Machine learning-based detection of subclinical and clinical ketosis in Holstein cows using sensor data during the...

Ketosis, a metabolic disorder in dairy cows, poses a risk of substantial economic losses, particularly when it progresses to clinical forms. Previous prediction models relied on smart farming data and binary classification, without incorporating risk factors such as calf birth weight. Therefore, we aimed to develop a multiclass classification model to differentiate non- (NK), subclinical (SCK), and clinical (CK) ketosis in Holstein cows by integrating behavioral indicators, cow-specific traits, and environmental variables. We hypothesized that integrating these diverse data sources would improve the ability of the model to accurately classify ketosis severity during the transition period. A total of 132 Holsteins were monitored for 21 d after calving using automatic monitoring (HR-TAG). Input features included activity, rumination time, calving age, calf birth weight, and calving season. Blood β-hydroxybutyrate concentrations were measured at eight time points, and cows were classified into NK (<1.2 mmol/L), SCK (1.2-2.9 mmol/L), or CK (≥3.0 mmol/L) groups based on the highest BHBA value recorded across the sampling period. Five machine-learning algorithms-K-nearest neighbors, decision tree, random forest, support vector machine, and extreme gradient boosting (XGBoost)-were trained on 70 % of the dataset and optimized using 10-fold cross-validation, and final model performance was evaluated on the remaining 30 % test set. XGBoost performed best, achieving an accuracy, sensitivity, specificity, F-measure, kappa, and an area under the curve of 0.959, 0.935, 0.966, 0.951, 0.918, and 0.950, respectively. Feature importance analysis identified calving age, calf birth weight, and calving season as key predictors for ketosis severity. These results demonstrate that sensor-based behavioral traits, together with cow-specific characteristics and environmental factors, enable accurate classification of ketosis severity and support the application of precision dairy technologies for early detection and tailored herd management.