The AI Stethoscope Heard a Murmur. A Fourth-Year Student Heard It Better.

A new NC State study just handed veterinary medicine a reality check on AI-enabled auscultation, and the results are messier than any pitch deck wants you to believe.

Every few months, a new device promises to make auscultation foolproof. Point it at a chest, let the algorithm listen, get an answer. It sounds like the kind of tool that could level the playing field for a nervous new grad standing over a wiggling Chihuahua at 8pm on a Friday. A new study out of North Carolina State University says not so fast.

Researchers put a popular AI-enabled digital stethoscope, the Core 500 from Eko Health, through its paces on 105 real patients: 54 dogs and 51 cats seen at the university teaching hospital between August and December of 2025. Every animal got the full workup, auscultation from a board-certified cardiologist, a cardiology resident, and a fourth-year veterinary student, plus a six-lead ECG and an echocardiogram when warranted. Then the humans and the algorithm were compared head to head.

The dog results looked, at first glance, respectable. The AI caught murmurs with 86.8 percent sensitivity and correctly flagged 33 of 38 true murmur cases. Fourth-year students matched that performance almost exactly. Not beat it. Matched it. A student a few months from graduation was just as reliable as the algorithm trained to replace second-guessing.

Cats told a different story, and not a flattering one for the machine. Of 22 cats with a real murmur, the AI stethoscope found exactly two. That is a 9.1 percent sensitivity rate on an animal that walks into exam rooms every single day. Veterinary students, meanwhile, caught close to two-thirds of those same murmurs using nothing but their ears and their training.

"This device essentially couldn’t find a murmur. A tool vets lean on for reassurance could let real disease go undetected, and you wouldn’t necessarily think about the fact that the diagnostic algorithms are designed based on humans."
  Jake Johnson, DVM, Cardiology Resident, NC State

That last point is the whole story. The AI wasn't trained on cat hearts or dog hearts. It was trained on people, and feline and canine cardiac physiology does not politely translate.

Arrhythmia detection went sideways too. The stethoscope never once called a dog's rhythm normal. Not a single dog, out of 54, got a clean bill of rhythm health from the algorithm. It did correctly catch all six true cases of atrial fibrillation, but it also slapped that same label on 22 dogs who did not have it. Three wrong calls for every one right one.

“We saw a similar pattern in dogs with arrhythmias,” Johnson says. “The stethoscope never once called a dog's rhythm normal, and its atrial fibrillation calls were wrong three out of four times. That said, the ECG tracing it captures is genuinely good quality. This tool works best as an adjunct, a quick ECG or a flag worth a second look, not a stand-alone diagnosis.”

Kursten Pierce, assistant professor of clinical sciences and board-certified cardiologist at NC State, says the study grew out of something veterinarians kept telling her: they were second-guessing their own trained ears because a device disagreed with them.

“Stethoscopes are universal instruments. Veterinarians use the same stethoscope as a human physician uses,” Pierce says. “These new AI-enabled stethoscopes are being adopted by many veterinarians, because they have a lot of very useful technical features, such as recording a heart murmur or an EKG. However, the diagnostic AI for these stethoscopes is trained on human data, not dog or cat data, and we were hearing from veterinarians who were second-guessing themselves based upon the stethoscope's findings.”

That is the part worth sitting with. A tool built to reduce uncertainty was, in practice, manufacturing it, especially for the clinicians with the least experience to fall back on.

JUST NOT THE WHOLE ANSWER.

None of this means AI-assisted cardiac tools belong in the trash. The ECG quality the device captured was genuinely strong, and high-grade murmurs, the ones most likely to signal real disease, were far more likely to get caught than soft ones. Murmur grade was the single biggest predictor of whether the AI got it right, and dogs with a grade three murmur or higher were over fifteen times more likely to be correctly flagged.

The lesson isn't that the technology is worthless. It's that it isn't a substitute for a trained ear, a full physical exam, and clinical judgment, especially in cats, where the miss rate is not a rounding error. It's a blind spot.

“We want to encourage veterinarians to rely on the expertise they've gained through their training and to understand what this tool can and cannot provide in a veterinary setting,” Pierce says, “so that we continue to provide the best possible care to people and their companion animals.”

The full study, “An artificial intelligence–enabled digital stethoscope demonstrates moderate murmur detection in dogs but not cats and unreliable arrhythmia classification in both species,” appears in the Journal of the American Veterinary Medical Association. Joshua Stern, associate dean for research and graduate studies, and Teresa DeFrancesco, professor of clinical sciences, also contributed to the research.

So the next time that little AI light blinks green or red on a screen in your exam room, trust it as a second opinion. Not the only one.

SOURCE: NC STATE COLLEGE OF VETERINARY MEDICINE  |  JAVMA, DOI 10.2460/javma.26.05.0353