Key result
Machine learning models match traditional risk scores for stroke and bleeding prediction after TAVR.
Why the study?
Patients with AF following TAVR remain at risk of ischemic stroke and bleeding, but traditional risk scores provide only modest predictions of these events.
Do machine learning models improve the prediction of ischemic stroke, bleeding, and net adverse clinical events compared to traditional risk scores in patients with atrial fibrillation after TAVR?
Observational (n=1,377)
Do machine learning models improve the prediction of ischemic stroke, bleeding, and net adverse clinical events compared to traditional risk scores in patients with atrial fibrillation after TAVR?
Machine learning models do not significantly outperform traditional risk scores like CHA₂DS₂-VA and HAS-BLED in predicting thromboembolic and bleeding events in patients with atrial fibrillation after TAVR.
ML models match traditional scores for TAVR-AF risk prediction; leaves open incremental value without prospective validation.
Patients with atrial fibrillation (AF) following transcatheter aortic valve replacement (TAVR) remain at risk of ischemic stroke (IS) and bleeding. However, traditional risk scores provide modest predictions of IS and bleeding in these patients. We aimed to develop machine learning (ML) models that predict IS, major gastrointestinal bleeding (MGIB), all clinically relevant bleeding (CRB), and net adverse clinical events (NACE) using data from patients in the ENVISAGE-TAVI AF trial. Ten ML algorithms were trained per outcome using nested cross-validation; the best-performing model (highest F1 score) was validated on a 25% holdout set. Model performance was compared with logistic regression models using CHA₂DS₂-VA or HAS-BLED. Among 1,377 patients, 41 had an IS, 83 had MGIB, 375 had CRB, and 255 experienced NACE. The predictive abilities of a linear discriminant analysis algorithm for IS (F1 score = 0.08) and CHA₂DS₂-VA (F1 score = 0.09) were similarly low, but numerically better than HAS-BLED (F1 score = 0.05). Prediction of MGIB was similarly low for a logistic-lasso algorithm (F1 score = 0.11), CHA₂DS₂-VA (F1 score = 0.09), and HAS-BLED (F1 score = 0.12). For CRB, the predictive performance of a Naïve Bayes algorithm (F1 score = 0.39) was similar to CHA₂DS₂-VA (F1 score = 0.38) and HAS-BLED (F1 score = 0.41). The predictive ability of a logistic regression algorithm for NACE (F1 score = 0.33) was numerically better than CHA₂DS₂-VA (F1 score = 0.22) or HAS-BLED (F1 score = 0.27). In conclusion, ML offered similar predictive ability to established risk scores for thromboembolic and bleeding outcomes among TAVR patients with AF.
No takes yet. Share an insight, caveat, or question.
Nicolas et al. (2025) conducted an observational in Atrial fibrillation following transcatheter aortic valve replacement (n=1,377). Machine learning models vs. Traditional risk scores (CHA₂DS₂-VA or HAS-BLED) was evaluated on Ischemic stroke, major gastrointestinal bleeding, all clinically relevant bleeding, and net adverse clinical events. Machine learning models offered similar predictive ability to established risk scores like CHA₂DS₂-VA and HAS-BLED for thromboembolic and bleeding outcomes among TAVR patients with atrial fibrillation.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: