Fourier-Transform Infrared (FTIR) spectroscopy and machine learning approaches to detect and quantify cross-contact of non-gluten and gluten-rich flours

Wheat, barley, and rye grains contain a type of family of proteins known as gluten that affects people with gluten-related disorders. In this study, FTIR coupled with supervised machine learning approaches was evaluated for detection and quantification of cross-contact between non-gluten (corn flour (CF)) and gluten-rich (wheat flour (WF), barley flour (BF) and rye flour (RF)) flours, at contamination levels of 0% - 10% (w/w) with 0.5% increment. For the detection and quantification of the CF contaminated with BF, WF, and RF respectively, partial least squares discriminant analysis and partial least squares regression were used to develop the models. Parameters from the confusion matrix show true positive rates (TPR) of 0.87500, 0.81250, 0.9333, and 1.0 respectively for BF, WF, RF and CF classes. The best result using the full wavenumber region emerged by the SGD1 + MC model with prediction's coefficient of determination (R²p) and root mean square error of prediction (RMSEP) to be 0.857, 0.906, 0.928 and 1.187, 0.971, 0.865 for CF contaminated with BF, WF and RF respectively. While the best result for the optimized model emerged by auto-scaled models with R²p and RMSEP to be 0.876, 0.808, 0.851 and 1.100, 1.349, 1.177 for CF contaminated with BF, WF and RF, respectively. The TPR, R²p, and RMSEP obtained indicate that FTIR spectroscopy with machine learning approaches has the potential to authenticate the cross-contact of non-gluten and gluten-rich flours within the defined contamination levels.