# Beyond Conventional Meta-Analysis: A Meta-Learning Model to Predict Cohort-Level Mortality After Transcatheter Aortic Valve Replacement (TAVR)

## TL;DR

This study uses machine learning to better predict mortality after a heart procedure, finding that factors like patient risk and time trends explain much of the variation.

## Contribution

Introduces a meta-learning model that outperforms traditional methods in predicting TAVR mortality and identifying key determinants.

## Key findings

- Meta-learning explained 65.3% of variability in TAVR mortality, a 46 percentage-point improvement over traditional methods.
- Key predictors included patient risk scores, procedure year, and diabetes prevalence, showing a strong temporal trend.
- The model highlights how medical practice evolution impacts outcomes, beyond patient-level factors.

## Abstract

Context and Objective: Post-Transcatheter Aortic Valve Replacement (TAVR) mortality exhibits extreme heterogeneity that conventional meta-analyses fail to explain, limiting the clinical utility of evidence synthesis and hindering accurate prognostic assessment. This study evaluated whether meta-learning, using aggregate data from the literature, can predict cohort-level mortality and identify its determinants, overcoming the limitations of traditional methods to provide a clearer understanding of the factors driving TAVR outcomes. Methods: A systematic review following PRISMA guidelines was conducted across five databases. Methodological quality was assessed with standardized tools (Risk of Bias 2, Newcastle-Ottawa Scale, Risk of Bias in Non-randomized Studies of Exposure). After performing conventional meta-analyses and meta-regressions, multiple machine learning models were trained using study-level characteristics as predictors. Advanced optimization with regularization and ensemble techniques was applied to develop a final, optimized model. Results: Fifty-eight studies, encompassing over 533,000 patients, were included. Traditional meta-analysis confirmed extreme heterogeneity (I2 = 76.7% in Random Clinical Trials, 96.8% in observational studies), with no explanatory power via meta-regression. The initial AdaBoost model achieved R2 = 0.191, outperforming 17 alternative algorithms. Advanced optimization developed a Blend_Optimized model that explained 65.3% of the variability (R2 = 0.653), marking a substantial 46 percentage-point increase. Interpretability analysis identified four dominant predictors: Society of Thoracic Surgeons Predicted Risk of Operative Mortality (R2 = 0.300), Recruitment Year (R2 = 0.212), % Transfemoral (R2 = 0.201), and % Diabetes (R2 = 0.175), revealing a potent temporal gradient reflecting the evolution of medical practice. Conclusions: Meta-learning significantly surpasses traditional methods in extracting systematic signals from heterogeneous evidence. This study demonstrates that, in addition to patient risk factors, a significant temporal gradient models technological evolution and learning curves. The methodology transforms seemingly unexplained heterogeneity into clinically interpretable patterns, demonstrating the potential of meta-learning as a complementary tool for evidence synthesis in interventional cardiology and opening avenues for applications in other complex cardiovascular fields. Important Limitation: This model predicts cohort-level outcomes and should not be used for individual risk assessment.