Machine Learning for Clinical Outcome Prediction After Transcatheter Aortic Valve Replacement: A Systematic Review...

Background: Machine learning (ML) models are increasingly used to predict adverse outcomes after transcatheter aortic valve replacement (TAVR), but their generalizability and clinical readiness remain uncertain. We evaluated predictive performance across post-TAVR outcomes, emphasizing validation quality, transportability, methodological bias, and clinical utility. Methods: PubMed/MEDLINE, Embase, Scopus, and Web of Science were searched from inception through 6 August 2026 for supervised ML studies reporting out-of-sample performance for clinically relevant post-TAVR outcomes. Overlapping source populations were audited before synthesis. Areas under the receiver operating characteristic curve (AUCs) were pooled on the logit scale using restricted maximum likelihood (REML) random-effects models with modified Hartung-Knapp inference and prediction intervals. Prespecified analyses examined validation maturity, predictor timing, prediction horizon, data architecture, and risk of bias. Within-study dependence was addressed using multilevel meta-regression with bias-reduced cluster-robust variance estimation (CR2); Bayesian analyses complemented the primary synthesis. Results: Sixty-seven full-text reports were included. Twenty independent source/report families contributed to mortality meta-analysis, yielding a pooled AUC of 0.791 (95% confidence interval [CI], 0.732-0.840; I² = 91.5%; prediction interval, 0.480-0.940). The high heterogeneity and wide prediction interval indicate that this pooled AUC is an average across heterogeneous settings rather than a transportable performance guarantee. Stronger external, temporal, geographic, or inter-centre evaluation showed lower discrimination than internal/resampled evaluation (0.754 vs. 0.814; p = 0.299), while preprocedural models yielded 0.767 versus 0.866 for peri-/postprocedural or mixed models (p = 0.083). High-risk studies showed higher apparent mortality discrimination than low/unclear-risk studies (0.895 vs. 0.747; p for interaction = 0.0009); this observational comparison involved only five high-risk mortality studies and should not be interpreted causally. The Bayesian-adjusted difference was +0.118 AUC units (95% credible interval [CrI], +0.037 to +0.179). Pooled AUCs were 0.760 for permanent pacemaker implantation (PPMI), 0.778 for broader conduction abnormalities, and 0.747 for paravalvular leak (PVL). No algorithm family showed reproducible superiority (p = 0.775). Stronger independent evaluation, calibration/Brier assessment, and decision-curve analysis were reported in only 18/67, 23/67, and 10/67 reports, respectively. Conclusions: ML shows promising post-TAVR discrimination, particularly for mortality, but substantial heterogeneity, limited independent validation, and sparse calibration and decision utility assessment constrain clinical translation. A high AUC alone should not be interpreted as evidence of transportability or clinical usefulness.