Wearable sensor shown to specifically quantify pruritic behaviors in dogs

Pruritus is a common presenting complaint for dog owners and often a primary determinant of successful management in cases of skin and ear disease. It can be manifested in a variety of ways including scratching, rubbing, paw licking, head shaking, and others. These behaviors may vary both in distribution and intensity depending on the cause and degree of pruritus and the degree to which it is expressed in individual pets. Measurement of pruritus in dogs is challenging because it relies on the observations of pet owners. Obvious flaws in owner assessment of pruritus include the amount of time an owner observes the animal each day and the variability of an individual’s interpretation of the severity of pruritus. When presented to a veterinary practice, communication of the level of pruritus may be further influenced by the veterinarian or the veterinary clinic environment itself. Previous efforts to provide more consistent and reliable assessment of pruritus have included both visual analog scales and numerical scales that are presented to owners. Visual analog scales have been an improvement over standard numerical scales, but interobserver repeatability is still problematic as it continues to rely on subjective observations of the pet owner. Assessment of visual analog scales and numerical scales using video of pruritic dogs shown to first year veterinary students demonstrated poor consistency and repeatability among observers when using these methods. This was especially prevalent with mid-range to lower levels of pruritus [1]. A more refined pruritus visual analog scale (PVAS) was developed by Hill et al. with more specific behavior descriptions to guide an owner’s rating [2].

Limited research has been conducted using wearable sensors in an attempt to measure canine pruritus. Studies reported to date utilized differences in overall activity between atopic and normal dogs as measured by differences in total piezo-electric voltage generated during specific periods of time [3, 4, 5]. Results often indicated significantly greater activity for atopic dogs over healthy dogs, with increased activity particularly marked overnight. Of the weaknesses cited was the assumption that increased activity in atopic dogs was due to pruritus without evidence that this was true. Additional studies verified the association between increased nocturnal activity and nocturnal pruritus by videotaping kenneled dogs while measuring activity in the same manner [5]. While significant, exact pruritic behaviors could not be specified. Authors of another study observed that activity monitor data and observed behavior did not always coincide indicating a lack of specificity [4]. Each of these studies suggested increased overall activity in pruritic atopic dogs and sleep disturbance characterized by the increased activity during night-time hours. Data collection in these studies relied on changes in electric voltage generated over either 15 s or 1 min blocks of time. Activity during these times generated electric voltage regardless of the type and was dependent only on movement of the sensor in any direction. Each of these studies also summarized activity during specified time periods. This summary was intended to provide an overall representation of activity and was chosen at times of day when activities such as exercise were least expected (e g. night time hours). Three-dimensional movement detection and continuous sampling were not possible with this system. One previously-reported study attempted to classify specific behaviors using a multidimensional sensor [6]. This study evaluated the ability of a collar-based sensor to correctly identify behaviors in a small population of dogs (n = 13). Specific focus on pruritic behaviors was not attempted and scratching was not one of the behaviors researched. In this study a confusion matrix of individual behaviors was presented but accuracy for behaviors was summarized in a proof of concept fashion with the accuracy and testing of individual behavior analysis not reported. Based on previous research, a clear objective evaluation system to quantify pruritus in dogs is currently lacking that would support veterinary professionals with information adequate to help with decisions on treatments needed or success of those prescribed.

The objective of this study was to validate the ability of a multidimensional high frequency sensor and an advanced computer analysis system, to specifically identify pruritic behaviors (scratching and head shaking).

The study reported here validated the use of an objective measurement tool for pruritus using a system that combines a wearable high frequency data collection sensor with newly developed computer algorithms designed to recognize specific behaviors. Previous studies using wearable sensors have been limited to differences in overall patient energy expenditures generated during activity [3, 4, 5]. As such, while energy expenditures may have been greater for pruritic atopic dogs, specific behaviors were neither identifiable nor attempted to be recorded. A recent attempt to identify specific behaviors in dogs using a small number of dogs (n = 13) did not specify accuracy for individual behaviors making it difficult to understand the positive predictive value of results [6].

A critical part of the correct identification of behaviors is the development of an algorithm classifier capable of identifying actions from the data collected. The ability of an algorithm classifier to properly identify behaviors is dependent on the ability of the development process to successfully identify a unique data “fingerprint” for each behavior. In developing these data fingerprints, the importance of the frequency at which data was collected became readily apparent. Sensor data sampling at too low of a frequency resulted in an inability to recreate the original signal and significant loss of detail. In contrast, high frequency sampling allowed the original signal to be readily observed and re-created without the loss of information. Similarly, multidimensional sampling provided more unique and distinguishable data to allow differentiation between behaviors. The more detailed data captured by high frequency and multidimensional sensor data collection are important drivers of performance in developing an algorithm classifier, and become the data input for algorithm development.

This highly detailed data, coupled with a genetic programming model of algorithm development allowed data generated by the sensor to be interpreted and correctly document behaviors with a high degree of accuracy.

In the study reported here, using a multidimensional high frequency sensing platform identified greater amounts of information about each activity but also provided much more detailed information for development of individual data fingerprints for each behavior. With more detailed data, a more accurate computer algorithm was developed to correctly identify the specific behaviors. Using this model, the identification and tracking of additional behaviors such as seizure or syncope activity should also be feasible, provided they can be documented in a significantly large population of dogs.

Pruritus is a complex sensation with a large variety of potential causes and ways it may be expressed. Scratching in dogs often occurs by the dog using a rear leg to scratch somewhere on its body. This behavior is a high energy repetitive action making it recognizable both to observers and the activity detected in the sensing platform. Similarly head shaking is also a high energy repetitive action that can be more readily detected. The extremely high specificity reported here (99%) for both of these behaviors documented that scratching or shaking, when noted by the sensor were very likely to have occurred. With a sensitivity of 76.85% and 82.16% for scratching and shaking, respectively, it is possible that that system may miss individual episodes of these behaviors. However, compared to current methods of owner assessment using the PVAS or other scale, or summaries of overall activity suspected to represent pruritus, this method far exceeds any other tool and is in fact an objective measurement. Ideally, this system should be implemented and continue to monitor an individual dog for weeks to months or longer. This would provide an individual baseline applicable to each dog allowing clinicians and researchers to identify differences noted during pruritic episodes and both before and after a particular treatment intervention.

Limitations of this study include the use of only scratching and head shaking as indicators of pruritus. Considering pruritus is often made up of multiple different behaviors, the addition of paw licking and chewing at a variety of sites would be valuable additions to the system. With the prevalence of paw licking especially in dogs with atopic dermatitis and canine adverse food reaction, however, these would provide a more comprehensive evaluation of pruritus as a whole and should be a focus of subsequent studies. Although owner assessments are highly variable and subjective, further research and clinical use are also warranted to correlate owner assessments of pruritus with sensor documentation.

Future research is needed to continue to develop this model to identify additional behaviors of both clinical and research importance. These may include seizures, drinking, urination and others. By being able to identify these behaviors, more objective data on frequency of these behaviors and subsequently their response to therapies could substantially aide the veterinary professional in patient management.

Wearable sensors and the machine learning process are an exciting new frontier that offers tremendous benefits in veterinary medicine. With advances in these fields a variety of more objective real-time information can be developed to provide more accurate and timely medical assessments and decisions. Since pruritus has a significant effect on quality of life for pets and their owners, this information has potential to allow veterinarians to identify pruritic episodes quickly, suggest therapies and examinations to owners when needed and assess progress of medications and treatment plans. It also would allow the veterinary team to illustrate benefits of these treatments to pet owners during veterinary visits.