Data-Driven Hybrid Modeling for Predicting Unknown Protein Adulteration in Food and Development of Detection Method

Abstract

To preemptively predict unknown protein adulterants in food products and curb the incidence of food fraud at its origin, data-driven hybrid model was developed by machine learning (ML). The RF-based model exhibited superior performance, achieving 96.2%, 95.1%, and 88.0% accuracy for identifying odorless, tasteless, and colorless adulterants, respectively. 51 potential adulterants were identified, from which two cost-effective substances were chosen for adulteration testing. Despite no discernible sensory distinctions, there was a quantifiable increase protein content in adulterated milk powders. To further combat fraudulent practices by unscrupulous traders, qualitative and quantitative adulteration detection models were established using FTIR coupled with ML. The RF model proved most effective, with precision, recall, F1 score, and accuracy metrics all exceeding 0.95. For quantitative adulteration models, the PLS model achieved the highest R2 of 0.994 and the lowest RMSE of 0.027. This research paves the way for a proactive strategy to combat food fraud.

Keywords: Protein fraud, Machine Learning, Prediction, Potential adulterants