Bias Patterns in the Application of LLMs for Clinical Decision Support: A Comprehensive Study

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems
• Natural Language Processing

Background:

• Large Language Models (LLMs) are increasingly explored for clinical decision support.
• Concerns exist regarding potential social biases in LLM outputs based on protected patient attributes.
• Understanding how LLM design choices impact bias is crucial for safe implementation.

Purpose of the Study:

• To assess the extent of social bias in LLMs concerning protected patient attributes.
• To investigate the influence of LLM architecture and prompting strategies on clinical decision support bias.

Main Methods:

• Evaluated eight popular LLMs (general-purpose and clinically trained) using clinical vignettes and question-answering datasets.
• Employed red-teaming strategies to analyze demographic impacts on LLM outputs.
• Compared prompting techniques like Zero-shot and Chain of Thought.

Main Results:

• Observed significant disparities across protected groups in LLM outputs.
• Larger models and medical fine-tuning did not guarantee reduced bias.
• Prompt phrasing critically influenced bias, while Chain of Thought prompts reduced biased outcomes.

Conclusions:

• LLMs exhibit notable social biases in clinical contexts, affected by architecture and prompt engineering.
• Rigorous evaluation and enhancement of LLMs are essential before clinical integration.
• Further scrutiny is needed to ensure equity in AI-driven healthcare applications.