An Explainable and Class-Balanced Gradient Boosting Framework for Heart Disease Prediction

This research presents an explainable, class-balanced gradient boosting framework for heart disease prediction that overcomes common limitations in existing models, such as restricted dataset sizes, class imbalance, and lack of clinical interpretability. The significant contributions of this work include: The development of domain-driven composite features (BloodPressure_BMI, Cholesterol_Age, and Health_Score) to capture complex interactions among cardiovascular risk variables. The implementation of minority class upsampling to reduce prediction bias and improve the identification of high-risk heart disease cases. The integration of SHAP-based explainability to provide transparent, clinically interpretable feature attributions at both the global and patient levels. Validated on a diverse dataset of 10,000 patient records, the framework achieves robust predictive performance with an accuracy of 90.08%, a macro-averaged F1-score of 0.90, and a ROC-AUC of 0.9644.