Jump to Main Content
A novel chemometric classification for FTIR spectra of mycotoxin-contaminated maize and peanuts at regulatory limits Part A Chemistry, analysis, control, exposure & risk assessment
- Kos, Gregor, Sieger, Markus, McMullin, David, Zahradnik, Celine, Sulyok, Michael, Öner, Tuba, Mizaikoff, Boris, Krska, Rudolf
- Food additives & contaminants 2016 v.33 no.10 pp. 1596-1607
- European Union, Fourier transform infrared spectroscopy, Fusarium verticillioides, absorption, aflatoxin B1, chemometrics, corn, data collection, decision support systems, deoxynivalenol, food safety, fungal diseases of plants, models, peanuts, plant pathogenic fungi, principal component analysis, rapid methods, spectral analysis
- The rapid identification of mycotoxins such as deoxynivalenol and aflatoxin B ₁ in agricultural commodities is an ongoing concern for food importers and processors. While sophisticated chromatography-based methods are well established for regulatory testing by food safety authorities, few techniques exist to provide a rapid assessment for traders. This study advances the development of a mid-infrared spectroscopic method, recording spectra with little sample preparation. Spectral data were classified using a bootstrap-aggregated (bagged) decision tree method, evaluating the protein and carbohydrate absorption regions of the spectrum. The method was able to classify 79% of 110 maize samples at the European Union regulatory limit for deoxynivalenol of 1750 µg kg –¹ and, for the first time, 77% of 92 peanut samples at 8 µg kg –¹ of aflatoxin B ₁. A subset model revealed a dependency on variety and type of fungal infection. The employed CRC and SBL maize varieties could be pooled in the model with a reduction of classification accuracy from 90% to 79%. Samples infected with Fusarium verticillioides were removed, leaving samples infected with F. graminearum and F. culmorum in the dataset improving classification accuracy from 73% to 79%. A 500 µg kg –¹ classification threshold for deoxynivalenol in maize performed even better with 85% accuracy. This is assumed to be due to a larger number of samples around the threshold increasing representativity. Comparison with established principal component analysis classification, which consistently showed overlapping clusters, confirmed the superior performance of bagged decision tree classification.