Comprehensive Summary
Zaka and colleagues conducted a meta-analysis to test whether machine learning offers better mortality prediction after transcatheter aortic valve implantation compared with traditional surgical risk measures. Across nine studies and 29,608 patients, the investigators evaluated 16 different machine learning pipelines which tested against established tools such as STS, EuroSCORE II, ACC-TAVI, FRANCE-2, and CoreValve scores. The pooled C-statistic for the top-performing ML models was 0.79 (95% CI 0.71–0.86), while the best traditional methods reached 0.68 (95% CI 0.61–0.76). Specifically, the support vector machines reached an AUROC of 0.94, neural networks reached 0.91, and random forests reached 0.86. However, only two studies included external validation, and all models showed high risk of bias due to missing data, inconsistent calibration reporting, and limited transparency in model construction. Overall, machine learning approaches outperformed traditional scores but showed major methodological gaps.
Outcomes and Implications
These findings suggest that current TAVI risk tools may misclassify mortality risk in patients, especially those with complex comorbidities. Machine learning models provide improved discrimination and can incorporate richer clinical data. However, because external validation is limited and calibration reporting remains inconsistent, machine learning models are not yet ready for immediate clinical adoption. Broader collaboration and standardized performance criteria will be needed before machine learning-based TAVI risk prediction can reliably influence bedside care.