Abstract 12307: Machine Learning Model Predicting Bleeding Risk in Patients With Coronary Artery Disease

Abstract

Background: Bleeding complication during antiplatelet therapy for coronary artery disease (CAD) is an adverse issue due to high risk of subsequent death and cardiovascular events. Although many bleeding risk scores have been proposed, simple and personalized risk assessment in routine clinical practice is challenging because of the heterogeneous and multifactorial pathophysiological condition. We aimed to develop a machine learning-based risk prediction for bleeding events in CAD patients treated with antiplatelet therapy. Methods: This study was a retrospective, observational cohort study using the Japanese nationwide electronic medical record database. Data from 49,324 CAD patients who underwent revascularization and received antiplatelet therapy between July 2008 through July 2022 were randomly assigned to the model derivation and validation sets. Data balancing technique: synthetic minority over-sampling technique was used to process the imbalanced data. Risk prediction model was developed using random forest (RF) after feature selection by the Boruta algorithm. Outcome was defined as bleeding events requiring hospitalization. Results: Of all, 1,684 (3.4%) bleeding events were observed during study period. The area under the curves of receiver operator characteristics of RF (95% confidence intervals) for bleeding were 0.977 (0.976-0.978) in derivation and 0.975 (0.972-0.977) in validation set. The performance metrics of the model were well acceptable (accuracy, 0.939; precision, 0.990; recall, 0.887; F1-score 0.936, respectively) in validation set. The sigmoid calibrated RF models resulted in well approximating observed bleeding events in the calibration plot and the Brier score of 0.042. Conclusions: This machine learning model could predict bleeding risk for CAD patients treated with antiplatelet therapy in real-world setting.