This study evaluates whether pre-implantation clinical parameters can predict new-onset left bundle branch block (LBBB) after transcatheter aortic valve implantation (TAVI). The authors conducted a single-center retrospective analysis of 469 out of 1113 consecutive TAVI patients over 15 years, excluding those with preexisting LBBB, prior pacemaker, or missing data, and trained multiple machine learning models alongside GPT-3.5 and GPT-4 using few-shot and chain-of-thought prompting; performance was assessed by accuracy, precision, and F1 given class imbalance. New-onset persistent LBBB occurred in 15.29% of patients, with XGBoost performing best among conventional models (accuracy ~0.85, precision ~0.79, F1 ~0.82) and GPT-4 with chain-of-thought achieving strong but slightly lower overall performance (accuracy ~0.73, precision ~0.76, F1 ~0.79), outperforming GPT-3.5; SMOTE did not improve results and TabPFN was not superior at this dataset size. The discussion emphasizes that using only pre-implantation variables enables true pre-procedural risk stratification (in contrast to studies relying on post-implantation features), highlights the promise of LLMs with prompt engineering for small imbalanced datasets, and notes limitations including single-center design, omission of some known predictors (e.g., calcium burden), and lack of external validation.