Deep learning-based detection of paroxysmal atrial fibrillation from sinus rhythm electrocardiograms and its association with stroke risk

Introduction Atrial fibrillation (AF) is associated with increased stroke risk and deep-learning model by electrocardiogram (ECG) enabled AF early detection even during normal sinus rhythm (NSR). However, impact of AF early detection on subsequent stroke is not well known. Purpose The primary aim was to develop and externally validate a deep learning model identifying paroxysmal AF (pAF) from standard 12-lead NSR ECGs. As a secondary objective, we investigated whether model output among patients without known AF/atrial flutter (AFL) was associated with subsequent stroke incidence. Methods We conducted a retrospective, multi-centre study using 12-lead ECG datasets from two independent institutions (n=171,100). For the positive cohort, NSR ECGs recorded within 14 days prior to an AF/AFL event were selected. For the negative cohort, patients without any AF/AFL history who had at least three NSR ECGs preceding the index recording were included. A binary classification model was trained on internal train and validation sets from Hospital A, and evaluated on an internal test set (n=11,172) and an external set from the MIMIC-IV ECG dataset (n=112,584). Among non-AF/AFL patients, subjects were stratified into positively- and negatively-predicted groups based on the model output at an optimal threshold, and stroke incidence was compared using Kaplan-Meier analysis with the log-rank test. Results The internal test set comprised 3,663 positive and 7,509 negative cases; the external validation set included 47,054 positive and 65,530 negative cases. The model achieved an AUROC of 0.853 (95% confidence interval [CI] 0.845–0.860) on the internal test set and 0.809 (95% CI 0.807–0.812) on the external validation set. At the optimal threshold, internal sensitivity and specificity were 71.8% and 82.5%, respectively, with corresponding external values of 82.6% and 61.4%. In the secondary analysis of 7,902 non-AF/AFL patients, Kaplan-Meier analysis over 12 months demonstrated higher cumulative stroke incidence in the positively-predicted group compared with the negatively-predicted group (57.46 % vs. 42.50 %, log-rank p<0.05) in the external validation set. Conclusion The model demonstrated robust performance across internal and external validation, effectively identifying pAF from NSR ECGs. Among patients without known AF/AFL, a positive model prediction was associated with elevated stroke risk, suggesting the model may identify hidden stroke risk in patients with undiagnosed AF/AFL.stroke predictionAF prediction