TITLE:
A Comparison of CNN and PLSR for Glucose Monitoring Using Mid-Infrared Absorption Spectroscopy
AUTHORS:
Baorong Fu, Yongji Meng, Xianwen Zhang, Zhushanying Zhang
KEYWORDS:
Mid-Infrared, Convolutional Neural Networks (CNN), Partial Least Square Regression (PLSR), Glucose
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.13 No.3,
March
28,
2023
ABSTRACT: With the development of mid-infrared (MIR)
photoelectric devices, mid-infrared spectroscopy has become
one of the important methods for non-invasive detection of blood glucose. The
mid-infrared region (4000 - 400 cm-1) has the well-known fingerprint region (1200 - 800 cm-1) of
glucose, which has clearer characteristic absorption peaks and better
specificity. There is a lot of molecular information about glucose in the MIR.
The non-invasive detection of blood glucose by mid-infrared spectroscopy needs
to achieve certain accuracy, and the quantitative model is an important factor
affecting the accuracy of glucose detection. In this paper, the samples of
imitation solution containing only glucose and the samples of imitation mixed
solution are taken as the research objects, and the mid-infrared spectral data
of the samples are collected. The full spectrum partial least squares
Regression (PLSR) model, SNV + Ctr-PLSR model, MSC + Ctr-PLSR model, and convolutional neural networks
(CNN) model of 3000 - 900 cm-1 band
were constructed. Full spectrum PLS model and CNN model of 1200 - 900 cm-1 band were constructed. The
experimental results show that the optimal model of the two bands is CNN, then
the correlation coefficient of prediction set (Rp) of 3000 - 900 cm-1 band is 0.95, and the root mean
square error of pre-diction set (RMSEP) value is 22.10. The Rp of 1200 - 900 cm-1 band is 0.95, and the RMSEP
value is 22.54. The research results show that CNN is a promising method, which
has higher accuracy than PLSR, and is especially suitable for modeling human
complex environment. In addition, the study provides a theoretical and
practical basis for CNN in feature selection and model interpretation.