Large Language Model-Assisted Research Question Development in Public Health: A Case Study in the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC)

Comprehensive Summary

This study, published in Public Health Nutrition, evaluated whether large language models can effectively assist in generating public health research questions. Researchers used ChatGPT-4 and the MixerBox Scholarly plugin to develop research questions related to revisions of the Special Supplemental Nutrition Program for Women, Infants, and Children food packages. Over three weeks, 90 questions were generated under two conditions: with and without being provided relevant literature. Six WIC experts evaluated the questions using the FINER criteria (Feasibility, Innovation, Novelty, Ethics, and Relevance). Results showed no significant difference between literature-fed and non-fed conditions. However, MixerBox-generated questions scored significantly lower than ChatGPT-generated ones. Ethics scores were rated higher than feasibility scores, and evaluator differences were significant. The study concludes that LLMs can generate acceptable-quality public health research questions, though further comparison with human researchers is needed.

Outcomes and Implications

This study highlights the growing role of artificial intelligence in public health research planning. LLMs like ChatGPT may help researchers, especially in resource-limited settings, efficiently develop structured and relevant research questions. If validated further, AI-assisted research design could streamline proposal development, reduce time burdens, and expand research capacity in programs such as WIC. However, differences in evaluator scoring and feasibility concerns suggest that human oversight remains essential.