Key Points
- Researchers from Zhejiang University of Science and Technology developed a novel method combining NIR spectroscopy with CNN to accurately predict resistant starch content in rice.
- The study used SHapley Additive exPlanations (SHAP) to identify the most influential spectral range (2000–2500 nm), improving both the accuracy and efficiency of the model while making the deep learning process more interpretable for researchers.
- The method’s compatibility with portable NIR instruments supports its use for rapid, on-site food analysis.
A new study published in the journal Food Chemistry by lead authors Qian Zhao and Jun Huang from Zhejiang University of Science and Technology unveil a new data-driven framework for predicting resistant starch content in rice (1). This study, which presents an improved method that combines spectroscopy and artificial intelligence (AI) technologies, highlights the current knowledge gap in interpretability concerns in food analysis studies.
What is resistant starch and why is it important?
Resistant starch is a chain of glucose that is found in a variety of foods, including oats and cashews. It is a better health alternative to regular starch because it has fewer calories (2). The result is that resistant starch helps consumers feel fuller while not ingesting as many calories (2). There are four types of resistant starch: Type 1, which is found in grains and seeds; Type 2, which is found in starchy foods like bananas; Type 3, which is found in potatoes and rice; and Type 4, which is manmade and formed through chemical processing (2). One of the main reasons why resistant starch is important for a person’s diet is that it provides sustenance to the good bacteria in their intestine, which helps stimulate the process of producing short-chain fatty acids (2).
Why is quantifying resistant starch difficult?
Currently, traditional methods struggle to quantify resistant starch because of several key issues. For one, traditional methods are often costly, so not all laboratories can use these methods. For two, the protocols required to use these traditional methods are inefficient, which makes them impractical for wider adoption (1). Therefore, there is a current instrumentation and method gap in this field.
What was the research team’s method proposed in this study?
In this study, the research team proposed a new method that could potentially ameliorate the concerns traditional methods bring. Their method involved using near-infrared (NIR) spectroscopy coupled with convolutional neural networks (CNN). CNN is a type of deep learning algorithm inspired by human visual perception (1). This hybrid model was further enhanced through data augmentation techniques that simulate variations in sample data to improve learning efficiency (1).
What was the result of their proposed method?
The result of using NIR spectroscopy in conjunction with CNN was that the researchers were able to predict resistant starch content more accurately. They achieved a correlation coefficient of prediction (Rp²) reaching 0.992 (1). This result demonstrated that the researchers’ method improved upon conventional learning models used in previous studies, including partial least squares regression (PLSR) and support vector machine regression (SVMR).
What made this approach unique?
The researchers’ approach was different from previous methods because it incorporated a game theory technique in their study. To help improve interpretability of the data, the researchers used SHapley Additive exPlanations (SHAP) to quantify the contribution of each input feature to the model’s output (1). SHAP analysis identified a narrow spectral window (in this case, 2000 to 2500 nm) as most influential in resistant starch detection (1). This is important because by focusing only on this critical spectral range, the process becomes faster, less computationally intensive, and more cost-efficient (1).
What are the important takeaways from this study?
Beyond academic validation, the method proposed in this study holds vast commercial and industrial potential. The optimized use of the 2000–2500 nm range opens the door to integration with portable NIR devices, allowing for rapid on-site resistant starch analysis in agricultural and manufacturing environments (1). Because portable instrumentation is a growing trend in this space, the method the researchers proposed in their study fits into the current market demand instrumentation-wise.
In addition, the study reinforces the ongoing trend in analytical science to integrate spectroscopic methods with machine learning. This combination is driving both performance and transparency improvements to current methods to help improve food analysis, as well as other applications (1).
To conclude their study, the research team stressed the importance of refining predictive models specifically tailored to the critical spectral window and advancing portable instrumentation. They argue that by doing so, critical improvements can be made to these novel methods, allowing them to be more widely implemented in the food industry (1).
References
- Zhu, Q.; Gao, Y.; Yang, B.; et al. Advanced Data-driven Interpretable Analysis for Predicting Resistant Starch Content in Rice Using NIR spectroscopy. Food Chem. 2025, 486, 144311. DOI: 10.1016/j.foodchem.2025.144311
- Gunnars, K. Resistant Starch 101 — Everything You Need to Know. Healthline.com. Available at: https://www.healthline.com/nutrition/resistant-starch-101 (accessed 2025-06-23).
