Evaluating gender bias in large language models in long-term care.

Background: Large language models (LLMs) are being used to reduce the administrative burden in long-term care by automatically generating and summarising case notes. However, LLMs can reproduce bias in their training data. This study evaluates gender bias in summaries of long-term care records generated with two state-of-the-art, open-source LLMs released in 2024: Meta's Llama 3 and Google Gemma.

Methods: Gender-swapped versions were created of long-term care records for 617 older people from a London local authority. Summaries of male and female versions were generated with Llama 3 and Gemma, as well as benchmark models from Meta and Google released in 2019: T5 and BART. Counterfactual bias was quantified through sentiment analysis alongside an evaluation of word frequency and thematic patterns.

Results: The benchmark models exhibited some variation in output on the basis of gender. Llama 3 showed no gender-based differences across any metrics. Gemma displayed the most significant gender-based differences. Male summaries focus more on physical and mental health issues. Language used for men was more direct, with women's needs downplayed more often than men's.

Conclusion: Care services are allocated on the basis of need. If women's health issues are underemphasised, this may lead to gender-based disparities in service receipt. LLMs may offer substantial benefits in easing administrative burden. However, the findings highlight the variation in state-of-the-art LLMs, and the need for evaluation of bias. The methods in this paper provide a practical framework for quantitative evaluation of gender bias in LLMs. The code is available on GitHub.