Can Wearables Classify Dog Gait? A 29-Dog Study

A source-backed Study Brief about inertial sensors, deep learning and the difference between testing familiar dogs and unseen dogs.

Research & Evidence | Study Brief | Updated September 1, 2026

The short answer

A 2026 Scientific Reports study showed that a research model could classify three groups in controlled gait recordings: healthy dogs, dogs with orthopedic disorders and dogs with neurological disorders. On a random split of the study data, the best multiclass accuracy was 0.960. When the researchers held out whole dogs, the best reported binary accuracy was 0.849, and the best multiclass accuracy was 0.800. For the study's orthopedic-versus-neurological task, the best unseen-dog accuracy was 0.893.

Those numbers describe one research model, one small group of dogs and specific walking and trotting protocols. They do not validate a consumer wearable, diagnose a dog at home or show that a dog is safe to exercise because a device produces a confident label.

This brief asks: How did the study use wearable inertial sensors and deep learning to classify canine gait, and what changed when the model met a dog it had not seen during training?

What the study measured

Palez and colleagues studied 29 dogs at the Department for Small Animal Medicine and Surgery at the University of Veterinary Medicine Hanover in Germany. The group included 17 healthy dogs, 6 dogs with orthopedic disorders and 6 dogs with neurological disorders. The clinical group labels came from general, orthopedic and neurological examinations. Some neurological cases also had magnetic resonance imaging to support the clinical assessment.

The recordings took place in a controlled testing area. Dogs moved along the same walkway while guided on a leash. The protocol included walking and trotting. The researchers used Xsens Movella DOT inertial measurement units, which recorded accelerometer and gyroscope signals at 120 Hz. They tested sensors at the top of the head or upper back, the lower back near the tail, the neck or collar, and a combination of all three positions.

The model received six signal channels, three from the accelerometer and three from the gyroscope. The study used a small one-dimensional convolutional network followed by fully connected layers. Across the model data, the researchers report 193 minutes of recordings, with 154 minutes used for training and 39 minutes for testing.

The paper asked three related classification questions:

1. Can the model separate healthy, orthopedic and neurological groups?
2. Can it separate healthy dogs from the two groups with diagnosed disorders combined?
3. Among the dogs with diagnosed disorders, can it distinguish the orthopedic group from the neurological group?

The third question is a research label in this paper. It is not a diagnosis that the model can make for a new dog in a living room.

Why the test split matters

The paper first used an 80/20 random split. This asks how well the model performs when the training and test data come from the same pool of recorded dogs. It is useful for measuring performance inside the study, but it does not answer how the model will behave with a completely new dog.

The researchers then used leave-one-dog-out cross-validation. In each round, all of one dog's recordings were held out for testing while the model trained on the other dogs. No time windows from the held-out dog entered the training set. That is a harder test of individual variation.

The paper's summary table reports these best sensor and movement combinations:

The table makes two points easier to see. First, accuracy and F1 belong to a named task and setup. Second, the result changes when the model has to generalize to a dog that was not in its training data.

What the model learned, and what it did not prove

The study supports a narrow finding: patterns in inertial signals can help a model classify the groups represented in this dataset. Sensor placement and movement protocol affected the result. The best combination for one task was not automatically the best combination for another.

The gap between the random split and the unseen-dog test is just as important as the high number. Dogs differ in size, body shape, gait strategy and ability to complete a protocol. A model that performs well on recordings from familiar dogs may need more data and broader testing before it works reliably on new dogs.

The study does not show that:

• a consumer activity tracker can identify orthopedic or neurological disease;
• a home recording can replace a veterinary examination;
• every breed, body shape or gait pattern is represented by this sample;
• a device output can tell a family whether exercise is safe; or
• the reported numbers form a leaderboard for pet wearables.

The authors describe the approach as having potential for clinical support. That is a research direction, not evidence that the model has been validated for household use. The study's dataset is not a public consumer benchmark. The authors state that access is available from the corresponding author on reasonable request.

A limitation that changes the reading

Only two dogs in the neurological group were able to trot. Several affected dogs could not complete the trotting protocol, so the trotting scenarios were materially imbalanced. That matters when the best numbers appear in a trotting row.

The sample was also small and collected in one controlled clinical environment. A leash, a familiar walkway, a fixed route and a particular sensor position are part of the measurement. They are not background details that can be removed when a result is carried into a home, park or clinic with different equipment.

The paper also reports a difference between within-sample performance and unseen-dog performance. That difference is consistent with a basic rule for reading animal-sensing research: the unit that matters is often the dog, not only the number of short signal windows produced by that dog.

Questions to ask about a future wearable claim

When a pet technology product reports a gait or health score, ask what the model was trained to classify. Then ask how many dogs supplied the data, which breeds and conditions were included, whether complete dogs were held out, what movement the dog had to perform, and whether the test used new dogs in new environments.

Accuracy alone cannot answer those questions. F1, recall, precision and confidence intervals also need their task, population and split. A score from a controlled research dataset is evidence about that dataset. It is not a general promise about every dog wearing a device.

A useful boundary for pet families

If you notice a new or worsening change in your dog's gait, treat the change as an observation to share with a veterinarian. A short recording may help you describe what you saw, but a research score cannot determine the cause. Do not use a device label to delay care or decide that exercise is safe.

For practical care information, see our guide to dog hock injuries. That is a different reader job from this study brief: care guidance should not be inferred from a model's research metric.

Bottom line

This study is useful because it tests a question that is easy to oversimplify. Wearable inertial sensors captured movement, and a deep-learning model found patterns that separated study-defined groups. The unseen-dog test then showed why the original sample and the evaluation design must stay attached to the result.

The responsible conclusion is modest. Canine gait models may help researchers quantify movement under defined conditions. The paper does not establish a home diagnostic tool, a universal gait score or a safe-exercise guarantee. For future claims, look for dog-level separation, new-dog validation, clear labels and a professional path for interpreting a real change in movement.

Sources and update note

The source article and its relevant tables were checked on September 1, 2026. This is a single-study brief, not a systematic review or meta-analysis. The article reports the study's sample, methods and metrics with their task and test context. Our interpretation is identified where it goes beyond the paper's direct result.

1. Palez, N., Straß, L., Meller, S. et al. "Canine gait analysis using inertial sensors and deep learning for orthopedic and neurological disorders." Scientific Reports 16, 13966 (2026). Nature record
2. Table 5: Multiclass classification performance for the generalization scenario
3. Table 6: Summary of the main classification results

This article is informational and does not provide veterinary diagnosis or treatment advice.