Office visits
98.1%
Matched the visit level the practice billed, on 945 charts with 376 held out.
Read the benchmark →CodeSight coding engine · August 2026
New model. New note type. New record.
We handed a brand-new CodeSightTM model a note type it had never seen. It came back with the highest accuracy we’ve ever published.
New result
99.2%
accurate
on a held-out sample, scored exactly once · August 2026
1,287
diagnostic ultrasound studies benchmarked in total
A new model
This result comes from a new, stronger CodeSight model making its public debut.
A new kind of note
Ultrasound reports are technical imaging documents — nothing like an office-visit chart. This was CodeSight’s first time coding them.
What “matched” means
Agreement with the codes the practice actually submitted — the same decision a working coder makes on that study, every time.
What was measured
- 1,287 studies Diagnostic ultrasound, from one de-identified outpatient vascular practice.
- Billed claims as ground truth “Matched” = the same codes the practice billed — not an independent judgement that those codes were correct.
- Locked before development The 99.2% comes from a held-out sample, sealed up front and scored exactly once, at the end.
The trajectory
Each benchmark builds on the last — same rules, published in full, every time.
Diagnostic ultrasound
99.2%
Matched the billed codes on a note type the model had never seen before — this page.
Procedures
A larger, substantially harder benchmark covering operative notes is underway. Same rules. No figures until it’s done.
Take it with you
The two-page flyer — the result, the trajectory, and what “matched” means — sized for print and sharing.
Method & limitations
- One practice, one specialty. De-identified claims and documents from a single outpatient vascular practice. Results may not transfer to another practice or another specialty.
- Development data is not an unbiased estimate. Most of the 1,287 studies were used to build and correct the system. Only a held-out sample, scored once against a configuration locked in advance, produced the 99.2% figure quoted above.
- Not peer reviewed. A manuscript is in preparation. The office-visit benchmark linked above is published with its full method.