Abstract
<title>Abstract</title> <p>Cardiac disease is one of the main reason of death globally and it is best to catch it early. Risk prediction becomes relevant to support timely clinical interventions. Predicting heart disease is, however, difficult because there is a complex relationship between physiological, symptomatic and demographic factors. While predictive modeling approaches have been used to improve prediction power, many current models lack interpretability and are not routinely tested for robustness, fairness, and prediction reliability. This paper introduces the leakage-free five-fold Out-of-Fold (OOF) stacked ensemble model called RXTGStack by leveraging RandomForest (RF), XGBoost (XGB), TabNet (TN) and GradientBoosting (GB), with LogisticRegression (LR) as a meta-learner. The framework includes Shapley Additive exPlanations (SHAP) for both global and local explanations with the aim of increasing model interpretability. To enhance predictive capability, feature engineering was performed by deriving clinically meaningful variables, including Body Mass Index (BMI), Pulse Pressure (PP), Mean Arterial Pressure (MAP), heart-rate reserve (HRR), rate-pressure product (RPP), and the cholesterol-age ratio (Chol_Age_Ratio). In addition, an empirical verification framework assesses local robustness under input perturbation, feature monotonicity, SHAP consistency, demographic fairness, and split-conformal prediction coverage. The proposed approach was evaluated on two independent cardiovascular disease datasets. On the clinically curated dataset (CVD-1K), it achieved 98.5% accuracy, 0.9994 ROC-AUC, 0.969 Cohen’s Kappa (κ), and 0.969 Matthews Correlation Coefficient (MCC). On the larger population-scaled data set (Cardio-68K), it reached an accuracy of 77.0%, ROC-AUC of 0.803, MCC of 0.539 and Cohen’s Kappa score of 0.539. SHAP analysis consistently identified the slope of the peak exercise ST-segment, chest pain type, resting blood pressure, and serum cholesterol is an influential predictors. The empirical verification analyses indicated prediction stability of 94.8–99.3% under the tested ±3% relative perturbation regime, a demographic parity difference of at most 0.026, and split-conformal coverage of 0.900 on the CVD-1K dataset and 0.899 on the Cardio-68K dataset against a target coverage of 0.90. Overall, the results suggest that RXTG-Stack not only achieves competitive prediction performance but also provides interpretable predictions and complementary verification analyses for cardiac disease prediction.</p>