Deep learning accurately predicts food categories and nutrients based on ingredient statements

Area of Science:

• Food Science
• Computational Biology
• Data Science

Background:

• Accurate food classification and nutrient analysis are crucial for understanding dietary intake and public health.
• Existing food databases often lack comprehensive data or require extensive manual curation.
• Developing automated methods for food attribute determination is essential for efficient data processing.

Purpose of the Study:

• To develop a novel, large-scale artificial intelligence (AI) dataset for food natural language processing (NLP).
• To evaluate the performance of AI models for food classification and nutrient estimation.
• To establish a foundation for AI-driven applications in food analysis and nutritional science.

Main Methods:

• Collection and modification of the USDA Branded Food Products Database (BFPD) to create the 134k BFPD dataset.
• Labeling the dataset with three food taxonomy levels and key nutrient values.
• Implementation and evaluation of the Multi-Layer Perceptron (MLP)-TF-SE model for food NLP tasks.

Main Results:

• The 134k BFPD dataset represents the largest collection of food types for AI-driven analysis to date.
• The MLP-TF-SE model achieved 99% accuracy in food classification.
• The model demonstrated high R-squared values for nutrient estimation, including 0.98 for calcium and 0.93-0.97 for calories, protein, sodium, and total lipids.

Conclusions:

• The developed AI dataset and MLP-TF-SE method offer a highly efficient approach to food classification and nutrient analysis.
• Deep learning techniques show significant potential for integration into various food classification and regression tasks.
• This work paves the way for broader AI applications in food science and nutritional monitoring.