A digital stethoscope promoted as an intelligent aid for detecting heart disease may be far less reliable in dogs and cats than in humans, according to a prospective study from North Carolina State University. The research found that an artificial intelligence–enabled device performed moderately well when identifying heart murmurs in dogs, but failed to detect most murmurs in cats and produced unreliable classifications of abnormal heart rhythms in both species. The findings highlight a growing challenge in veterinary medicine: technologies built around human medical data may generate convincing results while missing the biological differences that matter most in animal patients.
The study evaluated 105 companion animals treated at a university teaching hospital, including 54 dogs and 51 cats. Each animal underwent cardiac auscultation, the clinical process of listening to the heart with a stethoscope, at four thoracic locations. The examinations were performed using an EKO Core 500 digital stethoscope equipped with artificial intelligence software, while a cardiology resident, a board-certified veterinary cardiologist and a fourth-year veterinary student independently assessed the animals. The researchers compared the device’s classifications with the clinicians’ findings, supported by six-lead electrocardiography and echocardiography when clinically indicated.
The device’s performance differed sharply between species. Among the dogs, clinicians identified heart murmurs in 38 animals, representing 70% of the canine group. The AI software correctly detected 33 of those murmurs, producing a sensitivity of 86.8%. Sensitivity measures the proportion of animals with a condition that a test successfully identifies. However, the system’s specificity—the ability to correctly recognize animals without the condition—was only 56.3%. Its positive predictive value was 82.5%, meaning that most, but not all, dogs labeled as having a murmur were judged by clinicians to have one.
In practical terms, the AI stethoscope performed similarly to the fourth-year veterinary students when identifying canine murmurs. The agreement between the device and clinicians was not significantly different from the agreement between students and clinicians, with a Cohen’s kappa value of 0.447. Kappa is a statistical measure that evaluates agreement beyond what might occur by chance; values in this range generally indicate moderate agreement. The researchers also found that murmur intensity was the most important factor influencing whether the software detected a murmur. High-grade murmurs, classified as grade 3 or higher, were about 15 times more likely to be identified than softer murmurs.
The picture changed dramatically in cats. Clinicians found murmurs in 22 of the 51 cats, or 43% of the feline group, but the AI system recognized only two of them. That corresponds to a sensitivity of 9.1%, meaning the device missed 20 of the 22 murmurs identified by veterinary experts. The agreement between the AI stethoscope and clinicians was extremely low, with a kappa value of 0.081. Veterinary students performed substantially better, correctly identifying nearly two-thirds of the murmurs in affected cats. The result is especially significant because feline heart disease can remain clinically subtle, and a quiet or apparently normal examination does not necessarily exclude a serious condition.
The researchers say the poor feline performance may reflect fundamental differences between human, canine and feline cardiac sounds. Heart murmurs vary according to blood flow, heart anatomy, body size, chest conformation, respiratory pattern and the location at which the sound is heard. Cats can produce low-intensity or acoustically complex murmurs that may be difficult to distinguish from normal heart sounds, particularly when the animal is stressed or purring. An algorithm trained primarily on human recordings may not have encountered enough representative feline sounds to learn those patterns. Even a technically excellent sensor can therefore produce weak clinical results if its software has not been validated on the species and conditions in which it is used.
Arrhythmia classification was even more problematic in dogs. Clinicians determined that 24 of the 54 dogs had an abnormal heart rhythm, while six had atrial fibrillation, a rhythm disorder in which the heart’s upper chambers beat rapidly and chaotically. The AI stethoscope correctly flagged all six dogs with atrial fibrillation, giving it a sensitivity of 100% for that specific condition. Yet it never classified a single dog as free of an arrhythmia. The software labeled 22 additional dogs as having atrial fibrillation even though clinicians did not find the disorder in them. As a result, three out of four of its atrial-fibrillation calls were incorrect among the dogs it identified as positive beyond the confirmed cases.
That pattern illustrates why sensitivity alone cannot establish that a diagnostic tool is clinically dependable. A system that detects every true case but also produces many false alarms can overwhelm clinicians with unnecessary follow-up examinations and cause owners considerable anxiety. Conversely, a system that misses disease can provide false reassurance. In this study, the device’s electrocardiogram recordings were described as genuinely good quality, suggesting that the hardware was capable of capturing useful electrical signals even when the automated interpretation was unreliable. The distinction is important: recording data and interpreting data are separate technical tasks, and strong signal acquisition does not guarantee accurate diagnosis.
Only one cat in the study was judged by clinicians to have an arrhythmia, limiting the conclusions that can be drawn about the device’s feline rhythm classification. The overall sample was also collected at a single university teaching hospital, where animals may have a different disease profile from those seen in general veterinary practices. Nevertheless, the prospective design and direct comparison among an AI device, a student, a resident and a specialist provide practical evidence about how the technology performs during real clinical examinations. The authors emphasize that the results should not be interpreted as a rejection of digital stethoscopes or artificial intelligence, but as a warning against treating a universal instrument as universally validated.
For veterinarians, the researchers recommend using such systems as adjuncts rather than autonomous diagnostic authorities. A digital stethoscope can store auscultation recordings, provide a visual waveform, capture an electrocardiogram and draw attention to a possible abnormality. Those capabilities may be valuable for documentation, remote consultation, education and deciding which patients deserve closer examination. The final interpretation, however, must account for species-specific anatomy, the animal’s clinical history, the quality of the examination and the veterinarian’s own training. The North Carolina State team hopes the findings will encourage clinicians and students to understand what AI-assisted instruments can and cannot do before relying on their outputs in routine care. The study, published in the Journal of the American Veterinary Medical Association, makes a broader point relevant far beyond veterinary cardiology: artificial intelligence is only as dependable as the data used to train it and the clinical setting in which it is validated.
Subject of Research: Animals
Article Title: An artificial intelligence–enabled digital stethoscope demonstrates moderate murmur detection in dogs but not cats and unreliable arrhythmia classification in both species
News Publication Date: August 5, 2026
Web References: https://avmajournals.avma.org/view/journals/javma/aop/javma.26.05.0353/javma.26.05.0353.xml
References: Johnson J, Stern J, DeFrancesco T, Pierce K. “An artificial intelligence–enabled digital stethoscope demonstrates moderate murmur detection in dogs but not cats and unreliable arrhythmia classification in both species.” Journal of the American Veterinary Medical Association. DOI: 10.2460/javma.26.05.0353
Keywords: Veterinary medicine, artificial intelligence, digital stethoscope, dogs, cats, heart murmurs, arrhythmias, atrial fibrillation, diagnostic accuracy, veterinary cardiology
