Generalized regression estimator improved the accuracy rate of estimated dialysis accesses stenotic condition on in-vitro arteriovenous graft experimental model

Dialysis vascular accesses are critical for patients receiving hemodialysis treatment. However, dialysis access stenosis and further dysfunction are engendered by thrombosis or outflow (venous anastomosis site) stenosis and the progression of inflow (arterial anastomosis site) stenosis. Thus, any narrowed access causes vibrations, turbulent flow, and murmur sounds around stenosis sites. Auscultation and frequency-based techniques are employed to detect these sounds, and frequency components are also validated on the basis of the degree of stenosis (DOS). In this paper, a biophysical experimental model employing an in vitro arteriovenous graft model was established to produce various acoustic signals associated with single stenosis and multiple stenoses. By analyzing various combinations of stenoses, this paper selected suitable features of the frequency and power spectra using the Burg autoregressive method. A multiple regression model applying a higher number of explanatory variables and response variables, as a generalized regression neural network, was employed to identify DOS levels at inflow and outflow sites. The experimental results indicated that the proposed screening model provided a higher average hit rate of >90%, average true-positive rate of >90%, and true-negative rate of 100% in single and multiple stenosis screening, compared with the multiple linear regression model.