Abstract / Summary
Background: Pacing-induced cardiomyopathy (PICM) is a major complication of chronic right ventricular pacing and develops in up to 25% of patients. Reliable prediction of PICM may facilitate individualized pacing strategy selection; however, currently available approaches demonstrate limited discriminative performance. This study aimed to develop and validate machine learning (ML) models for predicting PICM in patients with atrioventricular block (AVB) underlying heart failure with mildly reduced or preserved ejection fraction. Methods: In this multicenter prospective study derived from the RISK-PICM registry (January 2002 to January 2026), 125 patients with AVB who underwent pacemaker implantation were analyzed. PICM was defined as a ?10% reduction in left ventricular ejection fraction (LVEF) resulting in an LVEF <50%. Four ML algorithms, including logistic regression, support vector machine, random forest, and XGBoost, were trained using stratified 4-fold cross-validation with hyperparameter optimization and model discrimination was evaluated. Results: PICM developed in 41 patients (32.8%) during a median follow-up of 4.8 years. Male sex, pre-implant minimum QTc interval, and post-implant maximum QRS duration were significantly associated with PICM in univariable Cox regression. XGBoost achieved the highest discriminative performance (AUROC 0.831; sensitivity 0.805; specificity 0.845). Feature importance analysis identified baseline LVEF and post-implant maximum QRS duration as the most influential features. Model-derived risk stratification significantly differentiated high-risk and low-risk groups on Kaplan-Meier analysis (P < 0.001). Conclusions: ML-based models, particularly XGBoost, demonstrated clinically meaningful performance for PICM and may support individualized pacing strategy selection in patients with AVB and underlying heart failure with mildly reduced or preserved ejection fraction.