Machine Learning with SHAP-Driven Interpretability Enhances Decision-Making in Coronary Bifurcation Percutaneous Coronary Intervention: A Prospective Study

Abstract

Background: This prospective registry-based cross-sectional study of 500 percutaneous coronary intervention (PCI) patients assessed lesion morphology, clinical and procedural determinants of coronary bifurcation complexity, and the utility of machine learning (ML) for lesion stratification. Objectives: Characterize bifurcation lesion classes (0: No bifurcation; 1: Simple; 2: Complex), identify key demographic, anatomical, and procedural predictors of complexity, and evaluate interpretable ML models for accurate classification. Methods: We analyzed patient demographics, comorbidities, angiographic features (e.g., side-branch stenosis, bifurcation angle, calcification), and procedural outcomes, selecting ten critical complexity drivers. Four ML approaches — k-nearest neighbors (KNN), support vector machines (SVM), ensemble trees, and probabilistic classifiers — were optimized via hyperparameter tuning and feature-selection methods, with SHapley Additive exPlanations (SHAP) values quantifying feature importance. Results: The SHAP analysis identified side-branch stenosis, heavy calcification, and dual-stent technique as top predictors. Weighted KNN and medium-scale SVM achieved 89 - 92% accuracy, while ensemble models peaked at 97.8% using 10 - 15 features. Complex lesions (class 2) required dual-stent deployment more often (35% vs. 10%) and had lower post-PCI TIMI 3 flow (85% vs. 92%). Wrapper-based feature selection outperformed filter and embedded methods, reaching 96.8% accuracy. Conclusions: Integrating anatomical metrics, patient risk factors, and interpretable ML significantly improves PCI decision-making for bifurcation lesions, outperforming traditional systems and enabling personalized interventional strategies to optimize outcomes and resource allocation.