Interpretable machine learning for predicting in-hospital bleeding after PCI in patients with diabetes and acute coronary syndrome: a retrospective cohort study

To develop and validate interpretable machine learning models for predicting in-hospital bleeding after percutaneous coronary intervention (PCI) in high-risk patients with diabetes and acute coronary syndrome (ACS). This retrospective cohort study included 1,488 patients with diabetes and ACS who underwent PCI between 2021 and 2024. We employed LASSO and Boruta algorithms for feature selection from a comprehensive set of clinical variables. Five ML models were developed using a 70/30 training-validation split. The optimal model was selected based on performance metrics, including balanced accuracy, and was interpreted using SHapley Additive exPlanations (SHAP). In-hospital bleeding occurred in 80 patients (5.4%). In the validation set, the XGBoost model was selected as the optimal model, achieving the highest balanced accuracy (0.861) with a sensitivity of 0.76 and specificity of 0.962. SHAP analysis revealed that low hemoglobin, multivessel disease, and the absence of proton pump inhibitor therapy were the most significant predictors of bleeding. An interpretable XGBoost model accurately predicts in-hospital bleeding risk in patients with diabetes and ACS undergoing PCI. By providing patient-specific explanations for its predictions, the model enhances personalized risk stratification and provides a valuable tool to guide clinical decision-making regarding antithrombotic therapy.