Key result
Machine-learning models trained on routinely collected registry data provided only modest discrimination for in-hospital neurologic events after TAVR (logistic regression AUC 0.631; CatBoost AUC 0.600).
Why the study?
Reliable individual-level prediction of stroke and transient ischemic attack after transcatheter aortic valve replacement using routinely collected data remains limited.
Can machine-learning models accurately predict in-hospital neurologic events (stroke or TIA) in patients undergoing TAVR using multicenter registry data?
Cohort (n=20,740)
Yes
Can machine-learning models accurately predict in-hospital neurologic events (stroke or TIA) in patients undergoing TAVR using multicenter registry data?
Effect estimate: AUC 0.631 (95% CI 0.575-0.690)
Machine-learning models trained on routinely collected TAVR registry data provided only modest discrimination for predicting in-hospital stroke or TIA, highlighting the need for richer data to enable useful individual-level prediction.
Background Stroke and transient ischemic attack (TIA) are serious complications of transcatheter aortic valve replacement (TAVR), and reliable individual-level prediction from routinely collected data remains limited. We evaluated whether machine-learning models can predict in-hospital neurologic events using multicenter registry data. Methods We retrospectively analyzed an international TAVR registry. The primary outcome was registry-coded in-hospital stroke or TIA, defined by Valve Academic Research Consortium-2 criteria. Gradient-boosting (CatBoost) and regularized logistic-regression models were developed using 37 routinely collected clinical/anatomic, echocardiographic, and procedural/device variables and evaluated on a group-aware held-out test set and a temporal validation set (development 2005–2021; test 2022–2024). We assessed discrimination (area under the receiver-operating-characteristic curve [AUC] and precision–recall curve [PR-AUC]) and calibration using 1000 bootstrap resamples. Results Among 20,740 procedures from 16 centers, stroke/TIA occurred in 562 (2.7%). In the held-out test set, discrimination was modest (logistic regression AUC 0.631 [95% CI, 0.575–0.690]; CatBoost AUC 0.600 [95% CI, 0.550–0.652]), with similar results in temporal validation (AUC 0.654 and 0.629, respectively); PR-AUC was low (0.04–0.06). CatBoost did not outperform logistic regression, and echocardiographic and procedural/device variables provided only small, nonsignificant gains. The highest predicted-risk decile had an observed stroke/TIA rate of 6.8% versus 2.7% overall. Feature importance was distributed across multiple predictor domains. Conclusions Machine-learning models trained on routinely collected registry data provided only modest discrimination and population-level risk stratification for neurologic events after TAVR, with no meaningful gains from flexible algorithms or procedural variables. Clinically useful individual-level prediction will likely require richer data and standardized neurologic outcome ascertainment
No takes yet. Share an insight, caveat, or question.
Itelman et al. (2026) conducted a cohort in Transcatheter aortic valve replacement (TAVR) (n=20,740). Machine-learning prediction models (CatBoost and logistic regression) was evaluated on Registry-coded in-hospital stroke or transient ischemic attack (TIA) (AUC 0.631, 95% CI 0.575-0.690). Machine-learning models trained on routinely collected registry data provided only modest discrimination for in-hospital neurologic events after TAVR (logistic regression AUC 0.631; CatBoost AUC 0.600).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: