A Deep Learning Algorithm for Automated Cardiac Murmur Detection Via a Digital Stethoscope Platform

John S Chorba; Avi M Shapiro; Le Le; John Maidens; John Prince; Steve Pham; Mia M Kanzawa; Daniel N Barbosa; Brent E White; Jason Paek; Sophie G Fuller; Grant W Stalker; Sara A Bravo; Dina Jean; Subramaniam Venkatraman; Patrick M McCarthy; James D Thomas

We first compared algorithm output on the test set of 1774 recordings to their annotated ground truth. The cardiologists provided a murmur/no murmur vote and ordinal grade to each of those recordings while the algorithm responded murmur/no murmur for only those of sufficient signal quality. To assess performance of this 'poor signal' screening, which excluded 226 recordings from subsequent murmur detection, we plotted median annotator sound quality histograms faceted by algorithm output in Figure 1 . We observed that the 'poor signal' recordings have low annotator signal quality scores, showing the algorithm is not preventing analysis of potential murmurs when recordings are good. The algorithm was able to screen 1548 recordings or 87% of the test set, and thus would provide a useful clinical tool. All further performance analysis was based on this subset of the test set.