Pediatric triage in the emergency department: ten open-source LLMs tested against demographic and social biases

Changing only a child's origin, health coverage or social context changes the level of urgency assigned by certain models. Bigger or more medical does not mean less biased.

Hospitals are interested in large open-source language models for clinical decision support, because they can run locally. But are they fair? A team of researchers subjected ten models to a counterfactual audit on pediatric triage in the emergency department, according to the ESI (Emergency Severity Index) scale. From real clinical vignettes and textbook cases, they created pairs of medically identical cases, differing only in one variable: demographic, socio-economic, access to care, behavior, social context or organization of the department.

Among the models tested are Qwen2.5-7B and 14B, variants of MedGemma, MedLLaMA2-7B, GPT-OSS-20B and GPT-OSS-120B. Sensitivity to these changes varies greatly and does not necessarily decrease with size or medical specialization; several larger or medical models have even been shown to be more unstable. The best performer is a Qwen2.5-7B refined by QLoRA: its sorting level changes in 5.27% of cases, compared to 16.02% for the base model.

The authors note that the overall rates mask clinically important trends, such as the direction of the lag (sub-triage or over-triage). They argue for this type of audit, which is light and interpretable, before any clinical deployment.