AI Detects Pediatric Heart Murmurs with Cardiologist-Level Accuracy
Abstract
Introduction: Artificial intelligence based on deep learning has recently revolutionized diverse fields, leading to preternatural performance on perceptual tasks including image analysis and speech recognition. Deep learning has the potential to similarly transform cardiac murmur recognition through its integration with digital stethoscope technologies.
Hypothesis: A deep neural network can correctly flag pediatric heart murmurs in a noisy hospital setting with sensitivity and specificity comparable to a cardiologist.
Methods: A retrospective analysis of audio recordings from 54 patients collected at duPont Hospital for Children using an Eko CORE digital stethoscope was performed. Only patients with echo-confirmed pathologic heart murmurs and normal heart sounds were included (n=42), while patients with innocent murmurs and other abnormal heart sounds were excluded.
A deep neural network was trained on a separate murmur database and blinded to test data at the time of training. Recordings were analyzed by the network to predict probability of murmur presence and this probability was thresholded to generate an ROC curve.
Five pediatric cardiologists reported auscultation findings (pathologic murmur vs. no pathologic murmur) for each patient while blinded to echo findings and patient history. 95% confidence intervals for sensitivity and specificity relative to echocardiogram were computed via Wilson’s method.
Conclusions: This is the first head-to-head study demonstrating that a deep neural network can detect pediatric heart murmurs with comparable accuracy to a cardiologist. Limitations include small sample size and the retrospective nature of the analysis, hence further study is justified.
