PSXII-4 Using deep learning to evaluate the relationship between gait and dressage performance in sport horses.

Abstract In the equestrian sport of three-day eventing, a panel of competition officials evaluates the movement characteristics for each horse both prior to and during the competition. These evaluations, colloquially known as “jogs,” determine whether or not a horse is fit enough to complete the upcoming phases of competition. Yet, as the state of the art for these evaluations, and for clinical lameness exams, is observation by a skilled observer, this is not a repeatable or unbiased process. During an eventing competition, “jogs” are completed immediately before and after the first “phase”, which is Dressage. Dressage scores are of particular interest due to the fact that they are a continuous and numerical assessment, producing a score out of 100 reflecting the execution and quality of a predetermined set of movements completed by the horse and rider. The goal of this analysis is to better understand how competition affects gait quality and examine how gait quality affects dressage scoring, in the hope of introducing more objectivity to the soundness evaluation process. Utilizing an analysis pipeline previously constructed for evaluations of rodent gait mechanics, and a custom DeepLabCut labeling model trained on 3,120 individual video frames of horses in motion. For this experiment, we processed a total of 194 “jog” videos of individual horses at the trot, collected at a total of five venues across the southeast United States from 2019-2022. Our custom DLC model applies twenty-two anatomical keypoints representing notable locations on the equine skeleton and tracks these key points’ movement in pixels throughout the duration of the video. Using raw data generated from DLC, our custom pipeline calculates seven distinct gait parameters: Duty Factor, Fetlock Angle Range, Forelimb Swing Range, Hindlimb Swing Range, Stride Length, Forelimb Travel Distance, and Hindlimb Travel Distance, each standardized to account for bias resulting from venue, differences in horse size, and level of competition. A generalized linear mixed model (GLMM) compared each individual’s gait parameters assessed by video to their dressage score at the same competition, execued three-four days later. GLMM indicated a moderately strong positive correlation between stride length and dressage score (p=0.00213). The p-value for Duty Factor was also close to significance at p=0.06895. Further refinement of these results through implementation of a Generalized Additive Model (GAM) showed that all gait parameters’ relationships to Dressage Score were more accurately described by nonlinear models (with the exception of Stride Length, Forelimb Travel Distance, and Fetlock Angle Range, edf > 1.000). However, the GAM confirmed a strong correlation between stride length and dressage score when compared alongside other gait parameters’ effects on score (p=0.000361). GAM also indicates that there is a weaker, yet not insignifcant, correlation between an individual’s Duty Factor Average and their dressage score (p=0.06299), corroborating the aforementioned results. The findings suggest that gait quality and competition performance, specifically in dressage, are intimately linked. Further genetic analysis might allow for these advantageous gait qualities to be quantified.