Quantitative analysis of glucose in whole blood based on FT-Raman spectroscopy and back propagation artificial neural network

In this paper, the models of quantitative analysis of glucose concentration in the whole blood based on FT-Raman spectroscopy and back propagation-artificial neural network (BP-ANN) were established. The accurate values of the synaptic weights of the ANN were obtained by inverse delayed (ID) function model of neuron. All analysis were carried out by whole spectrum that pretreated by baseline correction, Savitzky-Golay smoothing and mean-centering, Savitzky-Golay derivative. And the dimension of spectra was reduced by the principal component analysis (PCA). The Levenber-Marquardt training algorithm and Log-sigmoid transfer function were adopted in the BP-ANN. The optimized number of hidden node, transfer functions of hidden layer and output layer were inputted into BP-ANN to establish the calibration model. The results showed that the correlation coefficients of calibration and prediction were 0.9955 and 0.9953, respectively, and the root mean square error were 0.0323 and 0.027, respectively. In this paper, the error of estimating glucose concentration between the ANN based on ID function model with 15 hidden neurons in hidden layer and conventional neuron model are 1.02 mg/dl and 5.48 mg/dl. The results demonstrated that this method is feasible, convenient, rapid and no pretreatment.