Does Patient Insurance Influence ChatGPT’s Recommendations? A Closer Look at AI Bias in Interventional Radiology

Monday, March 31, 2025

3:36 PM – 3:45 PM CT

Location: Music City Center, 205 A

Presenting Author(s)

Nick Dietrich, MS

Graduate Student
Institute of Biomedical Engineering, University of Toronto, Canada

Author/Co-author(s)

BS

Brett Stubbert, BS

Medical Trainee
Schulich School of Medicine & Dentistry, Western University, Canada

Purpose:
Patients are increasingly relying on large language models (LLMs) like ChatGPT for medical guidance. However, sociodemographic biases in these tools have become a concern. Insurance coverage, a key factor in healthcare decisions, may influence ChatGPT’s recommendations, but has not yet been investigated. This study aimed to evaluate the impact of different insurance types on ChatGPT's treatment recommendations for common conditions treated by interventional radiology (IR).

Materials and Methods:
Eight conditions commonly treated by IR were identified using clinical prevalence data. A standardized prompt was developed: "I have [condition] and I have [private/public/no insurance]. Which physician(s) should I seek for treatment?" and inputted into a custom data pipeline that used GPT-4o to generate outputs. To account for model randomness, each prompt was repeated 10 times, and the mean frequency of IR recommendations was calculated. A one-way ANOVA, followed by post-hoc analysis, was conducted to determine differences in outputs generated for each condition across insurance types.

Results:
ChatGPT recommended IR consultation in 174 out of 240 outputs (72.5%). Table 1 shows the frequency counts for each condition. ANOVA revealed significant differences in recommendations based on insurance type (F(2,237) = 11.97, p < 0.001). Post-hoc Tukey HSD test showed significant differences between private and uninsured groups (mean difference: -6.3, p < 0.001), and between public and uninsured groups (mean difference: -5.0, p = 0.004). However, there was no statistically significant difference between private and public insurance groups (mean difference: -1.3, p = 0.63).

Conclusion:
ChatGPT is less likely to recommend IR care for uninsured patients, suggesting potential bias in AI-driven treatment guidance. Further research should explore how such disparities may impact clinical outcomes as LLMs become more integrated into healthcare decision-making.