Abstract
<title>Abstract</title> <p>Major adverse cardiovascular events (MACE) remain common after successful primary percutaneous coronary intervention (PCI) for ST-elevation myocardial infarction (STEMI), and conventional regression-based models may miss non-linear biomarker-risk associations unless these are specified in advance. Apelin-12 has been associated with MACE after STEMI, but its role in interpretable machine-learning-based risk prediction remains unclear. In a public, prospectively collected cohort of 464 patients with STEMI treated with primary PCI, including 118 MACE events during the nominal follow-up period of up to 30 months, we compared logistic regression, random forest, and XGBoost models using 14 predictors selected by LASSO, with training and test sets imputed separately to avoid direct leakage across partitions. Test-set discrimination was similar across models (AUC 0.750–0.781; DeLong’s test, all P ≥ 0.29), and the machine-learning models did not significantly outperform logistic regression, consistent with evidence that gains from flexible algorithms are often limited in modest-sized structured clinical datasets. SHAP analysis of the XGBoost model ranked Δapelin-12 and age as the most influential predictors, with apelin-12 also among the top-ranked predictors. Restricted cubic spline and piecewise logistic regression supported non-linear associations for apelin-12 variables, particularly Δapelin-12, for which the estimated breakpoint (16.5%) closely matched a 20% cut-point reported in an earlier subgroup analysis of this same cohort; the corresponding apelin-12 breakpoint (0.43 ng/mL) did not closely match that earlier study’s 0.76 ng/mL subgroup threshold. A reduced XGBoost model excluding apelin-12 variables had lower test-set AUC than the full model (0.732 vs. 0.781), but the bootstrapped 95% confidence interval for this difference crossed zero (− 0.004 to 0.102). These findings support further evaluation of apelin-12 in interpretable machine-learning-based risk models after STEMI, but external validation and a more definitive assessment of incremental value are required.</p>