Reliability, agreement and variability of a markerless computer vision algorithm for equine gait analysis under field conditions.

Abstract

BACKGROUND: Computer vision-based algorithms offer accessible alternatives for equine gait analysis but require thorough assessment under diverse conditions. OBJECTIVES: To evaluate a proprietary vision-based algorithm's reliability in measuring vertical displacement signals (VDS) at the eye, withers and croup, alongside groundline estimation, for horses trotting on straight lines and circles under field conditions. STUDY DESIGN: Cross-sectional comparative study evaluating agreement, variability and reliability of a markerless computer vision algorithm. METHODS: We obtained 67 handheld iPhone recordings from 37 horses. A vision-based algorithm and independent manual annotation produced 2D anatomical keypoints on all frames of the recordings, which were processed to estimate a groundline and compute VDS and stride-based maxima (Maxdiff) and minima (Mindiff) vertical differences. Mean signed error (MSE), mean absolute error (MAE) and Bland-Altman plots were used to compare detected and annotated data. RESULTS: The frame level vertical keypoint accuracy was 4.5 mm (eye), 5.5 mm (croup) and 11.8 mm (withers), and the manual annotation error was averaged at 2.7 mm. At the stride level (n = 1556), the overall mean absolute errors (MAEs) for both Maxdiff and Mindiff were 4.3 mm. The eye keypoint exhibited the lowest errors (2.9 mm Maxdiff, 3.0 mm Mindiff), while the withers error was 5.5 mm for both Maxdiff and Mindiff, and the croup showed 4.3 mm (Maxdiff) and 4.4 mm (Mindiff). Trial-level (n = 67) analysis, with below optimal number of strides per trial in this study, revealed lower overall absolute differences (Eye: 2.3 mm, Withers: 3.7 mm, Croup: 2.7 mm), indicating consistent performance across multiple strides. Subjective lameness scoring aligned with objective measures with some variation. MAIN LIMITATIONS: Groundline estimation accuracy was stress-tested on treadmill data in another study. Further clinical comparison with established gait analysis systems is recommended. CONCLUSIONS: The algorithm robustly measured vertical displacements under varied conditions.