Introduction: Automated analysis of exercise electrocardiograms (ECGs) offers significant clinical value but is challenged by motion artefacts, muscle noise, and baseline drift. Premature atrial contractions (PACs) are of particular interest during exercise, where physiological stress may unmask arrhythmic burden. While deep learning approaches perform well for PAC detection in resting and ambulatory settings, PACs are difficult to identify during exercise due to their morphological similarity to normal beats, especially in single-lead recordings. To date, no validated method exists for automated PAC detection in exercise ECGs.
Methods: We developed a deep learning–based approach for PAC detection from single-lead exercise ECG signals, using a CNN–BiLSTM U-Net architecture with self-attention. Ground truth labels were generated by two annotators using predefined guidelines developed in consultation with a cardiologist; ambiguous cases were reviewed with the cardiologist.
The dataset was derived from the UK Biobank exercise cohort and divided into a training set (n = 695) enriched for ectopic activity and a held-out test set (n = 1410) of participants without cardiovascular disease, reflecting a real-world distribution with low PAC burden. Model performance was evaluated using beat-level classification metrics and clinically relevant agreement in PAC counts at the participant level, with inter-participant 5-fold cross-validation and independent testing on the held-out cohort.
Results: On the held-out test set, the model achieved a beat-level PAC F1 score of 0.74, with precision of 0.82 and recall of 0.68 within a multi-class classification framework. Bland–Altman analysis demonstrated good agreement between predicted and ground truth PAC counts per participant, with a mean bias of −0.19 and limits of agreement ±3.00 beats.
Conclusion: This study demonstrates the feasibility of high-precision PAC detection in noisy, single-lead exercise ECGs. The proposed approach enables scalable and reliable estimation of PAC burden, supporting its use for population-level phenotyping.