Title: Gender-Biased Diagnostic Substitution in Large Language Model Medical Triage: Identical Symptoms, Disparate Urgency Levels

Abstract: This study examines whether large language models generate divergent medical triage outcomes for the same neurological symptoms when the patient’s gender and age are the only variables changed. We utilized three model architectures—Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini—to evaluate a standardized clinical profile characterized by persistent headaches, blurred vision, morning nausea, and visual disturbances. The experiment covered seven demographic scenarios: two genders (male, female) across three age brackets (25, 38, and 65), alongside a gender-neutral baseline. Each model was tested 30 times per condition, totaling 630 trials.

The results reveal a profound, systemic disparity in triage recommendations based on gender. Young female patients were assigned significantly lower emergency room (ER) referral rates compared to their male counterparts of the same age. Specifically, Gemini referred 0% of young women to the ER versus 23.3% of young men; Claude referred 6.7% of women versus 96.7% of men; and GPT referred 6.7% of women versus 66.7% of men (all p < 0.001). However, this gap vanished at age 65, with all models treating patients equally regardless of gender.

The root cause was identified as diagnostic substitution. The models tended to anchor on gender-linked diagnoses, frequently classifying young women with Idiopathic Intracranial Hypertension (IIH), a condition statistically more common in women of childbearing age. In contrast, young men were more often diagnosed with generic increased intracranial pressure involving space-occupying lesions. This diagnostic framing led the models to route female patients toward lower-urgency care, such as outpatient appointments, despite their symptoms being rated with comparable severity (7-9/10) to those of men. These findings indicate that clinical LLMs mirror established human biases by leveraging epidemiological priors to downplay triage urgency. Consequently, the study suggests that AI triage systems need to separate urgency assessment from probabilistic diagnostic assumptions. All code, prompts, and raw data are made publicly available.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC