JMIR Publications

ABSTRACT

Background:

Atrial fibrillation (AF) is a common arrhythmia associated with an increased risk of stroke and heart failure. To improve prevention and reduce complications, recent studies have utilized deep learning models to identify at-risk individuals early from the normal sinus rhythm (NSR). However, studies using mobile electrocardiogram (mECG) in outpatient, real-world settings remain underexplored.

Objective:

We develop and validate deep learning models using real-world limb lead mECG database to predict short-term AF occurrence from NSR recordings.

Methods:

The mECG data were collected from the real-world users of commercially available handheld mECG devices capable of capturing six limb leads. AF occurrence was defined as the presence of an AF event occurred within a predefined time window (7, 14, or 31 days) from the NSR recorded date. Transformer-based prediction models were developed with a multistage training, including self-supervised pretraining and domain adaptation, utilizing both the open large-scale clinical 12-lead ECG database and the proprietary real-world mECG database. The models were validated on an internal real-world cohort and further examined on an external cohort via the time-to-event analysis.

Results:

223,004 mECGs were acquired from 7,439 users. There were 18,949, 25,216, and 33,524 AF incidences within the 7, 14, and 31-day time windows. The models were then trained with 787,257 12-lead ECGs and 97,447 mECGs, respectively, and achieved areas under the receiver operating characteristic curves of 0.793, 0.785, and 0.787 for 7-, 14-, and 31-day predictions on the internal cohort. In the external cohort (n=144), the 31-day model perfectly stratified all the new-onset AF events, showing significantly different survival function between the positively and negatively predicted groups (p < 0.05).

Conclusions:

Our findings elucidate the feasibility of deep learning–based AF risk prediction using single NSR recordings from mobile devices, highlighting the potential for remote AF management in real-world population.