Machine-learning-based identification of arrhythmia-driving regions in ventricular tachycardia

Electroanatomical 3-D mapping provides bipolar electrograms (EGMs) to guide ablation for scar-related ventricular tachycardia (VT). However, identifying arrhythmia-driving regions remains challenging and operator-dependent. We propose and evaluate a supervised machine-learning pipeline that combines point-wise EGM features with explicit 3-D neighborhood aggregation to prioritize potential ablation target regions. We retrospectively analyzed 86,765 bipolar EGMs from 14 VT patients and constructed a curated dataset of 16,078 EGMs, including 1,729 EGMs from clinically defined arrhythmogenic (ARR) regions. We compared three feature sets: point-wise baseline features (FS 1), baseline plus 5-mm neighborhood aggregation (FS 2), and a compact feature subset (FS 3). Three classifiers (sup- port vector machine, k-nearest neighbors, and bagged trees) were evaluated using repeated stratified cross-validation on the curated dataset and leave-one-patient-out (LOPO) full-map validation. In repeated cross-validation, the best FS 1 model achieved 62.1% sensitivity, 98.8% specificity, and 92.8% accuracy. Using FS 2 improved performance to 92.6% sensitivity, 99.6% specificity, and 98.5% accuracy. In LOPO full-map validation, specificity remained high (97.5%), whereas sensitivity decreased under realistic class imbalance. These results suggest that explicit 3-D neighborhood context improves the discrimination between ARR and NON-ARR regions and may support prioritizing potential ablation target regions in VT.