Abstract
INTRODUCTION: Heart disease remains a leading cause of global morbidity and mortality, motivating the development of predictive models that are both accurate and clinically interpretable. We introduce the Interpretable Ensemble Learning Framework (IELF), which integrates Explainable Boosting Machines (EBM) with XGBoost, SHAP-based explanations, and LIME for enhanced local interpretability. METHODS: IELF was evaluated on two benchmark datasets: Cleveland (n = 303) and Framingham (n = 4,240). Model assessment included 5-fold cross-validation, held-out test sets, calibration, subgroup analyses, and explanation stability evaluation using Kendall's τ and Overlap@10. RESULTS: IELF achieved robust discrimination (AUC 0.899, accuracy 88.5% on Cleveland; AUC 0.696, accuracy 82.6% on Framingham) with balanced precision-recall profiles. Compared with EBM, IELF significantly improved recall, F1, and AUC on the Framingham dataset (p < 0.05), while differences versus XGBoost were less consistent. IELF produced transparent feature rankings aligned with established cardiovascular risk factors and stable explanations across folds. DISCUSSION: IELF is, to our knowledge, the first framework to combine EBM and XGBoost with SHAP and LIME under strict nested cross-validation and calibration procedures. Although headline accuracies are lower than some recent >97% reports, IELF was developed under stricter methodological controls that enhance reproducibility, interpretability, and clinical reliability. These findings position IELF as a trustworthy benchmark for translational AI in cardiovascular risk prediction, complementing high-accuracy but less transparent models.