EntSun News

Perch Publishes Open Test Set That Catches AI Scribes Inventing Drug Names

EntSun News/11105629
The public test set puts Perch's own results next to a commercial veterinary AI scribe's, on the same clips, including a case where the other product charted a drug that does not exist.

NEW YORK - EntSun -- Perch, the veterinary AI scribe built by Talcoe LLC, today published an open test set against itself, built to catch one specific failure: an AI scribe writing a medication into a chart that nobody said. The test set includes Perch's own results next to a miss from a commercial veterinary AI scribe run on the same clips.

The test set holds 70 short clinical dictations: 45 ordinary cases and 25 traps written to bait a false drug name. The dictations carry 84 drug mentions. Talcoe ran the same 70 clips through Perch and through a commercial veterinary AI scribe.

In one clip, the other product in the test charted Roflexacin, a drug that does not exist, next to the Baytril that was actually dictated.

More on EntSun News
An earlier Perch build had its own miss, on a separate test set: it charted Zycortal, a real but wrong drug, on one of its own trap clips. The build running today declines to chart anything on that clip rather than guess. The other product's repair layer also recovered atenolol from a badly garbled dictation, on a clip where one of Perch's two speech configurations missed it.

Across the full run, Perch invented zero drug names on its side of the clips.

One limit applies. The audio is synthetic text to speech voices, one voice per clip, not a real clinician in a real exam room. This is Perch's own test, and real speech runs harder than synthetic speech. Talcoe is publishing it as a reproducible, shared test set, not an accuracy claim about any product.

The clips, written dictations and scoring method are posted at https://github.com/perchvet/perch-trapset. A longer account of the run is posted at https://perch.talcoe.com/news/invented-drug-names-test.

More on EntSun News
"We built this test to check whether an AI scribe can put a drug into a medical record that nobody said, and we ran it on our own product first," said Mohammed Usama, cofounder of Talcoe LLC. "The clips, the written dictations, and the scoring method are all public, so a vet or a reporter can check the finding rather than take our word for it."

About Perch: Perch is a veterinary AI scribe built by a two person team at Talcoe LLC in New York. A vet talks through a visit, and Perch drafts the note in SOAP, DAP or BIRP format while the vet is still speaking. It runs in the browser, works alongside AVImark, and is free to start with a 30 day pilot.

Contact
Mohammed Usama, Cofounder, Talcoe LLC
***@talcoe.com


Source: Talcoe LLC

Show All News | Disclaimer | Report Violation

0 Comments

Latest on EntSun News