AnzarSeha

Data provenance

By Dr Rida Akodad · Publication director · Updated 21 August 2026

The provenance of a data point is the verifiable answer to a simple question: where did this number come from? It names the document, rule or person the value derives from, and keeps that link all the way into the exported dataset.

Definition

Provenance (sometimes called lineage) describes the origin and journey of a value. In clinical research a well-traced value carries four pieces of information: its origin, its justification, the date it was fixed, and who validated it.

The four possible origins

OriginWhat it meansVerifiability
Registry fieldThe value already existed, captured in structured formDirect — trace back to the record
Deterministic ruleComputed: a BMI from weight and height, a stage from criteriaReproducible — the rule can be replayed
Model proposalExtracted from free text by a model, citing its source sentenceIndirect — the quoted sentence remains available
Human decisionA reviewer adjudicated, writing down whyDocumentary — the justification is kept

A value with no declared origin is not thereby wrong. It is simply unverifiable, which — in a study meant for publication — amounts to the same thing once the question is asked.

Provenance versus audit trail

The two notions are often blurred in discussion and are cleanly distinct in practice.

  • Provenance attaches to *the value*: it answers "where did this 31.2 come from?".
  • The audit trail attaches to *the study*: it answers "what happened, and in what order?".

A study can have an impeccable audit log and values with no provenance — the log will say a value was entered at a given time by a given person, without saying where that person got it. The reverse is equally possible. Both are necessary; neither substitutes for the other.

Example

A clavien_max variable (highest postoperative complication on the Clavien-Dindo classification) across a sixty-record cohort:

  • For 41 records the origin is a deterministic rule: no complication recorded in the registry, so the value is 0, computed.
  • For 14 records the origin is a model proposal, with the sentence from the operative report quoted alongside, then confirmed by a reviewer.
  • For 5 records the origin is a human decision: the report was ambiguous, and the reviewer adjudicated, writing the justification.

That distribution is itself methodological information. It can be published, it can be audited, and it lets a reviewer concentrate attention on the nineteen records that required judgement.

Related standards

ICH E6(R3) establishes the principle that reported data must be reconcilable with its source. The FAIR principles treat provenance as first-rank metadata, a condition of reuse.

Sources and standards