From adults to children
The team's earlier work found adult emergency recommendations shifting with a patient's label. This study asked the same question in pediatric emergencies, where a caregiver usually speaks for the child. An ensemble of ten models read 500 validated standardized cases and 500 real clinical scenarios, with sociodemographic labels attached to the child or to the caregiver, and its recommendations were compared with a physician-derived ground truth.
points in suspicion of maltreatment for cases labelled Black and unhoused
points in additional investigations for the same label
points in urgent interventions, each against white or high-income cases
Same case, different care
The largest deviations came with labels of socioeconomic adversity, such as unstable housing or low family income, and intersection with Black race intensified them. Cases labelled Black and unhoused received more recommendations for urgent interventions (+10.5 percentage points), additional investigations (+14.1) and suspicion of maltreatment (+26.6) than white or high-income cases, each with adjusted P < .001 and without clinical justification.
Some extra vigilance toward real risk factors can be clinically reasonable. The magnitude and consistency of the models' recommendations still ran well above the physicians' ground truth, especially for low-income and immigrant groups.
Suspicion of maltreatment: recommended 26.6 percentage points more often for cases labelled Black and unhoused than for white or high-income cases (adjusted P < .001), even without clinical justification.
How many percentage points more often the ten-model ensemble recommended each step for cases labelled Black and unhoused than for white or high-income cases; adjusted P < .001 for each. The abstract gives the differences only, not the rates behind them, and does not say whether the label sat on the child or on the caregiver. Tap or hover a row for its reading. Data from the abstract of the paper.
The caregiver's label
The models were also sensitive to the caregiver's demographics, as expected. Caregiver labels moved the child's recommendations slightly less than the child's own labels, yet still with significant variations and in the same directions.
The label describes the child. Labels of socioeconomic adversity, such as unstable housing or low family income, moved the recommendations most.
A label on the caregiver moved the child's care in the same direction as a label on the child.
Bias, vigilance, or both
The design cannot fully separate appropriate clinical sensitivity from bias, and the authors call for further work to tell them apart. They advise caution with model recommendations that draw on sociodemographic identifiers, especially when they rest on limited early clinical information, and suggest guideline-based safeguards and smaller, context-specific models.
A child's label, or the caregiver's, can change the care a model recommends.
Across more than 3.7 million outputs from ten models on 1,000 pediatric emergency cases, sociodemographic labels shifted recommendations beyond the physicians' ground truth, most of all for socioeconomic adversity, and intersection with Black race intensified the shifts. Clinicians using these tools in pediatric care should be cautious with sociodemographic content in prompts.
Mahmud Omar, Reem Agbareia, Razi Abu Salah, Nicola Luigi Bragazzi, Donald U. Apakama, Benjamin S. Glicksberg, Bruce D. Gelb, Emma Holmes, Girish N. Nadkarni, Eyal Klang
Icahn School of Medicine at Mount Sinai, Hadassah Medical Center, LMU Munich
Pediatrics 157(6), e2025073479 (2026)