Abstract 4146190: Multinational Clinical and Community-based Validation of PRESENT-SHD, an Ensemble Deep Learning Approach for the Detection of Multiple Structural Heart Disorders Using Electrocardiographic Images

Background: Screening for structural heart disorders (SHDs) requires cardiac imaging, which has limited accessibility and requires clinical expertise. We report the multinational validation of PRESENT-SHD, an artificial intelligence-based approach for detecting multiple SHDs from a single image of a 12-lead electrocardiogram (ECG) across clinical settings in the US and a community-based cohort from Brazil. Methods: Using paired ECG images and transthoracic echocardiograms (TTEs) from the Yale New Haven Hospital (2015-2023) from 110,228 patients, we developed PRESENT-SHD, an ensemble deep learning approach combining CNN probabilities for individual SHDs with a person’s age and sex, as a single screen for multiple SHDs. SHD was defined as any LV systolic dysfunction (LVEF < 40%), moderate-or-severe left-sided valve disease, or severe LVH (IVSd > 15mm and LV diastolic dysfunction). We validated PRESENT-SHD across patients from 4 distinct hospitals and an outpatient medical network (Northeast Medical Group [NEMG]) in the US and in a large prospective population-based cohort study, the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil), who underwent protocolized ECGs and TTEs concurrently. Results: We evaluated the model on a single ECG drawn from 8921 individuals at Bridgeport Hospital, 2507 at Greenwich Hospital, 17,856 at Lawrence + Memorial Hospital, 3638 at Westerly Hospital, 21,055 at NEMG, and 3014 in ELSA-Brasil. SHD prevalence ranged from 20-23% at the hospital sites and 22% at the outpatient NEMG. It was 3% in the population-based ELSA-Brasil. In the hospital-based sites, PRESENT-SHD had an AUROC ranging 0.847-0.899, with sensitivities and specifities of 81-85% and 72-80%, respectively. In the outpatient NEMG, the performance was also consistent (AUROC, 0.815 [95% CI, 0.808-0.822], sensitivity 61%, specificity 84%). In ELSA-Brasil, the model had an AUROC of 0.853 (95% CI: 0.808-0.890) with sensitivity and specificity of 63% and 90%, respectively. Model performance was consistent across demographic subgroups. Conclusion: PRESENT-SHD maintained excellent performance in identifying SHD on ECG images across diverse populations and clinical settings in the US and Brazil. This demonstrates strong potential as a scalable and accessible tool for SHD screening using digital or printed versions of 12-lead ECGs.