Estimation of Vascular Stenosis and Blood Flow Rate Using Machine Learning Classification of Blood Sounds

Phonoangiography (PAG) is a promising noninvasive approach for quantifying the blood sounds (bruits) heard via auscultation. In this work, we developed a vascular phantom for in vitro reproduction of bruits under varying flow rates and degree of stenosis in large arteries. Phantoms used blood-mimicking fluid and scattering media to enable the use of clinical ultrasound for validation of hemodynamic flows. Acoustic recordings at multiple sites using clinical stethoscopes were analyzed to extract descriptive features with low computational complexity. Machine learning models using an optimal feature subset achieved up to 100% accuracy in stenosis classification; many classifiers performed very well, suggesting generalizability of these features for stenosis detection. Additionally, delta feature analysis improved stenosis detection while reducing the need for precise recording site selection. Beyond stenosis classification, PAG features were also found to successfully predict flow rate, highlighting vascular screening feasibility even in the absence of significant stenosis. These results demonstrated the feasibility of PAG-based screening for vascular access dysfunction, offering a cost-effective, real-time alternative to conventional imaging for early stenosis detection and hemodynamic evaluation.Clinical Relevance-This work further studied the application of signal processing and machine learning to phonoangiography. This may enable point-of-care or at-home detection of hemodynamically significant vascular stenosis via auscultation.