Sociodemographic Variability in Pediatric Emergency Decisions by AI.

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems
• Pediatric Medicine

Background:

• Large language models (LLMs) offer potential for clinical decision-making.
• Sociodemographic disparities in LLM recommendations for pediatric care are not well understood.
• Ensuring equitable AI in healthcare is critical.

Purpose of the Study:

• To investigate sociodemographic variations in LLM clinical recommendations for pediatric cases.
• To identify potential biases in LLM outputs related to patient and caregiver demographics.
• To inform the safe and equitable deployment of LLMs in pediatric settings.

Main Methods:

• Analysis of an ensemble of 10 large language models.
• Evaluation using 500 standardized and 500 real-world pediatric clinical scenarios.
• Assessment of over 3.7 million model-generated recommendations.

Main Results:

• LLMs exhibited significant recommendation deviations for cases with socioeconomic adversity (e.g., low income, unstable housing).
• Disparities were amplified for intersectional groups, notably Black children, with increased recommendations for urgent interventions and maltreatment suspicion.
• LLMs showed sensitivity to caregiver demographics, mirroring trends observed for child factors.

Conclusions:

• LLMs demonstrate differential sensitivity to sociodemographic factors, necessitating further research to distinguish clinical sensitivity from bias.
• Caution is recommended when interpreting LLM outputs incorporating sociodemographic data, especially with limited clinical information.
• Integrating safeguards and developing context-specific models may mitigate biases for safer pediatric care.