A Structured Explainability Framework for Heart Disease Risk Prediction Based on SHAP-Nomogram Integration

Cardiovascular disease is the leading cause of death globally, and traditional risk prediction models generally suffer from the limitations of being a 'black box.' To improve the interpretability of heart disease risk prediction while maintaining high predictive accuracy, this study proposes a structured interpretability framework that deeply integrates SHAP with nomograms. This framework enables 'global-local-combined' model interpretation, provides a global ranking of feature importance, and generates individualized risk visualization nomograms, automatically identifying key feature interaction combinations. Experimental results show that this framework achieves both high predictive performance (test set AUC 0.9439, accuracy 0.8949) and excellent interpretability in heart disease prediction tasks, with stability superior to LIME. It can identify four clinically relevant feature interaction combinations, clarify core predictive features such as ST_Slope, quantify individual risk contributions, and intuitively present them through nomograms. Ablation studies confirm the necessity of each module.