Explainable Ensemble Learning for Cardiovascular Risk Stratification A Multi-Hospital Stacking Approach with SHAP-Based Clinical Decision Support
TL;DR
This paper presents an explainable stacking ensemble framework for binary heart disease classification and three-tier risk stratification using multi-hospital cardiac data using a novel Dual-Balanced SMOTE strategy that simultaneously addresses class imbalance and gender bias.
Abstract
Cardiovascular disease (CVD) remains the leading cause of mortality worldwide, necessitating accurate and interpretable risk stratification for clinical decision support. This paper presents an explainable stacking ensemble framework for binary heart disease classification and three-tier risk stratification using multi-hospital cardiac data. The approach integrates five heterogeneous base learners — Random Forest, XGBoost, Gradient Boosting, Extra Trees, and Support Vector Machine — with a Logistic Regression meta-learner trained via 5-fold cross-validation. A clinically motivated preprocessing pipeline incorporates eight engineered interaction features, RobustScaler normalization, threshold-optimized prediction (threshold = 0.82), and a novel Dual-Balanced SMOTE strategy that simultaneously addresses class imbalance and gender bias. Evaluated on the Cleveland UCI Heart Disease dataset (303 patients, 20% holdout, n = 61), the model achieves 96.72% accuracy, 0.9935 AUC-ROC, and zero false negatives at the optimal threshold. SHAP (SHapley Additive exPlanations) provides global and per-patient interpretability, identifying chest pain type (cp_4), the age–oldpeak interaction, and exercise-induced angina as the most influential predictors. The system is deployed as an end-to-end Flask web application (CardioRisk AI) enabling real-time clinical inference with SHAP waterfall visualizations and gender fairness evaluation confirming an F1 gap below 5% between male and female subgroups.