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

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Healthcare Informatics

Background:

• Large language models (LLMs) are increasingly used to automate administrative tasks in long-term care, such as summarizing patient records.
• However, LLMs can perpetuate biases present in their training data, potentially impacting healthcare equity.

Purpose of the Study:

• To evaluate gender bias in summaries of long-term care records generated by recent open-source LLMs.
• Specifically comparing Meta's Llama 3 and Google Gemma against older benchmark models.

Main Methods:

• Generated gender-swapped versions of 617 long-term care records.
• Produced summaries using Llama 3, Gemma, T5, and BART.
• Quantified counterfactual gender bias using sentiment analysis, word frequency, and thematic patterns.

Main Results:

• Benchmark models showed some gender-based variations.
• Llama 3 demonstrated no significant gender-based differences in summaries.
• Google Gemma exhibited the most pronounced gender bias, with male summaries focusing more on health issues and women's needs being downplayed.

Conclusions:

• Gender bias in LLM summaries, particularly downplaying women's health issues, could lead to disparities in care service allocation.
• While LLMs offer administrative benefits, their varying performance necessitates rigorous bias evaluation.
• The study provides a practical framework for quantitatively assessing gender bias in LLMs.