Provenance

The record of where an observation came from. The Exam Finding List specification states the requirement directly: findings “can be generated by LLM from an existing report or generated live during exam time; each Observation should include some kind of provenance marker.”1

Provenance matters because sources are not interchangeable. The deck lists AI validation as a stated ACR priority, specifically the ability to “correlate AI vs radiologist Observations,” which is only possible if the two are distinguishable.2 The lineage FHIR samples make the distinction with FHIR status: an AI-generated finding carries preliminary and the radiologist’s carries final, with otherwise identical structure.3

A second kind of provenance operates within a report. Every extracted finding carries a verbatim report_text quote, validated as verbatim rather than paraphrased, so a consumer can see the sentence the structure came from.4 That quote is what the reviewing tools display and what the evaluation design keys on.

Synonyms and near-synonyms

  • Provenance marker is the specification’s phrase for the field.
  • Source and attribution are informal equivalents.
  • status in FHIR is the field the lineage samples use to carry it.
  • Verbatim quote is evidence for one observation, not a statement of who produced it.
  • Crosswalk provenance is a different concept: mapping direction, scope, and strength when translating codes between terminologies.5

Identifier form

None. No provenance field is specified beyond the requirement that one exist.

Where it is used

Exam Finding List, Observation, and Report extraction platform.

Conflicts

There is no agreed provenance field. The specification requires “some kind of provenance marker” without naming one, the sample Exam Finding Lists have no provenance field at all, and the lineage samples overload FHIR status for the purpose. A formal provenance model is not covered by any current plan.

Footnotes

  1. imaging-problem-list README ↩

  2. Open Imaging Data Model 2026 Status Update, January 2026 ↩

  3. AI-generated finding Observation ↩

  4. Initial extraction plan ↩

  5. med-ontology-lookup product roadmap ↩