An Interpretability Heart Disease Prediction Model Based on Stacking Ensemble with SHAP

Abstract

In the era of big data, the healthcare sector has witnessed exponential growth in data accumulation. This research employed ensemble learning techniques to explore the correlations between various health indicators and heart disease, identifying key features that significantly influence disease prediction and assisting medical professionals in understanding the pathogenesis of heart disease. By constructing a two layer stacking ensemble model that integrates the strengths of four base learners—Naive Bayes, Decision Trees, CatBoost, and Gradient Boosting Trees, the model achieves enhanced prediction accuracy and stability. To address the challenges of high complexity and poor interpretability in ensemble models, we introduce the SHAP technique to visualize the decision - making logic of the ensemble model. Experimental results demonstrate that the stacking ensemble model achieves 86.69% accuracy, 87.14% weighted average precision, 86.69% weighted average recall, and 86.91% weighted average F1 score. Unlike single learners that prioritize high precision at the expense of recall, the stacking model strikes an optimal balance between both metrics. Furthermore, we identify key predictors: maintaining an average sleep duration of 7 - 8 hours significantly reduces the risk of heart disease, while advanced age and poor health status increase susceptibility.