XGBoost Model With CMR Features for Prognostic Assessment in Patients With ST-Segment Elevation Myocardial Infarction

BACKGROUND: Accurate prognostic models for ST-segment elevation myocardial infarction (STEMI) are essential to guide clinical practice. OBJECTIVES: This study sought to construct predictive models integrating demographic, clinical, and cardiac magnetic resonance (CMR) imaging variables to forecast long-term major adverse cardiac and cerebrovascular events (MACCEs). METHODS: Patients with STEMI underwent CMR imaging within 1 week and 1 month after primary percutaneous coronary intervention. Twenty-four demographic, clinical, and CMR variables were used to construct 4 predictive models (naive Bayes, logistic regression, k-nearest neighbors, and XGBoost) for forecasting MACCEs during long-term follow-up. Model interpretability was assessed using Shapley values. RESULTS: A total of 483 patients were included (median age: 59.6 years, IQR: 54.0-65.0 years; median follow-up: 89.3 months; IQR: 60.3-115.4 months). During follow-up, 98 of 483 patients (20.3%) experienced MACCEs. The XGBoost model demonstrated superior predictive performance compared with the other approaches. Key CMR predictors included microvascular obstruction, left ventricular ejection fraction recovery, infarct size, intramyocardial hemorrhage, infarct core T1, and remote myocardium T1. Among clinical features, Killip class, body mass index, and age were most influential. Remote myocardium T1 was inversely correlated with left ventricular ejection fraction recovery at 1 month after percutaneous coronary intervention (R = -0.34; 95% CI: -0.43 to -0.27; P < 0.01). CONCLUSIONS: An XGBoost model integrating clinical and CMR features effectively predicted long-term MACCEs in patients with STEMI. Microvascular obstruction emerged as the most important CMR-based prognostic factor.