Abstract 4371172: Artificial Intelligence-based screening for Hypertrophic Cardiomyopathy from Single-lead Electrocardiograms: A Multinational Development and Validation Study

Abstract

Background: Hypertrophic cardiomyopathy (HCM) is a leading cause of sudden death in young and middle-aged adults and frequently goes undiagnosed due to the absence of accessible and scalable screening strategies. Artificial intelligence (AI) applied to single-lead electrocardiograms (ECGs) from portable devices offers a promising approach for large-scale screening. However, noisy signals can substantially compromise diagnostic accuracy. Aim: We developed and validated a noise-adapted AI-ECG model specifically designed to detect HCM from noisy single-lead ECGs. Methods: We developed an AI-ECG model using lead I from 160,396 unique 12-lead ECGs of 85,967 individuals in the Yale New Haven Health System (YNHHS) and augmented the ECG signal with real-world noises to develop noise-resilient models. A held-out test including 38,426 ECGs from unique individuals (mean age of 53.9 ± 19.3 years, 20,309 [53%] women) was used for internal validation. There were 59 (0.2%) HCM cases, adjudicated by expert clinicians using cardiac magnetic resonance (CMR) imaging. External validation was performed in manually validated MIMIC-IV (n=995, 66 HCM cases) and the UK Biobank (n=57,963, 53 HCM cases). To assess model fairness, we conducted stratified analyses by age, sex, race/ethnicity, and key ECG features. Results: The model demonstrated robust discrimination in internal validation with an area under the receiver operating curve (AUROC) of 0.95 (95% CI 0.93-0.97), sensitivity of 0.90 and specificity op 0.90 ( Fig. 1 ). The performance remained consistent in external validation in MIMIC-IV (AUROC 0.92 [0.89-0.95]) and UK Biobank (AUROC 0.88 [0.82-0.93]). Stratified analyses demonstrated stable model performance across all demographic subgroups, including age (≥60,> 60 and <60), sex (men vs. women), and racial/ethnic subgroups. ( Fig. 2 ) However, diagnostic accuracy was lower in patients with left bundle branch block or atrial fibrillation. Conclusion: Our findings demonstrate that a noise-adapted AI-ECG model can effectively detect HCM from noisy, single-lead ECGs. This approach has the potential to enable scalable, automated, and accessible screening using wearable or portable ECG devices for the earlier identification of HCM in community settings.