Retrieval Shift as a Source of Demographic Bias in Medical RAG

Ensuring equitable clinical decision support across patient demographics is essential for the safe deployment of AI systems in healthcare. Prior work has extensively studied demographic bias in standalone LLMs, while the effect of retrieval on demographic bias in retrieval-augmented generation (RAG) systems remains largely unexplored. We construct a demographically perturbed dataset from the MedQA test set and evaluate three medical RAG systems alongside two standalone LLMs. Our experiments show that RAG systems are substantially more sensitive to demographic perturbations than standalone LLMs, indicating that retrieval significantly amplifies demographic instability. Further analysis shows that demographic perturbations substantially alter retrieved evidence, frequently replacing clinically relevant documents with demographically matched but clinically irrelevant literature. These retrieval shifts persist even without strong corpus-level demographic imbalance, indicating that the root cause lies in the retrieval representation itself. While classifier-based filtering partially mitigates this effect, post-hoc filtering alone is insufficient to prevent demographic bias. Our findings highlight the need for demographic-invariant retrieval representations for reliable clinical deployment of medical RAG systems.