Advanced data-driven interpretable analysis for predicting resistant starch content in rice using NIR spectroscopy.

Area of Science:

• Food Science and Technology
• Analytical Chemistry
• Biotechnology

Background:

• Resistant starch (RS) offers significant health benefits but traditional quantification methods are inefficient for large-scale use.
• Existing methods for RS analysis are often labor-intensive, expensive, and not suitable for real-time industrial applications.
• There is a need for rapid, cost-effective, and scalable analytical solutions for RS determination in food products.

Purpose of the Study:

• To develop and validate a data-driven framework for accurate and efficient resistant starch prediction.
• To integrate Near-Infrared (NIR) spectroscopy with advanced machine learning models for quantitative analysis.
• To enhance the interpretability of deep learning models in spectroscopic analysis for food quality assessment.

Main Methods:

• Utilized Near-Infrared (NIR) spectroscopy for spectral data acquisition.
• Developed a Convolutional Neural Network (CNN) model incorporating data augmentation for RS prediction.
• Employed SHapley Additive exPlanations (SHAP) to interpret the CNN model and identify critical spectral regions.

Main Results:

• The CNN model achieved exceptional prediction accuracy (Rp² = 0.992), surpassing traditional methods like PLSR and SVMR.
• SHAP analysis identified specific critical wavelengths (2000-2500 nm) contributing significantly to RS prediction.
• The optimized spectral range reduced data acquisition time and analytical costs.

Conclusions:

• The integrated NIR-CNN-SHAP framework provides a rapid, cost-effective, and interpretable solution for resistant starch quantification.
• This approach enhances data acquisition efficiency and simplifies operational complexity for food quality control.
• The study establishes a practical and scalable method for deploying NIR spectroscopy in industrial food production settings.