Providing dietary feedback is important for promoting healthy behaviors in weight management, but the rapid development of obesity and the shortage of medical nutrition human resources have limited this health service. The rise of large language models (LLMs) offers the possibility of using artificial intelligence (AI) to simulate the behavior of human dietitians. However, existing studies have only explored LLM performance when generating answers to common nutrition-related questions; the use of LLMs to generate situation-adapted dietary feedback in practical weight management scenarios still needs further research. In this study, we collected dietary records and dietary feedback from primary dietitians through an mHealth weight management application. We conducted topic modeling to generalize how dietitians deliver nutrition guidance in real-world dietary feedback scenarios. Combining the in-context learning capability of LLMs with real-world data, we proposed a synthetic data generation approach (HDI-SDG) and trained an LLM for dietary feedback with the synthetic data (LLMDF-EXP). Experiments on automatic and manual evaluation of LLMDF-EXP and an LLM trained directly with the real-world data as well as generalized LLMs illustrated that LLMDF-EXP performed most similarly to human experts. Notably, there were no significant differences from human experts in terms of professionalism (p-value = 0.510) and usefulness (p-value = 0.498). The study highlights that integrating LLMs with real-world data in health management processes can enhance the situational adaptability of LLMs in practical health management environment applications.
Dai et al. (Thu,) studied this question.