Large language models such as GPT-4 are seen as promising tools in health care, from automating administrative tasks to supporting clinical decisions. However, they also risk perpetuating biases and giving incorrect diagnoses, with a direct impact on patient care. This study assessed whether GPT-4 shows racial and gender biases in four possible uses: medical education, diagnostic reasoning, treatment planning and assessments of patients.
The researchers used clinical cases from medical education resources and from published research on implicit bias in health care. When asked to create clinical cases, GPT-4 did not reflect the real demographic diversity of medical conditions, and consistently produced cases that stereotyped how conditions present in different groups. For the same clinical cases, its lists of possible diagnoses were more likely to include diagnoses that stereotype certain races, ethnicities and genders.
The treatment plans and assessments it produced also varied with patients’ demographic characteristics, including recommendations for more expensive procedures and differences in how patients were perceived. The authors conclude that tools like GPT-4 urgently need thorough and transparent bias assessments for each intended use before they are introduced into clinical care.
Learn more about this study here: https://doi.org/10.1016/S2589-7500(23)00225-X
Reference:
Zack, T., Lehman, E., Suzgun, M., Rodriguez, J.A., Celi, L.A., Gichoya, J., Jurafsky, D., Szolovits, P., Bates, D.W., Abdulnour, R.-E.E., Butte, A.J., & Alsentzer, E. (2024). Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study. The Lancet Digital Health, 6(1), e12–e22.