Personalized deep learning for short-term forecasting of impending atrial fibrillation from continuous wearable ECG signals

Abstract Continuous wearable electrocardiogram (ECG) monitoring is increasingly used for ambulatory arrhythmia surveillance, yet forecasting impending atrial fibrillation (AF) remains challenging because of inter-patient ECG variability. We investigated whether personalizing a global model by fine-tuning it on an individual’s ECG improves short-term AF forecasting. A global model trained on ICENTIA11K was compared with personalized models fine-tuned across three cohorts (ICENTIA11K, IRIDIA-AF, and MobiCARE), using 60-second ECG segments and a five-minute forecast horizon. We assessed how the amount of adaptation data affected performance and analyzed ECG features such as heart rate and RMSSD. Personalized models significantly outperformed the global model, with AUROCs of 0.711 vs. 0.614 (ICENTIA11K) and 0.686 vs. 0.585 (MobiCARE), and the benefits grew with more patient-specific fine-tuning data. While the global model’s accuracy rose as AF onset approached, personalized models in the two external cohorts showed distinct temporal dynamics, suggesting that they captured patient-specific cues less dependent on onset proximity. Pre-AF episodes showed elevated heart rate and RMSSD, and feature attributions highlighted clinically relevant precursors, including frequent premature atrial complexes (PACs) and short supraventricular tachycardias (SVTs). Adapting deep learning models with patient-specific wearable ECG data significantly enhances short-term AF forecasting, supporting timely preventive intervention and improved AF management in ambulatory monitoring.