The problem it solves
In a retrospective study, most capture time goes not into typing but into reading: opening two hundred reports, hunting for a stage, a date, a complication, and copying it across. The work is slow, repetitive, and prone to attention error rather than competence error.
A language model reads fast and does not tire. It carries a structural flaw in exchange: it produces a plausible answer even when the information is absent. In a clinical research task, an invented value that resembles a correct one is worse than no value at all, because it is undetectable downstream.
The three non-negotiable constraints
1. Mandatory citation
The model must not be allowed to answer "stage T2". It must answer "stage T2" together with the exact sentence from the document it draws it from. If no sentence can be quoted, the proposal is discarded rather than kept with low confidence.
This constraint changes the nature of the exercise. It does not remove model error — it can quote a sentence that does not say what it claims — but it makes the error verifiable in seconds by a reviewer, instead of requiring the whole document to be reread.
2. Human validation, without exception
The model proposes; it does not decide. A model-origin value stays a proposal until confirmed by an identified person. That person sees the value, the quoted sentence, and the context, and adjudicates.
3. Pseudonymisation before any transmission
Direct identifiers — name, record number, contact details, precise dates of birth — must be removed from the text before it leaves the institution's infrastructure. See pseudonymisation for what that operation guarantees, and what it does not.
Human in the loop: what the phrase covers
The phrase *human-in-the-loop* is used loosely for very different arrangements. Three levels deserve distinguishing, because they do not offer the same guarantees.
| Level | What the human does | What it is worth |
|---|---|---|
| Oversight | Inspects a sample after the fact | Weak — unsampled error passes through |
| Validation | Confirms every value before it becomes final | Strong — no value goes unseen |
| Adjudication | Resolves cases the model flags as ambiguous | Complementary — does not replace validation |
Only the second level supports claiming, in a methods section, that every value was verified. The other two describe partial controls and must be described as such.
What to report in a publication
Using a model in capture is a methodological characteristic, like a measuring instrument. Concealing it is an omission; describing it vaguely is one too. STROBE expects a description of how variables were measured, and this falls under that.
- Which variables were model-extracted, and which were not.
- The proportion of values of model, rule and human origin.
- The validation procedure: who reviewed, on what material, and by what rejection criterion.
- The rejection rate at review — an indicator of quality, not a confession.
- The pseudonymisation procedure applied before any external transmission.
How AnzarSeha addresses it
- A model-proposed value must carry the source sentence; without a citation it is discarded before reaching the reviewer.
- No model-origin value becomes final without human confirmation, record by record.
- Direct identifiers are stripped from the text before any external call, and a logged consent gate precedes every request.
- With no recorded consent decision, only the deterministic rule engine runs — the study remains possible, without a model.
- The origin of every value — registry, rule, model, verified — stays attached to it all the way into the export.
What the method does not allow
- It does not excuse you from reading. It turns exhaustive reading into targeted verification, which is a real gain and not an abolition.
- It does not invent absent information. If the stage is not in the record, no model will produce it — and if one does, the citation constraint must reject it.
- It is not suitable for the primary endpoint without reinforced review. The variable carrying the conclusion deserves double reading, whatever its origin.
- It does not settle the regulatory question. Sending clinical text, even pseudonymised, to an external service requires a compliance analysis specific to each institution — GDPR and, in Morocco, law 09-08 / CNDP.
Frequent questions
Does the model see patient names?
It must not. Direct identifiers — name, record number, contact details — are stripped from the text before any external call. That does not make the text anonymous: a detailed report can remain re-identifiable by cross-referencing, which is covered in pseudonymisation.
Can review be skipped if the model is good?
No, and better models do not change this. What is at stake is not the error rate but accountability: a value no identified person validated cannot be defended before a reviewer, whatever rate was measured.
What if the model and the reviewer disagree?
The reviewer decides, and the decision is recorded with its justification. The disagreement is itself useful information: a rising rejection rate on one variable usually signals a problem with that variable's definition, not with the model.
Will an ethics committee accept AI extraction?
That depends on the committee, the data and the infrastructure, and cannot be prejudged here. What helps a submission is describing the arrangement precisely: what is sent, after what processing, to which destination, and how each value is validated.