Predicting dairy cow locomotion ability based on kinematic 3D coordinates

• Early detection of gait abnormality is needed to prevent clinical cases of lameness. • Data augmentation was used to expand the sample size and create a balanced data set. • LSTM models were trained to predict dairy cow gait scores using 3D kinematic data. • Zero-mean, unit-standard deviation normalization outperformed min-max scaling. • The best model achieved a 0.96 accuracy, precision, recall, and F1 score. Our study addresses the challenge of early lameness detection in dairy cows. Traditional visual scoring methods, while non-invasive and cost-effective, require extensive training and are impractical for continuous monitoring. Our research proposes an automated alternative using kinematic data and machine learning. Kinematic data were collected multiple times from 12 Holstein dairy cows over four weeks. After data cleaning, a total of 73 passages were available for model training. A trained observer scored the gait of each passage using a numerical rating system (NRS) ranging from 1 (sound cow) to 5 (severe lame cow) with 0.5 intervals. Data augmentation was used to obtain balanced data sets by adding 1 %, 2.5 %, 5 %, 7.5 %, and 10 % Gaussian noise along with random shifting and followed by two data normalization strategies. The augmented data was split into training (75 %) and testing (25 %) sets. A long short-term memory neural network was trained and evaluated. The highest accuracy, precision, recall, and F1 score achieved on the test set was 0.96 (SD = 0.03) for all metrics. Models trained with data normalized to a mean of zero and standard deviation of one outperformed those using normalization to a range between zero and one. Future research should focus on expanding the range of locomotion scores, particularly covering the early stages of locomotion changes. This is necessary to enable earlier identification and treatment of cows with impaired locomotion ability before they develop lameness.