Abstract 4358082: Population Screening for LV Dysfunction Using Single-Lead ECG-AI: Insights from the RURAL Study

Background: Moderate to severe left ventricular dysfunction (LVD), defined by a left ventricular ejection fraction (LVEF) <40%, is a key marker of heart failure with reduced ejection fraction (HFrEF). Early detection is crucial to reduce heart failure burden, yet no effective screening tool exists at the population level. While we and others have developed electrocardiographic artificial intelligence (ECG-AI) models to detect LVD, they are not yet appropriate for widespread screening due to selection bias—most ECGs used in training were collected for clinical, not screening, purposes. Community-based validation is needed to assess model generalizability. Goal: This study aims to evaluate the performance of a previously developed ECG-AI model for left ventricular dysfunction (LVD) detection (Wake Forest’s ECG-AI model) among participants in the RURAL (Risk Underlying Rural Areas Longitudinal) Cohort Study. Methods: Wake Forest’s ECG-AI Model is a ResNet-based deep learning algorithm originally developed to detect both LVD and HF with preserved ejection fraction. Trained on over 1 million lead I ECGs from Wake Forest School of Medicine, it achieved an AUC of 0.89 for identifying LVEF <40% on a held-out test set. External validation using ~100,000 ECGs from University of Tennessee Health Science Center, Memphis, TN, showed an AUC of 0.90. In the RURAL longitudinal cohort, participants had both 12-lead ECGs and echocardiograms at enrollment. Without calibration or fine-tuning, we applied the model to lead I ECGs from RURAL and calculated the AUC for detecting LVEF <40%. Results: The analytical cohort included 1,276 RURAL participants from Alabama and Mississippi (65.5% women, 63.8% Black, mean age 48.5±11.0 years). Fifteen participants (1.2%) had LVEF<40%. The model achieved an AUC of 0.91 (95% CI: 0.80–1.00). At the optimal Youden Index threshold of 0.821, specificity was 95.4%, sensitivity 86.7%, negative predictive value 99.7%, and positive predictive value 18.3%. The model identified 71 participants as having LVEF<40%, including 13 true positives. Among the 58 false positives, 12 had mildly reduced LVEF between 40-50%. Conclusion: The Wake Forest ECG-AI Model showed strong performance in identifying moderate to severe LVD using only a 10-second single-lead ECG in a community-based setting. Further validation in larger cohorts is needed to support its use as a scalable, non-invasive population screening tool.