Statistical, Pure Optimization, Existing Hybrid Methods, and the Proposed Framework across the following criteria: Statistical Testing, Effect Size, Metaheuristic Optimization, and Interpretability Focus.

The main contributions of this work are as follows. First, to address data scarcity and class imbalance, a standard GAN with binary cross-entropy loss and Adam optimization is implemented, along with a Gaussian perturbation fallback when TensorFlow is unavailable. Second, to address feature redundancy, a hybrid feature selection strategy is proposed that combines statistical pre-filtering with HHO using a V-shaped transfer function. This dual-criteria approach aims to select features that are both statistically significant and clinically relevant. Third, to address model opacity, a multi-method explainability suite is integrated, including SHAP beeswarm and waterfall plots, PDPs, and odds ratios with 95% confidence intervals. A simple reconciliation protocol is provided for clinical users: if a PDP shows a non-linear trend, the SHAP explanation should be prioritized over logistic regression coefficients. Fourth, to support reproducibility and structured validation, the framework includes stratified cross-validation, fairness auditing, ablation studies, an external validation protocol for MIMIC-III, and documentation of key hyperparameters.

Protocol

Ethics statement, dataset, software, and data preparation
The findings of this study were based on the UCI Machine Learning Repository’s Heart Disease dataset. As this is a publicly accessible and de-identified resource, its use did not require ethical committee approval. The authors also verify the originality of this manuscript, confirming that it has not been previously published or submitted to other journals.

The Cleveland Heart Disease dataset was split into training and test sets at 80/20. The dataset typically contains 303 instances; therefore, approximately 242 samples were used for training, and 61 samples were retained as a clean test set. Synthetic samples generated by the GAN or Gaussian fallback method were added only to the training data to reduce the risk of data leakage. The final augmented training set consisted of approximately 242 real samples and 1,000 synthetic samples, giving 1,242 training samples. No fixed static validation set was used. Instead, stratified cross-validation was applied during model training, with each fold splitting the augmented training data into training and validation subsets.

The dataset was loaded into a Pandas DataFrame and inspected for missing values. Numerical features with missing values were handled using median imputation with the SimpleImputer class from scikit-learn using strategy = 'median'. Categorical features with missing values were handled using mode imputation with SimpleImputer using strategy = 'most_frequent'. Missingness mechanisms were documented by calculating the missing percentage for each feature using df.isnull().sum() / len(df). Non-random missingness patterns were assessed by comparing the mean values of other features between samples with and without missing data using t.-tests for numerical features and chi-square tests for categorical features. Missingness was then documented as Missing Completely at Random (MCAR), Missing at Random (MAR), or Missing Not at Random (MNAR), where applicable.

For datasets with substantial missingness, a sensitivity analysis was recommended by comparing median/mode imputation with Multiple Imputation by Chained Equations (MICE), using fancyimpute.IterativeImputer with max_iter = 10, and KNN imputation, using fancyimpute.KNN with k = 5. An accuracy difference of less than 0.03 was treated as indicating robustness to the imputation method¹². This sensitivity analysis was considered optional for the Cleveland dataset because of its limited missingness but was recommended for other clinical datasets with more than 5% missing values. Missingness patterns were also visualized using the missingno library by generating a missingness matrix heatmap with msno.matrix(df). Clustering of missingness patterns was used to identify whether missing values co-occurred systematically, which may indicate MNAR mechanisms requiring clinical expert input.

Numerical features were standardized using z-score normalization. StandardScaler from scikit-learn was fitted on the training data and then applied to both training and test sets. Categorical variables were encoded using one-hot encoding. Chest pain type (cp), which contains four categories, was converted into four binary indicator columns using pandas.get_dummies. Thalassemia (thal), which contains three categories, was converted into three binary indicator columns. Because the GAN generator and discriminator used a fixed input/output dimension of 13 features corresponding to the original dataset before one-hot encoding, synthetic samples were generated in the original 13-feature space and then passed through the same one-hot encoding pipeline as the real data. This preserved compatibility with the GAN architecture while allowing encoded features to be used for model training.

Mathematical definitions and quality metrics
The Fréchet distance was used to compare real and synthetic feature distributions. The Fréchet distance Fr(F, G) between two distributions F and G was defined as follows:

Fr2(F,G)=minX,YE|X-Y|2  (1)

where E represents the expectation, and the minimization is taken over all random variables X and Y having distributions F and G, respectively¹⁹.

Harris Hawk Optimization (HHO) was used as the metaheuristic optimization method. HHO is inspired by the cooperative hunting behavior of Harris hawks²⁰. The transition between exploration and exploitation phases was controlled by the escape energy E. In the exploration phase, where |E| ≥ 1, the update was defined as follows:

X(t+1) = X_rand (t) - r₁ | X_rand (t) - 2r₂X (t)|

In the exploitation phase, where |E| < 1, updates were determined by the escape energy E = 2E(1 − t/T) and jump strength J = 2(1 − r₅). In the soft besiege condition, where r ≥ 0.5 and |E| ≥ 0.5, the update was defined as follows:

X(t+1) = ΔX(t) - E|JX_rabbit  (t) - X(t)|

In the hard besiege condition, where r ≥ 0.5 and |E| < 0.5, the update was defined as follows:

X(t+1) = X_rabbit (t) - E|ΔX(t)|

Fairness was assessed using statistical parity and error-rate balance, following Hardt et al.²⁰ and Lima et al.²¹. Demographic parity difference, equalized odds difference, and age-based calibration error were used as fairness metrics.

ΔEO = max(|TPR_A - TPR_B |,| FPR_A - FPR_B |)

ΔBS=|BS_age<50 - BS_age>50 |

where BS is the Brier Score:

Dashboard interpretability was based on SHAP values, which are calculated using coalitional game theory18.

where Φ_i represents the SHAP attribution for feature i, F. f(S) represents the set of all features, and represents the model prediction for a subset of features S.

GAN-based data augmentation
To address data scarcity and class imbalance, a Generative Adversarial Network (GAN) was used to generate synthetic samples. The generator architecture was configured in TensorFlow/Keras. It accepted a 100-dimensional noise vector sampled from a standard normal distribution N(0,1), followed by dense layers with 128, 256, and 512 units using ReLU activation. The output layer contained 13 units, corresponding to the original feature dimension, and used sigmoid activation.

The discriminator architecture accepted a 13-dimensional feature vector as input. It consisted of dense layers with 512, 256, and 128 units using LeakyReLU activation with α = 0.2. The output layer contained one unit with sigmoid activation for binary classification of real versus synthetic samples.

The GAN was trained for 100 epochs using a batch size of 64. The Adam optimizer was used with a learning rate of 0.0002, β1 = 0.5, and β2 = 0.999. During each epoch, the discriminator was alternately trained on real and synthetic batches, and the generator was trained to fool the discriminator. After training, 1,000 random noise vectors were input into the generator to produce 1,000 synthetic samples, which were added only to the training set.

Mode collapse was monitored during GAN training by measuring the variance of each synthetic feature across 100 generated samples after every 10 epochs. If the variance of any feature dropped below 10% of the corresponding real-data variance for three consecutive checks, mode collapse was suspected. Mitigation strategies included reducing the learning rate to 1 × 10⁻⁴, increasing the batch size to 128, restarting training with different weight initialization, or replacing the standard GAN with Wasserstein GAN with Gradient Penalty (WGAN-GP), as described by Arjovsky et al.¹⁷. The implementation used a standard GAN with a Gaussian perturbation fallback to ensure synthetic data generation when TensorFlow was unavailable.

Synthetic data quality was evaluated by computing the Fréchet Distance between real and synthetic feature distributions using a custom implementation. A classifier, such as logistic regression, was also trained to distinguish real from synthetic samples; near-chance classification accuracy was treated as indicating high fidelity. Precision-Recall AUC was calculated, with values greater than 0.9 considered indicative of good distribution capture. Pearson correlations between feature pairs in the real and synthetic datasets were also compared, with differences below 0.05 treated as acceptable preservation of correlation structure.

When TensorFlow/Keras was unavailable or GAN training failed, a Gaussian perturbation fallback method was used. For each class, the mean (µ) and standard deviation (σ) of each feature were computed from the training set. Synthetic samples were then generated as follows:

Xsynthetic = µ + ε × σ × 0.05, where ε ~ N(0,1)

Class labels were generated in proportion to the original class distribution. This fallback was included to support reproducibility across environments without deep learning dependencies.

Hybrid feature selection
A two-stage hybrid feature selection strategy was applied. In the first stage, statistical pre-filtering was performed. For each feature xi in the feature set X, values were divided into two groups according to the binary outcome variable: G0 for y = 0, indicating no disease, and G1 for y = 1, indicating disease presence. Welch’s two-sample t.-test was performed using scipy.stats.ttest_ind with equal_var = False. Cohen’s d effect size was then computed as follows:

d = (mean1 − mean2) / pooled_std

where:

pooled_std = sqrt((std1² + std2²)/2)

The feature name, p-value, and Cohen’s d value were stored in a results table. Features were selected if they satisfied both criteria: p-value < 0.05 and |Cohen’s d| ≥ 0.5. The resulting feature set was defined as Xfiltered.

Welch’s t-test was used because it is appropriate for continuous numerical features such as age, thalach, and oldpeak. For binary categorical features such as sex and exang, the t.-test produces results comparable to a proportion test when comparing two groups. Multi-categorical features such as cp and thal were one-hot encoded, and each binary indicator was tested individually against the outcome variable. This approach was considered appropriate because the Cleveland dataset has more than 30 samples, the features were standardized before analysis, and equal_var = False accounts for unequal variances between groups. For features with severe violations of normality, the Mann-Whitney U test was considered as an alternative non-parametric test.

The threshold p < 0.05 followed conventional statistical significance, while |Cohen’s d| ≥ 0.5 corresponded to a moderate-to-large effect size. For datasets with small sample sizes or rare outcomes, bootstrap-based adjustment, Hedges’ g correction, or relaxed exploratory thresholds were recommended with expert clinical input. For example, 1,000 bootstrap resamples could be used to compute Cohen’s d confidence intervals, and Hedges’ g could be applied to correct for small-sample bias. Features with borderline statistics, such as p-values between 0.03 and 0.08 or |d| values between 0.4 and 0.6, were documented for possible clinical expert review before exclusion.

In the second stage, Harris Hawk Optimization (HHO) was applied to the statistically filtered feature set. The HHO population size was set to 20, and the maximum number of iterations was set to 50. Each solution was represented as a binary vector with length equal to the number of features in Xfiltered, where 1 indicated that a feature was selected and 0 indicated that it was not selected. Continuous HHO positions were mapped to binary vectors using the V-shaped transfer function:

T(x) = |tanh(x)|

The binary value was set to 1 if T(x) > 0.5 and 0 otherwise. The V-shaped function was selected because it supports balanced exploration and exploitation during binary conversion.

The fitness function for each solution was defined using logistic regression. A logistic regression model was trained using only the features selected by the binary vector, and 5-fold cross-validation was performed using cross_val_score from scikit-learn. Fitness was computed as follows:

fitness = 1 − mean accuracy

The population of hawk positions was initialized uniformly in the range [−1, 1] using numpy.random.uniform(−1, 1, (population_size, n_features)) with a fixed random seed of 42 for reproducibility. At each iteration, the fitness of all hawks was evaluated, the best hawk position was identified as the rabbit, and hawk positions were updated using the HHO exploration and exploitation equations based on escape energy. After convergence, the best-performing binary vector was selected as the final feature subset, Xfinal.

The selected features were recorded from a single HHO optimization run with a fixed random seed. For applications requiring higher statistical confidence, 30 independent runs with different random seeds were recommended, and consensus features appearing in at least 80% of runs could be selected. The reported implementation was based on a single representative run, as preliminary testing indicated consistent convergence.

Model training and optimization
Three models were considered: Logistic Regression, Random Forest, and a PSO-optimized Artificial Neural Network (ANN). Logistic Regression was trained using stratified 5-fold cross-validation to maintain class distribution. The regularization strength C was optimized using the search space C [0.001, 0.01, 0.1, 1, 10]. For each fold and each value of C, the model was trained on the training fold and evaluated on the validation fold. The value of C that maximized mean validation accuracy across folds was selected.

The Random Forest model was trained using hyperparameter tuning. The search space included max_depth = [5, 10, 15, None] and min_samples_split = [2, 5, 10]. Grid search with 5-fold cross-validation was performed using ROC-AUC as the optimization metric through GridSearchCV with scoring = 'roc_auc'. The selected Random Forest model used max_depth = 10 and min_samples_split = 5. Out-of-bag (OOB) score estimation was enabled using oob_score = True.

Overfitting was assessed by calculating the gap between training accuracy and OOB score:

overfitting_gap = training_accuracy − oob_score

An overfitting gap below 0.05 was considered indicative of good generalization, while a gap greater than 0.10 indicated the need to reduce max_depth or increase min_samples_split. With an OOB score of 0.9296 and typical training accuracy between 0.94 and 0.96, the gap was approximately 0.01–0.03.

An ANN classifier was also optimized using Particle Swarm Optimization (PSO). The ANN architecture consisted of an input layer, one hidden layer with 64 neurons using ReLU activation, and an output layer with one neuron using sigmoid activation. PSO was initialized with 50 particles and 50 iterations and was used to optimize the initial network weights. The ANN was then trained using standard backpropagation. Because PSO optimization was performed on the same training data without nested cross-validation, this component was treated cautiously. For future applications, nested cross-validation was recommended, with an outer 10-fold loop for evaluation and an inner 10-fold loop for PSO hyperparameter selection. A generalization gap below 0.08 was considered acceptable, whereas a gap above 0.15 indicated potential overfitting requiring model simplification.

Model evaluation
Model evaluation was performed using stratified 10-fold cross-validation on the final feature set, Xfinal. In each fold, Logistic Regression and Random Forest models were trained on the training data and evaluated on the validation data. Accuracy, Precision, Recall, F1-score, and ROC-AUC were calculated using classification_report and roc_auc_score from scikit-learn. The mean and standard deviation of all metrics were computed across the 10 folds.

The generalization gap was also calculated for each fold as follows:

generalization_gap = training_accuracy − validation_accuracy

The mean generalization gap across all 10 folds was reported. A mean gap below 0.08 was considered to indicate minimal overfitting, while a gap above 0.15 suggested overfitting and the need for regularization or reduced model complexity. Wilcoxon signed-rank tests were performed to compare the proposed framework against baseline methods across 10 folds using α = 0.01.

Explainability analysis
Explainability analysis was performed using model-specific and model-agnostic methods. For Logistic Regression, the final model was fitted and coefficient values were extracted for each selected feature. Odds ratios were calculated as exp(coefficient), and 95% confidence intervals were computed using the standard errors of the coefficients.

For Random Forest, Gini importance scores were extracted from the trained model using the feature_importances_ attribute and normalized to sum to 1. SHAP explanations were generated using the SHAP library. A KernelExplainer object was created using the trained model and a background dataset, such as 100 randomly selected training samples. SHAP values were computed for all instances in the test set using shap_values. Beeswarm summary plots were generated using shap.summary_plot, and bar plots of mean absolute SHAP values were generated using shap.bar_plot.

Partial Dependence Plots (PDPs) were generated for the top features identified by SHAP analysis. For each selected feature, a sequence of values spanning the feature range was created. Each value was substituted into the feature column while holding other features constant, and the average predicted probability was calculated across all instances. Feature values were plotted against average predictions using matplotlib. Confidence intervals of 95% were added using 100 bootstrap resampling iterations.

Individual Conditional Expectation (ICE) plots were generated for selected features by plotting prediction trajectories for individual instances as feature values changed. The PDP line was overlaid on the ICE plot. Explanation methods were compared by calculating Spearman’s rank correlation between Logistic Regression odds ratios and Random Forest SHAP values using scipy.stats.spearmanr. Discrepancies between explanation methods were documented for clinical interpretation. When SHAP and logistic regression coefficients conflicted, the PDP for that feature was examined. If the PDP showed a non-linear trend, the SHAP explanation was prioritized over the logistic regression coefficient because Random Forest can capture non-linear relationships that linear models cannot.

Framework generalization protocol for external validation using MIMIC-III
An external validation protocol was outlined for applying the framework to the MIMIC-III database. Access to MIMIC-III would require PhysioNet approval and completion of required human subjects training. The proposed cohort would include adult patients aged 18 years or older with first ICU admission and ICD-9 codes 410–414 for acute myocardial infarction or ICD-10 codes I20–I25 for ischemic heart disease. Exclusion criteria would include more than 30% missing values in target features, length of stay below 24 h, age above 90 years, prior cardiac surgery, or congenital heart disease.

The proposed outcome was major adverse cardiac events (MACE) within 72 h of admission, defined as a composite of in-hospital mortality, cardiogenic shock, or ventricular arrhythmia requiring intervention. Time-series features such as heart rate and blood pressure would be aggregated over the first 24 h of ICU stay using mean, median, minimum, maximum, and trend, where trend would be estimated as the slope from linear regression over time. Maximum heart rate would be used as the mapped equivalent of thalach.

Cleveland dataset features would be mapped to MIMIC-III variables. For example, thalach would be mapped to maximum heart rate recorded during the first 24 h of ICU stay, cp would be mapped to structured pain assessments and NLP-extracted chest pain mentions, and oldpeak would be mapped to ST-segment deviation from ECG reports. A mapping table would be created to document all feature alignments.

Before applying the full pipeline, NLP extraction for oldpeak would be validated on 100 randomly selected ECG reports. Precision, recall, and F1-score would be calculated against manual annotation by two clinicians. If F1-score was below 0.85, regex patterns would be revised or structured ECG data from chartevents would be used as an alternative. The preprocessing pipeline would then be repeated on extracted MIMIC-III data, the GAN would be retrained for augmentation, hybrid feature selection would be reapplied, models would be retrained, explanations would be generated, and performance metrics would be compared with the Cleveland dataset results.

Clinical dashboard implementation
A web-based clinical dashboard prototype was designed using a framework such as Flask or Django. HL7/FHIR API endpoints were planned for EHR integration, with authentication and authorization configured according to institutional security policies. Data mapping functions were designed to convert EHR data into model input format.

The user interface included three main views. The pre-screening view displayed patient demographics and calculated risk scores with color-coded risk levels. The decision-support view displayed a SHAP waterfall plot showing the top contributing factors for a specific patient. The intervention-planning view enabled what-if analysis by adjusting modifiable risk factors and displaying updated risk predictions. Export functionality was included to save reports as PDF files or integrate them with EHR documentation systems.

For future clinical deployment, dashboard usability evaluation was planned with at least five clinicians. The evaluation would use the System Usability Scale, with a target score greater than 68, task completion time, with a target reduction of at least 20% compared with EHR use alone, and 5-point satisfaction scales for explanation clarity and trust. This usability evaluation was planned as a future step and was not implemented in the current study.

Results

Experimental settings and performance metrics
All experiments were conducted in Python 3.9 using scikit-learn, TensorFlow, and SHAP libraries. Stratified 10-fold cross-validation was employed. The evaluation metrics included Accuracy, Precision, Recall, F1-score, and ROC-AUC.

Performance comparison with selected baseline methods

...

Method

Discussion

The framework outlined here provides a reproducible approach for developing interpretable heart disease prediction models. A prototype clinical dashboard integrating these explanations is shown in Figure 4, which implements a three-stage workflow: pre-screening, decision support with SHAP, and intervention planning [Figure 4 here]. Several critical steps require close attention to ensure successful execution of this framework.

As shown in Table 5,...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Cleveland Heart Disease Dataset	UCI Machine Learning Repository	https://archive.ics.uci.edu/ml/datasets/heart+disease	Benchmark heart disease dataset used for model development/evaluation
Django	Django Software Foundation	N/A	Alternative web framework for dashboard implementation
fancyimpute	fancyimpute developers	N/A	Optional MICE and KNN imputation sensitivity analysis
Flask	Pallets Projects	N/A	Web framework for dashboard implementation
HL7/FHIR API standard	HL7 International	N/A	Planned standard for EHR/dashboard integration
Keras	Keras developers	N/A	Neural network API used with TensorFlow/Keras for GAN architecture
matplotlib	matplotlib developers	N/A	Plotting library
MIMIC-III Database	PhysioNet	https://physionet.org/content/mimiciii/1.4/	Critical care database for planned external validation
missingno	missingno developers	N/A	Missingness matrix visualization
NumPy	NumPy developers	N/A	Numerical computing
pandas	pandas developers	N/A	Data manipulation
PhysioNet	PhysioNet	https://physionet.org/	Access platform/source for MIMIC-III
Python	Python Software Foundation	N/A	Version 3.9/3.9.7
scikit-learn	scikit-learn developers	N/A	Machine learning library, including preprocessing, model training, cross-validation, and metrics
SciPy	SciPy developers	N/A	Statistical testing, including Welch’s t-test and Spearman correlation
SHAP	SHAP developers	N/A	Explainable AI library
Statlog Heart Disease Dataset	UCI Machine Learning Repository	https://archive.ics.uci.edu/ml/datasets/statlog+(heart)	Benchmark heart disease dataset
TensorFlow	Google	N/A	Deep learning framework for GAN implementation
UCI Machine Learning Repository	University of California, Irvine	https://archive.ics.uci.edu/	Repository source for Cleveland and Statlog datasets