Invisible at the Edges: The Unborn and the Family in Clinical-AI Data

Family medicine takes the family and the life course as its unit of care, yet clinical AI for primary care is trained on coded electronic health record data built around the individual patient per encounter. We asked whether the two edges that family medicine most attends to, the time before birth and the kin alongside the patient, are present in that coded substrate. Using a Synthea cohort of 11,466 patients with SNOMED CT conditions and FHIR exports, where the generator knows every familial relationship (ground truth), we checked whether the unborn has its own record (E1), whether any mother-child link exists (E2), whether the family or household is represented (E3), and how kinship and identity change are stored (E4). Of 11,466 patients, 3,400 (29.7%) carried obstetric coding, yet the fetus had no record of its own, appearing only on the mother's chart (83 instances of "Fetus with chromosomal abnormality"). No mother-child link existed (0 patient links, 0 kinship resources); the family and household were absent (0 household resources); family history survived only disguised as the patient's own disease (60 records of "Familial Alzheimer's disease", 0 FamilyMemberHistory resources); and kinship survived only as a footnote (every patient stored a parent as a free-text string, none as a linked person; 27.5% carried an unresolved changed name). Absences replicated exactly in a seeded n=1,462 subsample. The blind spot is a property of the encounter-and-individual data model, not of incomplete recording; the open real-world data needed to measure it does not exist.