Improving Personalized Meal Planning with Large Language Models: Identifying and Decomposing Compound Ingredients.

Area of Science:

• Artificial Intelligence
• Computational Nutrition
• Food Science

Background:

• Accurate identification and decomposition of compound ingredients in meal plans are crucial for personalized nutrition and dietary analysis.
• Challenges exist in precisely breaking down complex food items for allergy management and nutritional evaluation.

Purpose of the Study:

• To evaluate the effectiveness of three large language models (LLMs) in decomposing compound ingredients into basic components within structured meal plans.
• To compare the performance of GPT-4o, Llama-3 (70B), and Mixtral (8x7B) in ingredient decomposition for nutritional analysis.

Main Methods:

• Generated 15 structured meal plans with compound ingredients using GPT-4o.
• Assessed three LLMs (GPT-4o, Llama-3 (70B), Mixtral (8x7B)) for their ability to identify and decompose compound ingredients.
• Mapped decomposed ingredients to the USDA FoodData Central repository via API for nutritional value aggregation.
• Evaluated accuracy using manual review by nutritionists, F1-scores, and statistical significance testing.

Main Results:

• Both Llama-3 (70B) and GPT-4o demonstrated superior performance over Mixtral (8x7B).
• Llama-3 (70B) achieved an average F1-score of 0.894 (95% CI: 0.84-0.95).
• GPT-4o achieved an average F1-score of 0.842 (95% CI: 0.79-0.89), while Mixtral (8x7B) scored 0.690 (95% CI: 0.62-0.76).

Conclusions:

• The open-source Llama-3 (70B) model outperformed the commercial GPT-4o, demonstrating superior ingredient decomposition capabilities.
• LLMs show significant potential for enhancing precision nutrition, enabling better meal customization and dietary practice promotion.
• Accurate decomposition of meal components by advanced LLMs can facilitate healthier, individualized dietary recommendations.