The automatic classification of canine state

Abstract

Osteoarthritis is a prevalent disease among domestic dogs which, even when well managed, often causes bouts of chronic pain and a lesser quality of life. Despite a lack of training dog owners are relied upon to recognise the signs of pain or illness in their animals. This often leads to treatment being sought later than would be ideal, resulting in the unnecessary and avoidable suffering of their dogs. This can be further compounded by the qualitative nature of lameness assessment performed by veterin arians. The difficulty of which is further exacerbated when symptoms are subtle, and the disease is in its early stages. This thesis investigates the use of remote, animal borne, tri-axial accelerometers to supplement the welfare information available to both caregivers and veterinarians. Published acceleration-derived measures, of both the time and frequency domains, common within human and non-human animal acceler ometer research, are assessed for their potential as daily and weekly identifiers of os teoarthritic lameness. The suitability of identified measures was evaluated using both Principal Component Analysis based feature selection and logistic linear models. The results of this process highlighted a potential link between both the level and entropy of an animals overall weekly activity with the occurrence of osteoarthritis. It also provided insight into areas of further development and established the complexity of the task of recognising lameness from acceleration data. A behaviour-based methodology was established hybridising techniques used across wildlife ecology deployments, existing veterinary assessment of lameness and, the assessment of human gait impacted by both physical illness and neurodegeneration. This led to the development of a method ology focussing on the identification of behaviours, starting with canine postural state, to provide context as to the daily activities of the subject. Two distinct approaches to postural recognition were assessed both employing machine learning techniques with a focus on the interpretability of results. The first, examined the identification of 6 pos tural transitions, similar to methods established in human accelerometer assessments, using linear discriminant analyses at 3 different sliding window lengths. The inclusion of an empirical cumulative distribution function representation was also assessed. The results suggested that the isolation of transitional periods from among non-transitional periods was difficult and there was high confusion between the transitions themselves. The second examined the identification of the postures themselves alongside the oc currence of locomotion during the standing posture. Linear discrimination analyses were once again used due to the interpretability of the method and the simplicity of its implementation. The effects of pre-processing techniques and differing posture group ings were also explored. The findings suggested a binary decision tree approach was the most effective mechanism and that the application of pre-processing techniques to clean data caused a distinct negative impact that requires forethought as to the poten tial costs and benefits of their use. Standing was the most easily identified, perhaps due to its prominence, and the further classification of locomotion from among stand ing periods was ineffective. To further supplement the postural methods of identifying osteoarthritis an investigation of the remote monitoring of circadian rhythm was estab lished. This is of interest due to prior results highlighting the potential relationship of activity entropy and level with lameness and the reports of sleep disruption by human chronic pain sufferers. Features relating to the length and frequency of both resting and active bouts were used in logistic regression models to establish their relationship to the presence or absence of osteoarthritis. Minor disruption was observed to the amplitude of activity frequencies within osteoarthritic dogs consistent with prior find ings. However, further work is needed to disentangle this effect from that of advanced age, a possible confound. The potential of remote sensing technologies is shown but further development of methodologies is required. A combination of the described approaches, with the refinements highlighted within this thesis, could further improve their efficacy and should be investigated. A behaviour based, transparent and fully in terpretable monitor of lameness, pain, and/or welfare could prove valuable to the early and effective treatment of canine osteoarthritis and should be pursued further