Abstract
Hypertrophic cardiomyopathy (HCM) is frequently underdiagnosed. Although deep learning (DL) models using raw electrocardiographic (ECG) voltage data can enhance detection, their use at the point of care is limited. Here we report the development and validation of a DL model that detects HCM from images of 12-lead ECGs across layouts. The model was developed using 124,553 ECGs from 66,987 individuals at the Yale New Haven Hospital (YNHH), with HCM features determined by concurrent imaging (cardiac magnetic resonance (CMR) or echocardiography). External validation included ECG images from MIMIC-IV, the Amsterdam University Medical Center (AUMC) and the UK Biobank (UKB), where HCM was defined by CMR (YNHH, MIMIC-IV and AUMC) and diagnosis codes (UKB). The model demonstrated robust performance across image formats and sites (areas under the receiver operating characteristic curve (AUROCs): 0.95 internal testing; 0.94 MIMIC-IV; 0.92 AUMC; 0.91 UKB). Discriminative features localized to anterior/lateral leads (V4 and V5) regardless of layout. This approach enables scalable, image-based screening for HCM across clinical settings.