A novel AI-ECG tool for detecting non-pulmonary vein trigger before atrial fibrillation ablation

Background Pulmonary vein (PV) triggers are the main sources initiating atrial fibrillation (AF), and PV isolation remains the cornerstone of rhythm control. However, non-pulmonary vein (non-PV) triggers are increasingly recognized as key factors—not only contributing to post-ablation recurrence but also guiding the selection of optimal ablation modalities, such as pulse field ablation or radiofrequency catheter ablation. Identifying patients with a higher likelihood of non-PV triggers before ablation could improve procedural strategy and rhythm outcomes. Purpose To evaluate whether an AI-based 12-lead ECG model can predict PV and non-PV triggers before AF ablation and to identify independent predictors of non-PV triggers. Methods A total of 1,546 patients who underwent AF catheter ablation with standardized trigger testing according to the University of Pennsylvania protocol were analyzed. For paroxysmal AF, 12-lead sinus rhythm ECGs obtained on the day of ablation were used, and for persistent AF, ECGs recorded within one year after direct-current cardioversion were included. AF triggers were classified as PV or non-PV according to their anatomical origin. An AI-ECG model integrating clinical and ECG waveform features (morphologic and frequency-domain characteristics) was developed using an 8:1:1 split for training, validation, and testing datasets. Results Among patients with non-PV triggers, the most frequent sites of origin were the superior vena cava (79.2%), left atrial posterior wall (55.3%), and interatrial septum (left atrial 24.3%, right atrial 11.2%), followed by the coronary sinus ostium (7.2%) and crista terminalis (2.4%). Among 1,546 patients (mean age 61 ± 11 years, 67% male), non-PV triggers were identified in 22% of cases. Patients with non-PV triggers were older (64 ± 10 vs. 60 ± 11 years, p < 0.01), more frequently female (41% vs. 29%, p = 0.02), and had longer QTc intervals (448 ± 25 vs. 435 ± 22 ms, p < 0.001) and higher AI-ECG scores compared with PV-trigger patients. The AI-ECG model showed good discrimination for predicting PV triggers (AUROC = 0.759) and maintained consistent performance across subgroups. In multivariable logistic regression, independent predictors of non-PV triggers included older age (OR 1.38 [1.12–1.70], p = 0.002), female sex (OR 1.42 [1.09–1.84], p = 0.009), prolonged QTc interval (OR 1.16 [1.08–1.26], p < 0.001), and a higher AI-ECG score for predicting non-PV trigger (OR 1.48 [1.23–1.79], p < 0.001). Conclusion A novel AI-derived 12-lead ECG model demonstrated the potential to distinguish PV from non-PV triggers prior to AF ablation. By combining AI-ECG analysis with simple clinical and ECG parameters, this approach may assist in tailoring individualized ablation strategies and is expected to contribute to improved rhythm-control outcomes in patients with AF.