Abstract Body (Do not enter title and authors here): Introduction
Personal electrocardiogram (ECG) devices have significant potential for monitoring abnormal heart rhythms outside clinical environments. However, the reliability of these devices heavily depends on their automated ECG analysis capabilities, which guide patients on whether medical consultation is necessary. Many consumer-grade devices, such as the Apple Watch ECG (AW-ECG), frequently classify atrial and ventricular ectopy as ‘inconclusive’.

Hypothesis/Objective
To evaluate the performance of a novel deep learning algorithm in detecting atrial fibrillation (AF), premature ventricular contractions (PVCs), and premature atrial contractions (PACs) from AW-ECG recordings.

Method
A database of 2,500 AW-ECG recordings was analysed, each annotated by four electrophysiologists, with a consensus annotation used as the ground truth. The algorithm's performance in detecting AF, PVC, and PAC was assessed by measuring sensitivity, specificity, and positive predictive value (PPV) against the electrophysiologists' consensus.

Results
For AF detection, the AI algorithm achieved a sensitivity of 94% (95% CI: 90, 96), specificity of 99% (95% CI: 99, 100), and PPV of 94% (95% CI: 90, 96). For PVC detection, the sensitivity was 84% (95% CI: 79, 88), specificity was 99% (95% CI: 99, 100), and PPV was 91% (95% CI: 87, 94). For PAC detection, the sensitivity was 71% (95% CI: 65, 77), specificity was 97% (95% CI: 97, 98), and PPV was 75% (95% CI: 69, 80).

Conclusion
The study highlights the potential of a novel AI algorithm to accurately detect AF, PVCs, and PACs from AW-ECG recordings. The integration of this algorithm into AW-ECG analysis could significantly improve the identification of atrial fibrillation and ectopic beats, thereby reducing the frequency of inconclusive interpretations and minimising the need for human review.