Abstract 9905: A Deep Neural Network Detects Elevated Right Ventricular Systolic Pressure Using Data Captured from a Digital Stethoscope

Abstract

Introduction: Pulmonary hypertension (PH) and elevated peak right ventricular systolic pressure (RVSP) are associated with poor outcomes in a number of cardiopulmonary diseases. PH often goes undiagnosed because right heart catheterization is invasive and expensive. RVSP measured by echocardiography has been used with some success as a non-invasive proxy, yet it is still expensive and operator-dependent, resulting in inadequate and/or delayed diagnosis of PH. We developed and evaluated an artificial intelligence (AI) tool to detect elevated RVSP, using phonocardiogram (PCG) recorded from digital stethoscopes. Methods: 6,062 15-second PCG recordings were taken using the Eko CORE and DUO digital stethoscopes from 797 patients with corresponding echocardiogram-measured RVSP. These, plus an additional 169,160 unlabelled PCG recordings with no corresponding echocardiogram, were used to train a deep convolutional neural network to identify an RVSP above 40mmHg. Mel spectrograms from the PCGs were randomly cropped into 5s segments and used to train the network. The networks were evaluated via 5-fold cross validation with the folds split by patient to prevent leakage between training and test sets. Results: Across the 5 cross validation folds, the model produced an average area under the ROC curve of 0.79 (range 0.76 - 0.85), sensitivity of 0.73 (range 0.67 - 0.80) and specificity of 0.74 (range 0.64 - 0.82) at detecting a peak RVSP >= 40 mmHg. Incorporating patient age, gender and body mass index into the network did not improve performance. Conclusions: We demonstrated the ability of an AI tool to detect elevated peak RVSP using PCG. Use of a digital stethoscope with this tool has the potential to be a low-cost, rapid and minimally invasive screen for elevated RVSP in a point-of-care setting, to increase early detection of PH.