Equine veterinary journal2026; doi: 10.1002/evj.70326

Context-aware machine learning for lameness side and severity prediction in trotting horses: A proof-of-concept study.

Gérard, M
·
Hanne-Poujade, S
·
Dubois, G
·
Hébert, C
·
Coudry, V
·
Jacquet, S
·
Bertoni, L
·
Tallaj, A
·
Mezghani, N
·
Chateau, H

Abstract: 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. Objective: To assess the ability of a context-aware machine learning approach to predict lameness side and severity at the trot using trunk-mounted IMUs. Methods: 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. Conclusions: 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.

Get Full Text

Publication Date: 2026-09-13 PubMed ID: 42733278DOI: 10.1002/evj.70326Google Scholar: Lookup

The Equine Research Bank provides access to a large database of publicly available scientific literature. Inclusion in the Research Bank does not imply endorsement of study methods or findings by Mad Barn.

Facebook X LinkedIn Email Copy

Summary
Cite This
Publication
Affiliations
Grants
References
Citations

Journal Article

Summary

This research summary has been generated with artificial intelligence and may contain errors and omissions. Refer to the original study to confirm details provided. Submit correction.

Overview

This study investigates using a context-aware machine learning model to predict which side and how severe a horse’s lameness is during trot, based on data from sensors attached to the horse’s body.
The model incorporates different examination contexts to better interpret movement asymmetry and attempts to provide more objective assessments than traditional subjective clinical evaluations.

Introduction and Background

Lameness in horses is typically assessed through visual examination by veterinarians during various conditions, such as on different ground surfaces or specific movement patterns (figures).
These visual evaluations are subjective and can vary depending on who examines the horse and under what conditions.
Objective measures exist, for example symmetry indices calculated from motion data, but interpreting these across multiple conditions is difficult because:

Multiple complementary indices are needed to capture the complexity of movement.
The thresholds for defining lameness vary depending on the context (e.g., type of ground, gait pattern).

Objectives

To assess if a context-aware machine learning method can predict both the side (left/right, forelimb/hindlimb) and the severity of lameness in trotting horses.
The study uses data from inertial measurement units (IMUs) mounted on the horse’s trunk to gather motion information.

Methods

Study Design: Retrospective observational study analyzing 625 horses who underwent routine clinical lameness evaluations.
Data Collection: Symmetry indices were derived from the motion of the head, withers (part of the back), and pelvis using trunk-mounted IMUs.
Context Incorporation: The model integrates contextual factors like ground surface and figure (the pattern of movement) using a technique called feature-wise linear modulation (FiLM), which adjusts features based on context.
Reference Labels: Clinical examination reports provided subjective lameness grades, aggregated into four ordinal severity levels (none to severe), excluding horses with bilateral lameness.
Model Output: For each data input, the model predicted lameness class (side and severity) and generated a confidence score reflecting prediction certainty.

Results

Accuracy:
- Exact match accuracy for forelimb lameness was 46%, ranging from 43% to 53% depending on context.
- For hindlimb lameness, accuracy was 47%, with a 45%-52% range across contexts.
Error Patterns:
- Most classification errors involved confusion between neighboring severity categories (e.g., mild vs moderate), indicating the model struggles with fine-grained severity distinctions.
- Misclassification of the lameness side (left vs right limb) was rare, occurring in less than 3% of cases, showing good lateralization performance.
Confidence Filtering:
- Predictions with higher confidence scores had improved accuracy, suggesting confidence can be used as a filter for more reliable predictions.

Limitations

The reference labels used for training and evaluation were subjective clinical grades, lacking systematic confirmation by more objective methods such as diagnostic analgesia (local anesthesia to confirm pain source) or repeated measurements.
The dataset came from a single clinic, limiting the generalizability of the findings.
Evaluation was done with a limited number of veterinary assessors, which may introduce evaluator bias.
The model architecture separated forelimb and hindlimb analyses rather than integrating them into a single comprehensive prediction framework.

Conclusions and Future Directions

This proof-of-concept study demonstrates the feasibility of a context-aware machine learning approach that leverages multiple IMU-derived gait indices and contextual examination data to predict both side and severity of lameness in horses.
Agreement with subjective clinical grading was moderate; however, most errors were close to the true severity level rather than large misclassifications, and errors in identifying the correct lame limb were uncommon.
Further research should involve:
- Using stronger reference standards like diagnostic analgesia or repeated objective gait analyses to provide more definitive labels.
- Testing the approach on independent and larger datasets from multiple clinics to improve model robustness and applicability.
- Exploring integrated models combining forelimb and hindlimb lameness predictions.
- Assessing real-world clinical relevance and potential to support veterinary decision-making in lameness diagnosis.

Cite This Article

APA

Gérard M, Hanne-Poujade S, Dubois G, Hébert C, Coudry V, Jacquet S, Bertoni L, Tallaj A, Mezghani N, Chateau H. (2026). Context-aware machine learning for lameness side and severity prediction in trotting horses: A proof-of-concept study. Equine Vet J. https://doi.org/10.1002/evj.70326

Publication

Equine veterinary journal

ISSN: 2042-3306

NlmUniqueID: 0173320

Country: United States

Language: English

Researcher Affiliations

Gérard, Mahaut

Ecole Nationale Vétérinaire d'Alfort, ACAP3, CHUV-Eq, Goustranville, France.
Applied artificial intelligence Institut (I2A), TELUQ University, Québec, Canada.
LIM Group, Nontron, France.

Hanne-Poujade, Sandrine