Race and gender disparities in mental health documentation in primary care: Findings from a video vignette experiment.

Unstructured clinical notes could be a key source for mental health detection and care in primary care settings. A growing use of LLMs to extract mental health indicators from unstructured clinical notes has the potential to aid in this effort. However, these technologies do not account for provider bias at the point of data entry. In this video vignette survey experiment study, we explore how mental health information in written clinical notes differ in frequency, content, and quality by patient race-gender presentation. Leveraging the strengths of an experimental design with 111 practicing healthcare providers in the US, we find substantive qualitative differences in clinical notes and significant quantitative differences in the assessment of patient mental health based on the race-gender presentation of the patient. Specifically, providers documented fewer mental health concerns and fewer requests or recommendations for mental health follow-up care for Black patients. They also assessed that White patients were in significantly worse mental health compared to their Black counterparts. Providers also included more justifying language for alcohol use and inquiries about drug use for male patients compared to female patients. These differences in the source data for LLM extraction have implications for health equity, making it so these new innovations do not avoid existing bias in mental health detection and care. Further use of these technologies should carefully consider how to account for and address this bias to improve upon existing mental health care disparities.

Duke Scholars

Author Fatima Fairfax  

Author Julius Middleton Wilder Medicine, Gastroenterology