Mitigating Unfairness Risk Generated by Large Language Models in Healthcare

Read full text Summarise with VIRAT
Authors: Shuyue Song
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Large Language Models (LLMs) are increasingly being adopted in healthcare, specifically to synthesize data related to electronic health records (EHRs), which is emerging as a means of moderating data availability constraints. However, general LLMs exhibit severe social biases in medical tasks under Direct prompting. In this study, the Markov chain Monte Carlo (MCMC) algorithm is introduced into GPT-4’s prompt to mitigate the unfairness risk of general LLMs in medical text generation tasks, which is known as MCMC with GPT-4. In addition, we demonstrate bounds on the bias of the MCMC with GPT-4 estimator with respect to GPT-4’s representation and the true distribution, respectively, where GPT-4’s choices are regarded as elements of the MCMC algorithm. Accordingly, we show that the social bias perpetuated in the GPT-4’s medical text generation task is efficiently mitigated under MCMC with GPT-4 compared to the Direct Prompting experimentally.

Fields of Study & Subjects

Ethics and Social Impacts of AI Artificial Intelligence in Healthcare and Education Topic Modeling

Publication details

Journal / source
Macquarie University
Publication date
25 March 2026
Language
EN
Document type
Thesis