Gender-Dependent Diagnostic Substitution in LLM Medical Triage: Same Symptoms, Unequal Urgency

2026-06-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Clinical Care & Medical Practice · Depth: Expert, quick

Summary

A study investigating large language models (LLMs) in medical triage found significant gender-dependent disparities in emergency room (ER) referral rates for identical neurological symptoms. Using Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini, researchers presented a standardized symptom profile (persistent headache, blurred vision, morning nausea, visual disturbances) to patients across various age and gender demographics. Young women received substantially lower ER referral rates than age-matched men (Gemini: 0% vs. 23.3%; Claude: 6.7% vs. 96.7%; GPT: 6.7% vs. 66.7%, all p < 0.001). This disparity vanished at age 65. The primary mechanism identified was diagnostic substitution, where models preferentially diagnosed young women with Idiopathic Intracranial Hypertension (IIH) and men with generic increased intracranial pressure, routing women to lower-urgency outpatient care despite comparable symptom severity.

Key takeaway

For AI Scientists and Research Scientists developing clinical LLMs, this study highlights a critical need to mitigate gender-dependent diagnostic bias. Your models, like Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini, may replicate human biases, leading to unequal triage urgency. You must decouple urgency assessment from probabilistic diagnostic priors and rigorously test for demographic disparities, especially concerning conditions with epidemiological links to specific groups, to ensure equitable patient care.

Key insights

LLMs exhibit systemic gender bias in medical triage, driven by diagnostic substitution that lowers urgency for young women.

Principles

• LLMs replicate human clinical biases.
• Epidemiological priors can suppress triage urgency.
• Decouple urgency assessment from diagnostic priors.

Method

Standardized symptom profiles were presented to Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini across age and gender conditions (630 trials) to assess triage recommendations.

In practice

• Test clinical LLMs for demographic biases.
• Review LLM diagnostic substitution patterns.
• Implement urgency-diagnosis decoupling.

Topics

• LLM Bias
• Medical Triage
• Diagnostic Substitution
• Gender Disparity
• Clinical AI
• Healthcare AI

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• LLMs show bias in opioid prescribing — LLMs show biases in opioid prescribing recommendations, disproportionately affecting marginalized patient groups.
• Please don’t trust your chatbot for medical advice — LLMs are unreliable for medical advice, frequently generating misinformation with overconfidence and lacking clinical reasoning.
• Auditing Multimodal LLM Raters: Central Tendency Bias in Clinical Ordinal Scoring — Multimodal LLMs exhibit a central tendency bias in clinical ordinal scoring, compressing predictions towards the scale's middle.
• v287: Proceedings of CHIL 2025 — The proceedings highlight diverse advancements in machine learning applications across healthcare domains.
• Gender Disambiguation in Machine Translation: Diagnostic Evaluation in Decoder-Only Architectures — Decoder-only MT models exhibit gender bias, but post-training improves contextual awareness and reduces masculine "Prior Bias."
• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.