Multiethnic Validation of Artificial Intelligence-Enhanced Electrocardiographic Image Analysis in Detecting Cardiac Structural and Functional Abnormalities: A UK Biobank Study

Background Although artificial intelligence–enhanced electrocardiography (AI-ECG) has shown promise in detecting cardiac abnormalities, large-scale validation against cardiac magnetic resonance (CMR) parameters in a multiethnic general population remains limited. Methods In 38,804 UK Biobank participants with paired ECG and CMR data, we evaluated six AI-ECG models: four targeting functional abnormalities (left and right ventricular dysfunction [QCG-LVD and QCG-RVD], global longitudinal strain [ECG-LVGLS and ECG-RVGLS]) and two targeting structural abnormalities (left ventricular hypertrophy [AI-ECG-LVH] and left atrial enlargement [AI-ECG-LAE]). Abnormalities were defined as the top 1% extreme values of each CMR-derived parameter. Results The AI-ECG models demonstrated robust diagnostic performance. For left ventricular dysfunction, the AUCs were 0.887 (QCG-LVD) and 0.896 (ECG-LVGLS); for right ventricular dysfunction, the AUCs were 0.778 (QCG-RVD) and 0.825 (ECG-RVGLS). In a sub-cohort with available corresponding CMR data (n=21,267), AUCs were 0.824 for AI-ECG-LVH and 0.883 for AI-ECG-LAE. Subgroup analyses showed consistent performance across demographic and clinical groups, with enhanced accuracy observed in older individuals, males, and those with hypertension. Conclusion In this large-scale, population-based study, AI-ECG demonstrated strong performance in detecting structural and functional cardiac abnormalities defined by CMR. These findings support the potential utility of AI-ECG as a scalable screening tool for cardiovascular disease in general populations.