Machine Learning-Driven Probability of Permanent Pacemaker Implantation After Transcatheter Aortic Valve Replacement

Background/Objectives: Permanent pacemaker implantation (PPI) remains one of the most common complications following transcatheter aortic valve replacement (TAVR). Identifying patients at increased risk for post-procedural conduction disturbances is clinically important for procedural planning and patient management. The aim of this study was to develop and evaluate a machine learning-based model for predicting the risk of PPI after TAVR. Methods: This prospective study was conducted between 2019 and 2025, and included 179 patients with severe aortic stenosis who underwent TAVR. Patient eligibility was determined by a multidisciplinary Heart Team based on clinical, echocardiographic, and imaging criteria. The primary endpoint was PPI occurring during hospitalization or within 30 days after the procedure. Statistical analyses were performed using RStudio (v. 2024.09.1+394)and Python (v.3.12.3), including comparative tests for continuous and categorical variables, receiver operating characteristic analysis to assess model performance, and SHapley Additive exPlanations (SHAP) to evaluate feature importance and model interpretability. Results: A total of 179 patients undergoing TAVR were included in the analysis. PPI occurred in 17 patients (9.5%) within 30 days after the procedure. A machine learning model was developed to predict post-TAVR PPI. The model demonstrated good predictive performance, with an overall accuracy of 0.944 and a weighted F1-score of 0.947. The confusion matrix showed that the model correctly classified 155 patients without PPI and 14 patients with PPI, with only a small number of false predictions. Explainability analyses using SHAP and permutation feature importance revealed that anatomical and procedural variables had the greatest impact on model predictions. The most influential predictors included valve size, right coronary sinus diameter, prosthetic valve diameter, and mean aortic annulus diameter. In contrast, baseline clinical variables such as left ventricular ejection fraction, previous myocardial infarction, and mean transaortic gradient showed a comparatively lower contribution to the prediction of PPI after TAVR. Conclusions: This study demonstrates that machine learning models can effectively predict the risk of PPI after TAVR. Anatomical characteristics of the aortic root and prosthesis-related parameters were the main determinants of PPI, whereas baseline clinical variables had a lower impact. The use of explainable artificial intelligence methods, such as SHAP analysis, may improve risk stratification and support procedural planning in patients undergoing TAVR.