For end-stage renal disease patients undergoing hemodialysis, thrombosis caused by stenosis hinders the long-term use of vascular access. However, traditional spectral bruit analysis techniques for detecting the severity of vascular access stenosis are not robust. Accordingly, the present study proposes an automated method for mimicking a trained practitioner in performing the auscultation process. In the proposed approach, the bruit obtained using a standard phonoangiographic method is transformed into the time-frequency domain, and two spectro-temporal features, namely the auditory spectrum flux and the auditory spectral centroid, are then extracted. The distributions of the two features are analyzed using a multivariate Gaussian distribution (MGD) model. Finally, the distribution parameters of the MGD model are used to detect the presence (or otherwise) of vascular access stenosis. The validity of the proposed approach is investigated using the phonoangiography signals obtained from 16 hemodialysis patients with straight arteriovenous grafts over the upper arm region. The results show that the MGD covariance matrix coefficient of the auditory spectral centroid feature yields an accuracy of 83.87 % in detecting significant vascular access stenosis. Thus, the proposed method has significant potential for the applications of vascular access stenosis detection.

No abstract