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

Technical Summary: Gender-Dependent Diagnostic Substitution in LLM Medical Triage

Problem Statement
This study investigates whether Large Language Models (LLMs) integrated into healthcare applications produce disparate medical triage recommendations for identical clinical presentations when only the patient's stated gender and age vary. The research addresses a critical requirement for clinical AI: demographic parity in urgency assessment. Specifically, it tests the hypothesis that identical neurological symptoms should yield equivalent urgency recommendations regardless of gender, unless gender is clinically relevant to the urgency of the action (distinct from the diagnosis itself). The work is motivated by documented human clinical biases where women's neurological and cardiac symptoms are frequently attributed to benign or psychosomatic causes, leading to delayed or less aggressive workups.

Methodology
The authors conducted a counterfactual patient variation (CPV) experiment using three state-of-the-art, deployment-tier LLM families: Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini. These specific models were selected as they represent the most likely candidates for integration into production symptom-checking tools due to their cost and latency profiles.

• Experimental Design: The study utilized a single, standardized neurological symptom profile describing signs of raised intracranial pressure (persistent headache unresponsive to analgesics, blurred vision, morning nausea, and visual disturbances).
• Conditions: The symptom text was identical across 30 trials per condition, with only demographic identifiers varying. The conditions included three age groups (25, 38, 65) crossed with two genders (male, female), plus a gender-unspecified baseline (age 38). This resulted in 630 total trials (210 per model).
• Output Constraints: Models were prompted via API (temperature 0.3) to return structured JSON containing a diagnosis, severity score (1–10), urgency classification (Emergency/Urgent/Routine), recommended action (ER/Doctor appointment/Self-care), and differential tests.
• Analysis: The primary outcome was the Emergency Room (ER) referral rate. Secondary analysis focused on the specific diagnoses generated. Statistical significance was determined using Fisher's exact tests with Bonferroni correction, and effect sizes were calculated using Cohen's h.

Key Findings
The study identified a stark, systemic gender disparity in triage urgency for young patients, which was driven by a mechanism termed diagnostic substitution.

1. Disparity in ER Referrals: Young women received significantly lower ER referral rates than age-matched men, despite identical symptoms and comparable severity scores (7–9/10).

   • Age 25:
     • Gemini: 0% ER for women vs. 23.3% for men ($p < 0.001$).
     • Claude: 6.7% ER for women vs. 96.7% for men ($p < 0.001$).
     • GPT: 6.7% ER for women vs. 66.7% for men ($p < 0.001$).
   • Age 65: The disparity disappeared entirely across all models, with both genders receiving 90–100% ER referral rates.

2. Mechanism: Diagnostic Substitution: The models did not simply assign lower urgency to the same diagnosis for women. Instead, they altered the diagnosis itself based on gendered epidemiological priors.

   • Young Women: The models overwhelmingly anchored on Idiopathic Intracranial Hypertension (IIH), a condition statistically more common in women of childbearing age. Consequently, they routed these patients to outpatient appointments, assuming IIH is managed non-emergently.
   • Young Men: The models prioritized space-occupying lesions (masses, tumors) in the differential diagnosis, triggering high-urgency ER referrals to rule out these conditions.
   • Clinical Error: The authors note that while IIH is a valid differential for young women, the symptoms described (progressive visual disturbance, unrelenting headache) constitute red flags for raised intracranial pressure requiring urgent evaluation regardless of whether the etiology is IIH or a mass. The models correctly assessed severity but incorrectly decoupled urgency from the diagnosis.

3. Age Modulation: The bias was strictly age-dependent. At age 65, the models ceased anchoring on IIH for women (consistent with IIH epidemiology dropping post-menopause) and shifted to age-appropriate differentials (e.g., temporal arteritis, cerebrovascular disease) that triggered urgent pathways for both genders.

Contributions

1. Systemic Disparity Demonstration: The paper provides empirical evidence of significant gender disparities in triage urgency across three leading LLM families, with effect sizes reaching Cohen's h > 2.2.
2. Mechanism Identification: It identifies diagnostic substitution as the shared mechanism, where models select a gender-linked diagnosis (IIH) that inadvertently suppresses urgency, rather than applying a crude "women are lower urgency" heuristic.
3. Epidemiological Priors vs. Clinical Urgency: The study demonstrates that the bias is driven by statistical clinical priors learned from training data. The models replicate human clinical errors where a gender-associated diagnosis is used to prematurely close the diagnostic loop, reducing the perceived need for emergent care.
4. Open Science: The authors release all code, prompts, and raw results to facilitate further auditing.

Significance and Claims
The paper claims that clinical LLMs replicate documented human clinical biases not through explicit instruction, but by learning and applying epidemiological priors that are clinically inappropriate in the context of urgency assessment. The authors argue that:

• Urgency must be decoupled from diagnosis: For symptoms indicating raised intracranial pressure, the need for emergent evaluation should be determined by symptom severity, not the probability of a specific diagnosis like IIH.
• Audit requirements must evolve: Auditing AI triage systems requires analyzing the diagnostic reasoning (the differential diagnosis), not just the final action (ER vs. Doctor). A model can produce high severity scores while still routing patients to inferior care pathways via diagnostic substitution.
• Epidemiological accuracy is not clinical safety: A model correctly identifying that IIH is more common in young women is not sufficient; it fails if this statistical likelihood leads to a reduction in care urgency for a condition that requires immediate attention.

The authors conclude that while the models are not "sexist" in a crude sense, they have learned a specific, clinically grounded pattern that produces worse outcomes for women by prioritizing statistical likelihood over the immediate necessity of urgent investigation.