Feature-weighted calibrated tabular explainable AI for heart disease prediction

Abstract

Cardiovascular disease (CVD) remains a leading cause of mortality and morbidity globally. Early detection and stratification of CVD risk using machine-learning models can improve patient outcomes; however, existing models often suffer from limited interpretability, small sample sizes, class imbalance, and uncalibrated probability outputs. This paper proposes FWC-TabXAI++, a feature-weighted calibrated tabular explainable AI framework for heart disease prediction. The method employs clinically constrained data preprocessing, second-order feature-interaction augmentation, mutual-information-based feature weighting, a heterogeneous ensemble of standard methods, validation-informed probability blending, and threshold selection. Its output includes probability estimates assessed for calibration, SHAP values, and LIME-based global and local explanations. The framework was tested on a cleaned clinical dataset comprising 836 records. FWC-TabXAI++ achieved an accuracy of 91.07%, a precision of 92.22%, a recall of 91.21%, a specificity of 90.91%, an F1-score of 91.71%, and an ROC-AUC and PR-AUC of 94.01% and 93.24%, respectively, and achieved the best overall balance across performance metrics. The model also exhibited a low Brier score of 9.08%, and bootstrap confidence intervals supported the internal stability of the proposed framework. SHAP and LIME analyses identified ST-segment slope and exercise-induced angina as dominant predictive contributors and provided patient-centred explanations. These results suggest that FWC-TabXAI++ offers a robust and interpretable approach for CVD risk prediction. This research supports healthcare providers and policymakers by enabling transparent, data-driven heart disease risk assessment using interpretable machine learning. The framework may improve early screening, patient prioritisation, and clinical decision support while maintaining clinician oversight.