Accelerating the Classification of NOVA Food Processing Levels Using a Fine-Tuned Language Model: A Multi-Country...

Area of Science:

• Nutrition Science
• Computational Biology
• Public Health

Background:

• Ultra-processed foods (UPFs) consumption is rising globally, linked to increased noncommunicable disease risks.
• Current methods for classifying food processing levels (e.g., NOVA system) are manual, labor-intensive, and time-consuming.
• There's a need for efficient, scalable methods to monitor UPF content in national food supplies.

Purpose of the Study:

• To develop and validate an automated method for classifying food processing levels using transformer-based language models.
• To assess the accuracy and generalizability of the automated classification across different national food databases (Canada, Argentina, US).
• To provide a cost-effective tool for policymakers to monitor and regulate UPFs.

Main Methods:

• Utilized transformer-based language models (e.g., BERT) fine-tuned on ingredient list text from food labels.
• Applied the models to national food databases from Canada, Argentina, and the US.
• Compared the performance of language models against traditional machine learning models using nutrient data and bag-of-words approaches.

Main Results:

• Achieved high overall accuracy (F1 score of 0.979) in classifying Canadian food products, outperforming traditional models.
• Demonstrated high prediction accuracy (0.98) for processed and ultra-processed foods across most food categories.
• Confirmed the effectiveness and generalizability of the automated strategy for Argentina and US food databases.

Conclusions:

• Transformer-based language models provide an accurate and efficient automated solution for classifying food processing levels.
• The developed method offers a scalable and cost-effective approach for monitoring UPFs in global food supplies.
• This automation supports public health initiatives aimed at regulating UPFs and mitigating associated noncommunicable disease risks.