Detection of dextran, maltodextrin and soluble starch in the adulterated Lycium barbarum polysaccharides (LBPs) using Fourier-transform infrared spectroscopy (FTIR) and machine learning models

Due to the similar chemical structures and physicochemical properties, it is challenging to distinguish dextran, maltodextrin, and soluble starch from the polysaccharide products of plant origin, such as Lycium barbarum polysaccharides (LBPs). Using the first-order derivatives of Fourier-transformed infrared spectroscopy (FTIR, wave range 1800–400 cm(−1)), this study proposed a two-step pipeline to identify dextran, maltodextrin, and soluble starch from adulterated LBPs samples qualitatively and quantitatively. We applied principal component analysis (PCA) to reduce the dimensionality of FTIR features. For the qualitative step, a set of machine learning models, including logistic regression, support vector machine (SVM), Naïve Bayes, and partial least squares (PLS), were used to classify the adulterants. For the quantitative step, linear regression, LASSO, random forest, and PLS were used to predict the concentration of LBPs adulterants. The results showed that logistic regression and SVM are suitable for classifying adulterants, and random forests is superior for predicting adulterant concentrations. This would be the first attempt to discriminate the adulterants from the polysaccharide's product of plant origin. The proposed two-step methods can be easily extended to other applications for the quantitative and qualitative detection of samples from adulterants with similar chemical structures.