Sociodemographic Variability in Pediatric Emergency Decisions by AI

Large language models (LLMs) have the potential to support clinical decision-making in pediatric settings. However, whether they exhibit sociodemographic differences in clinical recommendations for similar clinical presentations is unknown.

We analyzed sociodemographic variations in pediatric emergency recommendations from an ensemble of 10 LLMs, evaluating 500 validated standardized cases and 500 real clinical scenarios, totaling more than 3.7 million model outputs.

Significant deviations emerged, particularly for cases labeled with socioeconomic adversity, such as unstable housing or low family income. Although increased vigilance toward certain risk factors might be clinically reasonable, the magnitude and consistency of model recommendations were notably high compared with the physician-derived ground truth, especially for low-income and immigrant groups. Intersectionality involving Black race consistently intensified these differences. For example, cases labeled Black unhoused received substantially higher recommendations for urgent interventions (+10.5 percentage points [pp]; adjusted P < .001), additional investigations (+14.1 pp; adjusted P < .001), and suspicion of maltreatment (+26.6 pp; adjusted P < .001), even without clinical justification, compared with white or high-income cases. The LLMs also demonstrated clinical sensitivity to caregiver demographics, as expected. However, caregiver factors were associated with different recommendation patterns to a slightly lesser degree yet still showed significant variations and similar trends as child factors.

This suggests the models demonstrate differential sensitivity to sociodemographic factors that warrants further investigation to distinguish appropriate clinical sensitivity from potential bias. We suggest caution when interpreting LLM recommendations that incorporate sociodemographic identifiers, especially when based on limited early clinical information. Integrating explicit guideline-based safeguards and developing smaller, context-specific models may reduce these biases, ensuring safe and clinically appropriate pediatric care.