Predicting Significant Stenosis of Arteriovenous Access Through Wavelet Transform and Machine Learning on Sounds...

For arteriovenous (AV) access stenosis, physical auscultation findings are subjective and nonquantitative. Here, we sought to predict significant stenosis in patients with arteriovenous fistulas (AVFs) based on sounds recorded through an electronic stethoscope, and their evaluation using a deep learning approach.

From January 2023 to December 2023, we enrolled, after informed consent, 30 end-stage renal disease patients with dysfunction of AV access receiving endovascular treatments. Sounds were recorded at 3 sites: the anastomotic site (point 1), arterial puncture site (point 2), and venous puncture site (point 3), each for 10 sec long before and after percutaneous transluminal angioplasty (PTA). The severity of stenosis was confirmed by angiography. Those sound signals were Wavelet transformed before fed into a three-dimensional Mesh component. Then, those data were used to train a deep learning network to achieve the purpose of automatic identification of stenosis. We applied a convolutional neural network (CNN) architecture to construct a model for predicting significant stenosis of AV access.

Sounds were labeled as "significant stenosis" (>50%) or "nonsignificant stenosis" (≤50%). Firstly, we use data from 10 patients to train the model. Initial training results showed that signals from the arterial puncture site had better accuracy and lower error rates. Secondly, we fed data from the remaining 20 patients for model testing. A total of 40 audio testing data samples (all recorded at the arterial puncture site), which consisted of 20 significant stenosis cases and 20 nonsignificant stenosis cases. The model's sensitivity, specificity, and accuracy were found to be 100%.

Wavelet transform and CNN model was able to satisfactorily predict the presence of severe stenosis of AV access requiring PTA. The model is potentially useful in AVF surveillance. Further study is needed to predict cases of less severe stenosis of AV access.