Leakage-free evaluation of heart disease prediction models under nested cross-validation: discrimination, calibration, and decision-curve analysis

Cardiovascular disease remains a leading cause of mortality, and simple clinical risk models can support triage and follow-up testing. I re-analyze the processed UCI Heart Disease cohorts (Cleveland, Hungarian, Switzerland, and Long Beach VA) and evaluate representation learning with reconstruction ICA (RICA) alongside baseline and PCA pipelines. To avoid optimistic bias, every step -imputation, scaling, feature learning, and hyperparameter selection- is performed within a nested cross-validation framework. Performance is summarized with discrimination (ROC-AUC, PR-AUC), threshold-based metrics (F1, precision, recall), calibration (Brier score, reliability), and decision-curve analysis (net benefit). Across the combined cohort (N = 920), the best pipeline achieved ROC-AUC ≈ 0.89 with stable outer-fold variability, while maintaining good calibration and consistent net benefit across clinically relevant thresholds. These results emphasize that robust validation and clinically oriented reporting provide a more reliable basis for decision support than single-split headline scores.