STEMI-OP in-hospital mortality prediction algorithms: Frailty-integrated machine learning in older patients undergoing primary PCI.

Abstract

Despite advances in medical care, older patients with ST-elevation myocardial infarction (STEMI) undergoing primary percutaneous coronary intervention (PCI) currently face high in-hospital mortality rates. Traditional prognostic models, primarily developed in Caucasian populations with fewer older participants and using classical statistical approaches, may not perform well in Southeast Asian settings. This study explores the need for artificial intelligence-based risk assessment models-the STEMI-OP algorithms-designed explicitly for STEMI patients aged 60 and older following primary PCI in Vietnam. Machine learning (ML) models were developed and validated using pre- and post-PCI features, with advanced feature selection techniques to identify key predictors. SHapley Additive exPlanations and Causal Random Forests were employed to improve interpretability and causal relationships between features and outcomes, highlighting the key factors, including the Killip classification, the Clinical Frailty Scale, glucose levels, and creatinine levels in predicting in-hospital mortality. The CatBoost model with ElasticNet regression for pre-PCI prediction and the Random Forest model with Ridge regression post-PCI prediction demonstrated significantly superior performance compared to traditional risk scores, achieving AUC values of 92.16% and 95.10%, respectively, outperforming the GRACE 2.0 score (83.48%) and the CADILLAC score (87.01%). By incorporating frailty and employing advanced ML techniques, the STEMI-OP algorithms produced more precise, personalized risk assessments that could enhance clinical decision-making and improve outcomes for older STEMI patients undergoing primary PCI.