The transformer-based DL model achieved macro-average AUC of 0.915 for localizing PVC origins, supporting its clinical utility in pre-ablation planning.
Does a transformer-based deep learning model using standard 12-lead ECGs accurately localize the origin of idiopathic PVCs compared to ablation-confirmed origins?
A transformer-based deep learning model can accurately localize idiopathic PVC origins from standard 12-lead ECGs, offering a reliable non-invasive tool for pre-ablation planning.
Absolute Event Rate: 0% vs 0%
Abstract Background Accurate localization of premature ventricular contractions (PVCs) is critical for successful catheter ablation, particularly for origins in the ventricular outflow tract (OT) or other sites. Traditional 12-lead ECG interpretation often requires expert input and may yield inconsistent results. Transformer-based deep learning (DL) offers a reproducible, non-invasive solution using ECG data alone. Objective To develop and validate a transformer-based DL model using 12-lead ECGs to localize PVC origins into four anatomical categories: left ventricular outflow tract (LVOT), right ventricular outflow tract (RVOT), left ventricular non-OT (LV Non-OT), and right ventricular non-OT (RV Non-OT). The model performs two binary tasks: (1) OT vs. non-OT and (2) left vs. right ventricular origin. Methods This retrospective study included 584 patients (mean age 46.5 ± 14.3 years; 255 men, 329 women) who underwent catheter ablation for idiopathic PVCs. After ECG quality review, 470 standard 10-second 12-lead ECGs were selected. Two transformer-based models were independently trained using patient-level splits (70% training, 10% validation, 20% test). Final localization was derived by combining output probabilities. Model performance was evaluated against ablation-confirmed PVC origins (Figure 1). Results PVC origins were classified as follows: RVOT (295, 50.5%), LVOT (93, 15.9%), anterolateral mitral annulus (37, 6.3%), posteromedial papillary muscle (21, 3.6%), RV inflow (16, 2.7%), LV summit (14, 2.4%), anterolateral papillary muscle (13, 2.2%), and others (95, 16.3%). The model achieved AUCs of 0.98 for LVOT, 0.91 for RVOT, 0.84 for LV Non-OT, and 0.93 for RV Non-OT, with a macro-average AUC of 0.915 (Figure 2). Performance remained consistent across internal validation, supporting its clinical utility. Conclusion This study presents a clinically validated transformer-based DL model capable of accurately localizing the origin of PVCs using standard 12-lead ECGs. The model consistently achieved high diagnostic performance and may serve as a valuable non-invasive tool to support pre-ablation planning in patients with ventricular arrhythmias
Baek et al. (Thu,) reported a other. The transformer-based DL model achieved macro-average AUC of 0.915 for localizing PVC origins, supporting its clinical utility in pre-ablation planning.