BRENNEN

Industries / Health / Veterinary

Exams and diagnosis

Examining an animal, judging its pain and weight, and reasoning from its signs to a diagnosis, in person or by video.

5

Milestones achieved, from 1987 to 2023.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1987

    A computer lists possible diagnoses

    Maurice White of Cornell University published the full list of signs and diseases behind Consultant, a computer database that takes the signs a vet enters and lists the diseases that could cause them, for dogs, cats, horses, cattle, sheep, goats and pigs. Cornell still offers it free online as Dr. White's Cornell Consultant.

    vet.cornell.eduebi.ac.uk

    Credit

    Maurice E. White, Cornell University College of Veterinary Medicine

  2. 2004

    Body fat scored by hand

    A University of Tennessee team measured body fat in 23 healthy pet dogs with a DEXA scan, a low-dose X-ray scan that measures fat, and compared it with a body condition score, a rating a vet gives by looking at and feeling the dog. The score tracked the scan closely, with an r² of 0.92 by the authors' figures (1 would be a perfect match), so a rating that needs no equipment can stand in for a scan.

    ebi.ac.uk

    Credit

    D. I. Mawby, J. W. Bartges, D. P. Laflamme and colleagues (Journal of the American Animal Hospital Association, 2004)

  3. 2007

    A pain score for dogs

    The short form of the Glasgow Composite Measure Pain Scale rates six behaviors, from crying out and attention to the wound to response to touch and posture, for a maximum of 24 points. Vets at Glasgow, University College Dublin and North Carolina scored 122 dogs after surgery with it, and 6 out of 24 became the point at which to give pain relief. A 2021 review found it in 114 published studies.

    Lessonthe same review found studies that changed the scale or its cut-off and many that reported their design poorly, which makes their results hard to compare.

    api.crossref.orgapi.crossref.org

    Credit

    J. Reid, A. M. Nolan, J. M. L. Hughes, D. Lascelles, P. Pawson and E. M. Scott (Animal Welfare, 2007); review by Testa and colleagues (Frontiers in Veterinary Science, 2021)

  4. 2019

    Reading pain on cats' faces

    The Feline Grimace Scale scores five features of a cat's face, such as ear position and tightening around the eyes, from 0 to 2 each. In the study that built it, with 35 client-owned cats and 20 controls, observers agreed well (an intraclass correlation of 0.89, where 1 is perfect agreement), and a total above 0.39 out of 1 picked out pain with 90.7% sensitivity and 86.6% specificity. By 2023 the scale had a phone app that supplied 1,092 of the images the same group used to train software to score it, which reached 95.5% accuracy in the group's own test.

    pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov

    Credit

    M. C. Evangelista, P. V. Steagall and colleagues, Université de Montréal

  5. 2023

    Prescribing after a video call

    New "under care" guidance from the UK's Royal College of Veterinary Surgeons (RCVS) came into force on 1 September 2023. Whether a vet must examine an animal in person before prescribing a prescription-only medicine is now left to the vet's judgment, as long as the practice can examine the animal in person at any hour. For pets, a physical exam is still required in all but exceptional cases before prescribing antibiotics, antifungals, antiparasitics or antivirals, and the antiparasitic part was delayed to 12 January 2024.

    rcvs.org.ukvettimes.com

    Credit

    Royal College of Veterinary Surgeons (RCVS)

Watch for

  1. AI vet chat ranks first

    Vet-AI, the company behind the Joii Pet Care app, says UK vets rating answers blind gave its own language model 81% for clinical accuracy on 48 simulated pet-owner conversations, against 69% for Gemini and 50% for ChatGPT. The company does not name the vets or define how accuracy was scored.

    Achieved whenthe conversations, scoring guide and reviewers are published and a group with no tie to Vet-AI reruns the comparison.

    vetclick.com

    Claimed by

    Vet-AI, September 2025

  2. Full marks on own questions

    OpenVet says its AI answered all 600 multiple-choice questions on a benchmark it built from public veterinary literature and AI-generated questions, including some with radiographs. In the same release it states that it did not take the North American Veterinary Licensing Examination (NAVLE) and that earlier wording suggesting otherwise was imprecise.

    Achieved whenthe question set is published and an outside group reproduces the score, or the system is scored on an exam set by an outside body.

    einpresswire.com

    Claimed by

    OpenVet, April 2026

Open

  1. 01

    Test pet symptom checkers

    A public set of at least 300 short cases about sick or injured dogs and cats, written the way an owner would describe them, each labeled with an urgency level agreed by at least three emergency vets: go now, see a vet within 24 hours, or watch at home. A test harness runs every case through free symptom checkers and chat tools, such as PetMD's checker, VetHelpDirect, Petriage, Joii and general AI chat models, and publishes how often each one gives the agreed urgency and how often it sends a "go now" case home.

    The brief
    Credit · unclaimed
  2. 02

    Score AI on case reports

    A set of 300 cases taken from open-access veterinary case reports, each cut off before the diagnosis so that it holds only the history, exam findings and test results a vet would have had. Ask AI models for a ranked list of likely diagnoses and the next tests to run, and score whether the confirmed diagnosis appears in the top five. Use reports published after each model's training cutoff, so the model cannot have read the answer.

    The brief
    Credit · unclaimed

Open 01 · the brief

Test pet symptom checkers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of at least 300 short cases about sick or injured dogs and cats, written the way an owner would describe them, each labeled with an urgency level agreed by at least three emergency vets: go now, see a vet within 24 hours, or watch at home. A test harness runs every case through free symptom checkers and chat tools, such as PetMD's checker, VetHelpDirect, Petriage, Joii and general AI chat models, and publishes how often each one gives the agreed urgency and how often it sends a "go now" case home.

Why it is still open

A 2020 review found no peer-reviewed studies of animal symptom checkers. The one outside study found since, published in 2026, ran a single chatbot once on each of 13 owner-style scenarios, and its authors say the result should not be generalized. The only comparison across several tools was run by Vet-AI, which sells one of them.

ajol.infothesai.org

Built with what exists

cases written from textbook presentations and from the leading causes in the VetCOT trauma registry (see Emergency care); the CHART reporting guideline for studies of chatbot health advice; model APIs and open-weight models such as Mistral Large 4 (see the Note "A trillion open parameters"); a model that returns probabilities for fixed answers, which suits three urgency levels (see the Note "A thousand questions, one pass"); Cohen's kappa, a standard measure of how well raters agree beyond chance; a Python test harness.

  1. First step

    Write 50 cases covering bite wounds, vehicle strikes, common poisonings and signs vets treat as emergencies, such as a male cat straining to pass urine, and ask three emergency vets to label them on their own.

  2. Done looks like

    300 labeled cases released under an open license, the vets' agreement with each other reported, and at least six tools scored on how often they under-call and over-call urgency, rerun every six months.

Open 02 · the brief

Score AI on case reports

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A set of 300 cases taken from open-access veterinary case reports, each cut off before the diagnosis so that it holds only the history, exam findings and test results a vet would have had. Ask AI models for a ranked list of likely diagnoses and the next tests to run, and score whether the confirmed diagnosis appears in the top five. Use reports published after each model's training cutoff, so the model cannot have read the answer.

Why it is still open

Published tests of AI models in veterinary medicine use multiple-choice exam questions, such as the 250 questions from a Hong Kong qualifying exam scored in 2025. A new set, PetQA (September 2026), holds about 18,800 Korean questions on dogs and cats with long answers written by vets, not cases with a confirmed diagnosis. Company claims rest on question sets nobody else can see, and a 2026 review of 98 diagnostic AI studies in dogs and cats found that only 13 shared their data.

api.crossref.orgarxiv.orgresearchsquare.com

Built with what exists

the Europe PMC search API to find case reports under a CC BY license, a script to split each report at the diagnosis, open-weight models such as Mistral Large 4 (see the Note "A trillion open parameters") alongside hosted models, and two vets to check the answer key.

  1. First step

    Pull 500 candidate case reports with a CC BY license from Europe PMC and record the species, publication date and final diagnosis of each.

  2. Done looks like

    300 cases across dogs, cats, horses and cattle, top-one and top-five accuracy for at least five models, the vets' agreement on the answer key, and all prompts, answers and scores published under an open license.