Investigation of Bias Using ChatGPT-4 in Pain Management and Patient Profiling

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems
• Health Equity

Background:

• Large language models (LLMs) like ChatGPT (cGPT) are increasingly used in clinical settings.
• Concerns exist about LLM bias in pain management due to documented disparities.
• Patient demographics like race, socioeconomic status, and substance use history can influence care.

Purpose of the Study:

• To assess if cGPT-4 generates consistent and equitable pain management recommendations.
• To investigate the impact of altering patient demographic variables on cGPT-4 outputs.
• To evaluate the potential for LLMs in supporting equitable healthcare decisions.

Main Methods:

• Six common pain complaint scenarios were created using cGPT-4.
• Scenarios were systematically modified for diverse demographics (race, housing, language, opioid use disorder history).
• Three investigators inputted 60 prompts, comparing outputs for agreement using Fleiss' κ and Gwet's AC1 statistics.

Main Results:

• High overall agreement among investigators for ED (82%) and discharge (78%) medications.
• Demographic factors (race, language, socioeconomic status) generally did not alter recommendations.
• Patients with opioid use disorder history consistently received opioid-sparing regimens, indicating safety responsiveness.

Conclusions:

• cGPT-4 frequently generated consistent and equitable pain management plans.
• Observed bias in opioid use disorder scenarios aligned with clinical best practices.
• LLMs, with monitoring and refinement, can support equitable healthcare decision-making.
• Continued evaluation and prompt engineering are crucial for minimizing bias and maximizing LLM utility.