In-depth Analysis, Model Optimization, and Interpretability of Heart Disease Risk Factors Using Machine Learning and Stacking Ensembles

Heart disease, as a major threat to human health, imposes critical significance on early prediction for clinical prevention and control. To enhance prediction accuracy, this study leverages the Heart Failure Prediction Dataset from Kaggle, addressing data imbalance via SMOTE combined with Tomek Links algorithm, normalizing dimensions and handling outliers through Z-score standardization and capping method. Logistic Regression, Random Forest, XGBoost, and Stacking ensemble learning models are constructed, with their prediction performances systematically compared. Key influencing factors including ST-segment slope, chest pain type, and exercise-induced angina are identified via Random Forest feature importance analysis. Using SHAP values to interpret model decision logic reveals a strong correlation between abnormal ST-segment slope and heart disease incidence. The Stacking model demonstrates optimal comprehensive performance, achieving an accuracy of 0.9057 and an AUC value of 0.9581, outperforming single models. This study provides efficient modeling tools and interpretable risk assessment basis for early heart disease prediction, facilitating the formulation of clinical precision prevention strategies and public health interventions.