Sociodemographic bias in LLMs' clinical decision-making for dizziness

Area of Science:

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

Background:

• Large language models (LLMs) are increasingly used in clinical decision support.
• Concerns exist regarding potential sociodemographic bias in LLM outputs.
• Assessing bias in LLM recommendations for dizziness is crucial for equitable healthcare.

Purpose of the Study:

• To evaluate if LLM recommendations for dizziness vary based on patient descriptors and clinical detail.
• To quantify sociodemographic bias in LLM-generated clinical decisions.
• To explore methods for mitigating bias in AI-driven clinical support.

Main Methods:

• A cross-randomized in-silico vignette study using 100 synthetic emergency department dizziness cases.
• Vignettes were presented in neutral form and with 33 sociodemographic descriptor variants.
• Twelve instruction-tuned LLMs evaluated cases, answering five clinical decision questions, with 2,040,000 responses generated.

Main Results:

• Sociodemographic descriptors significantly influenced LLM recommendations, particularly for mental health referrals in ambiguous cases.
• Referral likelihood decreased for specific groups, including Black transgender women and Black patients experiencing homelessness.
• Increased neuroimaging recommendations were noted for low-income descriptors, though effects were smaller.

Conclusions:

• LLM clinical recommendations demonstrate variability based on sociodemographic descriptors, especially under diagnostic uncertainty.
• Incorporating more detailed clinical information can attenuate these disparities.
• Structured inputs for AI systems show promise in mitigating bias in clinical AI.