Using kinematic to quantify gait attributes and predict gait score in dairy cows.
Authors: C. Julliot and G. M. Dallago and A. Nejati and A. B. Diallo and E. Vasseur
Date: 2023-06-01
Status: Published
Early identification of gait abnormalities could reduce lameness in dairy cows, but current assessment methods are laborious, making their use prohibitive for continuous monitoring. The objective of this pilot study was to predict the gait scores of dairy cows based on the gait attributes obtained using kinematic data. Data were collected from 12 Holstein cows between January 18 and February 12, 2021, using reflective markers attached to 20 anatomical locations. Cows were walked multiple times in a 7-m passageway corridor with 3 cameras on each side, totaling 69 passages of 3 steps long each. Five gait attributes of distance, duration, and velocity were calculated using 3D coordinates of the hoof markers. Range of motion was measured based on the angle between the stifle, hock, and fetlock markers of the rear legs. A trained observer scored cows’ gait on a 5-point numerical rating system (NRS), with scores ranging from 1 (sound) to 5 (severe lameness). Passages with NRS scores of 1.5 (n = 1) and 4 (n = 2) were removed since there were insufficient samples to represent these scores. The data were split into training and validation sets following a 70:30 ratio stratified by the distribution of the NRS scores. The machine learning algorithms RPART, GBM, XGBM, RF, and SVM with a radial basis kernel were trained using leave-one-out cross-validation. In addition to the original data, weighted classification and synthetic minority over-sampling technique (SMOTE) were also tested due to uneven NRS distribution. The models were evaluated according to their accuracy, sensitivity, specificity, F1 score, and balanced accuracy on the validation split. The best model was the XGBM trained using the original data, which achieved an overall accuracy of 0.66 (95% CI = 0.53–0.78). Conversely, the weighted RPART classification model had the lowest overall accuracy of 0.44 (95% CI = 0.31–0.57). The insights from this pilot study contribute to developing an automatic monitoring system to identify and treat cows that have impaired locomotion but are not yet clinically lame, allowing for improved welfare and profitability of dairy cows.