Machine Learning for Predicting Zearalenone Contamination Levels in Pet Food.

Area of Science:

• Food safety
• Analytical chemistry
• Computational toxicology

Background:

• Zearalenone (ZEN) mycotoxin contamination in pet food poses significant health risks to animals.
• Early detection methods are essential for ensuring pet food safety and regulatory compliance.

Purpose of the Study:

• To develop a rapid, cost-effective electronic nose (E-nose) and machine learning approach for predicting ZEN contamination levels in pet food.
• To classify pet food samples as above or below the Chinese regulatory limit of 250 µg/kg for ZEN.

Main Methods:

• Analysis of 142 pet food samples using liquid chromatography-tandem mass spectrometry (LC-MS/MS) for ZEN quantification.
• Utilized an "AIR PEN 3" electronic nose with 10 metal oxide sensors to capture volatile compound profiles.
• Applied various machine learning algorithms (linear regression, k-NN, SVM, Random Forest, XGBoost, MLP) for classification based on E-nose data.

Main Results:

• The Multi-Layer Perceptron (MLP) algorithm achieved 86.6% accuracy in differentiating ZEN-contaminated samples.
• Other machine learning models demonstrated moderate accuracies ranging from 77.1% to 84.8%.
• An ensemble model combining all classifiers reached a peak accuracy of 90.1% for ZEN contamination screening.

Conclusions:

• The E-nose coupled with machine learning offers a promising, rapid, and economical method for screening ZEN mycotoxin in pet food.
• This approach can be effectively implemented for market entry surveillance to ensure pet food safety.