EQUAINE: Asking machines to understand equine data and keep learning from it

Abstract

The use of horses in our society comes with a responsibility regarding their health and welfare. Amongst performance-impeding issues, lameness and respiratory pathology are particularly concerning health and welfare. Their early detection relies on regular monitoring of locomotion and physiological statuses. The combined progress in miniaturized sensors and artificial intelligence (AI) now offers new opportunities and perspectives for automating this monitoring, as data quantifying phenomena imperceptible to the human senses can now be collected and processed. This thesis exploits these two technological advances to develop quantification tools based on sensor signals processing by AI models. The work is divided into three parts. The first part explores the ability of AI models to identify the context of a measurement (surface type, lameness status) or to estimate different parameters that were previously difficult to obtain (limb loading, dynamic respiration rate). It shows that a model trained on data from a single limb sensor can accurately classify the type of terrain onto which a horse walks or trots, even at low sampling rates. It also demonstrates that a combination of sensors placed on the head, withers, and pelvis can discriminate between sound and lame strides, with performance varying according to whether the lameness is stemming from a front or hind limb. Furthermore, it shows that vertical ground forces can better be estimated from upper-body sensors compared to limb sensors, that and dynamic respiratory rate can be computed from downsampled audio signals, paving the way for less computationally expensive tools. The second part of this thesis focuses on the interpretability of the developed models. The analysis of which features were recurrently selected to differentiate between terrain types shows that not only are the longitudinal accelerations at the limb levels important but also the mediolateral ones, suggesting an interesting axis of research on the link between surface and orthopaedic injuries. An explainable AI approach applied to lameness detection confirms that the models rely on kinematic landmarks similar to those used by veterinarians, while identifying additional movement phases not perceptible to the naked eye, thus enhancing the potential for transparency and clinical acceptability of these tools. The third section addresses the need to validate models using field data. A new data collection platform, dedicated to harness racing horses, is presented. It focuses on the simultaneous collection of locomotion and physiological data, while offering valuable real-time feedback to trainers. An audio dataset comprising raw signals, rarely shared in this field, is also published to support future efforts by the scientific community. Overall, this thesis demonstrates that AI models, combined with explainable AI methods, are valuable tools for characterising equine locomotion and respiration from sensor data, while also identifying new research axes. Furthermore, through the development of data collection tools and the publication of ecologically valid datasets, it contributes to grounding these approaches in real-world applications.