Predicting Aflatoxin M₁ in Raw Milk Using Machine Learning and Basic Measurements

Abstract

Aflatoxin M₁ (AFM₁) is a carcinogenic and teratogenic mycotoxin that may be present in raw milk. Therefore, continuous monitoring of AFM₁ levels is essential to ensure dairy safety and regulatory compliance. Although laboratory-based analytical techniques such as ELISA and LC-MS/MS offer high accuracy, their cost, sample preparation requirements, and dependence on specialized personnel make them less practical for high-frequency or large-volume screening in dairy processing facilities. This creates a need for complementary, cost-effective prescreening approaches. This study proposed a qualitative AFM₁ prediction method based on routinely measured physicochemical indicators of raw milk, combined with machine learning algorithms. Five classical machine learning models were evaluated under a binary classification framework to determine whether AFM₁ levels exceed the regulatory threshold. Experimental results show that the multilayer perceptron achieves an accuracy and negative-sample recall rate above 80%, demonstrating the potential of machine learning as an effective prescreening tool for AFM₁. The findings provide a feasible direction for supporting rapid, economical, and large-scale monitoring of raw milk safety.