Long-term continuous monitoring of wearable ECG signals for atrial fibrillation detection.

Atrial Fibrillation (AFib) is the most common sustained cardiac arrhythmia and is associated with substantial morbidity and mortality, including increased risk of stroke and heart failure. Accurate detection of AFib in long-term electrocardiographic (ECG) monitoring is essential for timely diagnosis and early clinical intervention, particularly in ambulatory settings using wearable devices. In this study, we evaluate the performance of a lightweight convolutional neural network (CNN) for automated AFib detection, initially trained on the CinC2021 dataset and subsequently fine-tuned with real-world wearable ECG data. The model was validated across three independent datasets: InCorDB, CODE15, and the TRAdA cohort, with the latter representing long-term ECG monitoring from wearable devices. The TRAdA cohort comprised 173 subjects (87 with AFib and 86 healthy subjects) with continuous ECG recordings obtained over 24 h in Phase 1 (56 subjects) and 7 days in Phase 2 (117 subjects). In total, the dataset included 165,175 annotated ECG segments, including 24,252 AFib episodes. The fine-tuned model demonstrated superior performance compared to the original model on the TRAdA cohort, indicating improved adaptation to wearable ECG signals. On the Phase 2 dataset, the fine-tuned model achieved higher specificity (0.971 vs. 0.914), F1-score (0.891 vs. 0.777), area under the receiver operating characteristic curve (AUC; 0.992 vs. 0.988), and accuracy (0.969 vs. 0.924), highlighting its enhanced ability to detect AFib in extended real-world monitoring scenarios. These findings demonstrate that the domain adaptive fine-tuning significantly enhances AFib detection in long-term wearable ECG monitoring. Our findings support the feasibility of lightweight deep learning models for reliable AFib screening in real-world ambulatory settings.