Using Large Language Models to Generate Dietary Feedback Similar to Human Experts in Weight Management: Experiments...

Area of Science:

• Artificial Intelligence in Healthcare
• Digital Health and Nutrition
• Behavioral Science and Weight Management

Background:

• Dietary feedback is crucial for weight management, but resource limitations hinder its provision.
• Large language models (LLMs) show potential for simulating dietitian behavior in providing nutrition guidance.
• Current research on LLMs for dietary feedback is limited to general nutrition questions, not practical weight management scenarios.

Purpose of the Study:

• To investigate the use of LLMs for generating situation-adapted dietary feedback in weight management.
• To develop and evaluate a novel synthetic data generation approach for training LLMs in this domain.
• To assess the performance of a trained LLM against human experts and other LLM approaches.

Main Methods:

• Collected real-world dietary records and feedback from dietitians via an mHealth application.
• Employed topic modeling to understand dietitian feedback patterns in real-world scenarios.
• Developed the HDI-SDG approach to generate synthetic data for training an LLM (LLMDF-EXP) for dietary feedback.

Main Results:

• LLMDF-EXP, trained with synthetic data, demonstrated performance most similar to human experts.
• Evaluations showed no significant differences between LLMDF-EXP and human experts in professionalism (p=0.510) and usefulness (p=0.498).
• The proposed method outperformed LLMs trained directly on real-world data and generalized LLMs.

Conclusions:

• Integrating LLMs with real-world data through synthetic data generation enhances situational adaptability in health management.
• LLMDF-EXP represents a significant advancement in AI-driven dietary feedback for practical weight management.
• This approach offers a scalable solution to address the shortage of medical nutrition resources.