Using single corn kernel classification, pooling, and simulation to address skewness and clustering in aflatoxin distribution

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 highly skewed and clustered, i.e. the number of aflatoxin-contaminated corn kernels is disproportionately low and those kernels tend to aggregate spatially as hot-spots instead of being uniformly dispersed. These inherent characteristics of aflatoxin may render the bulk sample unrepresentative, causing inaccurate estimation of the overall aflatoxin concentration in a lot and eventually incorrect decision to accept or reject the entire lot. One solution to this problem is non-destructively classifying aflatoxin level on the single kernel level. In Chapter 2, a custom-built UV-Vis-NIR spectroscopy system was built to scan single corn kernels in motion. A random forest model was trained to use the collected spectrum to classify aflatoxin level in single kernels with 86% sensitivity and 97% specificity, proving that our spectroscopy system is capable of differentiating contaminated kernels from uncontaminated ones with high accuracy. Furthermore, three spectral regions (around 390, 540, and 1,050 nm) were identified as crucial to accurate classification, which provided insight on fine-tuning the classification model. Adopting the single kernel classification method results in the demand for single kernel aflatoxin testing. As the prevalence of aflatoxin contamination in corn is relatively low, pooling was proposed as a more efficient and potentially more economical approach to aflatoxin screening. In Chapter 3, four pooling strategies (one-dimensional pooling by row, one-dimensional pooling by column, two-dimensional pooling, Shifted Transversal Design) were compared in terms of test accuracies and cost-saving effect. The results show that pooling will not cause false negative results when the pool size is smaller than an assay-related threshold. It may reduce the reagent cost by up to 80% when the aflatoxin prevalence is low (< 13% - 21%, depending on the sampling strategy). This study offered practical guidance on when and how to conduct pooling to reduce cost in medium-scale experiments such as aflatoxin screening. An alternative to addressing the skewness and clustering problem is improving the sampling plan. In Chapter 4, a Monte-Carlo simulation model was designed and constructed to simulate probe sampling for aflatoxin-contaminated grains and predict acceptance probability. The effect of four parameters on acceptance probability were evaluated, including aflatoxin concentration, number of probes, sampling strategies (simple random sampling or SRS, stratified random sampling or STRS, systematic sampling or SS), and clustering level. Through sensitivity analysis, natural factors (aflatoxin concentration and clustering level) have a larger impact on acceptance probability than human factors (number of probes, sampling strategy). The three sampling strategies resulted in similar sampling performance (0.8%-3.5% absolute marginal change) at the 20 ppb threshold, except that SS causes large variance when hot-spots exist. Increasing the number of probes from 5 to 100 can improve aflatoxin detection (1.8% increase in absolute marginal change), especially when clustering is severe (> 100 kernels/cluster source). To improve sampling accuracy while saving on cost, one may consider using an auto-sampler to efficiently take a large number of samples.U of I OnlyAuthor requested U of Illinois access only (OA after 2yrs) in Vireo ETD syste