AI-based prediction of left bundle branch block risk post-TAVI using pre-implantation clinical parameters.

Abstract

Background and aims: Transcatheter Aortic Valve Implantation (TAVI) has revolutionized the treatment of severe aortic stenosis. Although its clinical efficacy is well established, the development of new-onset left bundle branch block (LBBB) following TAVI remains a frequent and concerning complication. This study aims to develop pre-implantation predictive models for new-onset LBBB after TAVI using both conventional machine learning (ML) algorithms and Large Language Models (LLMs).

Methods: Of the 1113 patients who underwent TAVI over a 15-year period, 469 were included after excluding those with preexisting LBBB, pacing rhythm, or missing relevant data. Pre-procedural clinical parameters - such as valve type, valve size, patient demographics, and comorbidities - were analyzed. The dataset was split into training and testing sets. Several ML algorithms were employed, and performance was evaluated using accuracy, precision, and F1 score. Additionally, LLMs (GPT-3.5 and GPT-4) were assessed using Few-Shot and Chain of Thought (CoT) prompting.

Results: New-onset persistent LBBB occurred in 15.29% of patients. Among ML models, XGBoost performed best. GPT-4 with CoT prompting demonstrated superior predictive performance compared to both conventional ML and GPT-3.5.

Conclusions: The current study establishes a predictive model leveraging pre-implantation parameters to anticipate the occurrence of new-onset left bundle branch block (LBBB) post-Transcatheter Aortic Valve Implantation (TAVI).