Abstract 12064: Accurate Detection of Severe Aortic Stenosis Using Wavelet Analysis and Machine Learning for Heart Murmur

Introduction: Auscultation is a well-accepted screening method for valvular disease. However, it is not adequate for the screening because the detecting rate of valvular disease by non-cardiologist is about 43% only. Hypothesis: Artificial intelligence (AI)-assisted heart murmur analysis could detect valvular disease following a heart murmur record using a digital stethoscope. Methods: The current study included 71 patients admitted to our hospital to undergo transcatheter aortic valve replacement due to aortic stenosis (AS). We recorded preoperative and postoperative heart murmur using digital stethoscope Eko. Three of 71 cases were diagnosed with severe AS due to prosthetic valve dysfunction. After denoise processing, an acoustic analysis was done to set objective parameters of the murmur. Machine learning was performed to detect severe AS by utilizing the characteristic parameters of severe AS as essential features. A total of 142 sound sources were acquired, 100 cases were used as training data, and the remaining were used as test data to verify the detection accuracy of severe AS. Results: The prolonged duration of the first sound and the peak frequency of systolic murmur with 300 Hz or higher were important for detecting severe AS. The detection accuracy of severe AS was 98% in the support vector machine model. All were detected in PVD cases, but not in the low flow & low gradient cases. Conclusions: It was suggested that severe AS could be detected by AI-assisted heart murmur analysis. We are developing the auscultation device, which is designed to be easy for elderly outpatients, to detect severe AS and to classify the severity of AS with differences in the peak frequency of the murmur.