Gradient Boosting Techniques in a Risk-Aware and Explainable Machine Learning Framework for Heart Disease Prediction

Abstract

The complicated connection between medical risk factors and the serious ramification of misdiagnosis highlights the essential challenge of detecting cardiovascular disease in its early stages. Although most examinations to date have focused on accuracy-centric evaluation, which may not absolutely account for clinical safety criteria, machine learning models have proven potential. In this paper, we present a risk-sensitive and interpretable machine learning framework for heart disease prediction using state-of-the-art gradient boosting models. The framework encompasses robustness analysis, uncertainty estimation, mutual information-based feature selection, SMOTE for class imbalance handling, data processing, and explainable artificial intelligence. An XGBoost vs. LightGBM comparison is used to assess different gradient boosting paradigms. A new Medical Risk Score is also proposed to penalize false negative predictions. The proposed framework, as validated by the experimental results, enhances clinical interpretability and diagnosis reliability, thereby making it a suitable tool for real-world healthcare decision support systems.