Development of deep learning models for predicting atrial fibrillation occurrence using real world hand-held mobile electrocardiograms

Abstract Background/Introduction Atrial fibrillation (AF) are major risk factors for stroke and other cardiovascular complications, making early detection crucial for preventive management. While handheld electrocardiogram (ECG) devices enable convenient self-monitoring, their potential for early prediction of arrhythmic events remains underexplored. Purpose We investigated the feasibility of deep learning models for predicting AF occurrence using commercially available handheld mobile ECG recordings composed of 6 limb-lead ECGs acquired from real-world users. Methods We developed deep learning models to analyze normal sinus rhythm (NSR) mobile ECGs and predict future AF occurrences. We utilized a limb lead (I, II, III, aVR, aVL, and aVF) mobile ECG database collected from commercially available handheld devices between 2023 and 2024. AF occurrence was labeled based on whether AF was detected within 7, 14, and 31 days after NSR signal acquisition, using device-generated diagnostic results. Mobile ECGs that did not exhibit NSR or lacked subsequent recordings within the specified time frames were excluded. For model training, we utilized a vision transformer model pretrained on a public 12-lead ECG dataset, MIMIC-IV-ECG, via Masked Autoencoder and finetuned it on the limb lead mobile ECG database. Two input configurations were considered: one using all limb leads and another using only lead I, resulting in six trained models. Results A total of 304,716 mobile ECGs were collected from 7,816 users. After labeling and exclusion, 127,146, 138,141, and 150,291 NSR recordings remained for the 7-day, 14-day, and 31-day prediction settings, respectively. Among these, 16.5%, 20.1%, and 24.5% of recordings were labeled as AF occurrences within the corresponding time frames. The area under the receiver operating characteristics curves (AUCs) for the limb lead models were 0.784, 0.774, and 0.773 for the 7-day, 14-day, and 31-day settings, while the AUCs for the lead I models were 0.723, 0.720, and 0.671. The limb lead models consistently outperformed the lead I models across all the time frames (p 0.001, DeLong’s method). Conclusion(s) This study demonstrated the feasibility of deep learning-based AF occurrence prediction using ECG recordings from commercially available mobile ECG devices in the general population, irrespective of hospital visit. It potentially enables proactive AF risk assessment and timely clinical intervention, even at public level. Predictive performance improved as the time frame shortened. The limb lead models showed higher predictability than the lead I models, suggesting that predicting AF occurrence may require additional information beyond rhythm irregularity within a single lead, potentially leveraging multi-lead inputs for more comprehensive signal interpretation.