Invisible Bias in GPT-4o-mini: Detecting Disparities in AI-Generated Patient Messaging.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Large language models (LLMs) show increasing promise for improving patient care.
• However, biases within AI training data can lead to harmful disparities in medical communications.
• Evaluating these biases is crucial for safe AI implementation in medicine.

Purpose of the Study:

• To develop and propose a novel method for evaluating bias in LLMs.
• To assess AI-generated patient communications for disparities using synthetic patient data.
• To identify specific patient characteristics that may influence AI communication quality.

Main Methods:

• Utilized GPT-4o-mini to generate patient communications from synthetic medical records.
• Systematically created diverse synthetic patient data, including demographic and historical information.
• Employed GPT-4o-mini to score AI-generated communications on empathy, accuracy, clarity, professionalism, respect, and encouragement.
• Analyzed scoring disparities linked to patient attributes to detect potential biases.

Main Results:

• Patient sex and religious preference demonstrated a statistically significant impact on the quality scores of AI-generated communications.
• Identified specific patient history components that correlated with disparities in AI communication.
• The proposed method successfully detected potential biases in LLM-generated patient interactions.

Conclusions:

• The study presents a novel methodology for bias detection in LLMs using synthetic data and AI-based scoring.
• Findings indicate that LLMs like GPT-4o-mini may exhibit biases related to patient demographics and personal history.
• Further research is recommended to refine evaluation criteria and assess a broader range of LLMs for clinical application.