Summarized by Masters of Longevity from UT Southwestern.
Study tests how large language models used for emergency triage change patient prioritization when clinical notes include demographic, psychiatric, or substance use details that may carry stigma.
Key Takeaways
- Three open-source LLMs initially identified higher-risk ED patients at rates above chance before stigmatizing details were added.
- Adding stigmatizing information like frequent ED visits or psychiatric history sometimes caused models to deprioritize patients who later deteriorated.
- In some cases nearly 10% of patients were incorrectly downgraded after stigmatizing language was added to records.
