A Lightweight Cost-Sensitive Explainable Ensemble Framework for Early Heart Disease Risk Prediction

Abstract

Cardiovascular disease is still one of the leading causes of death, and hence, the early prediction of risk is a very important task in preventive medicine. Although recent studies have shown encouraging results in the application of machine learning algorithms to the prediction of heart disease, it has been noticed that most of the algorithms are more concerned with accuracy-driven optimization than the concerns of safety and false negatives. In medical decision support systems, false negatives are more harmful.This paper presents a light-weight and interpretable machine learning approach for the early risk prediction of heart disease based on structured clinical data. Various models such as Logistic Regression, Random Forest, XGBoost, and stacking ensemble classifiers are compared based on clinically meaningful evaluation metrics such as accuracy, pre- cision, recall, F1-score, and ROC- AUC. The experimental results indicate that ensemble classifiers perform better than individual models, and the unoptimized StackingClassifier performs the best (Recall: 0.8807, F1-score: 0.8930, AUC: 0.9147). Cost-sensitive and threshold-optimized stacking further enhances the recall to 0.9266. To improve the transparency and clinical trust, SHAP and LIME are combined to offer global and local explanations. The findings point out ST depression, maximum heart rate reached, type of chest pain, cholesterol, and exercise-induced angina as the important risk factors. The proposed approach shows that simple and interpretable ensemble models can provide accurate heart disease risk predictions.