Comparative Performance of Machine Learning and Traditional Risk Scores in Predicting Adverse Events After...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Patients with atrial fibrillation (AF) post-transcatheter aortic valve replacement (TAVR) face significant risks of ischemic stroke (IS) and bleeding.
• Existing risk scores (CHA₂DS₂-VA, HAS-BLED) have limited predictive power in this specific patient population.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting IS, major gastrointestinal bleeding (MGIB), clinically relevant bleeding (CRB), and net adverse clinical events (NACE) in TAVR patients with AF.
• To compare the performance of ML models against traditional risk scores.

Main Methods:

• Ten ML algorithms were trained and validated using data from the ENVISAGE-TAVI AF trial.
• Model performance was assessed using the F1 score and compared with logistic regression models utilizing CHA₂DS₂-VA and HAS-BLED scores.
• A 25% holdout set was used for final model validation.

Main Results:

• ML models demonstrated predictive abilities comparable to CHA₂DS₂-VA and HAS-BLED for IS, MGIB, CRB, and NACE.
• Predictive performance for IS and MGIB was generally low across all tested models and risk scores.
• For CRB and NACE, ML models showed similar or numerically better predictive performance than traditional scores.

Conclusions:

• Machine learning models offer similar predictive capabilities to established risk scores for thromboembolic and bleeding outcomes in TAVR patients with AF.
• Further research may be needed to improve prediction accuracy for these complex outcomes in this patient group.