Abstract 17094: An Ensemble Deep Learning Model to Automate Screening For Multiple Structural Heart Diseases on 12-Lead Electrocardiograms

Abstract

Introduction: As echocardiographic screening is limited by access to technology, structural heart disease (SHD) is often diagnosed after the development of clinical symptoms. To enable automated and accessible screening of SHD, we developed and validated a deep-learning model for 12-lead electrocardiograms (ECGs) for various screening populations. Methods: We used 12-lead ECGs with paired TTEs at the Yale New Haven Hospital (2015-2021) to develop convolutional neural networks for detecting SHD. SHD was defined as the presence of any one of the following on a transthoracic echocardiogram (TTE) performed within 30 days of the ECG: hypertrophic cardiomyopathy (IVSd > 1.5cm and diastolic dysfunction), LV ejection fraction < 40%, or severe left-sided valvular disease (aortic/mitral stenosis or regurgitation). We developed an ensemble XGBoost model based on predictions for individual SHDs and patient age and sex as a single screen across all SHDs. We also simulated the model performance across cohorts with varying disease prevalence. Results: The model was developed in 456,927 ECGs/118,623 individuals (66.5±16.0 years, 45.6% women, 16.5% Black) and validated in 13,181 ECGs/13,181 individuals (65.1±17.3 years, 49.5% women, 14.5% Black). In the held-out validation set, the ensemble XGBoost model achieved an AUROC of 0.92 (95% CI: 0.91-0.94) and an AUPRC of 0.64 (95% CI: 0.57-0.70), with a sensitivity of 85% and specificity of 88%. The AUROCs for the individual disease models ranged from 0.72-0.95. (Panel A) In the test set with 10% disease prevalence, the ensemble model had a PPV 44% and an NPV of 98%. (Panel B) In simulated cohorts with 5% and 20% disease prevalence, the model had reached NPV & PPV of 99% & 27%, and 98% & 64%, respectively. (Panel C) Conclusion: We developed a novel ensemble deep-learning model for detecting several SHDs directly from ECGs with high PPV and NPV. This approach represents a robust, scalable, and accessible modality for automated SHD screening.