To examine whether large language models (LLMs) exhibit U.S.-centric bias in clinical decision support by disproportionately adhering to U.S. neuroradiology guidelines over equally authoritative international ones.
Approach:
Key Findings:
Significant variations in guideline adherence were noted, with adherence rates ranging from 58% to 86%.
Geographic bias in LLMs could misalign clinical workflows with local practice standards.
No prior studies have investigated geographic bias in LLMs for clinical decision support.
Interpretation:
The study highlights the potential for geographic bias in LLMs, which may affect the appropriateness of clinical recommendations based on local guidelines.
Limitations:
The study did not involve real patient data or human subjects.
Only guidelines published in English or French were considered.
Conclusion:
The findings suggest a need for further investigation into the geographic bias of LLMs in medical decision-making.
