Auditing Sex/Gender Disparities in Emergency Triage with LLM-based Paired Comparisons

A research paper from arXiv introduces a novel, domain-agnostic method utilizing Large Language Models (LLMs) to identify and quantify sex/gender disparities within documented clinical decision-making. This approach involves training an LLM to mimic observed decisions and then analyzing sex-swapped patient profiles where only the sex is altered. Initial application to emergency triage data from Bordeaux University Hospital and validation on MIMIC-IV datasets revealed that otherwise identical presentations were more likely to receive a lower triage priority when sex/gender was modified.

Why it matters

This research is strategically important because it provides a new, scalable method for identifying systemic biases in critical decision-making processes, particularly in domains like healthcare. Understanding and mitigating such biases is crucial for maintaining equity, improving operational integrity, and ensuring compliance with ethical and regulatory standards.

Key insights

• A new domain-agnostic method employs LLMs to quantify sex/gender-related asymmetries in clinical decision-making.
• The method trains an LLM to emulate existing clinical decisions and then evaluates hypothetical 'sex-swapped' pairs.
• This approach maintains all clinical content constant, varying only the sex/gender attribute for comparison.
• Applied to over 140,000 emergency admissions at Bordeaux University Hospital, it identified disparities.
• Methodological portability was demonstrated by successful testing on the MIMIC-IV dataset, which represents a different language, population, and healthcare system.
• Fine-tuning Mistral NeMo 12B for triage prediction and using Mistral Small 24B for pair generation were key technical components.
• Preliminary findings indicate that identical clinical presentations were more likely to receive a lower triage priority under certain sex/gender assignments.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2511.17124