Accurate prediction of glucose concentration and identification of major contributing features from hardly distinguishable near-infrared spectroscopy

Abstract

In this study, we tackle the accurate prediction of glucose aqueous concentration from hardly distinguishable near-infrared (NIR) spectroscopy. We adopted several machine learning approaches for the spectral analyses and identified important features learned by each model. The models we investigated include Partial Least Squares Regression (PLSR), Support Vector Machine Regression (SVMR), Random Forest Regression (RF), Extra Trees Regression (ETR), eXtreme Gradient Boosting (Xgboost), and hybrid Principal Component Analysis-Neural Network (PCA-NN) methods. From 47 different glucose aqueous concentrations which cover the range of 40–500 mg/dl, we measured 564 near-infrared (NIR) absorbance spectra samples with wavelength range 900 nm–2200 nm. Then the spectra samples were randomly split into 80% for the training set and 20% for the testing set. In our test, we found that the models SVMR, ETR, and PCA-NN reach extremely good performance, which had correlation coefficient R > 0.99 and determination of coefficient R2 > 0.985. To explore the robustness of each machine learning approach, we extracted their high-weighting features and examine their distribution. We found that having large overlapping of the high-weighting features learned by the model when trained by different data sets may be an indication of model stability. In addition, our analysis came up with the essential region of features to disentangle the hardly distinguishable signal. Our study demonstrates a robust machine learning models for the prediction of glucose aqueous concentration in an in-vitro setup using near-infrared spectroscopy.