Abstract 4146352: Development and Validation of Machine Learning-based Ischemic Outcome Prediction Using Clinical and Genetic Data in Patients with Percutaneous Coronary Intervention

Introduction: Identifying patients at risk for ischemic events after percutaneous coronary intervention (PCI) relies on traditional analysis of limited clinical and imaging variables. Machine learning (ML) has shown promise in effectively predicting cardiovascular risk in population studies. While existing ML models mainly predict mortality and incorporate clinical variables, there is a lack of tools that have utilized genetic data and that predict ischemic events. Aims: This study aims to develop and validate a ML model incorporating genotyping and clinical data to enhance prediction of ischemic outcomes for PCI patients utilizing large prospectively derived diverse datasets. Methods: Patients from the TAILOR-PCI trial (n=5302) were utilized for model development. 50% of the sample was utilized for Boruta feature selection and 50% for training and testing using cross validation. Features included demographics, medical history, medications, PCI characteristics, and genetic data (specifically, CYP2C19 *2, *3, *17 alleles). The primary endpoint was a composite of cardiovascular death, myocardial infarction, stroke, stent thrombosis, and severe recurrent ischemia at 12 months. Multiple ML classification algorithms, including Support Vector Machine (SVM) polynomial, Random Forest, Light gradient boost, XG Boost, among others, were benchmarked for their prediction performance on rare events. The top performing classifiers were externally validated on an independent dataset from the PRECISION PCI study (n=3,745). Results: Mean participant age of the training set was 64.2 ± 11.0 years, with 75.4% being male. During follow-up of 12 months, among 4,572 patients in the entire cohort 343 (7.5%) met the primary outcome. The SVM polynomial model demonstrated the highest area under the curve (AUC) of 0.67 for predicting the primary outcome with test dataset. The sensitivity, specificity, precision, and recall were 0.87, 0.28, 0.07, and 0.87 respectively (Figure). Peripheral arterial disease, body mass index, and age were among the top variables by feature importance. Conclusion: ML models incorporating both clinical and genetic data are feasible and highly promising in predicting major adverse cardiac events that may help guide use of anti-platelet drug therapy. The AUC values are reasonable given imbalances and misclassifications in datasets, and further model optimization with prospective utilization of the model will be paramount.