Researchers from Kinship’s Pet Insight Project team and the Waltham Petcare Science Institute report using deep learning technology and commercially available motion-sensing pet activity tracker, Whistle, to identify normal dog behaviours and activities, as well as those associated with common canine conditions. Results of the research are published in the journal Animals.
“Deep learning is a powerful technology that enables us to analyse enormous amounts of data to identify meaningful patterns in pet behaviour,” said Dr Aletha Carson, DVM, Sr Manager, Data and Clinical Research, an author on the manuscript. “With this research programme we used our data to build algorithms which allow us to objectively understand a pet’s behaviour in their home environment. A better understanding of day to day behaviours will allow us to identify potential signs of illnesses earlier than ever before and promote earlier treatment interventions.”
Research findings are a testament to the power of citizen science
To make the correlations between dog activities and motion sensor data, the researchers assembled machine learning training databases from more than 5000 videos of more than 2500 dogs submitted by pet owners and 11 million days of pet activity data collected from Whistle devices. They then developed a novel deep learning algorithm that can accurately categorise data from a collar-mounted sensor called an accelerometer into defined behaviours and activities. Just one second of data is sufficient for the algorithm to classify the behaviour.
To confirm and validate the algorithm’s accuracy in a real-world setting, the researchers then compared that data to pet activity reports from pet owners of 10,550 dogs, yielding data from 163,110 unique pet eating and drinking events. They found that the algorithm correctly identified eating (94 percent) and drinking (98.8 percent), and to a lesser extent could also identify more nuanced behaviours ranging from sniffing and scratching to rubbing and licking.
Ultimately, continuous monitoring of pet behaviour and activity could help pet owners identify pets with a range of possible conditions, including scratching, poor appetite, excessive weight, or osteoarthritis.
“This paper validates the accuracy of using behavioural ‘signs’ to detect potential health issues, based on real world data,” said Scott Lyle, Head of Pet Insight Project. “With the foundational algorithms built on the dataset, we can further our understanding of pet behaviour with devices like Whistle in seeking to advance individualised veterinary care. Across human and animal health with citizen science and behaviour monitoring devices we are constantly finding ways to advance health through science.”
