Detecting Sociodemographic Biases in the Content and Quality of Large Language Model-Generated Nursing Care:...

Area of Science:

• Artificial Intelligence in Healthcare
• Nursing Informatics
• Health Equity Research

Background:

• Concerns exist regarding sociodemographic bias in large language model (LLM)-generated nursing care plans.
• Lack of empirical evidence evaluating bias in LLM nursing recommendations.

Purpose of the Study:

• Investigate sociodemographic biases in LLM-generated nursing care plans.
• Assess implications for equitable nursing care.

Main Methods:

• Mixed methods simulation study using GPT-4 to generate 9600 nursing care plans.
• Quantitative analysis of plan content and expert evaluation of clinical quality for 500 plans.
• Varying patient profiles based on sex, age, income, education, and residence.

Main Results:

• LLM care plans showed systematic sociodemographic disparities in thematic content and clinical quality.
• Low income and education were associated with less favorable plan characteristics.
• Older patient profiles received more pain management but less nurse training recommendations.
• Expert review indicated high overall quality but noted differences in completeness, applicability, and safety based on urban residence.

Conclusions:

• LLMs reproduce sociodemographic biases in nursing care plans, risking reinforcement of health inequities.
• Empirical evidence demonstrates nuanced biases in LLM-generated nursing care plans.
• Highlights the need for human oversight to ensure AI advances equity in nursing.