Use of artificial neural networks in near-infrared spectroscopy calibrations for predicting glucose concentration in urine

Abstract

In this paper, Glucose concentration in urine sample were determined by means of Near-infrared spectroscopy combination with artificial neural networks (ANNs). Distinct ANN models were developed for the multicomponent analysis of biological samples. The developed prediction algorithms were tested on unseen data and exhibited high level of accuracy, comparable to the partial least squares (PLS) regression method. Near-infrared (NIR) spectra in the 10000~4000 cm-∈1 region were measured for human urine glucose with various concentrations at 37 °C. The most informative spectral ranges of 4400~4700 cm-∈1 was selected by principle component analysis (PCA) for glucose in the mixtures. For glucose, the correlation coefficient (R) of 0.999 and the root mean square error of prediction (RMSEP) of 20.61 mg/dl were obtained. The presented algorithms can be used as a rapid and reliable tool for quantitative analysis of glucose in biological sample, since the concentration can be determined based on NIR measurement. © 2008 Springer-Verlag Berlin Heidelberg.