Quantitative analysis of Raman spectra for glucose concentration in human blood using Gramian angular field and convolutional neural network

In this study, convolutional neural network based on Gramian angular field (GAF-CNN) was firstly proposed. The 1-D Raman spectral data was converted into images and used for predicting the biochemical value of blood glucose. 106 sets of blood spectrums were acquired by Fourier transform (FT) Raman spectroscopy. Spectral data ranging from 800 cm^-1 to 1800 cm^-1 were selected for quantitative analysis of the blood glucose. Data augmentation was used to train neural networks and normalize the Raman spectra. And, we applied principal component analysis (PCA) for dimension reduction and information extraction. The root mean squared error of prediction (RMSEP) are 0.06570 (GADF) and 0.06774 (GASF), the determination coefficient of prediction (R²) are 0.99929 (GADF) and 0.99925 (GASF), and the residual predictive deviation of prediction (RPD) are 37.56324 (GADF) and 36.43362 (GASF). GAF-CNN model performed better for predicting of glucose concentration. The GAF-CNN model can be used to establish a calibration model to predict blood glucose concentration.

A shallow convolutional neural network with elastic nets for blood glucose quantitative analysis using Raman spectroscopy

In this paper, a one-dimensional shallow convolutional neural network structure combined with elastic nets (1D-SCNN-EN) was firstly proposed to predict the glucose concentration of blood by Raman spectroscopy. A total of 106 different blood glucose spectra were obtained by Fourier transform (FT) Raman spectroscopy. The one-dimensional shallow convolutional neural network, with elastic nets added to the full connected layer, was presented to capture multiple deep features and reduce the complexity of the model. The 1D-SCNN-EN model has a better performance than conventional approaches (partial least squares and support vector machine). The root mean squared error of calibration (RMSEC), the root mean squared error of prediction (RMSEP), the determination coefficient of prediction (R_P²), and the residual predictive deviation of prediction (RPD) were 0.10262, 0.11210, 0.99403, and 12.94601, respectively. The experiment results showed that the 1D-SCNN-EN model has a higher prediction accuracy and stronger robustness than the other regression models. The overall studies indicated that the 1D-SCNN-EN model looked promising for predict the glucose concentration of blood by Raman spectroscopy when the sample size is small.