Using Machine Learning to Predict Adverse Events in Patients Undergoing Transcatheter Left Atrial Appendage Occlusion

BACKGROUND: Standard cardiovascular risk models generally have limited performance and cannot predict rare outcomes. Machine learning may overcome these limitations. OBJECTIVES: This study sought to determine how predictive performance varies for standard and machine learning methods in the National Cardiovascular Data Registry Left Atrial Appendage Occlusion (LAAO) Registry. METHODS: Logistic regression (LR), least absolute shrinkage and selection operator (LASSO), and eXtreme Gradient Boosting (XGBoost) were used to predict combined in-hospital major adverse events (MAEs) and individual events in patients undergoing transcatheter LAAO. Randomly selected 70% development and 30% validation cohorts were used for model creation and assessment with 16 variables from the previous LAAO risk model and an expanded set of 51 variables. RESULTS: The study included data from 81,703 LAAO procedures. The composite MAE rate was 1.39% (individual event rate 0.02% to 1.13%). XGBoost performed best for MAE using original model variables (validation AUC 0.648 [95% CI: 0.626-0.670] vs LR 0.630 [95% CI: 0.608-0.642] and LASSO 0.638 [95% CI: 0.626-0.670]). With expanded variables, XGBoost (AUC 0.653 [95% CI: 0.635-0.671]) performed marginally better than LASSO (AUC 0.644 [95% CI: 0.628-0.660]) for MAE, while LR performed poorly (AUC 0.515 [95% CI: 0.501-0.529]). Performance using all methods declined for infrequent events. XGBoost generally outperformed other methods for individual events, particularly with expanded variables, though not necessarily for rare events. Mortality prediction using XGBoost was incrementally better. CONCLUSIONS: In a nationwide LAAO cohort, XGBoost enhanced discrimination of composite MAE and several individual events. Prediction of rare mortality events was improved, although this was not consistently the case for other rare outcomes.