39 Using artificial intelligence to analyze horse gait parameters for genomics research in musculoskeletal traits

Arguably the most important factors considered when assessing elite performance horses, traits of “lameness susceptibility” and “gait quality” lack objective and quantitative measures. Popular qualitative approaches to determine lameness or gait quality do not provide accurate measures amenable to genetic study, nor can producers easily apply them. However, digital video analysis of gait utilizing recently developed artificial intelligence based approaches holds significant promise to meet these industry needs. It also provides quantitative endophenotypes, the comparison of behavior to phenotypes through genetic correlation, with sufficient precision for genome-wide association studies. Here we describe pilot work using consumer level digital video cameras to capture high-resolution and high speed videos of horses at the trot during mandatory veterinary inspection for Federation Equestre Internationale (FEI) level competitions. Twenty-three key skeletal landmarks, selected based on gait mobility, on each horse in 20 different frames of the video are labeled with DeepLabCut, a software package applying a deep neural network approach for object recognition in the frame image. Raw quantitative gait parameters such as stride length, stance time, and angular range of the fetlock joint were derived from the landmark tracking data by a robust data processing pipeline. The pipeline comprises many operations, including outlier removal due to occlusion, trajectory smoothing, model-fitting-based phase identification, and gait parameters calculation. The ease of video capture allowed for 1,597 samples of 945 horse and rider combinations from 5 FEI competitive levels within the same discipline of eventing to be captured from just 6 competitions around the southeast US, before closing of travel due to the pandemic. The current training run was applied to 1,280 frames from 64 videos with 23 landmark points plotted per frame. This training will be applied to our database of 1,533 additional videos allowing gait parameter analysis for horses exhibiting 3–5 trot strides in a single direction. Future work will incorporate additional trot strides, both directions of travel, and enlarging our sampling of horses. This database will be used in future applications in animal selection and lameness diagnosis, as well as in construction of quantitative performance traits suitable for genomics research.