At the core of Pet Insight Project is the principle that dogs show changes in behavior when they experience a change in wellness. That’s something both veterinary professionals and pet owners know and understand intuitively when something seems “off” about a dog, but we historically haven’t been equipped to objectively measure and evaluate these behavior changes.
Pet Insight Project was designed to turn behavior into a quantifiable vital sign by measuring the behavioral changes associated with changes in the wellness of participating dogs. This begins with the unique and critical use of participants’ Banfield medical records to provide context for the behavioral patterns detected by Whistle. This weekend we’ll be at the American Veterinary Medical Association Convention in Denver, Colorado talking to veterinary professionals about how medical histories can act as “maps” to uncover and measure behavioral change. Here’s a sneak peek of what we’ll share:
Medical Records as "Maps"
Banfield’s industry-leading electronic veterinary medical records system, called PetWare®, gives our research team a history of project participants’ well-being and provides guideposts for where to begin our exploration of behavioral changes. Each participant’s medical history is documented, organized, and securely stored in a standardized way, enabling our research team to identify all Pet Insight dogs that received a specific diagnosis or treatment and when it happened. For example, we can identify the 116 dogs that have already undergone a spay or neuter procedure while participating in the project.
We then use the medical records like a “treasure map” to orient our exploration of hidden behavioral patterns observed by the Whistle device. For example, by knowing which day the spay or neuter procedure occurred, we can determine the window of time to evaluate for behavior changes, telling us “where to dig” for insights. In this case, we looked at the dogs’ activity in the four weeks before the procedure compared to the four weeks following the procedure.
In these early stages of the project, daily minutes of activity and rest are the most quantifiable variables to evaluate behavioral change. Dogs have very different daily activity ‘baselines,’ so we can’t just compare a dog that usually gets 30 minutes of daily activity to one that normally gets 120 minutes as the first dog would always look lethargic next to the other even on the most active days. To ensure we’re making reasonable comparisons, we look at the relative changes in activity - i.e. how today’s activity compares to the average activity of that particular dog.
We can examine the changes in relative activity in the context of the medical event, in this case the spay/neuter procedure. Relative activity stays right around the baseline (100% of the dog’s activity averages) prior to the procedure (as you would expect). After the procedure, relative activity drops to ~20% of normal and remains below the pre-procedure average activity for over two weeks as the dog recovers.
The Banfield medical histories provide critical context to make sense of the drop in activity for these dogs, and the pairing of medical and behavior information enables direct comparison of hundreds (and eventually thousands) of dogs experiencing the same health event. In this case, we can clearly see the expected dip in activity as the dogs recover from a surgical procedure.
In summary, Pet Insight Project data and analysis will provide previously unseen behavioral information that will enable veterinary professionals and researchers to evaluate responses to treatment and inform normal recovery milestones for patients.
Zooming in from the pack to the individual
Our ultimate goal is to be able to not only measure, but also to interpret behavioral changes for an individual dog. While understanding how the ‘average’ dog’s behavior changes in response to a health event is useful, we must be able to understand what changes mean in the context of a unique individual to be able to inform care decisions. Interpreting behavior of an individual is complicated by the significant variation in many behaviors from day-to-day.
Each line in this graph represents the relative activity for each of the 116 dogs that had a spay or neuter procedure while participating in the project. While there is clearly a change after the procedure, the ‘spikiness’ of the dogs’ relative activity makes the trend much harder to see for any particular individual.
For example, we cannot set basic thresholds for changes in a single behavior (e.g. alert caregiver if a dog’s activity is more than 50% below baseline for 3 or more days), as many perfectly healthy dogs show that pattern (would be a ‘false positive’), while others that did have a health event don’t show that magnitude of activity decline (would be a ‘false negative’).
This is where using computational power and advanced algorithms becomes critical, as we need ‘super-human’ pattern recognition abilities to be able to consider huge amounts of data and detect similarities. We’ll need to consider the interactions of other behaviors we’re beginning to detect, such as eating, drinking, scratching, and resting, as well as the history of nuances and quirks in the behavior of that particular dog. The algorithms are able to not only consider the millions of data points collected from that individual, but also reference them against data collected from tens of thousands of other dogs and “see” patterns that act as signatures of medical conditions. While I am always impressed by the brain power of our data science team, using computers and algorithms unlocks a vast set of abilities to detect patterns that would otherwise be impossible as mere mortals.