Comparative machine learning highlights the value of calibration in lipid-based ischemic heart disease classification

Ischemic heart disease (IHD) is the leading cause of cardiovascular mortality in Pakistan, yet accessible tools for early biochemical risk stratification remain limited. This case–control study compared five supervised classifiers (Logistic Regression, Decision Tree, SVM (RBF), XGBoost, LightGBM) for IHD classification using routine lipid, anthropometric, and clinical variables, training on a stratified 70% set and evaluating on independent 20% testing and 10% validation partitions. XGBoost and LightGBM showed the most consistent discrimination across partitions: on testing, XGBoost reached accuracy 0.94 and ROC AUC 0.99, while LightGBM reached the highest accuracy (0.97) and recall (0.97); on validation, XGBoost achieved precision 1.00 and ROC AUC 0.99. SVM (RBF) showed comparably strong discrimination (testing accuracy 0.91, ROC AUC 0.97). A separability analysis showed BMI alone reached AUC 0.96 on the validation set; adding all features improved discrimination only modestly (ΔAUC ≤ 0.04) but improved calibration substantially (Integrated Calibration Index 0.32 to 0.02–0.03), indicating that high performance chiefly reflects the case–control contrast in BMI and LDL-C rather than latent multivariate structure. Given the case–control recruitment design, these metrics likely overestimate real-world screening performance and require prospective, population-based validation before clinical application. SHAP analysis identified BMI, LDL, VLDL, hypertension, and HDL as the most influential features. These findings support the potential of lipid-centered ML frameworks for early IHD risk stratification in resource-limited clinical settings, contingent on the validation noted above.