EXtreme Gradient Boosting (XGBoost), a machine learning model, outperformed more traditional methods for predicting composite major adverse events (MAEs) and many individual events in patients undergoing transcatheter left atrial appendage occlusion (LAAO), according to a recent study published in JACC: Advances.

Kamil F. Faridi, MD, MSc, et al., included 81,703 procedures from the ACC's LAAO Registry; 70% of procedures were used for development and 30% for validation. Investigators compared the discrimination of logistic regression (LR), least absolute shrinkage and selection operator (LASSO) and XGBoost in predicting combined in-hospital MAEs as well as individual events.

Using the original 16 model variables, XGBoost performed best in predicting composite MAE (validation AUC 0.648 [95% CI: 0.626-0.670] vs. LR 0.630 [95% CI: 0.608-0.642] and LASSO [95% CI: 0.626-0.670]). With an expanded set of 51 variables, XGBoost performed slightly better than LASSO in predicting MAE (AUC 0.653 [95% CI: 0.635-0.671] vs. AUC 0.515 [95% CI: 0.628-0.660]), and both XGBoost and LASSO performed far better than LR (AUC 0.515 [95% CI: 0.501-0.529]).

In predicting individual events, XGBoost generally outperformed other methods; however, performance was relatively poor for rare events. "LR, LASSO, and XGBoost models can all become unreliable for rare events regardless of the total cohort size or performance in the development cohort, indicative of overfitting," note Faridi and colleagues.

"XGBoost models can be considered specifically for performance assessment in the LAAO Registry," they write. "XGBoost should also be considered more broadly for a strategy for risk model development, though actual performance will vary in risk models for other outcomes and data sets, and clinical relevance of incremental changes needs to be considered."