EQUITRIAGE: A Fairness Audit of Gender Bias in LLM-Based Emergency Department Triage

The study uses MIMIC-IV-ED, a real clinical dataset, as its substrate, which means these aren't hypothetical vignettes but cases drawn from actual emergency department encounters. That methodological choice makes the bias findings harder to dismiss as artifacts of synthetic prompting.

The tension here is direct and uncomfortable. Coverage from May 3rd of the Harvard study showing LLMs outperforming emergency room physicians on diagnostic accuracy created real momentum for faster clinical deployment. EQUITRIAGE arrives two days later and complicates that picture considerably: accuracy on aggregate benchmarks can coexist with systematic demographic bias at the individual decision level. The ethical divergence benchmark covered around the same time, 'Same prompt, different morals' from The Decoder, showed frontier models encoding different value systems across moral domains. Triage is precisely the kind of high-stakes sequential decision context where those divergences compound, and the procedural faithfulness gaps documented in the arXiv step-execution paper from May 1st suggest models may also be inconsistently applying whatever clinical logic they do have.

Watch whether any of the five audited vendors, particularly Google or OpenAI, respond with model cards or updated system prompts specifically addressing gender-neutral clinical framing before hospital pilot programs expand past the proof-of-concept stage. A vendor response within 90 days would signal the audit has regulatory traction; silence would suggest deployment is outpacing accountability.