How FHIR R4 Graph Structure Shapes ML Pipeline Design for Patient Records

A machine learning engineer building a retrieval-augmented generation system over FHIR R4 patient records found that understanding the data's graph structure — not just retrieval mechanics — was the core design challenge. FHIR records are structured as Bundles containing independent, cross-referencing resource objects rather than narrative documents, meaning standard document-chunking approaches can produce technically functional but clinically incorrect results. Of FHIR R4's 140-plus resource types, roughly ten — including Condition, Observation, MedicationRequest, and Encounter — account for the vast majority of clinically relevant question-answering. In a 78-patient synthetic corpus, Observations outnumbered Conditions by more than 15 to 1, creating a retrieval bias toward lab values that requires deliberate mitigation. A key structural detail engineers must internalize is the CodeableConcept, where clinical codes are arrays that may reference multiple terminology systems simultaneously, making assumptions like always taking the first coding entry unreliable.