Inflow and outflow stenoses screening on biophysical experimental arteriovenous graft using big spectral data and bidirectional associative memory machine learning model

Long-term repeating traumatic puncture is required for dialysis therapy, which results in frequent thrombosis and graduate vascular access stenosis, such as inflow or outflow stenosis and coexistence of both. An arteriovenous graft has a higher patency rate than an arteriovenous fistula. This study intends to use the dual-channel auscultation-based non-invasive method to screen inflow and outflow stenoses. Frequency analysis is used to decompose phonoangiography (PAG) signals to frequency features using the different data length of acoustic data. Burg autoregressive method is employed to extract the key frequency parameters from sufficient spectral data, including characteristic frequencies and distinct peaks of power spectral densities (PSDs). In big data processing, PSDs and the degree of stenosis (DOS) have been validated to show a positive correlation with sufficient big spectral data. An intelligent machine learning model, bidirectional hetero-associative memory network (BHAMN), is carried out to identify the level of DOS at the inflow site, the mid-site, or the outflow site of a vascular access. The experimental results will indicate that the proposed intelligent machine learning model has higher hit rates.