Recent scientific research has demonstrated the potential utility of using artificial intelligence (AI) to analyze routinely collected dairy measurements to predict mycotoxin contamination and inform targeted testing. The survey results are Current research in food science.
Specifically, researchers developed a machine learning (ML) pre-screening approach to predict whether aflatoxin M1 (AFM1) in raw milk exceeds the EU regulatory standard of 0.05 micrograms per liter (μg/L). AFM1 is a carcinogenic mycotoxin that can be passed into milk when dairy cows consume feed contaminated with aflatoxin B1, which is then metabolized and excreted as AFM1.
Limitations of conventional clinical tests
Although sensitive detection methods such as ELISA and liquid chromatography-tandem mass spectrometry (LC-MS/MS) have been used to detect AFM1, this study highlighted that cost, timely sample preparation, and skilled labor requirements may limit high-throughput applications in large-scale dairy operations. In hopes of addressing these limitations, researchers investigated whether AI models could be used as a cost-effective tool to identify high-risk samples for further laboratory analysis.
Screen high-risk samples using regularly collected data
The model was trained on a dataset derived from dairy processor raw milk records collected over 20 years, measuring characteristics such as composition, microbiological indicators, and handling conditions. Important predictors routinely measured include protein, fat, lactose, total solids, nonfat milk solids, acidity, freezing point, relative density, milk temperature on receipt, total colony count, psychrophilic bacteria, somatic cell count, and viscosity. Geographical identifiers (i.e., country, province) were also included as categorical predictor variables to account for regional differences associated with feed contamination risk.
AI predicts 83% of samples contaminated with mycotoxins
AFM1 concentrations were labeled based on whether the sample contained levels above or below the EU threshold of 0.05 μg/L. Of the more than 40,000 samples in the dataset, approximately 500 samples exceeded the AFM1 level of 0.05 μg/L, requiring researchers to use statistical methods to balance the data.
Over 100 replicate experiments with a balanced dataset, the most successful model was able to accurately identify 83.2 percent of milk samples with AFM1 levels above 0.05 μg/L.
In an external validation experiment using an independent 2018 dataset, the model correctly identified 75.91 samples above the 0.05 μg/L limit, including 33 missing samples and 793 false positives. The dataset included 8,634 AFM1-positive samples and 137 AFM1-negative samples.
Looking for quick answers on food safety topics?
Try Ask FSM, our new smart AI search tool.
Please contact FSM →
The researchers said these results are consistent with unbalanced classification dynamics and are acceptable for risk-averse prescreening where sensitivity is a priority.
Industry impact: Use AI to enhance traditional detection workflows
Based on their findings, the researchers believe that ML models that leverage existing low-cost intake and quality measurements can be used to flag samples at high risk of mycotoxin contamination in real time. Applying this model as an early warning tool could allow processors to triage lots for confirmatory testing, focus laboratory resources on high-risk batches, and expand mycotoxin monitoring without wasting resources.
The researchers described the use of ML to enhance traditional detection workflows as “AI+” and noted that this approach could also be applied to other food safety hazards where routine plant measurements can help predict risk. However, more extensive multi-region validation and continuous model updates will be required before large-scale industrial applications.
