Abstract 4354222: Deep Learning Prediction of Left Atrial Structure and Function from 12-lead Electrocardiograms

Abstract

Introduction: Abnormalities in the function and structure of the left atrium, called atrial cardiopathy, are a precursor of atrial fibrillation (AF) and an important risk factor for several other cardiovascular outcomes, yet detecting abnormalities is challenging due to the cost and limited accessibility of high-quality cardiac imaging. Objective: To develop and validate a deep learning model (ECG-AI) that estimates left atrial (LA) structure and function from the resting 12-lead electrocardiogram and reliably predicts cardiovascular outcomes. Methods: We trained ECG-AI models on cardiac magnetic resonance imaging data from the UK Biobank (n=21,749) to estimate LA minimum volumes, maximum volumes and ejection fraction. Volumes were indexed to body surface area. Cox regression models adjusting for clinical risk factors evaluated associations of LA with incident AF, ischemic stroke and cardioembolic stroke in an external cohort of adults aged ≥ 65, the Cardiovascular Health Study. We compared prediction models for incident AF and cardioembolic stroke that included [1] age and sex only (base), [2] base + ECG-AI LA measures (ECG-AI), [3] CHARGE-AF risk score alone, and [4] CHARGE-AF + ECG-AI (combined). Results: ECG-AI estimates were moderately correlated with direct imaging measures (r=0.40-0.50) but demonstrated strong independent associations with outcomes in the external cohort and outperformed a conventional ECG measure of atrial cardiopathy, P terminal force in V1. In the Cardiovascular Health Study, each standard deviation increase in ECG-AI LA minimum volume yielded HRs of 1.44 (95% CI 1.35-1.47) for AF, 1.45 (95% CI 1.37-1.53) for ischemic stroke, and 1.66 (95% CI 1.47-1.86) for cardioembolic stroke, the hallmark complication of AF and atrial cardiopathy (Figure 1). In contrast, none of the ECG-AI measures was associated with large-artery atherosclerotic or small vessel stroke. In 5-year prediction models (Figure 2), the ECG-AI measures outperformed the CHARGE-AF risk prediction tool for both AF (delta AUC 0.03, 95% CI 0.01-0.05) and cardioembolic stroke (delta AUC 0.01, 95% CI -0.05-0.08). Conclusion: Our fully-trained ECG-AI tool estimates LA function and identifies individuals at elevated-risk for cardiovascular disease using only standard data from the inexpensive and widely-available 12-lead ECG.