Using 3D-kinematics in conjunction with machine learning approaches to predict dairy cow locomotor ability

Objectives

Lameness is a prevalent issue within the dairy industry that has serious financial and welfare implications. Traditional locomotion scoring systems conducted by a visual observer, such as a numeric rating system (NRS), provide a simple way to assess dairy cow gait; however, this method is prone to low reliability and is relatively subjective when compared to more automated approaches for assessing lameness. The objective of this validation study was to determine if kinematic data acquired for 20 specific joints on a cow when walking could be used with machine learning approaches to predict a locomotion score with high accuracy. Specifically, in continuation of previous work using similar kinematic data with a Convolutional Neural Network (CNN), we aimed to develop a model with Long Short-Term Memory (LSTM) architecture that could predict locomotion scores of dairy cows from kinematic data. The hypothesis was that kinematic data would reflect the gait quality of cows that was visually observed through locomotion scoring.

Materials and methods

The 3D-scaled coordinates of 20 specific marked joints on the cow were acquired through a kinematic system consisting of six video cameras and a motion tracking software. Seventy-four recorded passages, in which a cow walked the duration of a 7m designated walkway, from among 12 cows were retained for analysis and processed. Locomotion scoring was conducted according to a 5-point NRS scale with 0.5 intervals (9 gait classes) in which a score of 1 represented the highest quality locomotion, a score of 3 or higher represented clinical lameness, and a score of 5 represented severe lameness. A trained observer assigned scores for individual passages using corresponding video recorded through the kinematic system. Ultimately, across all passages, assigned scores represented 4 consecutive classes of gait quality ranging from scores of 2 to 3.5. [...]