Guided Reading of the Article: Guerra-Adames, A., Avalos-Fernandez, M., Dorémus, O. et al. Auditing sex/gender disparities in emergency triage with LLM-based paired comparisons. npj Digit. Med. (2026)

Biases are not exclusive to algorithms: they can also creep into human decisions, including those made by healthcare professionals, despite their experience and commitment to patients. In emergency medicine, where decisions often must be made quickly, it is essential to be able to detect potential disparities in patient care that are not justified by the patient’s clinical condition. But how can we identify these biases without resorting to testing methods (such as situational tests or on-site verification tests, empirical tests of discriminatory practices) that rely on comparing equivalent situations and whose application in a clinical setting raises ethical and practical difficulties?

This presentation aims to explore the original approach that uses AI to examine the decision-making process in hospital emergency departments. By training an AI model to replicate documented decisions and then comparing hypothetical scenarios that differ only in a single patient characteristic, such as sex, it is possible to investigate whether these differences can influence the course of action or the assigned priority level. This approach, on its own, does not allow us to conclude that biases exist in real-world practices, but it offers a way to detect clues, scrutinize the data, and guide future research.

The work presented in this article is based on the dissertation by Ariel Guerra-Adames, co-advised by Emmanuel Lagarde and Marta Ávalos. https://doi.org/10.1038/s41746-026-03090-7