In a promising new study, researchers from the University of Cambridge and leading veterinary centers in the United Kingdom have developed a machine-learning tool that could transform heart disease detection in dogs. This AI-driven approach can detect heart murmurs and monitor valve diseases, offering a faster, cost-effective alternative to traditional diagnostic methods. These heart issues are particularly common in smaller breeds such as King Charles Spaniels and Chihuahuas.

“Heart disease in humans is a huge health issue, but in dogs it’s an even bigger problem,” said first author Dr. Andrew McDonald from Cambridge’s Department of Engineering. “Most smaller dog breeds will have heart disease when they get older, but obviously dogs can’t communicate in the same way that humans can, so it’s up to primary care vets to detect heart disease early enough so it can be treated.”

Puppy hearts

Heart murmurs in dogs often signal serious conditions like myxomatous mitral valve disease (MMVD), which can affect the heart’s ability to regulate blood flow. If left unchecked, this condition can lead to congestive heart failure, but if it’s detected early, it can be managed effectively.

Traditionally, diagnosing and monitoring MMVD requires an echocardiogram. This is a complex ultrasound procedure typically performed only at specialized centers. The process is costly, time-intensive, and stressful for pets and owners, limiting accessibility.

“Diagnosing mitral valve disease currently relies on an ultrasound of the heart (echocardiogram), often performed at a referral center by a cardiologist.  This is expensive, especially for owners who have financial worries,” said McDonald for ZME Science.

The new AI model developed by McDonald and colleagues uses sound recordings from electronic stethoscopes to detect heart murmurs and estimate their severity. This approach could be integrated into regular veterinary practices, which means it can expand access to MMVD screening and enable veterinarians to catch early signs of disease before it progresses. McDonald, from Cambridge’s Department of Engineering, explained:

“It’s critical to start medication for dogs with mitral valve disease at the right time so that we can delay the onset of heart failure and prolong quality of life. However, we currently rely on detection and grading using an ultrasound of the heart (echocardiogram), which is expensive and time-consuming. Our technology could be used in general practice to identify dogs that are most in need of treatment.”

AI, coming soon to a veterinary center near you

The study involved over 750 dogs. Each underwent thorough physical exams and echocardiograms conducted by certified cardiologists. To train the model, researchers used recordings from electronic stethoscopes that captured audio of the dogs’ heartbeats. These recordings were then analyzed by a neural network, fine-tuned to identify and classify murmurs based on a grading scale ranging from “soft” to “thrilling.”

While similar to human-focused AI models, this approach required additional considerations.

“The recordings in humans and dogs showed a noticeable similarity, largely due to the comparable structure of their hearts. However, making a good quality heart sound recording from a dog can be more challenging than on a human. Unsurprisingly, dogs are much more likely to fidget and whine during recordings, which can add significant noise to the data,” the researcher told us in an email.

The AI model performed so well that researchers believe it could eventually match the diagnostic accuracy of cardiologists, improving murmur detection in general practice. This would be especially valuable for veterinarians, particularly students and recent graduates, who often struggle with confidence in stethoscope-based diagnosis.

“Studies have shown that veterinarians, particularly students and recent graduates, lack confidence in listening with a stethoscope. Our study showed that our algorithm performed favorably to expert cardiologist assessment, and we are planning a follow-on investigation to compare directly to general veterinarian assessment.”

A challenging AI application

MMVD, especially in its early stages, can require either monitoring or medication. The goal is to get this new AI-based diagnostic tool to identify dogs at critical stages and initiate early interventions. “Our technology could be used in general practice to identify dogs that are most in need of treatment,” said McDonald.

Building the tool, however, involved logistical challenges. Gathering data from hundreds of dogs required extensive collaboration with veterinarians across multiple centers. The researchers collected thousands of sound recordings and carefully matched each recording with echocardiographic findings to provide a comprehensive dataset.

McDonald acknowledged, “The most significant challenge was the time of our veterinary collaborators. Collecting and uploading thousands of heart sound recordings is a highly time-consuming task that requires careful planning and coordination. It had to be balanced around their ongoing clinical and teaching commitments.”

Once the data was ready, the team applied a “transfer learning” approach, adapting an algorithm initially trained to detect heart murmurs in humans. With additional tuning, the algorithm achieved high accuracy in dogs, accurately grading murmurs and matching up with cardiologists’ evaluations in nearly 87.9% of cases.

AI assistance

As is typically the case, the goal is not to have AI replace doctors, but complement their ability and reduce their workload.

“So many people talk about AI as a threat to jobs, but for me, I see it as a tool that will make me a better cardiologist,” said co-author Novo Matos. “We can’t perform heart scans on every dog in this country — we just don’t have enough time or specialists to screen every dog with a murmur. But tools like these could help vets and owners, so we can quickly identify those dogs who are most in need of treatment.”

In cases where faint murmurs or noisy recordings hinder detection, the algorithm can flag results for further review by a specialist.

“In cases where murmur detection is challenging (e.g. due to noise from the dog, or a faint sound), our algorithm can output a probability score to give an indication of the confidence in its decision. When the confidence is lower than a certain threshold, we could recommend a repeat recording is made or a veterinary expert is consulted.”

Can this technique become commonplace?

Researchers are confident this technique can be implemented in routine practice, but there’s still a lot of work to ensure it works properly.

In veterinary practice, false positives can lead to unnecessary treatments and added costs for pet owners. However, missed diagnoses may delay essential care, especially for early-stage MMVD cases. McDonald and his team envision using adaptable thresholds that practices can modify based on their patient populations and resources.

The team plans to continue this project, gathering more data and expanding it to more breeds and clinical environments.

“We are interested in transferring our human and canine algorithms to future applications in both other species and other sound-producing diseases. Collecting data from a wide variety of species and clinical environments will continue to improve the generalization of our algorithms.”

The study was published in the Journal of Veterinary Internal Medicine.