Abstract 4146678: Artificial Intelligence-enhanced Interpretation of Single-lead Electrocardiogram in Normal Sinus Rhythm to Detect and Predict Atrial Fibrillation and Flutter

Background: Portable ECG devices can help identify symptomatic atrial fibrillation/flutter (AF) episodes outside healthcare settings. However, most individuals remain undiagnosed due to the transient and often asymptomatic nature of AF. Here, we developed an AI-ECG model to detect concomitant paroxysmal AF and predict its future risk using lead I ECG recordings in normal sinus rhythm (NSR), with special adaptation for noisy acquisition on portable devices. Methods: The model was developed using 1.6 million ECGs at Yale (2000-2015). ECGs in NSR associated with an ECG with AF rhythm from 30 days before to any time after were considered NSR with concomitant AF. A convolutional neural network was trained using lead I ECG to detect concomitant AF among NSR ECGs, augmented using random Gaussian noise during training to make them resilient to noisy acquisition on portable devices. We evaluated the model's probability as a predictor for incident AF (defined by AF ECGs or clinical diagnosis), in (1) a separate set of individuals in a temporally distinct population at Yale (2016-2023) and (2) the UK Biobank (UKB) (2014-2021), excluding those without a history of AF. We constructed age- and sex-adjusted Cox proportional hazards models with model probability as a key independent variable. Results: The AI-ECG model achieved an AUROC of 0.821 (95% CI, 0.818-0.824) for detecting concomitant AF based on a single-lead ECG in NSR with a sensitivity of 75%, a specificity of 75%, a PPV of 28%, and an NPV of 96%. In the predictive assessment, 312,533 individuals (52 [IQR, 32-66] years, 54% women) at Yale, and 41,711 individuals (65 [59-71] years, 52% women) in UKB had no known AF. In these groups, 3.1% at Yale and 1.4% in UKB had incident AF over median follow-ups of 3.1 [1.6-4.5] years and 3.0 [2.1-4.5] years, respectively. AI-ECG had high discrimination for future AF risk (Harrell’s C-statistic, 0.77 in Yale and 0.68 in UKB). After accounting for age and sex, a positive AI-ECG was associated with a 2-fold higher hazard of AF (adjusted HR: 1.93 [1.84-2.02] in Yale and 2.34 [1.98-2.76] in UKB). Conclusion: We developed an AI-ECG model capable of detecting concomitant paroxysmal AF and predicting incident AF using a single-lead ECG noted to be in NSR. With the growing availability and accessibility of portable ECGs, this model enables a scalable screening and risk stratification strategy for AF, especially in older adults at increased risk of stroke, in the community.