Comparative Performance of Machine Learning and Traditional Risk Scores in Predicting Adverse Events After Transcatheter Aortic Valve Replacement in Patients With Atrial Fibrillation

Comprehensive Summary

In this study, Nicolas et. al. developed and validated a machine learning (ML) model for predicting adverse thromboembolic and bleeding events in patients with atrial fibrillation (AF) following transcatheter aortic valve replacement (TVAR). The performance of this model was then compared to that of two widely utilized risk scores: CHA₂DS₂-VA and HAS-BLED. Data from 1,377 patients from the ENVISAGE-TAVI AF trial was used and four clinical outcomes were monitored, including clinically relevant bleeding (CRB), ischemic stroke (IS), major gastrointestinal bleeding (MGB), and net adverse clinical events (NACE). 10 different ML models were used to compare the performance between ML models and traditional models. As a result, ML models offered similar predictive ability to conventional risk scores for thromboembolic and bleeding outcomes in TVAR patients with AF. Through Shapley Additive exPlanations (SHAP), it was found that the most impactful variables for the outcomes of CRB, IS, MGB, and NACE are the absence of a history of previous valve replacement before index procedures, patient history of stroke and heart failure, the treatment arm of edoxaban, and the treatment arm of edoxaban as well as sex, respectively.

Outcomes and Implications

Since the risk prediction between ML and traditional models was similar for thrombotic and bleeding outcomes, the development of ML based tools for the improvement of risk assessment and prediction of outcomes remains an emerging field. Specifically, the ML model may provide insight into pre-TVAR planning and post-TVAR management for patients with AF through further development and external validation.