Multi-mycotoxin detection using fluorescence spectroscopy and machine learning

Mycotoxins are a major concern in the agrifood industry, affecting food safety and human health, while also showing a significant economic impact. Their detection typically requires complex chemical analyses that are expensive, costly and time consuming. To enable an easier and faster detection, several optical spectroscopy methods have been investigated. The state-of-the-art spectroscopic technologies typically focus on the sensing of individual mycotoxins, not taking the full toxicity into account in the case of the co-occurrence of multiple mycotoxins. This work shows how fluorescence spectroscopy combined with machine learning algorithms allows multi-mycotoxin detection in maize. The best performing multi-label classification achieved a classification accuracy of 73% for aflatoxin, 91% for deoxynivalenol, 86% for zearalenone and 96% for fumonisin, when considering threshold concentrations of 3.5 μg/kg, 1000 μg/kg, 55.0 μm/kg and 1000 μm/kg, respectively. Furthermore, the most important wavelengths for the classification were identified using a new approach, called Information Elimination Approach, demonstrating that the models learned from toxin-specific fluorescence bands, thus increasing trustworthiness. This research, for the first time to the authors’ knowledge, presents a successful simultaneous detection of multiple co-occurring mycotoxins with fluorescence spectroscopy at concentrations relevant to the European legislation limits, thus paving the way for an enhanced food safety. • Multi-mycotoxin detection in maize using fluorescence spectroscopy is demonstrated. • Machine learning achieves up to 96% accuracy in toxin classification. • A novel explainability approach identifies the fluorescence key spectral features. • ML models is demonstrated to learn from toxin-specific key spectral features.