Using machine learning and precision livestock farming technology for early indication of health status in preweaned dairy calves

The objective of this study was to determine which algorithm, using machine learning techniques, accurately identifies calves destined to be positive for Bovine Respiratory Disease (BRD) status. We followed the health of 106 calves for 53±2 days using manual health scoring. Calves wore an accelerometer on the rear leg which recorded daily lying time, lying bouts, step counts, and an activity index. Calves were offered up to 10 L/d milk replacer by an automated milk feeder which recorded daily milk intake, drinking speed, and feeder visits. Of these calves, 54/106 were diagnosed with BRD based on two abnormal scores from the Wisconsin Health Scoring, and 3 cm2 of lung consolidation (day 0). First, we evaluated the potential of Ridge Classifier, linear SVM with Stochastic Gradient Descent learning (SGD), Gaussian Naïve Bayes, Decision trees, Adaboost classifier, and a K-Nearest Neighbor’s algorithm (KNN) for accuracy to classify data as positive or negative for BRD status using different window sizes (3 to 14d). The KNN and Decision Tree algorithms were the most accurate and had high precision and recall for BRD labelling correctly using 14d window size. For experiment II, two accurate algorithms were tested to identify calves destined to be positive to BRD in 7d prior to diagnosis using PLF variables. We found that KNN was highly accurate (80%) at classifying data as pre-sick up to -3d prior to BRD diagnosis. In summary, automatically collected behaviours and the use of KNN were found to have the potential to identify BRD in calves.