Abstract 4142809: Artificial Intelligence Enabled Prediction of Future Structural Heart Disease and Cardiovascular Risk from Single-lead Electrocardiograms

Background: Portable devices capable of recording single-lead electrocardiograms (ECGs) can enable large-scale community-based cardiovascular risk assessment. We developed a noise-adapted ensemble deep learning algorithm for lead I ECGs as the sole input and predicted the risk of major adverse cardiovascular events (MACE) and development of new-onset structural heart disorders (SHDs) across multinational cohorts, spanning a US health system and community-based cohorts in UK and Brazil. Methods: We included adults with outpatient ECGs in Yale New Haven Health System (YNHHS) and population-based cohorts of UK Biobank (UKB) and Brazilian Longitudinal Study of Adult Health (ELSA-Brasil). Among individuals with ECGs without cardiomyopathy or heart failure at baseline, we isolated lead I ECGs and deployed a noise-adapted AI-ECG model trained to identify cross-sectional SHD (LV ejection fraction < 40%, moderate/severe left-sided valve disease, severe LV hypertrophy [IVSd > 15mm and LV diastolic dysfunction]). Across cohorts, we evaluated the association of the baseline AI-ECG probability with risk of MACE using adjusted Cox proportional hazard models and assessed discrimination using Harrel’s c-statistic. In YNHHS, among those with serial echocardiograms, we evaluated the risk of new-onset SHD. Results: There were 169,381 patients with outpatient ECGs at YNHHS (age 55 years [IQR, 40-68], 98023 women [58%]), 42,741 UKB participants (65 years [59-71], 21,795 women [52%]), and 13,454 ELSA-Brasil participants (56 years [41-69], 7,348 women [55%]) with baseline ECGs. A total of 12,483 patients had MACE and 8595 developed SHD in YNHHS over 4.5 years (2.6-6.6). In UKB and ELSA-Brasil, 768 and 388 participants had MACE over 3.1 years (2.1-4.5) and 4.2 years (3.7-4.5), respectively. Each 0.1 increment in the model output portended a 15-35% higher hazard of MACE across cohorts, independent of age, sex, and comorbidities (aHR: YNHHS 1.20 [1.19-1.21], UKB 1.15 [1.09-1.20], ELSA-Brasil 1.35 [1.28-1.44]). Among patients with serial echocardiograms in YNHHS, a positive AI-ECG screen portended a 2-fold higher risk of new-onset SHD (aHR: 2.09 [1.98-2.20]). Conclusions: In geographically and clinically distinct cohorts, a noise-adapted AI model with lead I ECGs as the only input defined the risk of MACE and new-onset SHD. This represents a scalable strategy that can use portable and wearable devices with ECG capabilities for cardiovascular risk stratification in the community.