Conceptualizing bias in EHR data: A case study in performance disparities by demographic subgroups for a pediatric obesity incidence classifier

Editorial reading guide · Respect

What should we ask when an AI tool is said to be fair for children?

Some models performed slightly better for Black children and Medicaid patients, contrary to a simple expectation that underrepresented groups always fare worse. Predictive comorbidities and unequal case distributions helped explain this pattern.

Study context

Researchers used Children’s Hospital of Philadelphia records of newly obese children and matched controls to compare four classifiers and demographic performance patterns.

Age / stage: Pediatric patients in a US hospital system; exact age span not specified in this curation

Read with three questions

1. How did subgroup performance differ from the usual assumption about disadvantage?
2. Which recorded conditions drove the classifications?
3. Why does a descriptive association fail to explain the causes of an individual child’s weight?

Keep the limits in view

The study describes associations rather than causes and depends on one US health system’s records. Different disease prevalence, documentation, features, and geographic settings could change the direction of disparities, and restricted data limit reproduction.