Patients suffering from end-stage renal disease usually receive dialysis therapy. The arteriovenous (AV) shunt is a vital vascular access for achieving sufficient blood flow in the vein during hemodialysis and can be either an arteriovenous fistula (AVF) or an arteriovenous graft (AVG). However, the lumens of dialysis accesses are frequently narrowed by thrombosis, resulting in stenosis at the venous anastomosis site or progression of inflow stenosis at the arterial anastomosis site. A narrowed vascular wall will produce abnormal physical stress and cause turbulent flow and high blood pressure. Hence, murmur sounds will occur around the stenotic site. The auscultation method is a noninvasive technique to detect these sounds as phonoangiograph (PAG) signals. The empirical mode decomposition (EMD) method is used to decompose intrinsic fast and slow oscillation components from PAG signals. In this study, the slow oscillation component containing key spectral energy distributions (< 100, 100–300 Hz, and > 300 Hz) will be used to identify the normal and abnormal conditions for further stenosis level assessment. After extracting key frequency-based features using EMD, 1D convolutional and pooling processes, feature patterns can be extracted from spectral patterns, which can distinguish distinct feature patterns and reduce the dimensions of feature patterns and the number of feature datasets for training the classifier. Next, a convolutional neural network-based classifier is used to assess the stenosis levels at a near venous anastomosis site. Its model can solve nonlinear mapping applications and nonlinear separable classifications, including the normal condition, AVG stenosis, and AVF stenosis. The experimental tests with cross-validation will indicate that the proposed method provides a promising result in clinical trials, including mean recall (%), mean precision (%), mean accuracy (%), and a mean F1 score of 95.35 %, 89.49 %, 86.82 %, and 0.9232, respectively, by offering an automatic procedure for AV shunt stenosis assessment without the need for manual feature extraction and classification.
All Science Journal Classification (ASJC) codes