Early warning predictions of contaminants in animal feed using machine learning

Contaminated feed can lead to negative health effects in both animals and humans, as well as result in economic losses. In order to support risk-based monitoring of animal feed, it is essential that the potential presence of food safety hazards can be predicted in an accurate and timely manner. This study used machine learning to predict the presence of four contaminant groups in animal feed: mycotoxins, heavy metals, dioxins and pesticides. Emphasis was placed on a holistic design, using a broad range of external drivers that may affect the presence of the contaminants as model input. Data included historical feed safety monitoring data related to a wide range of different animal feeds, countries, and years (2010-2023) as well as a variety of socioeconomic and weather indicators related to the country of origin. The CatBoost algorithm was the best performing machine learning model and was used to predict the probability that a contaminant in a feed product from a specific country exceeds a predefined threshold (based on European legal limits or guidance values). The average sensitivity was 83% and average specificity was 73% across the four contaminant groups on a year not seen during model training. The results of the model predictions and data descriptions of the monitoring data were incorporated in a decision support system, which has been tested for use in practice. This system offers risk managers better insights in the potential presence of chemical contaminants in animal feed and supports them in the decision-making to safeguard food safety.