Researchers trained a deep learning model on 12-lead electrocardiograms to predict left atrial structure and function, according to a study published in Nature Communications on July 31, 2026. The neural network was developed using electrocardiograms paired with 21,749 cardiac magnetic resonance scans from the UK Biobank database.
Abnormal atrial structure and function, known as atrial cardiopathy, typically precedes atrial fibrillation and downstream cardiovascular complications. Detection is currently limited by the high cost and limited accessibility of cardiac imaging. The authors presented the model as an inexpensive, accessible tool that evaluates standard electrocardiograms to identify individuals at high risk for complications.
In two external validation cohorts, model-derived measures of atrial cardiopathy were strongly associated with new-onset atrial fibrillation, heart failure, and ischemic stroke. These associations remained significant after adjusting for clinical risk factors and biomarkers. The predictive magnitudes were comparable to or greater than those of direct imaging measures and clinical risk factors.
Predictive Performance
The study showed that cardioembolic stroke risk, a hallmark complication of atrial fibrillation, increases 66 percent per standard deviation of predicted left atrial volume. In exploratory analyses, the deep learning model predicted cardiac monitor-detected atrial fibrillation more accurately than a clinical risk prediction tool or levels of the biomarker NT-proBNP.
Jennifer A. Brody and Vidhushei Yogeswaran contributed equally as first authors alongside investigators from the University of Washington, University of Minnesota, Johns Hopkins Hospital, Wake Forest University, and the University of California-San Francisco. The research received financial support from institutes within the National Institutes of Health, including the National Heart, Lung, and Blood Institute.
