Leveraging deep learning for detecting atrial fibrillation and flutter and predicting thromboembolic events from images of 12-lead electrocardiograms in sinus rhythm

Background Artificial intelligence-enhanced interpretation of electrocardiography (AI-ECG) can detect atrial fibrillation and flutter (AF) from sinus rhythm ECGs (srECG). However, these algorithms rely on ECG signals, often unavailable at the point of care. We sought to develop a broadly accessible AI model that could detect risk of concomitant AF directly from widely available ECG images. Additionally, we sought to assess whether such an AI-ECG model could serve as a digital biomarker for thromboembolic events. Objective We developed and externally validated an AI-ECG algorithm to detect concomitant AF from images of 12-lead ECGs in sinus rhythm. We further used the model’s output for predicting thromboembolic events. Methods We developed a computer vision deep learning model using 4 million ECGs from 691,463 patients at a large, academic US hospital. For model development, concomitant AF with an srECG (AF-srECG) was characterized if the patient had an ECG with AF within 30 days prior or at any point thereafter. The model was trained on various lead layouts to ensure it remains layout-agnostic during deployment. The model was externally validated in four US community hospitals (N=171,189), one US outpatient network (N=28,254), and the population-based UK Biobank (UKB) cohort (N=42,739). We assessed the model’s performance in predicting thromboembolic events. For benchmarking, the AI-ECG’s performance for predicting thromboembolic events was compared with CHA2DS2-VASc. Results The AI-ECG demonstrated an area under the receiver operating characteristic curve (AUROC) of 0.802 (95% CI, 0.794–0.809) for detecting AF-srECG in the held-out test set, with the AUROC ranging from 0.798–0.874 in clinical external validation cohorts, and 0.802 (0.787–0.818) in UKB. A positive AI-ECG screen was associated with a higher hazard of thromboembolic events (HR, 3.59 [2.99–4.32] in clinical sites and HR, 8.06 [3.56–18.28] in UKB). For predicting thromboembolic events, the Harrell’s C-statistic was comparable for AI-ECG and CHA2DS2-VASc in clinical sites (0.764 [0.741–0.787] vs. 0.800 [0.779–0.821]). However, AI-ECG outperformed CHA2DS2-VASc in UKB (0.744 [0.714–0.775] vs. 0.570 [0.539–0.600]). Conclusion We report a novel AI-ECG algorithm that can detect concomitant AF among those in sinus rhythm, and predict future thromboembolic outcomes from images of 12-lead ECGs. The algorithm that relied on a single ECG either matched or exceeded the performance of traditional risk scores that require evaluation of multiple risk factors. The AI-ECG algorithm has the potential to enhance the detection, prediction, and risk stratification of AF at the point of care using a single ECG image.