AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Artificial Intelligence-Based Electrocardiographic Screening Detects Subclinical Atrial Fibrillation Earlier Than Standard 12-Lead Interpretation
By Chaudhry WR, Jawed I, Kiyani M et al.
Introduction: Atrial fibrillation (AF) remains the commonest sustained cardiac arrhythmia with a notable contribution to stroke, heart failure, and cardiovascular death. Paroxysmal and subclinical AF often go undetected on standard 12-lead ECG. According to a public health paper, a resting 12-lead electrocardiogram (ECG) equipped with artificial intelligence (AI) deep learning algorithms could be used for the detection of atrial fibrillation (AF) during sinus rhythm. Objective: We hypothesized that an AI ECG algorithm would be better able to identify subclinical AF on sinus rhythm ECGs with superior sensitivity and earlier detection than standard clinician ECG interpretation in a prospective community cohort. Methods: A planned observational cohort study recruited adults (n=22,418; mean age: 61.3±14.2 years; 51.4% female) known to have atrial fibrillation across 12 community health systems during 2022-2025. All candidates underwent a resting 12-lead ECG, followed by continuous remote cardiac monitoring (14-day patch monitor), the reference standard for atrial fibrillation diagnosis. A validated convolutional neural network (CNN) AI-ECG model was used to calculate an AF probability score from baseline ECGs. The clinician's standard interpretation of ECGs served as the control. We computed diagnostic accuracy metrics, including AUC, sensitivity, specificity, positive predictive value, and negative predictive value. The time-to-diagnosis sample was analyzed using the Kaplan-Meier method. Analyses were by subgroup: age, sex, and CHA ₂ DS ₂ -VASc score. Results: Over a mean follow-up of 18.6 months, subclinical AF was detected by a 14-day monitor in 2,847 participants (12.7%). The AUC was 0.89 (95% CI 0.87-0.91) for the AI-ECG model and 0.61 (95% CI 0.58-0.64) for standard clinician interpretation (p<0.001). The sensitivity was 81.3% versus 38.4%. The specificity was 84.9% versus 83.6%. The PPV was 52.3% versus 26.1%. The NPV was 95.8% versus 90.2%. The median time to diagnosis of AF was 38 days (IQR 19-67) in the AI-guided arm versus 126 days (IQR 64-218) in the standard arm (log-rank p<0.001). In high-risk participants (CHA ₂ DS ₂ -VASc ≥2), AI-ECG was 87.4% sensitive. The AI model was associated with 3.1 times the odds of detecting atrial fibrillation (AF) before the first clinical event (stroke or cardiovascular hospitalization) (OR 3.14; 95% CI 2.61-3.78; p<0.001). Conclusion: An ECG screening system powered by artificial intelligence detects subclinical atrial fibrillation with accuracy and speed that exceed those of standard clinical interpretation of sinus rhythm tracings. By integrating AI-ECG assessment into routine health care workflows at the community level, earlier anticoagulation and stroke prevention may occur in atrial fibrillation high-risk individuals.