Key Points
- A recent study demonstrated that NIR spectroscopy combined with PLS regression and artificial neural networks (ANN) can rapidly and non-destructively detect harmful fumonisins (FB1 and FB2) in maize.
- The study analyzed 150 maize samples from Portugal’s Tagus Valley using NIR spectral data and found strong predictive performance, with ANN models achieving a correlation coefficient of 0.99 for calibration and 0.95 for validation.
- Researchers concluded that NIR spectroscopy is a promising tool for real-time food safety monitoring in maize supply chains, with future work focused on improving model calibration and adapting it to a broader range of maize varieties and conditions for practical, widespread use.
In a recent study, researchers from the National Institute for Agricultural and Veterinary Research (INIAV) in Portugal explored how to improve food safety monitoring in the maize supply chain. This study, which was published in the journal Food Chemistry: X, explored the value of using near-infrared (NIR) spectroscopy and advanced chemometric algorithms to rapidly and non-destructively detect harmful fumonisins in maize (1).
What are fumonisins?
Fumonisins are mycotoxins commonly found in nature (2). They are known to pose serious risks to human and animal health (1,2). Fumonisins, particularly fumonisin B1 (FB1) and fumonisin B2 (FB2), are secondary metabolites produced by Fusarium fungi that commonly infect maize crops worldwide (1,2). These mycotoxins are associated with diseases such as esophageal cancer and neural tube defects in humans, as well as leukoencephalomalacia in horses and pulmonary edema in swine (1). Controlling their presence in food is critical for public health and agricultural trade. In the European Union, the maximum legal limit for fumonisins in unprocessed maize is 4000 μg/kg (1).
What did the researchers do in their study?
The main focus of this study was to evaluate whether NIR spectroscopy can be an effective method for fumonisin screening. To do so, the research team collected 150 maize samples from multiple farms in Portugal’s agriculturally rich Tagus Valley region between 2018 and 2020 (1). These maize samples were taken from harvester tanks during field harvesting, and the reason for doing so was to ensure depth-specific sampling and representative collection. Each sample was dried, milled uniformly to a 1 mm particle size, and stored at −20 °C before fumonisin extraction and spectral analysis (1).
What were the predictive models used in this study?
Two predictive models using chemometric techniques were used in the study. One of the models used partial least squares (PLS) regression and the other used artificial neural networks (ANN). Both methods were applied to raw NIR spectral data to predict concentrations of FB1, FB2, and their sum (FB1 + FB2) in maize flour samples (1).
The researchers found that PLS regression yielded strong calibration results. The coefficients of determination (R²) were 0.90 for FB1, 0.98 for FB2, and 0.91 for the combined FB1 + FB2 model (1). The ratio of prediction to deviation (RPD), which is a measure of model performance, ranged from 2.8 to 3.6, indicating satisfactory to excellent predictive power (1). Meanwhile, the ANN models showed even stronger results for predicting combined fumonisin levels, achieving a correlation coefficient (R) of 0.99 during calibration, with a root mean square error (RMSE) of just 131 μg/kg (1). In validation, the ANN model achieved R = 0.95 and RMSE = 656 μg/kg, which demonstrates the method’s value (1).
What were the key findings of the study?
Based on the results of their study, the researchers found that raw NIR spectra could differentiate between fumonisin-contaminated and uncontaminated samples based on peak positions and intensities (1). As a result, the researchers showed that NIR spectroscopy could be used to rapidly screening for mycotoxins. The researchers also showed that the PLS models were particularly effective in detecting FB2, whereas the ANN models demonstrated their effectiveness in modeling the total fumonisin content (1).
What are the next steps in this work?
The researchers discussed in their article that now that their methods showcased their potential in evaluating mycotoxins, the models need to be refined further. To do so, they recommend increasing calibration accuracy and extending model robustness across different maize varieties and environmental conditions (1). Doing so will ultimately make the model better for real-world applications.
If successfully implemented, predictive NIR models could be used to screen harvested maize in real time, helping processors avoid the use of contaminated batches and reducing the risk of fumonisin exposure to consumers (1). The ability to rapidly assess mycotoxin content without the need for complex sample preparation or costly laboratory instruments could also be transformative for small-scale farmers and producers in low-resource settings (1).
References
- Carbas, B.; Sampaio, P.; Barros, S. C.; et al. Rapid Screening of Fumonisins in Maize Using Near-infrared Spectroscopy (NIRS) and Machine Learning Algorithms. Food Cont. X 2025, 27, 102351. DOI: 10.1016/j.fochx.2025.102351
- Kamle, M.; Mahato, D. K.; Devi, S.; et al. Fumonisins: Impact on Agriculture, Food, and Human Health and their Management Strategies. Toxins (Basel). 2019, 11 (6), 328. DOI: 10.3390/toxins11060328
