Explainable hybrid models for cardiovascular disease detection and mortality prediction

Abstract

The impact of cardiovascular diseases (CVDs) is devastating, with 20.5 million deaths annually. Early detection and prediction tools exist, but current approaches struggle to balance predictive performance with clinical interpretability. In this work, a two-stage machine learning (ML) framework is proposed for heart disease detection and mortality prediction in heart failure patients. Logistic regression (LR), random forest (RF), and gradient boosting (GB) models were trained using the publicly available heart failure datasets, and their performance was compared, then a stacked ensemble approach was employed to enhance prediction accuracy. Model interpretability was achieved through Shapley additive explanations (SHAP), which provide global feature rankings and specific patient attributes, supporting explainable artificial intelligence (XAI) in clinical practice. The GB model achieved the highest performance in the first stage with a receiver operating characteristic area under the curve (ROC AUC) of 96% and an accuracy of 89% on internal testing, while external validation confirmed strong generalization (ROC AUC of 94%). In the second stage, stacked ensemble model was employed and achieved marginal improvements. Two interactive web applications were developed to enable real-time predictions with SHAP visualizations. The results demonstrate that combining high-performance ML models with interpretable outputs can significantly improve trust in real-world healthcare environments.