A large language model analysis of global inequities in precision medicine research on diabetes

Although nearly 80 % of patients with diabetes live in low- and middle-income countries, it is currently unknown what proportion of precision medicine research is based on these populations. Manual screening of literature is time consuming and resource intensive. Our objective is to characterize the proportionality of diabetes burden and precision medicine research across ten geographic regions using a scalable large language model (LLM) enabled workflow.

An electronic search of PubMed identified titles and abstracts of studies related to precision medicine in diabetes from 2010 to 2023 (n = 129,154). Two reviewers independently labelled a random sub-sample and classified their source populations, and whether these were primary studies of precision medicine in diabetes. Using this labeled data (n = 2196), we developed prompts and selected hyperparameters for GPT-4o. We then used GPT-4o to classify the remaining studies and estimated the ratio of research output to disability adjusted life years [DALY] from the Global Burden of Disease [GBD] study 2021.

Of the 15,507 studies identified as precision medicine in diabetes, 33.8 %, 20.9 % and 14.3 % were from North America, Western Europe, and East Asia respectively. The number of studies was the most proportionate to disease burden for North America (0.95 per 1000 DALYs) and Western Europe (0.78 per 1000 DALYs), and the least proportionate for Southeast Asia, South Asia, and Sub-Saharan Africa (0.02 each per 1000 DALYs).

Future research investments into omics-based research should prioritize regions outside Western Europe and North America for achieving global equity in diabetes care.