Medical Student University of Western Ontario London, Ontario, Canada
Background: Atrial fibrillation (AF) recurs after pulmonary vein isolation (PVI) in about 30% of patients, often because gaps in the ablation line allow pulmonary vein reconnection, although extrapulmonary triggers are also possible. If a gap is detected during a repeat procedure, it can be eliminated with additional ablation; however, when no gap is found, further lesions are unlikely to improve outcomes and are associated with lower rates of long‑term AF suppression. At present, there is no way to identify PVI gaps before an invasive EP procedure. Our goal is to develop a convolutional neural network that can detect presence of PVI and predict the presence of gaps on repeat procedures.
METHODS AND RESULTS: In this retrospective observational study, we included 176 patients who underwent an initial PVI and repeat ablation procedure between January 1, 2012 to December 31, 2023 at London Health Sciences Centre in Ontario. Information regarding patient demographics and ablation lesion sets were collected from chart review. 816 intra-procedure 12-lead ECG clips were extracted and used to train a convolutional neural network. The model was trained on ECGs before repeat ablation plus patients’ demographics to determine whether differences in p-wave morphology could indicate PV isolation. To further understand the significant patient characteristics, a predictor of post-PVI gap recurrence was trained using a random forest algorithm. The AI model was able to predict whether the ECG came from before or after PVI procedure with a ROC-AUC of 0.72, indicating its ability to identify a signal of PVI on a surface ECG. Upon review of repeat procedures, a gap in the PVI lesion sets was found in 100 of the 176 cases. The model was able to predict the presence of a gap with an AUROC of 0.75 when both demographic and ECG factors were included in the training. The AUROC dropped to 0.53 when trained solely on demographics, indicating that the ECG contains most predictive information The most significant demographic predictors of an extrapulmonary cause for recurrence were the length of time between initial and repeat procedures (OR=0.795), followed by left atrial volume index (OR=1.28) and patient age (OR=0.615).
Conclusion: The results suggest that our AI model can detect a small signal of acute PVI on a surface ECG. Evaluation of repeat ablation procedures indicates that a similar AI model can predict gaps in PVI on surface ECG; however, the signal is weak.
