New insights into raw milk adulterated with milk powder identification: ATR-FTIR spectroscopic fingerprints combined with machine learning and feature selection approaches

It is common that unscrupulous merchants attempt to gain profits by raw milk adulterated with high-temperature processing and long storage time of milk powder, which damages consumer rights and interests. In this study, we propose a practical, reagent-free and rapid method to identify raw milk adulterated with milk powder using ATRFTIR spectra coupled with machine learning-based approach. Four distinct machine learning algorithms were performed to determine adulterated milk in authentic pure milk based on full spectrum and three interest of ATRFTIR characteristic regions (1800-580 cm(-1), 3700 -2700 cm(-1) and 1800-580 cm(-1) +3700 -2700 cm(-1)). The results indicated that SVM and PLS-DA yielded higher average accuracy, sensitivity, specificity, negative predictive value and area under the receiver operating characteristic curve, compared with RF and SIMCA. In particular, SVM and PLS-DA based on ATR-FTIR feature region (1800 -580 cm(-1)) provided near-perfect classification performance for adulterated milk with low computational intensity and good robustness. The satisfactory performance of rapid, reagent free and low-cost infrared spectroscopy with various machine learning algorithm and feature selection approach, confirm its potential for on-line monitoring and intelligent detecting of milk adulteration.