Machine learning prediction of mechanical dilatation in transvenous lead extraction for cardiac device-related infections: insights from a high-volume centre

Background: Transvenous lead extraction (TLE) requires high expertise in mechanical dilatation (MD), with double the risk of death and failure in low-volume centres. No automated tools support decision-making. Objective: We aimed to develop an ML-based model to predict the need for MD in patients undergoing TLE for infections. Methods: Five ML models (k-nearest neighbours, support vector machine, decision tree, random forest, and gradient boosting) were developed on a retrospective cohort of patients who underwent TLE at our centre. Each patient was described by 21 features (14 clinical, 7 device-related). Models were trained to distinguish MD from manual traction (MT). Five-fold nested cross-validation assessed performance and identified the best classifier. Feature importance analysis highlighted the most influential factors in model decisions. Results: Data for model training were extracted from our 25-year TLE database (June 1998–March 2023), including 491 patients (77.8% male; age 69.7 ± 12.8 years) and 938 leads (ICD 21.2%; pacing 78.8%; indwelling time 61 ± 60 months). MT was used in 27.5% and MD in 72.5% of cases. In total, 393 patients were used for training and 98 for testing. According to nCV, the Gradient Boosting Machine performed best, with 89% accuracy (SD 2%), 95% sensitivity (SD 3%), 73% specificity (SD 8%), and 92% AUROC (SD 1%). The most relevant features were lead dwelling time, ejection fraction, creatinine, ICD presence, prior cardiac surgery, and permanent atrial fibrillation. Conclusions: ML tools showed reliable performance in predicting MD need in TLE procedures, supporting planning and referral to high-volume centres.