Spectral classification of aflatoxin and fumonisin contamination in corn at single kernel level

Aflatoxin contamination in corn has posed a serious health risk to consumers. To detect aflatoxin contamination in a lot, one may adopt the bulk sampling approach to estimate the overall aflatoxin concentration in the lot, compare it against the aflatoxin action level, and decide whether to accept or reject the whole lot. The challenge with bulk sampling is that aflatoxin distribution is nonhomogeneous and the level of contamination is skewed. The number of highly mycotoxin contaminated corn kernels is disproportionately low compared to the kernels that are safe to eat. Those highly contaminated kernels tend to aggregate spatially as clusters instead of being uniformly dispersed. These inherent characteristics of mycotoxin may render the bulk sample unrepresentative, causing inaccurate estimation of the overall mycotoxin concentration in a corn lot and eventually creating incorrect decisions to accept or reject the entire lot. One solution to this problem is to non-destructively classify aflatoxin levels at the single kernel level. In Chapter 3, a custom-built UV-Vis-NIR spectroscopy system, introduced in Cheng at al., (2019), was used to scan single corn kernels in motion and create classification algorithms to identify and classify aflatoxin- and fumonisin-contaminated kernels. After analysis, for aflatoxin, the best performing algorithm was a stochastic gradient boosting model with an accuracy of 0.83 (Sensitivity (Sn) = 0.75, Specificity (Sp) = 0.83), for both training and testing sets. For fumonisin, the penalized discriminant analysis outperformed the rest of the algorithms with a training accuracy of 0.89 (Sn = 0.87, Sp = 0.88), and testing accuracy of 0.86 (Sn = 0.78, Sp = 0.87). These results prove that our spectroscopy system can differentiate contaminated kernels from uncontaminated ones with high accuracy. In addition, it improves the foundations for single kernel classification of aflatoxin and fumonisin in corn, which can be applied to high throughput screening. An alternative to addressing the skewness and clustering problem is practical strategies that physically removes mycotoxin contamination, such as spectral sorting. However, as shown in Chapter 3, spectral classification methods require the inclusion of spectral data in combination with wet chemistry to generate a robust classification algorithm that could sort kernels that are contaminated with mycotoxin. In Chapter 4, we explore a new approach to train a spectral classification sorting system to provide a cost-effective solution for mycotoxin mitigation. The goal is to calibrate a LED-based spectral sorter, introduced in Stasiewicz et al., (2017), based on high-risk visual features associated with mycotoxin contamination. Samples obtained from Ghana (n = 76) were sorted and bulk aflatoxin and fumonisin measurements were performed using ELISA. After sorting, the mean aflatoxin contamination in accepted fraction was 8.3 ppb (range: 0.22 – 40 ppb), while for the rejected fraction, the mean was 37 ppb (range: 1.8 – 73 ppb). For fumonisin, the accepted fraction showed a mean contamination level of 0.05 ppm (range: <2.5 x 10-3 – 0.66 ppm), while the rejected fraction showed a mean of 1.9 ppm (range: 0.15 – 6.2 ppm). For aflatoxin, in 73 of 76 accepted streams (96%), contamination levels were reduced in comparison to rejected streams. While for fumonisin, all sample showed reduction. These results demonstrate that spectral sorting strategies based on visually high-risk kernels can reduce aflatoxin and fumonisin concentrations. Adopting improved single kernel classification methods could help improve identification of mycotoxin in corn. However, these models need to be calibrated to have an impact in samples that are different from the allocated training set. Therefore, it is crucial to establish a model that can be sensitive to aflatoxin contamination and be able to generalize predictions for accurate classification. In Chapter 5, the best performing algorithm from Chapter 3 was used to classify contaminated kernels in other types of corn from Ghana. GBM algorithm introduced in Chavez et al., (2021), was trained on the normalized data. There were 501 kernels in the training set, where 124 kernel had ≥15 ppb aflatoxin and 377 had <15 ppb aflatoxin. The model had a training specificity of 98.9% and a training sensitivity of 95.2%. The model was tested with a different set of kernels to evaluate whether the model was robust enough to classify unknown data. The testing specificity (77.7%) and testing sensitivity (86.7%) were lower than those from the training set. The results show that Stochastic gradient boosting classification accuracy is reasonably high, and could offer an opportunity for sorting strategies.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Food Science & Human Nutrition

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Chavez Viteri, Ruben Alexander

Contributors dc:contributor

• Stasiewicz, Matthew J.
• Miller, Michael J.
• Feng, Hao
• Rausch, Kent D.

• Aflatoxin
• Fumonisin
• Spectral Classification
• Spectral Sorting
• Corn
• Mycotoxin
• Mycotoxin reduction
• Mycotoxin sorting
• Classification Algorithms

Statement dc:rights

• Copyright 2022 Ruben Chavez Viteri

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/117623

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F117623