Deep Learning for Detection of Pulmonary Hypertension and Reduced Left Ventricular Ejection Fraction Using a Combined Digital Stethoscope and Three-lead Electrocardiogram

Pulmonary hypertension (PH) and low left ventricular ejection fraction (EF) are significant cardiovascular conditions associated with increased morbidity and mortality but often remain underdiagnosed due to the need for specialized imaging such as echocardiography or invasive right heart catheterization. Early detection tools could enable timely intervention and improved patient outcomes.

This prospective, observational study aims to determine whether acoustic heart sounds (phonocardiograms, PCG) and three-lead electrocardiograms (ECG) recorded with the Eko CORE 500 digital stethoscope can identify patients with PH or low EF (defined as EF ≤ 40%) when compared with echocardiographic or right heart catheterization findings as the reference standard. The study will enroll adult patients undergoing clinically indicated transthoracic echocardiography or right heart catheterization at outpatient sites.

Participants will complete a single study visit, lasting approximately 20 minutes, during which heart sounds and three-lead ECG signals will be recorded at four standard auscultation sites (aortic, pulmonic, tricuspid, and mitral) while seated. Each recording lasts approximately 15 seconds. If a participant has undergone a 12-lead ECG within 30 days of their echocardiogram or right heart catheterization, de-identified ECG data will also be included for comparison purposes. Poor-quality recordings will be repeated once before moving to the next auscultation site. No results from the CORE 500 device or developed algorithms will be shared with participants or entered into the medical record.

De-identified demographic data collected will include age, race/ethnicity, and sex. Clinical data will include past medical history, relevant laboratory results (such as BNP or NT-proBNP), electrocardiographic findings, and echocardiographic measurements including tricuspid regurgitant jet velocity, pulmonary artery pressures, chamber size, and left ventricular ejection fraction.

Data will be analyzed by Eko Health, Inc. using machine learning techniques, including transformer-based models implemented in Python with PyTorch. Models will initially be pre-trained on unlabeled data and then fine-tuned on labeled data, optimizing performance using the Adam optimizer and binary cross-entropy loss. Algorithm performance will be evaluated based on sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and other diagnostic metrics. Confidence intervals for sensitivity and specificity will be calculated to assess the statistical reliability of results.