The note is the unit of work.

Clinical documentation · Research direction

The problem

A clinician finishes a visit and then spends a second shift writing it up. The note feeds billing, handover, audit and the next visit. It is also where most errors hide, because it is written late and from memory.

The question

Can a system draft a note from what was actually said and done, so that a clinician reviews and signs instead of authoring from scratch, without ever inventing a finding?

The note is the unit of work

On why documentation, not diagnosis, is where software helps first.

Ask a clinician what they would give up first and many will say the typing. The visit is the work they trained for. The note is the work that follows them home.

A clinical note does five jobs at once. It is a memory aid for the writer. It is a message to the next clinician. It is the basis for what gets billed. It is a legal record. And it is, increasingly, the raw material that other systems read. Most tools serve one of these jobs and damage another.

Drafting a note from a conversation looks easy because the output is fluent. That fluency is the danger. A model can write a convincing sentence about a symptom nobody mentioned. In a chat window that is an embarrassment. In a chart it is a false record that someone may act on next month.

So the first design decision is provenance. Every sentence a system proposes should point at where it came from: a moment in the audio, a line in the chart, an order that was placed. If it cannot point, it should not be written, or it should be shown to the clinician as a question and not as a fact. The reviewer then does the job people are actually good at, which is noticing what is wrong, and does not have to re-author what is right.

The second decision is how to measure. Counting how much a draft was edited rewards drafts that are short and bland. We would rather count the kinds of mistakes. Did the draft invent something. Did it drop something that mattered. Did it put a true thing in the wrong place. Those are different failures and they need different fixes.

The third decision is who holds the pen. The clinician signs. The system proposes. That boundary should be visible in the interface and in the log, so that nobody, including the clinician on a tired Friday, can mistake a draft for a decision.

We have not yet run this with clinicians. What we can publish today is the method: how we would test it, what we would count, and what we would refuse to claim.

Test design

• Every sentence in a draft traces to a source span in the transcript, the chart or an order. A sentence with no source is flagged, not smoothed over.
• Omissions are measured as seriously as inventions. A note that drops the allergy discussion fails.
• Edit distance is a weak signal. We care which edits clinicians make and why.

Protocol: Source-traced drafting test