A new study from Heilongjiang Bayi Agricultural University pioneers rapid, non-destructive detection of illicit food additives using deep learning and near-infrared spectroscopy.
Wheat flour is a popular food item that is produced and shipped worldwide. As a result, it is susceptible to food fraud, meaning that illegal additives are added to it before being marketed to consumers. Because of this trend, it is important that there are methods in place to ensure wheat flour remains of high quality.
A recent study published in the journal Applied Food Research explores this topic (1). In their study, the research team from Heilongjiang Bayi Agricultural University (Daqing, China) has proposed a different method for detecting multiple illegal additives in wheat flour. By combining near-infrared (NIR) spectroscopy with deep learning and chemometric analysis, the researchers have developed a rapid, non-destructive, and highly accurate model to identify the presence of azodicarbonamide (ADA), talcum powder (TPD), and gypsum powder (GPD) (1).
Wheat is a cereal grain that is found in numerous food products worldwide. When grounded into flour, it is used to produce food items such as pasta and cereal. There are several classifications of wheat flour, including red wheat, white wheat, spring wheat, and winter wheat (2). The two main varieties are “soft” wheat, which has a low gluten content, and “hard” wheat, which has a high gluten content (2).
The researchers sought to improve on existing methods for detecting adulteration in wheat flour. Substances like ADA, which are used industrially as a blowing agent in plastics, and powders like talcum and gypsum can introduce harmful contaminants into the food chain. Traditional laboratory analyses, while precise, are often slow, expensive, and unsuitable for large-scale routine screening (1).
In their study, the researchers built a detection model that was designed to have the ability to identify all three additives in one test run. The model integrates long short-term memory (LSTM) neural networks, which are a type of deep learning algorithm known for handling sequential data, with partial least squares (PLS) regression, which is a powerful statistical tool for quantitative analysis (1). To further improve the performance of their model, the Bayesian optimization algorithm was employed to fine-tune the LSTM parameters, allowing the system to automatically identify optimal spectral features (1).
The researchers compared the LSTM-PLS model to other wavelength selection approaches, including competitive adaptive reweighted sampling (CARS) and genetic algorithm (GA) methods. The researchers found when they tested their model that it could achieve coefficients of determination (R²P) of 0.9828, 0.9771, and 0.9765 for ADA, TPD, and GPD, respectively, along with root mean square errors of prediction (RMSEP) of just 0.0008%, 0.2915%, and 0.2822% (1).
The model also showcased stability and high predictive accuracy, with the residual prediction deviation (RPD) values reaching 7.4067, 6.4020, and 6.2159 (1).
Using wheat flour from a farm in Daqing City, the researchers created 480 samples, including both single and mixed adulterated sets, with concentrations ranging from trace levels of 0.002% ADA to 20% TPD and GPD (1). Each sample was homogenized using a turbo oscillator to ensure even distribution of the additives, which is a critical step for accurate spectroscopic measurement.
By coupling near-infrared (NIR) spectroscopy with deep learning-based feature extraction, the research team has demonstrated that advanced data-driven approaches can transform food quality testing. Unlike conventional techniques that rely on chemical reagents or time-intensive chromatography, NIR spectroscopy offers a non-destructive, rapid, and environmentally friendly alternative (1). The addition of LSTM networks allows the model to handle the complex, overlapping spectral signatures that characterize adulterated samples, producing a level of accuracy suitable for industrial applications (1).
There are a couple of key takeaways from this study. First, the framework presented here has the potential to be used to detect adulteration in other food products. Second, through the use of artificial intelligence (AI), interpreting spectroscopic data could be automated, which allows for handheld sensors to screen for adulteration (1).
By demonstrating that LSTM-PLS modeling can accurately detect multiple contaminants in complex matrices, Liu and colleagues have provided the food industry with a powerful new tool to ensure product integrity and protect public health.
“Combining NIRS with LSTM offers a fast, reliable method for detecting adulterants in mixed wheat flour, advancing rapid quality evaluation of agricultural products,” the authors concluded in their study (1).
