Lameness detection in horses is a critical challenge in equine veterinary practice, particularly when symptoms are mild. This study aimed to develop a predictive system using a support vector machine (SVM) to identify the affected limb in horses trotting in a straight line. The system analyzed data from inertial measurement units (IMUs) placed on the horse’s head, withers, and pelvis, using variables such as vertical displacement and retraction angles. A total of 287 horses were included, with 256 showing single-limb lameness and 31 classified as sound. The model achieved an overall accuracy of 86%, with the highest success rates in identifying right and left forelimb lameness. However, there were challenges in identifying sound horses, with a 54.8% accuracy rate, and misclassification between forelimb and hindlimb lameness occurred in some cases. The study highlighted the importance of specific variables, such as vertical head and withers displacement, for accurate classification. Future research should focus on refining the model, exploring deep learning methods, and reducing the number of sensors required, with the goal of integrating these systems into equestrian equipment for early detection of locomotor issues.

Locomotor injuries in horses are a major cause of underperformance and serious welfare issue. Veterinarians typically investigate horses’ lameness through visual examination at separate gaits (walk, trot, gallop). To evaluate lameness objectively, Inertial Measurement Units (IMU) based systems have been developed. It is necessary to accurately identify the gait of each stride as vertical displacement symmetry is assessed at a defined gait, essentially trot. This study aimed to classify gaits into 6 classes and to assess the training sample size required to maximize the performance. Unlike previous methods, we used raw IMU data without manually preselecting specific signal segments. Seven sensors were strategically placed on the limbs, head, withers, and pelvis of horses. 1440 horses were used in our unsupervised model and the gait of 110 horses was labelled using IMU data for our supervised models. We divided the 6 gaits classification task into two subtasks: a four-gaits classification and a gallop-specific classification. In the first subtask, we compared the performance of a machine learning (XGBoost), a deep learning (LSTM) and a transfer learning (ENCOD-CNN) model, depending on the labelled training sample size. Our results show that the transfer learning approach outperformed the other models, achieving test accuracy of 91.9%. Our gallop classification task achieves 97.1% accuracy and the total pipeline reaches 91.2% accuracy. Beyond improving gait classification in a real clinical setting, this research demonstrates the potential of transfer learning for time-series datasets and provides a quantitative assessment of the required labeled sample size for effective implementation.

Et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Background: In clinical practice, lameness evaluation often relies on visual assessment under different examination conditions, including ground and figures, and remains subjective. Although objective symmetry indices provide quantitative measures of gait asymmetry, their interpretation across multiple conditions remains challenging because several complementary indices are required and their thresholds vary with context. Objectives: To assess the ability of a context‐aware machine learning approach to predict lameness side and severity at the trot using trunk‐mounted IMUs. Study Design: Retrospective observational study. Methods: A neural network incorporating contextual parameters (ground, figure) via feature‐wise linear modulation (FiLM) was trained on symmetry indices derived from head, withers and pelvis motion in 625 horses examined during routine clinical evaluations, after exclusion of bilateral lameness. Reference lameness grades were extracted from clinical examination reports for each recording and aggregated into four ordinal severity levels. The model generated a predicted lameness class and an associated confidence score. Results: Exact accuracy reached 46% (43%–53% across contexts) for forelimbs and 47% (45%–52%) for hindlimbs. Most errors occurred between neighbouring ordinal categories, while left–right limb inversion remained below 3%. A limited subset of high‐confidence predictions showed improved accuracy. Main Limitations: Reference labels were based on subjective clinical grading, without systematic confirmation by diagnostic analgesia or repeated objective measurements. Additional limitations include the single‐clinic dataset, limited number of evaluators, and separate forelimb and hindlimb models. Conclusions: This proof‐of‐concept study shows that a context‐aware approach can integrate multiple IMU‐derived indices across examination conditions to predict lameness side and severity. Agreement with subjective clinical grades was moderate, and most errors occurred between adjacent severity categories, whereas left–right limb inversion was uncommon. Further validation using stronger reference standards and independent datasets is required before clinical relevance can be assessed.

Locomotor injuries in horses impair both welfare and performance. Traditional visual assessments of equine locomotion by veterinarians are inherently subjective. To provide objective evaluation, systems using Inertial Measurement Units (IMUs) have been developed, enabling quantitative analysis by calculating trunk symmetry indices (head, withers, pelvis) during trot. Stance temporal parameters provide valuable information for locomotion analysis, but their accurate estimation requires precise detection of stance-related events. Gait events are typically identified using signal processing techniques applied to data from limb-mounted IMUs, often placed on the cannon bones. Despite their potential, these sensors are not well tolerated by all horses and the associated signal processing methods often fail to generalize across diverse locomotion patterns. A large amount of data is gathered by veterinarians in uncontrolled environments, opening the door for deep learning methods, but the annotation process needed for supervised learning remains time-consuming and costly. The aim of this study was to segment horse stances using trunk-mounted IMU data and to introduce two novel semi-supervised time-series segmentation methods, specifically tailored for contexts with limited strongly labeled data and abundant weakly labeled data. The best performing pipeline achieved errors of -0.1±1.6% of the stride duration for foot-on and 0.2±2.6% for heel-off events of the right forelimb, and 0.2±1.5% for foot-on and -0.4±1.4% for heel-off events of the right hindlimb. Beyond improving stance segmentation in real clinical settings, this work explores two semi-supervised methods for time-series segmentation.