Abstract 9715: Feasibility of Using Machine Learning to Establish a Risk Prediction Model for Sudden Cardiac Death in Patients with Hypertrophic Cardiomyopathy

Abstract

Introduction: Risk stratification for sudden cardiac death (SCD) is a great challenge in managing hypertrophic cardiomyopathy (HCM). Hypothesis: This study attempted for the first time to build a machine learning (ML) model for SCD risk prediction. Methods: The present study consisted of 1631 consecutive adult HCM patients without a history of SCD events. Extreme gradient boosting (XGBoost) was applied for ML model construction. We compared the predictive performance of the ML model with that of conventional models by calculating C-statistics and interpreted the ML model output using the SHapley Additive exPlanation (SHAP) value. Results: Fifty patients reached SCD endpoints during a median follow-up of 4.6 years. The C-statistic of the ML model was 0.745 [95% confidence interval (CI), 0.723-0.766], which was significantly higher than that of the model recommended in the 2014 European Society of Cardiology guidelines (C-statistics =0.609; 95% CI 0.585-0.623) and the model recommended in the 2020 American College of Cardiology/American Heart Association guidelines (C-statistics =0.628; 95% CI 0.604-0.652). The ML model incorporated 12 features as predictors, including 5 conventional SCD risk markers (NSVT, quantification of LGE, family history of SCD, LVEF 50%, unexplained syncope), WBC, creatine, uric acid, TBIL and cholesterol, NT-proBNP and family history of HCM. Using SHAP values, we quantified the contribution of each feature within the ML model to the final prediction and visualized the relationship between features and SCD risk. Conclusion: Our study constructed an ML model for the prediction of SCD risk in patients with HCM and provided a clinically intuitive interpretation. The ML model showed better discrimination performance than conventional models, demonstrating the feasibility of ML in SCD risk prediction for patients with HCM.