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

Having evaluated the algorithm on individual recordings, we can measure its performance as a patient screening tool for valvular heart disease by comparing predictions to echocardiograms. First, we consider aortic stenosis (AS). From the 373 patients represented in the test set, we exclude 60 patients with valve replacements and 11 with insufficient echocardiogram reports. Of the remaining, we group 40 having aortic stenosis of severity at least moderate-to-severe as cases and 103 patients without structural heart disease (i.e. no valve disease greater than trivial severity, and no congenital disease), as controls. As mentioned previously, the severity threshold for disease was chosen to include cases which would require evaluation for mechanical intervention. Figure 3 illustrates the flow of study participants. Importantly, whereas for the murmur detection algorithm the "gold standard" is simply the clinical interpretation (and likely more subject to inter-observer variability), for valvular heart disease screening, the "gold standard" is the echocardiogram. Since an aortic stenosis murmur is generally loud, we defined a positive test as one where a murmur is detected at either the aortic or pulmonic positions. A negative test is one where no murmur is detected at the aortic and pulmonic positions. A small fraction of subjects in both patient groups (i.e. cases and controls) had recordings limited by a 'poor signal' classification and were removed from the analysis if necessary.