Employing machine learning to assess the accuracy of near-infrared spectroscopy of spent dialysate fluid in monitoring the blood concentrations of uremic toxins

Abstract

Hemodialysis (HD) removes nitrogenous waste products from patients’ blood through a semipermeable mem- brane along a concentration gradient. Near-infrared spectroscopy (NIRS) is an underexplored method of monitoring the concentrations of several molecules that reflect the efficacy of the HD process in dialysate samples. In this study, we aimed to evaluate NIRS as a technique for the non-invasive detection of uremic solutes by assessing the correlations between the spectrum of the spent dialysate and the serum levels of urea, creatinine, and uric acid. Blood and dialysate samples were taken from 35 patients on maintenance HD. The absorption spectrum of each dialysate sample was measured three times in the wavelength range of 700-1700 nm, resulting in a dataset with 315 spectra. The artificial neural network (ANN) learn- ing technique was used to assess the correlations between the recorded NIR-absorbance spectra of the spent dialysate and serum levels of selected uremic toxins. Very good correlations between the NIR-absorbance spectra of the spent dialysate fluid with serum urea (R=0.91) and uric acid (R=0.91) and an excellent correlation with serum creatinine (R=0.97) were obtained. These results support the application of NIRS as a non-invasive, safe, accurate, and repetitive technique for online monitoring of uremic toxins to assist clinicians in assessing HD efficiency and individualization of HD treatments.