Key result
A convolutional neural network identifies atrial pacing sites from paced P-waves with 0.88 average AUROC.
Why the study?
In patients with recurrent AF after PV isolation, trigger origins are often uncertain, and localising ectopic sources via surface ECG could guide ablation strategies.
Does a deep learning model trained on paced P-wave morphologies accurately identify the origin of atrial ectopy in patients with atrial fibrillation?
Observational (n=78)
No
Does a deep learning model trained on paced P-wave morphologies accurately identify the origin of atrial ectopy in patients with atrial fibrillation?
Answers are grounded in this paper and cite their sources.
A deep learning model trained on paced P-waves can identify the origin of atrial ectopy with good discrimination, potentially offering a non-invasive tool to guide targeted ablation in atrial fibrillation.
Background Atrial ectopic beats are established as the most important triggers of atrial fibrillation. Pulmonary vein (PV) origin is by far the most common and is the primary target of ablation. However, non-PV foci can also initiate and sustain arrhythmia. In patients with recurrent AF after successful PV isolation, the trigger is often uncertain, making an effective ablation strategy challenging. The localisation of the dominant source of atrial ectopy using surface ECG could guide ablation strategies in patients with AF. We hypothesise that P-wave morphology during intracardiac pacing from known sites can serve as training data for a deep learning (DL) model to identify the ectopy origin. Purpose To develop and validate a DL model capable of predicting atrial pacing sites from surface ECG P-waves. Methods We conducted a single-centre prospective analysis of patients undergoing first-time catheter ablation for atrial fibrillation. During the electrophysiology procedure, 30 pulses of low-rate pacing were delivered from eight predefined atrial sites (four pulmonary veins, superior vena cava, coronary sinus ostium, left atrial posterior wall, and left atrial appendage) with simultaneous 12-lead ECG recording. The P-wave, defined as the 200 ms window starting from the stimulus, was analysed in leads I, II, III, aVF and V1. For noise filtering, we applied discrete wavelet transform and reconstruction. A convolutional neural network (CNN) was selected and trained using 5-fold cross-validation, while the final model testing was conducted using patient-wise leave-one-out cross-validation. Model performance was evaluated using accuracy, F1 score, and class-wise area under the receiver operating curve (AUROC). Results We analysed paced ECG segments from 8 atrial locations in a total of 78 patients, a total of 17621 P-waves. The CNN model demonstrated strong discrimination across atrial regions, with an average per-class AUROC of 0.88 (ranging from 0.83 for the posterior wall to 0.97 for the coronary sinus ostium). This suggests that paced P-wave morphology contains sufficient spatial information to define the source location. The average F1 score was 0.58, while the overall accuracy of 0.56. Calibration was favourable with an expected calibration error of 0.19 and a Brier score of 0.64. Conclusion A CNN trained on paced P-waves demonstrated accurate site localisation and good calibration. This approach may enable non-invasive identification of ectopic foci, potentially guiding targeted ablation and improving the efficiency of repeat procedures. Validation in spontaneous ectopy and multi-centre settings is warranted.
No takes yet. Share an insight, caveat, or question.
Komlosi et al. (2026) conducted an observational in Atrial fibrillation (n=78). Convolutional neural network (CNN) analysis of paced P-wave morphologies was evaluated on Prediction of atrial pacing sites from surface ECG P-waves (AUROC 0.88). A convolutional neural network trained on paced P-waves accurately identified atrial pacing sites with an average per-class AUROC of 0.88 and an overall accuracy of 0.56.
Synapse has enriched 10 related papers. Consider them for comparative context: