Start typing to search this publication.
techcirkle
Open menu

Subscribe to techcirkle

Get new posts delivered straight to your inbox.

Provenance Is the Schema Decision That Unlocks Clinical AI

Clinical schemas that record only whether data exists cannot safely accept AI-generated documentation. Why provenance must be a first-class concept.

techcirkle
5 min read
Cover image for Provenance Is the Schema Decision That Unlocks Clinical AI

Most clinical data models answer one question about any given fact: is it there or not? Patient has a penicillin allergy. Patient takes 10mg of something. Row exists, therefore true.

hero
hero

That model has worked for a long time because it matched reality: a human clinician sat at a keyboard and typed it, so the author was implicit and singular.

Two things broke that assumption at roughly the same time, and both require the same fix.

Break one: data from other organisations

Once records arrive from systems you do not control, "is it there" stops being sufficient.

An incoming record says 10mg. Yours says 20mg. Neither is obviously stale. Your schema has one row for that medication, so something has to give — and every available option is bad:

  • Overwrite — you destroyed clinical information that may have been correct

  • Ignore — you are knowingly out of date

  • Newest wins — clinical timestamps record when something was documented, not when it was true. A note written today may describe a change from last year.

There is no correct resolution available, because the schema cannot represent the actual situation, which is: two sources assert different things and a clinician needs to decide.

Break two: writes from AI systems

Ambient documentation tools listen to a consultation and draft the note. They are the first healthcare AI capability clinicians broadly want, and they are arriving whether or not your schema is ready.

The drafted note is not the valuable part. The valuable part is the structured extraction alongside it — diagnoses, medication changes, orders, follow-up intent.

To accept those writes safely, you have to be able to say: this fact was asserted by a machine, with this confidence, at this time, and no clinician has confirmed it yet.

A schema that records only presence cannot express that. So the fact either goes in indistinguishable from a clinician assertion — which is unsafe, because a language model produces fluent, plausible, confidently wrong output — or it does not go in at all, which means you cannot ship the feature.

Both breaks have the same root cause and the same fix.

alt
alt

Provenance as a first-class concept

Every clinical fact needs to carry, at minimum:

Attribute

Why

source

Which organisation and system asserted this

asserted_at

When the fact was claimed to be true

recorded_at

When it entered the system — often much later

confidence

For machine-derived facts

confirmation_status

Unconfirmed, clinician-confirmed, clinician-rejected

confirmed_by / confirmed_at

Audit trail for the confirmation

The consequential change is not the columns. It is that a clinical fact becomes a set of assertions rather than a single value. One row per active medication becomes many rows, each attributed, with a resolution policy applied at read time.

That is a materially different data model, and it changes every read path in the application. Which is exactly why retrofitting it is one of the more painful migrations in health tech, and why it is worth doing before you strictly need it.

What this buys you

Reconciliation becomes representable. You no longer have to resolve a clinical disagreement silently. Keep both assertions, attribute each, surface the conflict to a clinician where it matters.

AI writes become safe. Machine-extracted facts sit in the same structure, marked unconfirmed. They are visible, queryable, and permanently distinguishable from clinician assertions.

Legacy note extraction becomes viable. This one is worth dwelling on, because the economics changed dramatically.

Converting a decade of narrative clinical notes into coded structured data used to mean clinical coders working by hand — expensive enough that most organisations simply did not, and accepted that historical data was searchable only as text. Language models can now do that extraction at a fraction of the cost, which turns an inert data asset into a usable one.

The risk is equally clear. An extraction error in clinical data is a potential patient safety issue, and language models are exceptionally good at being confidently wrong.

With provenance, the safe pattern is available: extracted facts live in a provisional layer, never silently merged into the authoritative record. Confidence scores drive review queues. Low-confidence extractions go to human coders. Extracted facts stay visibly distinguishable from clinician-asserted facts, permanently, everywhere they surface in the UI.

Without provenance, your only options are a bulk import into the primary record — a liability that will surface at the worst possible time — or not doing it at all.

Implementation notes

A few things worth getting right the first time.

Separate asserted_at from recorded_at everywhere. They diverge routinely in clinical data, and any logic that conflates them will eventually overwrite current information with older facts that happened to be documented more recently.

Make confirmation status explicit rather than inferred. Do not infer confirmation from "a clinician viewed this screen." Viewing is not confirming, and the audit trail needs to reflect an affirmative action.

Surface unconfirmed facts differently in the UI, always. Not just in a review queue — everywhere the fact appears. If an unconfirmed machine-extracted diagnosis renders identically to a clinician-asserted one in the problem list, you have built the unsafe version with extra steps.

Never let a provisional fact become authoritative through inaction. Timeouts that auto-confirm are exactly the wrong pattern here.

Version your resolution policy. How you collapse multiple assertions into a displayed value will change. Record which policy version produced a given view, or you will not be able to explain historical displays during an audit.

Full guide covering FHIR posture, migration sequencing and certification scope: EHR vs EMR: What the Difference Means for Your Build.

Frequently Asked Questions

Why is presence-only modelling insufficient for clinical data?

Because it cannot represent disagreement. When an external record conflicts with yours, a schema holding one value per fact must resolve it silently — by overwriting, ignoring or applying timestamp logic — and every one of those options loses clinically relevant information.

Why can timestamps not resolve conflicting records?

Because clinical timestamps typically record when a fact was documented rather than when it was true. A note written today may describe a medication change from a year ago, so newest-wins logic can overwrite current data with older clinical reality.

What does a clinical fact look like with provenance?

A set of attributed assertions rather than a single value, each carrying its source, assertion time, recording time, confidence where applicable, and confirmation status. A resolution policy collapses them into a displayed value at read time.

How does provenance make AI-generated documentation safe?

It lets machine-extracted facts enter the system marked as unconfirmed and attributed to their source, remaining permanently distinguishable from clinician assertions. Without it, such facts are either indistinguishable from clinician input or cannot be stored at all.

Can legacy free-text notes be safely converted using language models?

Yes, provided extracted facts stay in a provisional layer rather than being merged into the authoritative record, with confidence scores driving human review queues. Bulk importing model output directly into the primary record creates a liability that surfaces later.

Why retrofit provenance before it is needed?

Because it converts each clinical fact from one row to many and changes every read path in the application. Doing it as part of a planned design is straightforward; doing it under pressure to ship an AI feature is one of the harder migrations in health tech.

Subscribe to techcirkle