Using Large Language Models to Audit Model Healthcare Biases.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Healthcare Informatics

Background:

• Large language models (LLMs) show promise in healthcare but exhibit demographic biases.
• Manual bias auditing is impractical due to large data volumes.
• LLMs can potentially audit other models for bias, but their effectiveness varies.

Purpose of the Study:

• To evaluate how LLM size and prompting strategies impact bias detection.
• To compare the bias detection performance of different LLMs using a healthcare dataset.
• To identify cost-effective LLM solutions for bias auditing in AI.

Main Methods:

• Utilized the Stanford Healthcare red-teaming dataset with prompts, outputs, and bias labels.
• Tested GPT-3.5-turbo, GPT-4o, llama3.3, and o1-mini for bias detection capabilities.
• Employed prompting techniques, including Thread of Thought, to assess their influence on bias detection.

Main Results:

• Smaller models like o1-mini achieved higher precision and F1 scores than GPT-4o.
• Self-critiquing features in larger models did not significantly improve bias detection.
• Thread of Thought prompting substantially enhanced bias detection across all tested models.

Conclusions:

• Smaller LLMs can be effective and cost-efficient for bias detection, particularly when precision is key.
• Prompting techniques are crucial for improving LLM-based bias auditing.
• The choice of LLM for bias detection should align with specific metric priorities.