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 ....

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 wer....

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