Abstract 4146487: Prediction of New-onset Structural Heart Disease and Cardiovascular Risk Using Ensemble Deep Learning Applied to Electrocardiographic Images

Background: Despite the availability of effective pharmacological and procedural interventions, risk stratification strategies for structural heart disorders (SHDs) remain elusive. We developed an ensemble deep learning model for ECG images that predicts the development of future SHD and the risk of major adverse cardiovascular events (MACE). Methods: In Yale New Haven Health System (YNHHS), UK Biobank (UKB), and the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil), we included adults without baseline cardiomyopathy or heart failure. In YNHHS, where serial echocardiogram data were available, we identified the first transthoracic echocardiogram (TTE) with SHD (LVEF < 40%, moderate-to-severe left valve diseases, or severe LVH [IVSd > 15mm and LV diastolic dysfunction]). Across all cohorts, we followed individuals till the first MACE (HF/acute MI/stroke/all-cause death). We developed an ensemble convolutional neural network model that detects cross-sectional SHD from images of ECGs. We assessed the association of this model’s output probability with future risk of SHD or MACE using adjusted Cox models and discrimination using Harrel’s c-statistic. Results: Among 213,652 YNHHS patients, 11,695 developed SHD and 21,929 had MACE over 4.5 years (IQR 2.5-6.6). In UKB and ELSA-Brasil, among 42,147 and 13,454 people, 768 and 338 had MACE over 3.1 (2.1-4.5) and 4.2 (3.7-4.5) years. In YNHHS, a positive AI-ECG screen portended 2.5-fold higher risk of developing SHD (age-, sex-adjusted HR: YNHHS, 2.46 [95% CI, 2.35-2.57]) in those with serial TTEs. The association was consistent after accounting for comorbidities (aHR: 2.42 [2.32-2.54]) and the competing risk of death (aHR: 2.27 [2.17-2.38]). In YNHHS, model discrimination for new-onset SHD was 0.811 (0.807-0.814). Across cohorts, a positive AI-ECG screen portended a significantly higher risk of MACE (aHR: YNHHS, 1.86 [1.80-1.93]; UKB, 2.00 [1.62-2.46]; ELSA-Brasil, 3.01 [2.16-4.20]). Model discrimination for MACE was 0.763 (0.760-0.766) in YNHHS, 0.667 (0.646-0.688) in UKB, and 0.723 (0.695-0.751) in ELSA-Brasil. Conclusion: Across multinational cohorts, an ensemble AI model applied to a 12-lead ECG image identified those at elevated risk of SHD and MACE. This approach represents a digital biomarker to enable scalable cardiovascular risk stratification using ECG printouts or digital images, especially in low-resource settings.