The following is a summary of “Deep Learning Algorithms to Detect Murmurs Associated With Structural Heart Disease,” published in the October 2023 issue of Cardiology by Prince et al.
Accurate identification of structural heart disease through cardiac auscultation remains inconsistent among healthcare providers, potentially resulting in overlooked treatments. The application of machine learning in cardiac auscultation holds promise to address this issue, yet despite growing interest, few algorithms have transitioned into clinical use. Here, the researchers evaluated a novel suite of Food and Drug Administration (FDA)-cleared algorithms, trained using deep learning on a comprehensive dataset of over 15,000 heart sound recordings.
Their validation involved testing these algorithms on a dataset comprising 2,375 recordings from 615 distinct subjects collected in real clinical settings using commercially available digital stethoscopes. Board-certified cardiologists meticulously annotated these recordings and paired them with echocardiograms, serving as the gold standard. In a clinical simulation, the study group compared the algorithm’s performance against 10 clinicians using a subset of the validation database. The algorithm consistently and accurately detected structural murmurs, achieving a sensitivity of 85.6% and specificity of 84.4%. When focusing solely on clearly audible murmurs in adults, the algorithm demonstrated even greater accuracy with a sensitivity of 97.9% and specificity of 90.6%. Furthermore, the algorithm precisely distinguished between systolic and diastolic murmurs by reporting their timing within the cardiac cycle. Notably, despite optimizing acoustics for clinicians, the algorithm substantially outperformed the clinicians, with an average clinician accuracy of 77.9% compared to the algorithm’s accuracy of 84.7%.
Their findings reveal the algorithms’ remarkable accuracy in identifying murmurs associated with structural heart disease. Notably, these results underscore the stark contrast between the consistent performance of the algorithm and the significant variability observed among clinicians. These outcomes strongly suggest the potential benefits of integrating machine learning algorithms into clinical practice to enhance the detection of structural heart diseases, thereby facilitating improved patient care.