Artificial Intelligence Assisted Curation of Population Groups in Biomedical Literature

Abstract

Curation of the growing body of published biomedical research is of great importance to both the synthesis of contemporary science and the archiving of historical biomedical literature. Each of these tasks has become increasingly challenging given the expansion of journal titles, preprint repositories and electronic databases. Added to this challenge is the need for curation of biomedical literature across population groups to better capture study populations for improved understanding of the generalizability of findings. To address this, our study aims to explore the use of generative artificial intelligence (AI) in the form of large language models (LLMs) such as GPT-4 as an AI curation assistant for the task of curating biomedical literature for population groups. We conducted a series of experiments which qualitatively and quantitatively evaluate the performance of OpenAI’s GPT-4 in curating population information from biomedical literature. Using OpenAI’s GPT-4 and curation instructions, executed through prompts, we evaluate the ability of GPT-4 to classify study ‘populations’, ‘continents’ and ‘countries’ from a previously curated dataset of public health COVID-19 studies.

Using three different experimental approaches, we examined performance by: A) evaluation of accuracy (concordance with human curation) using both exact and approximate string matches within a single experimental approach; B) evaluation of accuracy across experimental approaches; and C) conducting a qualitative phenomenology analysis to describe and classify the nature of difference between human curation and GPT curation. Our study shows that GPT-4 has the potential to provide assistance in the curation of population groups in biomedical literature. Additionally, phenomenology provided key information for prompt design that further improved the LLM’s performance in these tasks. Future research should aim to improve prompt design, as well as explore other generative AI models to improve curation performance. An increased understanding of the populations included in research studies is critical for the interpretation of findings, and we believe this study provides keen insight on the potential to increase the scalability of population curation in biomedical studies.