AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Novel Machine-Learning Models for 1-Year Mortality Prediction After Transcatheter Aortic Valve Replacement and Comparisons with Established Risk Scores
By Lo SW, Hoyek K, Tereshchenko LG et al.
Background: Accurate 1-year mortality prediction after transcatheter aortic valve replacement (TAVR) is critical for patient selection and futility assessment. The EuroSCORE II and STS/ACC TVT, developed primarily for short-term surgical outcomes, consistently demonstrate suboptimal discrimination for 1-year post-TAVR mortality. Machine learning (ML) may improve risk stratification by capturing complex interactions, yet systematic benchmarking against established scores within a unified framework is lacking. Research Question: Do ML models achieve better discrimination and predictive accuracy for 1-year all-cause mortality after TAVR compared with EuroSCORE II and STS/ACC TVT? Methods: Consecutive patients with native AS undergoing TAVR evaluation (2019-2020) at a single center were retrospectively studied. The primary endpoint was 1-year all-cause mortality. Using demographics, comorbidities, and echo/CT parameters, we trained 7 survival ML models (random survival forest [RSF], XGBoost survival, gradient boosting survival analysis [GBSA], Lasso-Cox, Elastic Net-Cox, Cox proportional hazards, and support vector survival machine) and 2 binary classifiers (XGBoost, Lasso logistic regression). Internal validation used 10-fold cross-validation across multiply imputed datasets. Discrimination was assessed by 1-year time-dependent AUC (tdAUC); predictive accuracy by Index of Predictive Accuracy (IPA; defined as 1 – [model Brier score / null model Brier score], where the null model assigns the marginal event rate to all patients). Confidence intervals were derived from bootstrap (500 replicates). Performance was benchmarked against EuroSCORE II, and STS/ACC TVT. Results: Among 1,035 patients (age 79 ± 9 years; 41.5% female; 1-year mortality 6.9%), 4 ML models demonstrated significantly better discrimination than both EuroSCORE II (tdAUC 0.576) and STS/ACC TVT (0.556): RSF (tdAUC 0.780; 95% CI, 0.729-0.832), XGBoost survival (0.768; 0.714-0.821), GBSA (0.757; 0.695-0.819), and XGBoost binary classification (0.735; 0.673-0.796; all adjusted p < 0.05) ( Fig. 1 ). RSF achieved the best predictive accuracy (IPA 0.078) ( Fig. 2 ), indicating a 7.8% reduction in prediction error relative to the null model. The most influential covariates associated with 1-year mortality in the RSF model are illustrated in Fig. 3 . Conclusions: Four ML models, RSF, XGBoost survival, GBSA, and XGBoost binary classification, demonstrated significantly higher discrimination than both EuroSCORE II and STS/ACC TVT for 1-year post-TAVR mortality, with RSF also achieving the highest predictive accuracy. These findings support the use of ML-based risk stratification incorporating laboratory, imaging, and cardiopulmonary history and presentations to enhance patient selection, futility assessment, and shared decision-making in TAVR candidates.