Bias Evaluation and Mitigation in Retrieval-Augmented Medical Question-Answering Systems

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Health Informatics

Background:

• Retrieval-Augmented Generation (RAG) enhances medical question-answering (QA) systems by integrating external knowledge.
• Standalone large language models (LLMs) can exhibit inaccuracies, while RAG systems may inadvertently perpetuate demographic biases (race, gender, socioeconomic factors).

Purpose of the Study:

• To systematically evaluate demographic biases in medical RAG pipelines.
• To assess the effectiveness of bias mitigation strategies in improving fairness and accuracy.

Main Methods:

• Evaluation across multiple QA benchmarks (MedQA, MedMCQA, MMLU, EquityMedQA) using demographic-sensitive queries.
• Implementation and comparison of bias mitigation techniques: Chain-of-Thought, Counterfactual filtering, Adversarial prompt refinement, and Majority Vote aggregation.

Main Results:

• Significant demographic disparities were identified in retrieval consistency and answer correctness.
• Majority Vote aggregation demonstrated improvements in both accuracy and fairness metrics.

Conclusions:

• Medical RAG systems exhibit notable demographic biases.
• Fairness-aware retrieval and prompt engineering are crucial for developing equitable medical QA systems.