Integrating lipid-related composite indices and explainable machine learning for coronary heart disease risk assessment.

Background: Composite indices integrating inflammation and lipid metabolism have emerged as promising markers for coronary heart disease (CHD), yet their comparative performance and discriminative ability for identifying CHD status remain incompletely understood. Methods: In this hospital-based study, 270 patients were enrolled, including 99 with CHD and 171 without CHD. Exposures included C-reactive protein (CRP) and composite indices (TG/HDL, LDL/HDL, AIP, CRP/HDL, and CRP/TG). Logistic regression, restricted cubic spline (RCS), and subgroup analyses were used to evaluate associations with CHD. Machine learning models were developed using significant predictors, and model performance was assessed by AUC, calibration, and decision curve analysis. SHapley Additive exPlanations (SHAP) were applied to interpret model outputs. Results: After multivariable adjustment, TG/HDL (OR = 2.74, 95% CI: 1.10-7.10), LDL/HDL (OR = 3.01, 95% CI: 1.21-7.81), and AIP (OR = 6.59, 95% CI: 1.61-28.51) were associated with increased odds of CHD, whereas CRP and CRP-based indices were not. RCS analyses indicated no significant nonlinearity, suggesting monotonic associations. Subgroup analyses showed generally consistent results across key strata. In classification modeling, ensemble tree-based methods performed best, with random forest and XGBoost achieving the highest discrimination ability (AUC = 0.748). SHAP analysis identified age and lipid-related composite indices as the primary contributors to CHD classification. Conclusion: Lipid-related composite indices, particularly TG/HDL, LDL/HDL, and AIP, are robust markers associated with CHD status and can be effectively integrated into machine learning models for individualized CHD classification.