Aflatoxin (AFL), a type of mycotoxin, is produced in cereal grains, such as corn, as secondary metabolites from certain types of fungi on plants. Corn is susceptible to AFL toxigenic strains, which pose a significant health, economic, and safety risk to humans and livestock when they consume contaminated products (Munkvold et al., 2019). The economic impact of AFL has been estimated to be between $418 million to $1.66 billion for all stakeholders in the US agricultural industry and can infiltrate the supply chain in corn-based commodities (Wu, 2006; Mitchell et al., 2016). AFL is primarily produced from the fungal strains Aspergillus flavus and A. parasiticus via the polyketide pathway (Sweeney and Dobson, 1998). Contamination can happen during any stage, from in the field pre-harvest, during the growing season, harvest, and in post-harvest storage (Payne and Widstrom, 1992).

Extensive literature has been published on environmental conditions conducive to producing AFL in corn, with specific conditions that favor production and development (Diener et al., 1987; Payne and Widstrom, 1992; Cotty and Jaime-Garcia, 2007; Windham et al., 2009). Three specific factors are needed to create the right conditions for a pathogen to invade plants, create disease, and produce mycotoxins. This is known as the traditional balanced triangle between (1) the pathogen and pest, and (2) host, and (3) environmental conditions (Medina et al., 2017; Perrone et al., 2020). A. flavus and parasiticus have been documented to cause infection under drought conditions in dry, hot weather ranging from 29 to 35°C (Schindler et al., 1967; Payne, 1998). Additionally, AFL infection will likely develop when these high temperatures continue through the nighttime period without proper cooldown (Diener et al., 1987; CAST, 2003). Corn is susceptible to AFL infection through the ear silks, with stress conditions at pollination increasing the chance of plant disease (Marsh and Payne, 1984; Damianidis et al., 2018). Furthermore, soil is often the reservoir for Aspergillus, while insect vectors, direct contact, or dust can transmit spores (Winter and Pereg, 2019). The distribution and growth of AFL in soil depend on many factors, including geographical region, soil type, water retention rate, climatic conditions, crop rotation, and insect presence (Zhang et al., 2017; Winter and Pereg, 2019). Elevated soil temperatures have been positively correlated to AFL contamination and the degree to which insect activity impacts AFL content regionally; however, more literature is necessary regarding soil properties and how they influence fungal growth (Bilgrami and Choudhary, 1998; Payne, 1998). Temperature and rainfall conditions in the principal corn-growing states in the US are typically sufficient to slow the growth of A. flavus and parasiticus, avoiding significant AFL accumulation (Munkvold, 2014). However, in drought and high-temperature years, AFL contamination has been documented in Iowa (Lillehoj et al., 1976; Schmitt and Hurburgh, 1989; Mitchell et al., 2016). These AFL challenges exist in the Corn Belt region of the US; climate change patterns with temperature increases will likely increase the AFL concentration in corn in the US (Wu et al., 2011; Yu et al., 2022).

The Food Safety Modernization Act (FSMA) legislation warrants stakeholders to be preventative versus reactive with food and feed safety events, including mycotoxin outbreaks of AFL in corn (Grover et al., 2016). Therefore, AFL prediction and risk assessment systems that alert stakeholders of possible outbreaks are essential. Wu et al. (2011) stated, “Quantitative, site-specific risk assessments or predictive models for mycotoxin accumulation could contribute significantly to management efficiency in maize.” While many efforts have been undertaken to predict AFL by leveraging climate and weather data and interactions with crop developmental phases, the models are often based on generating new datasets, in vitro data, or conducted in other regions of the world (Johansson et al., 2006; Probst and Cotty, 2012; Leggieri et al., 2015; Battilani et al., 2016; Smith et al., 2016). These models are generally not applicable to the US corn growers, grain handlers, processors, and end-users due to differences in geographical location, management practices, weather, and predictions of mycotoxin contamination with decreased accuracy levels (de Schrijver et al., 2021; Castano-Duque et al., 2022). The European models established a general framework for a US-specific model with mycotoxin corn predictions (Battilani et al., 2013; Van der Fels-Klerx et al., 2019). Castano-Duque et al. (2022) developed the first US machine-learning models using feature engineering in combination with gradient boosting machine (GBM) learning and Bayesian networks to predict AFL contamination in Illinois-grown corn concerning weather and plant-related parameters such as vegetative index, aflatoxin risk index (ARI), and climate zones.

With Iowa being the top corn-producing state in the US, thoughtful, comprehensive, and strategic management solutions, such as predicting AFL contamination on an annual and localized basis created for grain processors and handlers, allow for appropriate decision-making in a preventative versus reactive manner (USDA-NASS, 2023a,b). The development of prediction models can enable early action to prevent or hinder mycotoxin development through integrated solutions that are controllable such as early harvest of at-risk grain, isolation of contaminated grain, application of fungicides, drying to lower storage moistures, and strategic marketing to more tolerant end users (Fumagalli et al., 2021). For grain elevators, handlers, and processors, prediction models enable proactive planning for handling, storing, and marketing grain with differing risk levels and facilitates strategic sampling and testing (Fumagalli et al., 2021). These machine-learning models can guide rapid decision-making and diversion necessary before the point of first receipt at the elevator to improve the overall safety and profitability of the US corn supply without compromising the profitability of individual grain businesses (Mitchell et al., 2016; Castano-Duque et al., 2022).

The main objective of this study aimed to evaluate the performance of AFL prediction with GBM models and feature engineering in Iowa corn with two risk thresholds: 20-ppb and 5-ppb. Historical climate data, soil property data, and historical Iowa AFL data collected in 2010, 2011, 2012, and 2021 were used in the GBM model. The combination of historical climate, weather, soil property data, and AFL contamination data in Iowa helped determine indicators of risk preharvest. AFL risk predictions from the Iowa-centric model provide a baseline for indicating disease in the corn crop, paving the way for further development of proactive actions and decisions that grain supply chain stakeholders can adopt for AFL mitigation control.

AFL is of great concern to the US corn industry, as it poses a significant food and feed safety risk to humans and livestock due to the detrimental effects of being a known class 1a carcinogen (Eaton and Gallagher, 1994; CAST, 2003; Mitchell et al., 2016; Yu et al., 2022). Additionally, AFL has significant cost implications for agricultural and food economies in the US and even globally due to the far-reaching contribution of US corn (Mitchell et al., 2016). In this case study, an Iowa-centric model was developed to predict AFL contamination using historical corn contamination, meteorological data, and soil property data in the state producing the largest amount of corn in the US (USDA-NASS, 2023a,b). This research follows Castano-Duque’s model that predicted AFL and FUM contamination in Illinois corn, a neighboring state to Iowa (Castano-Duque et al., 2022). The GBM model was selected for analysis due to the nature of the data, with GBM performing better for AFL contamination in the Illinois-centric model (Castano-Duque et al., 2022). The AFL risk values were set using US FDA regulations in FSMA, where corn entering general commerce has an action level of 20-ppb (FDA, 2000). The risk level was reduced to 5-ppb in this study to compare with global standards that pose more stringent AFL regulations (Wu and Guclu, 2012; EFSA, 2013; Wu, 2015). The AFL-GBM had an overall accuracy of 96.77% for the 20-ppb risk threshold and 90.32% for the 5-ppb risk threshold. Using GBM, predictors that significantly influenced the model could be determined. The predictor analysis indicated several meteorological events and soil properties before planting and during corn growth that strongly influenced predictions of AFL during harvest. The ability to predict AFL contamination while corn is in the field signifies preventative versus reactive management of mycotoxin outbreaks, following FSMA’s overall goal for food and feed safety (Grover et al., 2016; King and Bedale, 2017).

The AFL-GBM model with a threshold of 20-ppb had adequate overall accuracy; however, the balanced accuracy was 50.00% for high and low contamination events. With only 4 years of historical Iowa AFL contamination data, only 2.30% of high contamination levels were above the regulatory limits of 20-ppb in the full historical database. When the risk threshold values were reduced to 5-ppb for AFL’s high and low contamination levels, the balanced accuracy was increased to 64.88%. The increased balanced accuracy was due to the enhanced amount of AFL contamination events in Iowa at 7.10%, therefore, including more observations of high AFL contamination due to the lower threshold. Compared to the published Illinois-centric AFL-GBM, which had a balanced accuracy of 61% for high, 54% for medium, and 60% for low contamination levels (Castano-Duque et al., 2022), the Iowa-centric AFL-GBM balanced accuracy for 20-ppb threshold was reduced (50%, Table 3) due to a lower incidence of AFL contamination events and a reduced overall total amount of observations available for the training data set. The reduced sensitivity with the GBM model for both thresholds indicates it has a low power to predict high AFL contamination events. Future research is needed to fine-tune the model to enhance the sensitivity. Cheng et al. (2019) suggest enriching the sample set with higher AFL contamination observations or adjusting the algorithm to penalize high false negative rates for improving the overall balanced accuracy of predictive models (Krawczyk, 2016). These results differ significantly from European models that show >75% general accuracy for corn in multiple regions (Battilani et al., 2013; Leggieri et al., 2021). If the risk value were reduced, as shown above, the model would have a higher specificity because the model would have the ability to learn from more balanced data (Cooper, 1990; Friedman, 2001; Natekin and Knoll, 2013).

The Iowa-centric GBM model included 70 predictors, with 26 having non-zero influence for AFL at 20-ppb and 53 for 5-ppb, respectively. The highest influence for AFL-GBM was the vegetative index in August for both risk thresholds (Supplementary Table S2). Like Castano-Duque et al. (2022)’s Illinois-centric model, weather data was acquired to help perform feature engineering with mechanistic mathematical equations to determine AFL production from Aspergillus growth. The vegetative index, a satellite-acquired data type, also known as the NDVI, helps distinguish visible red and near-infrared reflectance bands, allowing for the identification of vegetation, soil, water, and other features (Gro-Intelligence, 2019). NDVI reports plant greenness by indirectly measuring chlorophyll content and photosynthetic activity and behaving as a proxy of plant health, biomass, and yield (Wang et al., 2016; Gro-Intelligence, 2019). The vegetative index in August had the highest relative influence in the model for both risk thresholds; therefore, the results agreed that plant greenness in August was a significant determinant of AFL contamination at the time of harvest (Castano-Duque et al., 2022; Supplementary Figure S1). In August, on average, corn in Iowa should be approximately 8 feet tall with reasonably high vegetative indices due to sufficient plant greenness (Gro-Intelligence, 2019; USDA-NASS, 2023a,b). Suppose the NDVI index is low; this could be a diagnostic for drought and other crop stressors that might not be visible; this event could potentially predict AFL contamination due to fungal outbreaks during pre-harvest (Kerry et al., 2017; Gro-Intelligence, 2019). The vegetative index in July was also in the top five influential factors for AFL-GBM for the Iowa-centric model with the 20-ppb threshold and the seventh for the 5-ppb threshold, which is comparable to the Illinois-centric model (Castano-Duque et al., 2022). Therefore, NDVI may enhance AFL prediction preharvest in the late summer months in the Midwest Corn Belt due to the potential presence of detecting corn plant stress (Wang et al., 2016; Kerry et al., 2017).

Another top influential factor in the model was ARI in May, the second most influential feature for determining AFL contamination at the end of the growing season for the 20-ppb risk threshold (Supplementary Figure S2). During May, on average, in Iowa, corn is in the vegetative growth stage (USDA-NASS, 2023a,b). Higher ARI in this month was linked to the prediction of high AFL contamination levels; thus, it agrees with Yu et al. (2022) that warmer weather early in the planting and growing season leads to higher AFL contamination levels during harvest (Castano-Duque et al., 2022). For the 5-ppb risk threshold, ARI in July and May were the top influential factors. These findings agreed with the Illinois-centric model for early months having a high relative influence on predicting AFL contamination; therefore, agronomic practices that happen when corn is not in the field, such as tilling and drilling, should be further researched to determine if fungal growth for AFL is being harbored in soil residues (Accinelli et al., 2008; Abbas et al., 2009; Herrera et al., 2023). Tillage practices are an essential pre-planting factor for determining Aspergillus spores in leftover crop stover; the chances for infection are greater if leftover stover is left on the soil (Payne et al., 1986; Herrera et al., 2023). Even though it is wise from a conservation standpoint to conduct no-till practices to maintain soil resources, it may enhance AFL contamination (Abbas et al., 2009). Furthermore, alternative cover crop management may be allelopathic to A. flavus and parasiticus and should be further researched to create pest management practices that could be conducive to lower AFL contamination risk while maintaining no-till practices (Abbas et al., 2009; Damianidis et al., 2018).

Additionally, latitude as a predictor in the 20-ppb risk threshold AFL-GBM showed a high influence on contamination levels at the end of the year (Supplementary Table S2). Although high AFL contamination in Iowa was rare (2.30%), these events primarily occurred in the southern portion of the state at lower latitudes. This has been documented in previous studies, where the areas of Iowa that had the highest contamination values nearing the 20-ppb threshold were the Southwest and South-Central Crop Reporting Districts in Iowa (Mitchell et al., 2016). Similar results for AFL contamination at lower geographic latitudes were found in 1983 and 1989 in Iowa (Schmitt and Hurburgh, 1989; Russell et al., 1991). A unique finding in this study is that when the risk index values for AFL were reduced from 20-ppb to 5-ppb for high and low contamination events, latitude was removed from the top 20 influential predictors of the GBM model (Supplementary Table S2). It is hypothesized that the difference between latitude being in the top 20 influential factors for the model with the 20-ppb AFL regulatory limits is common cause versus special cause variation (MacGregor and Kourti, 1995). The 20-ppb model includes latitude as a factor for AFL contamination as a special cause variation; thus, the variation is unusual and unexpected, pointing to unique weather events such as drought or other crop stressors in the lower latitude in Iowa for higher AFL contamination values (MacGregor and Kourti, 1995). The 5-ppb model excludes latitude, as it is seen as a common cause variation, where it is expected and a consistent range of values with no pattern (MacGregor and Kourti, 1995). This finding agrees that low levels of AFL are commonly found in Iowa, even during years that are not conducive to producing AFL fungi (Zuber and Lillehoj, 1979; Munkvold et al., 2019).

A new addition to the Iowa-centric model was soil property predictors, which were indicated as a potentially influential factor for predicting AFL preharvest in the Illinois-centric model (Castano-Duque et al., 2022; Supplementary Table S1) and have been used in Europe-centric models (Leggieri et al., 2021). The top two influences for the 20-ppb threshold for soil properties were bulk density (db) (g/cm³) and saturated hydraulic conductivity (Ksat), 5 cm depth (μm/s⁻¹). For the 5-ppb threshold, the top two influences for soil properties were pH and rock fragments (% by volume from 0 to 50 cm) (Supplementary Table S2). Ksat was third in relative influence compared to the 20-ppb risk threshold (Figure 2). Ksat measures the water flow rate through saturated soil at a given hydraulic gradient. It relates to the water-holding capacity, the level of soil compaction, soil db, and texture and influences the propensity for soils to become water-logged in moist conditions (Vauclin et al., 1994; Brady and Weil, 2002; Libohova et al., 2018). Ksat and db are directly correlated, as both predictors affect water movement in the soil (Laboski et al., 1998; Rahimi et al., 2011; Figure 1). High db indicates low soil porosity and compaction, which may restrict root growth, air, and water movement (Blake, 1965; Supplementary Figure S3). Furthermore, soil pH is considered a crucial soil predictor as it broadly influences many soil processes, including nutrient and micro-nutrient availability, species richness of fungi, plant growth, acidification processes, cation exchange capacity, redox potential, and plant diseases (Thomas, 1996; Winter and Pereg, 2019; Baltensweiler et al., 2020). A. flavus and parasiticus have an optimum growth pH between 3.5 and 8 (Wheeler et al., 1991; Winter and Pereg, 2019). Soil pH is correlated with db, indicating an increase in pH with increased db (Li et al., 2020). To our knowledge, no studies have been published regarding the relationship between Ksat, db, pH, and A. flavus or parasiticus fungal growth for AFL production. Therefore, our results indicate that soil types and textures may be one of the main drivers in determining whether AFL in Iowa will be a high or low contamination year (Brady and Weil, 2002; Kerry et al., 2022).

These findings for the influential factors for both risk thresholds showcase that AFL contamination consists of a multi-traffic network, requiring interactions among the fungi, corn plant, and environmental conditions consisting of climate, weather, and soil properties (Klich, 2007; Abbas et al., 2009). This case study aimed to determine a baseline for interactions of these influential factors at two AFL risk thresholds of 20- and 5-ppb, respectively. Providing an Iowa-centric model that can predict AFL risk values pre-harvest to stakeholders in the context of the largest corn production region in the US enables early actions to prevent or hinder at-risk corn for individuals conducting hazard analyses and risk assessments. Stakeholders can take preventative or mitigative procedures that are controllable such as isolating AFL-contaminated grain, increasing fungicide application, early harvest, increased drying for lower storage moistures, and strategic marketing to end-users with higher tolerances such as beef cattle (Widstrom et al., 2003; Fumagalli et al., 2021). For grain elevators, handlers, and processors, the Iowa-centric model can be used in a proactive mycotoxin monitoring program, paving the way for strategic and targeted sampling and testing of corn (Park and Troxell, 2002; Whitaker, 2003).

Further research is warranted to understand what may cause corn plants to become susceptible versus non-susceptible to AFL infection based on the top influential factors provided by the GBM model. From a practical and management perspective, AFL contamination in corn can only be controlled to a certain degree. Stakeholders cannot fix uncontrollable factors for AFL contamination, such as certain climate or weather parameters; however, agronomic or management practices can mitigate AFL contamination for controllable influences, such as the management practices listed above (Fumagalli et al., 2021).

In conclusion, this study demonstrated that developing a predictive model for AFL in Iowa corn with historical contamination data, meteorological data, and soil properties had high accuracy at 96.77% for 20-ppb and 90.32% for 5-ppb, respectively. Comparable to the Illinois-centric model, the GBM model performed well by predicting low levels of AFL contamination; however, the balanced accuracy for high contamination values of AFL was reduced due to the rarity of high contamination events in Iowa. The vegetative index in August significantly influenced AFL risk for both thresholds, indicating that August is environmentally and ecologically important due to drought concerns. Additionally, soil property predictors, such as Ksat, pH, and db, may influence AFL contamination levels preharvest. Future work will finetune the Iowa model from county to macro-scale Crop Reporting District to enhance the data included to improve balanced accuracies and detect high AFL contamination levels. Applications of how to utilize the Iowa-centric model by stakeholders will be developed in collaboration with USDA-ARS.

The datasets presented in this article are not readily available because they are confidential, as requested by the Iowa Department of Agriculture and Land Stewardship. Requests to access the datasets should be directed to GM, gamosher@iastate.edu.

EB-S performed data collection, preparation, contribution to biological relationships with the model, and manuscript preparation. LC-D performed data analysis, imputation, feature engineering, modeling, contribution to biological relationships with the model, and manuscript preparation. PO and EW provided soil property parameters, biological relationships with the model, and manuscript preparation. KR, CH, GM, and EB performed initial planning and support for collaborative work. All authors contributed to the article and approved the submitted version.

Portions of this project were funded by USDA/NIFA award 2022-690008-36645.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.