Development of machine learning algorithms to estimate maximum residue limits for veterinary medicines

Area of Science:

• Veterinary pharmacology
• Food safety science
• Computational toxicology

Background:

• Maximum-residue limits (MRLs) are crucial for ensuring human food safety from veterinary drug residues.
• Current MRLs are often species-specific, leaving gaps for under-represented food commodities.
• Establishing new MRLs involves extensive and costly animal studies.

Purpose of the Study:

• To develop and validate a machine learning (ML) approach for predicting unestablished MRLs.
• To assess the reliability of ML models in estimating MRLs for food commodities lacking regulatory limits.
• To explore the potential of ML in reducing the burden of new MRL determination.

Main Methods:

• Utilized classification methods designed for imbalanced data, including resampling techniques.
• Analyzed international MRL data across various countries.
• Evaluated seven ML classifiers: support vector classifier, multi-layer perceptron (MLP), random forest, decision tree, k-neighbors, Gaussian NB, and AdaBoost.

Main Results:

• The neural network multi-layer perceptron (MLP) classifier demonstrated superior performance, achieving an accuracy greater than 99% with markers and approximately 88% without.
• ML algorithms were successfully applied to predict unestablished MRLs for under-represented food commodity groups.
• The study established a data-mining method for predicting MRLs.

Conclusions:

• Machine learning offers a reliable and efficient method for predicting unestablished maximum-residue limits (MRLs).
• This approach can significantly decrease the reliance on live animal testing and reduce research costs.
• The findings represent a novel application of ML in regulatory food animal medicine, enhancing food safety protocols.