Artificial Intelligence in Risk Stratification and Outcome Prediction for Transcatheter Aortic Valve Replacement: A Systematic Review and Meta-Analysis.

Background/

Objectives

Transcatheter aortic valve replacement (TAVR) has been introduced as an optimal treatment for patients with severe aortic stenosis, offering a minimally invasive alternative to surgical aortic valve replacement. Predicting these outcomes following TAVR is crucial. Artificial intelligence (AI) has emerged as a promising tool for improving post-TAVR outcome prediction. In this systematic review and meta-analysis, we aim to summarize the current evidence on utilizing AI in predicting post-TAVR outcomes.

Methods

A comprehensive search was conducted to evaluate the studies focused on TAVR that applied AI methods for risk stratification. We assessed various ML algorithms, including random forests, neural networks, extreme gradient boosting, and support vector machines. Model performance metrics-recall, area under the curve (AUC), and accuracy-were collected with 95% confidence intervals (CIs). A random-effects meta-analysis was conducted to pool effect estimates.

Results

We included 43 studies evaluating 366,269 patients (mean age 80 ± 8.25; 52.9% men) following TAVR. Meta-analyses for AI model performances demonstrated the following results: all-cause mortality (AUC = 0.78 (0.74-0.82), accuracy = 0.81 (0.69-0.89), and recall = 0.90 (0.70-0.97); permanent pacemaker implantation or new left bundle branch block (AUC = 0.75 (0.68-0.82), accuracy = 0.73 (0.59-0.84), and recall = 0.87 (0.50-0.98)); valve-related dysfunction (AUC = 0.73 (0.62-0.84), accuracy = 0.79 (0.57-0.91), and recall = 0.54 (0.26-0.80)); and major adverse cardiovascular events (AUC = 0.79 (0.67-0.92)). Subgroup analyses based on the model development approaches indicated that models incorporating baseline clinical data, imaging, and biomarker information enhanced predictive performance.

Conclusions

AI-based risk prediction for TAVR complications has demonstrated promising performance. However, it is necessary to evaluate the efficiency of the aforementioned models in external validation datasets.