Prediction of Pulmonary Vein Isolation and Gap Recurrence on 12-Lead ECG Using Deep Learning

Pulmonary vein isolation (PVI) is key to atrial fibrillation (AF) ablation, but arrhythmia often recurs due to conduction gaps permitting pulmonary vein (PV) reconnection. Currently, gap identification requires invasive remapping. We evaluated whether deep learning applied to surface electrocardiograms (ECGs) could (i) detect the electrophysiologic signature of PVI and (ii) predict PV reconnection at redo ablation. We retrospectively studied 176 patients (2012–2023) who had initial PVI and repeat ablation. A total of 865 10-second 12-lead ECGs were extracted from GE MUSE and CardioLab systems, segmented into 1-2 second clips, and used to train ResNet-based convolutional neural networks. Separate models were developed for: (i) PVI detection (pre-vs. post-ablation ECGs) and (ii) gap prediction using pre-redo ECGs. Demographic features were tested alone and in multimodal fusion with ECGs. Performance was evaluated using Receiver Operating Characteristic (ROC) curves, sensitivity, and specificity with stratified cross-validation. Gradient-weighted class activation mapping (Grad-CAM) assessed feature importance. The best PVI detection model distinguished ECGs before and after PVI with the area under the ROC (AUROC) = 0.879. Grad-CAM localized attention to the diastolic period and P-wave morphology. For gap prediction, the model trained on pre-redo ECGs achieved an AUROC of 0.819 (sensitivity 77.5%, specificity 75.8%). Adding demographics improved the AUROC to 0.830 (sensitivity 84%, specificity 72%), whereas demographics alone performed no better than chance (AUROC = <50%). Feature importance analysis highlighted P-wave onset and offset, inter-ablation time interval, left atrial volume index, age, and left ventricular ejection function as the strongest contributors in gap prediction. Deep learning identifies a consistent ECG biosignature of acute PVI and predicts PV reconnection before redo ablation with moderate accuracy, primarily using P-wave morphology. These models may inform patient selection, procedural planning, and counselling in patients with recurrent AF after PVI. Arrhythmia recurrence post-pulmonary vein isolation (PVI) are commonly caused by conduction gaps in the PV, and repeat isolation procedures lead to better arrhythmia-free survival compared to if recurrences are due to extrapulmonary triggers with chronically isolated veins. Currently, there are no non-invasive methods to detect conduction gaps after PVI. Our deep learning models can predict the presence of conduction gaps from non-invasive surface electrocardiograms prior to repeat procedures with moderate accuracy. We identified important ECG and demographic features in predicting gap presence. Ultimately, we provide potential methods of risk-stratifying patients and selecting candidates for repeat procedures that can be accessed in outpatient settings.