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

The algorithm presented here is intended as a decision support tool for clinicians, and our data support its use in detecting murmurs attributed to valvular heart disease. To put this in perspective, in the general elderly population, where the prevalence of surgically intervenable aortic stenosis is approximately 5% 27 , a negative test, carrying a negative likelihood ratio of 0.032, will nearly rule out the diagnosis, reducing its probability to <0.2%. Conversely, applying the positive likelihood ratio of 4.37 in the setting of a positive result will increase the disease probability to 19%, though this would include non-surgical VHD disease given our study design. In both cases, however, these changes will almost assuredly affect clinical management. Moreover, it is likely that the overall accuracy to make a clinical diagnosis of VHD would be higher when combining the provider's interpretation of the heart sounds along with the algorithm results 28 . We anticipate that such a tool would be particularly useful in a hurried setting like an emergency department, where minimizing the time to diagnostic test results, as well as the strain on providers, is particularly important. While an emergent environment was not explicitly captured in our test set, we purposefully captured heart sounds in a real-world clinical setting to enhance the generalizability of our findings. Importantly, the algorithm tested here is designed solely for the evaluation of the presence of a cardiac murmur in single recordings. Auscultatory findings, however, are much richer than simply the presence or absence of a murmur. Indeed, classical teaching of aortic stenosis includes nonmurmur characteristics, such as the "softness" of the A2 component of the second heart sound, and the timing of the peak of the systolic murmur, as indicators of severity 13 . It is reasonable to conjecture that an extended algorithm, which includes more known signs of VHD in its decisionmaking process would therefore have better performance as a complete disease screening test. This is an area of active investigation and development for our group as low inter-observer reliability in the identification of other murmur characteristics, or even patients with specific diseases, creates many challenges for curating richly labeled datasets large enough for training comprehensive heart sound models. This, however, underscores the need for more investigation in this area. Though the test set populations may well represent the U.S. population, they may not reflect populations in developing countries, where the prevalence and etiology of VHD are different. As these populations would likely very much benefit from a potentially low-cost support tool such as this, further investigation in these populations are warranted.