Answers are grounded in this paper and cite their sources.
Coronary artery disease (CAD) is commonly confirmed using invasive or imaging-based investigations, motivating transparent prediction from routinely collected clinical information. We analysed 920 records from the Cleveland, Hungarian, Switzerland, and VA Long Beach cohorts of the UCI Heart Disease database. The binary endpoint was absence versus presence of angiographic heart disease. Missing continuous and categorical variables were imputed within training folds, together with fold-contained encoding, scaling, and probability calibration. A stratified split produced 644 development and 276 untouched evaluation records. The primary ensemble averaged sigmoid-calibrated probabilities from K-nearest neighbours, logistic regression, and a support-vector classifier. It achieved accuracy 0.8297, sensitivity 0.8824, specificity 0.7642, precision 0.8232, F1-score 0.8517, AUROC 0.8840, average precision 0.8867, and Brier score 0.1325. A focused comparison of Extra Trees, gradient boosting, histogram gradient boosting, weighted blending, and probability stacking showed that the strongest deterministic and probabilistic models differed: SVC achieved accuracy 0.8333, whereas a constrained blend achieved AUROC 0.8883 and Brier score 0.1299. Permutation importance and Monte Carlo SHAP explanations identified chest-pain type, exercise-induced ST depression, exercise-induced angina, and thallium-test results as influential contributors. Leave-one-cohort-out AUROC ranged from 0.7152 to 0.9011. The combined framework integrates calibrated ensembling, uncertainty quantification, patient-level explanation, and cohort-aware validation for non-invasive CAD prediction.
No takes yet. Share an insight, caveat, or question.
Mithila et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: