Is it possible to vaccinate AI against bias? An exploratory study in epilepsy

Area of Science:

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

Background:

• Large language models (LLMs) are increasingly utilized for clinical decision support.
• Concerns exist regarding their potential to perpetuate socioeconomic biases.
• The efficacy of simple prompt-based interventions to mitigate these biases is not well-established.

Purpose of the Study:

• To evaluate if a prompt-based 'inoculation' can reduce bias and enhance accuracy in LLM recommendations.
• To assess the impact of instructing LLMs to disregard clinically irrelevant socioeconomic information.

Main Methods:

• An experimental study using publicly available LLM interfaces (memory disabled).
• Two fictional epilepsy vignettes (diagnostic and therapeutic) with varied socioeconomic (SES) descriptors were used.
• Each vignette was presented 10 times per condition to 6 frontier LLMs, generating 480 responses.
• Accuracy and bias were assessed by comparing responses to evidence-based guidelines.

Main Results:

• Base diagnostic accuracy was 36% with a 45-percentage point SES bias gap; inoculation improved accuracy to 55% and reduced bias to 27 percentage points.
• Base treatment accuracy was 51% with a 25-percentage point SES bias gap; inoculation improved accuracy to 63% and reduced bias to 8 percentage points.
• Intervention effects varied significantly across LLMs, with some showing bias elimination and others paradoxical worsening.

Conclusions:

• A simple prompt-based intervention can reduce socioeconomic bias and improve accuracy in LLM clinical recommendations.
• Prompt engineering offers a practical strategy for mitigating AI bias in healthcare.
• Model-specific performance variations necessitate ongoing oversight and complementary bias-mitigation strategies.