Mould infested maize poses a severe problem for farmers, food producers and for consumers worldwide. Mycotoxins are secondary metabolites produced by certain mould species that contaminate food and feed. The consumption of these toxins may cause serious health problems for humans and animals. The trichothecene deoxynivalenol (DON) constitutes one of the most commonly occurring Fusarium toxins encountered in maize and requires improved methods to limit its entrance into the food and feed system. While a variety of chromatographic and mass spectrometry methods for the identification of such toxins have been established, these are considered time-consuming, cost-intensive and require highly qualified personnel. Alternatively, optical techniques, such as mid-infrared spectroscopy offer rapid detection of fungal infections in cereals and other commodities with minimised sample preparation and analysis time. The present study demonstrates a rapid fungal contamination detection strategy in maize taking advantage of IR-spectroscopy combined with advanced machine learning algorithms. The developed method represents an advancement for the analysis of differences in protein and carbohydrate content revealed in the associated IR-spectra related to the amount of toxin contamination at the European Union (EU) regulatory limits for DON in maize (i.e. 1,250 μg/kg). The employed maize varieties are naturally infected samples or have been infected with Fusarium verticillioides, Fusarium graminearum or Fusarium culmorum. Sieved maize samples at the EU regulatory limit were correctly classified using machine learning approaches, therefore enabling the differentiation between DON-contaminated and non-contaminated maize samples. Specifically, a variety of machine learning methods, including Adaptive Boosting (AdaBoost), Random Forests, Support Vector Machine (SVM) and Multilayer Perceptron (MLP) demonstrated excellent classification and validation performance using the obtained IR-spectra. As a result, 183 maize samples of different varieties and infection levels were accurately classified. 94% of the non-contaminated samples and 91% of the contaminated samples were correctly classified using an MLP classification approach.