Data-driven pipeline modeling for predicting unknown protein adulteration in dairy products

Area of Science:

• Food science and technology
• Analytical chemistry
• Machine learning applications

Background:

• Food fraud, particularly protein adulteration, poses a significant threat to food safety and consumer trust.
• Current detection methods often identify known adulterants, leaving a gap for unknown contaminants.
• Proactive prediction of potential adulterants is crucial for effective food fraud prevention.

Purpose of the Study:

• To develop data-driven models for the preemptive prediction of unknown protein adulterants in food.
• To identify and evaluate machine learning algorithms for optimal performance in detecting food fraud.
• To implement predictive models for identifying potential adulterants and validating their impact.

Main Methods:

• Utilized three machine learning (ML) algorithms, including random forest (RF), to build predictive models.
• Trained and validated models on datasets for identifying odorless, tasteless, and colorless adulterants.
• Applied optimal models for external prediction, identifying 51 potential adulterants and conducting adulteration tests on selected candidates.

Main Results:

• The random forest model demonstrated superior performance, achieving high accuracies: 96.2% for odorless, 95.1% for tasteless, and 88.0% for colorless adulterants.
• External prediction identified 51 potential protein adulterants.
• Adulteration tests on milk powder showed no significant sensory difference but an increase in protein content, indicating successful undetected adulteration.

Conclusions:

• Machine learning models, particularly random forest, offer a powerful tool for the proactive detection of unknown protein adulterants.
• The study successfully predicted potential adulterants and demonstrated their ability to increase protein content without altering sensory properties.
• This research provides a proactive strategy to effectively combat food fraud at its source.