Predictors of Paravalvular Leakage After Transcatheter Aortic Valve Replacement in Patients With BAV

Comprehensive Summary

Mao et al. investigated whether paravalvular leak after transcatheter aortic valve replacement in patients with bicuspid aortic valve stenosis could be predicted using preprocedural data. The study analyzed 1,080 patients from multiple centers to develop a machine-learning model and tested it in an independent cohort of 109 patients. Using a random forest approach, the final model incorporated seven predictors, dominated by corrected calcification burden at different levels of the aortic root, annular and outflow tract geometry, and whether postdilation was performed. In the derivation cohort, the model showed excellent discrimination for identifying at least mild paravalvular leak, with an area under the curve of 0.982. Performance remained high in external validation (AUC 0.975). A logistic regression model built from similar predictors performed comparably (AUC 0.978 in derivation and 0.988 in validation), suggesting that much of the predictive signal arises from clinically interpretable anatomic features rather than complex model behavior.

Outcomes and Implications

From a clinical standpoint, these findings are most relevant during preprocedural planning for bicuspid TAVR, where anticipating leak risk can influence valve sizing, deployment strategy, and the anticipated need for postdilation. The dominance of calcification-related variables aligns with everyday experience in structural practice and provides reassurance that the model is capturing meaningful anatomy rather than spurious correlations. However, discrimination approaching 1.0 warrants caution, particularly in a retrospective dataset and when the endpoint is mild paravalvular leak rather than moderate or severe regurgitation. As such, the model is best viewed as a quantitative adjunct to expert CT interpretation rather than a substitute for procedural judgment.