Capitalizing on Artificial Intelligence to Capture Undiagnosed Structural Heart Disease from Electrocardiograms (CACTUS)

Structural heart disease (SHD), including valvular heart disease, right or left ventricular dysfunction, and various cardiomyopathies, is a significant cause of cardiovascular morbidity and mortality. Despite the benefits of early treatment, SHD remains underdiagnosed. Transthoracic echocardiography (TTE), a standard diagnostic method, has limited use in population screening and in emergency departments (ED) due to cost and expertise requirements. Patients from disadvantaged backgrounds, often relying on EDs for primary care, are particularly affected. Traditional ECG interpretation fails in accurately detecting SHD, necessitating a novel approach. To address this limitation, we created a deep learning model to interpret raw ECG waveforms, named EchoNext, that generates a composite prediction for SHD with high accuracy to determine which ECGs should be followed up by TTE. EchoNext is the first successful, unified model detecting all components of SHD from an ECG waveform. The model was built using our database of 12 million ECGs and 1.5 million echocardiograms from 8 hospitals, 6 echocardiography labs, and 190 clinics serving a diverse population disproportionately at risk for SHD. The EchoNext model has been retrospectively validated at 11 hospitals across 4 health systems that are geographically distinct and serve diverse populations. This project will determine the accuracy of EchoNext in the detection of undiagnosed SHD among patients presenting to six emergency departments. Our specific aims are: 1) Perform a large, prospective, observational study to validate the accuracy of EchoNext in detecting undiagnosed SHD, 2) Determine the impact of demographics/socioeconomic status on TTE referral rates and if EchoNext accuracy varies in disadvantaged populations, and 3) Assess AI model performance stability over 4 years across various health system populations. This project aligns with NHLBI priorities by leveraging emerging opportunities in data science to create a novel diagnostic strategy for SHD through the use of deep learning, investigate factors contributing to health disparities, and study real-world performance to optimize future clinical deployment of AI models. EchoNext may enable effective, equitable diagnosis of SHD to improve cardiovascular outcomes for at-risk patients. Project Number: 1R01HL177055-01 | Fiscal Year: 2025 | NIH Institute/Center: National Heart Lung and Blood Institute (NHLBI) | Principal Investigator: Pierre Elias | Institution: COLUMBIA UNIVERSITY HEALTH SCIENCES, NEW YORK, NY | Award Amount: $704,611 | Activity Code: R01 | Study Section: Special Emphasis Panel[ZRG1 HSS-D (90)] View on NIH RePORTER: https://reporter.nih.gov/project-details/1R01HL17705501