Investigation of Bias Using ChatGPT-4 in Pain Management and Patient Profiling
J Emerg Med. 2026 Jan 30;84:28-36. doi: 10.1016/j.jemermed.2026.01.011. Online ahead of print.
ABSTRACT
BACKGROUND: Large language models, such as ChatGPT (cGPT), are being integrated increasingly into clinical workflows and medical education. However, concerns persist regarding their susceptibility to bias, especially in high-stakes areas like pain management, when disparities across race, socioeconomic status, and substance use history are well documented.
OBJECTIVE: This study investigated whether cGPT-4 generates consistent and equitable pain management recommendations when patient demographic variables are altered.
METHODS: Using cGPT-4, researchers created six clinical scenarios representing common pain complaints (e.g., migraine, chest pain, and deep vein thrombosis). Each scenario was systematically modified to reflect diverse patient demographic characteristics, including race, housing status, language proficiency, and history of opioid use disorder. Three investigators input a total of 60 prompts into cGPT-4 and compared outputs for agreement using Fleiss' κ and Gwet's AC1 statistics.
RESULTS: Overall agreement across investigators was high (i.e., 82% for emergency department (ED) medications and 78% for discharge medications). When using cGPT, demographic factors such as race, language, and socioeconomic status often did not alter recommendations. However, patients with history of opioid use disorder consistently received different pain management suggestions-typically opioid-sparing regimens-indicating cGPT's responsiveness to clinically relevant safety concerns. Scenario-specific variation was observed, particularly in cases of migraine and sciatica.
CONCLUSIONS: cGPT-4 often produced consistent and equitable pain management plans across diverse patient profiles. Although bias was observed in opioid use disorder-related scenarios, it seemed aligned with clinical best practices. These findings suggest that, when properly monitored and refined, large language models can support equitable decision making in health care. Continued evaluation and prompt engineering are critical to minimizing unintended bias and maximizing utility in clinical settings.
PMID:41875512 | DOI:10.1016/j.jemermed.2026.01.011