Predicting Equine Body Condition from 3D Surface Geometry

Sammanfattning

Equine obesity is a growing welfare concern. The Body Condition Score (BCS) is an established nine-point scale for assessing equine body fat. Typically, BCS is assessed subjectively by a human rater and it is therefore, known to be subject to significant inter-rater variability. The VAREN model is a biologically accurate parametric 3D horse model that encodes individual equine morphology through 39 objective shape parameters, representing a plausible basis for automated and objective BCS prediction. This study investigates two main questions: a) To what extent can machine learning models estimate the Body Condition Score (BCS) of a horse using only objective 3D shape parameters? b) Which algorithmic approach yields the highest robustness against data imbalance and clinical variability? A dataset of 208 horses across 23 breeds was used across four phases of modelling, progressing from baseline 80/20 evaluation through repeated cross-validation, forward sequential feature selection, and anatomical stacking ensembles. Models were evaluated using Mean Squared Error (MSE), Mean Absolute Error (MAE), and clinical tolerance within +-0.5 and +-1.0 BCS units. The strongest performance was achieved when reducing to seven optimal shape parameters (Feature Selection, Phase 3), where the Voting Ensemble yielded a test MSE of 0.444 and correctly predicted BCS within +-1.0 units for approximately 86% of horses. Linear models, particularly PLS and Ridge regression, demonstrated the most consistent robustness, while ordinal regression achieved near-perfect tolerance within +-1.0 BCS at 97.6%. The results provide a proof of concept for objective automated BCS estimation from 3D equine morphology. The observed performance ceiling is most plausibly attributed to label noise, dataset size, and mixed-breed composition rather than a fundamental limitation of the approach. Future work should prioritise larger, breed-stratified datasets and more objective ground truth measures to build on this methodological foundation.