Machine Learning Models Integrating CYP2C19 Genotypes for Long-Term Risk Prediction and Personalized Antiplatelet Therapy Selection in Myocardial Infarction

Abstract: Machine learning (ML) models integrating genetic and clinical data show promise for personalizing antiplatelet therapy after myocardial infarction (MI). This study aimed to develop an ML model using clinical features and CYP2C19 genotype to predict long-term major adverse cardiac events (MACEs) and guide P2Y12 inhibitor selection. In a prospective observational study of 218 patients with MI undergoing percutaneous coronary intervention, with up to 9-year follow-up for MACEs, we trained and evaluated multiple models incorporating clinical, genetic, and angiographic variables. Uplift modeling principles were applied to assess treatment heterogeneity, and feature importance was analyzed using Shapley Additive Explanations (SHAP). The optimal model (CatBoost with SHAP-based feature selection) achieved an area under the receiver operating characteristic curve of 0.721. Discrimination remained stable under bootstrap resampling and across both infarction presentations. Key prognostic predictors included age, comorbidity index, number of significant coronary lesions, P2Y12 inhibitor type, stent type, and CYP2C19 loss-of-function/gain-of-function variants. Notably, CYP2C19 variants were significant MACE predictors, while drug-eluting stents and ticagrelor were associated with lower predicted long-term risk. These findings suggest that an ML framework incorporating CYP2C19 genotype can stratify long-term MACE risk. This data-driven approach points to the prognostic utility of genetic testing and supports the potential role of ticagrelor, particularly in genetically defined high-risk patients with MI, for optimizing secondary prevention.