Implicit Geographic Inference in LLM Medical Triage: Language-Driven Disparities in Emergency Recommendations

Technical Summary: Implicit Geographic Inference in LLM Medical Triage

Problem Statement
This study investigates a critical assumption in the deployment of Large Language Models (LLMs) for healthcare: that clinical recommendations are driven solely by the medical content of the input rather than superficial features like the language of the query. The authors hypothesize that models may use the input language as a proxy for the patient's geographic location, subsequently applying region-specific healthcare norms (e.g., "clinic-first" pathways vs. "defensive medicine") to identical clinical scenarios. This creates a potential disparity where multilingual users, immigrants, or expatriates receive inappropriate triage advice based on the language they choose to speak, rather than their actual location or medical need.

Methodology
The researchers conducted a controlled experiment using Gemini 3.5 Flash (May 2026 release) with a temperature of 0.3.

• Stimulus: A single, clinically ambiguous neurological symptom profile (persistent headache, blurred vision, morning nausea, visual disturbances in a 38-year-old) was presented.
• Variables: The prompt was authored manually in six languages (English, Spanish, Simplified Chinese, Hindi, Japanese, Arabic) to ensure native phrasing conventions were met.
• Experimental Conditions:
  1. Baseline: Prompts in 6 languages with no location context ($n=30$ per language).
  2. US Location Anchor: The same prompts with an appended sentence specifying the patient is in the United States.
  3. Reverse Anchor: An English prompt with appended sentences specifying locations in Tokyo or Mumbai.
  4. Back-Translation Control: The Japanese prompt was translated to English by the model and then triaged to rule out comprehension gaps.
• Output Metrics: The model was forced to output structured JSON including a severity score (1–10), urgency classification, and a specific action recommendation (Emergency Room vs. Doctor Appointment vs. Self-care).
• Scale: A total of 450 API calls were executed.

Key Findings

1. Language-Driven Disparities: Despite assigning nearly identical severity scores (7.7–8.0/10) across all languages, the model's recommended actions varied drastically.
   • English and Arabic prompts resulted in Emergency Room (ER) recommendation rates of 30%.
   • Japanese and Hindi prompts resulted in 0% ER recommendations.
   • Chinese and Spanish fell in between (20% and 13.3%, respectively).
2. Mechanism Validation (Geographic Inference):
   • US Anchor: Adding a US location to non-English prompts caused ER rates to surge (e.g., Japanese jumped from 0% to 46.7%; Hindi from 0% to 73.3%).
   • Reverse Anchor: Adding a Tokyo location to an English prompt reduced the ER rate from 30% to 6.7%.
   • Back-Translation: Translating the Japanese prompt to English resulted in an ER rate of 36.7%, statistically indistinguishable from the English baseline. This confirmed the model understood the symptoms equally well in Japanese; the 0% ER rate was due to inferred Japanese healthcare norms, not comprehension failure.
3. Implicit Reasoning: The model's stated reasoning (in the JSON output) described the same clinical risks (e.g., intracranial pressure) regardless of language. The geographic inference was implicit, influencing the final action without being explicitly stated in the rationale.

Contributions
The paper makes three primary contributions:

1. Demonstration of Systematic Bias: It provides empirical evidence that a production LLM generates systematically different medical triage recommendations based solely on prompt language, independent of clinical content.
2. Mechanism Identification: It identifies and validates "implicit geographic inference" as the causal mechanism, proving that language acts as a proxy for location which triggers region-specific healthcare norms.
3. Open Science: The authors release the complete dataset, experiment code, and raw results to facilitate reproduction and further study.

Significance and Claims
The authors argue that while these models may function correctly for monolingual users located in the country associated with their language, they fail for immigrants, expatriates, tourists, and multilingual users whose language does not match their geographic reality.

The paper concludes that the solution is not to remove geographic awareness from models, but to decouple it from language. The authors propose that developers should explicitly anchor the patient's location in the system prompt (e.g., "The patient is located in [Country]. Use [Country]'s healthcare system norms...") or prompt the user for their location before generating a recommendation. This ensures that language determines how the model communicates, not what it recommends.

Limitations
The authors modestly note several limitations: the study tested only one model (Gemini 3.5 Flash) and one specific symptom profile; the prompts were not independently verified by medical professionals in every language; and the sample size ($n=30$) limits statistical significance for mid-range comparisons (e.g., English vs. Spanish). They emphasize that without human clinician baselines, they cannot determine which language's recommendation is "correct," only that they differ.