Prediction of incident atrial fibrillation from a 10-second single-lead ECG using a neural network

Introduction Artificial intelligence has enabled atrial fibrillation (AF) risk estimation from sinus rhythm electrocardiograms (ECGs). Although short single-lead recordings have shown associations with future AF in selected clinical and screening settings, evidence from large service-based registries using a single 10-second ECG recording remains limited. Purpose To evaluate the performance in subgroups of a neural network applied to a 10-second single-lead ECG for prediction of incident AF in a large registry of telehealth ECGs. Methods A total of 541,516 tele-ECG recordings from 127,988 individuals without AF at baseline were available from a large telehealth-based 12-lead ECG cohort in routine clinical care. To ensure patient-level analysis and avoid repeated measurements, the earliest eligible 10-second ECG per participant meeting predefined quality criteria was selected, yielding 106,862 individuals for analysis. Non-AF-participants had a pre-specified minimum follow-up of 3 years (median 4.0 years); incident AF was identified through follow-up cardiologist annotations over a maximum follow-up of 7.4 years. Discrimination was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC). Predefined subgroup analyses were performed according to sex, age groups, and baseline rhythm characteristics. Results Among 106,862 individuals (median age 56 years; 63.6% women), 6,048 developed AF during follow-up, with a median time to AF of 1.75 years after baseline. Overall discrimination by the neural network was good (AUROC 0.797; 95% CI 0.791–0.803; AUPRC 0.243; Figure 1). Predictive performance was higher in women than in men and remained robust across age groups, although AUPRC values were lower in younger individuals (Figure 2A). Reduced discrimination was observed in the presence of atrioventricular block, bundle branch blocks, sinus bradycardia, and sinus tachycardia (Figure 2B). Conclusions A single 10-second single-lead ECG recording provides sufficient information to predict incident AF with moderate discrimination in a registry of telehealth ECGs. Although performance varies by sex and rhythm context, the model maintains robust predictive ability across age and clinically relevant subgroups. These findings support the feasibility of scalable AI-based AF risk stratification using short ECG recordings.