Abstract
This study assessed five supervised machine learning (ML) models. Automated devices that continuously captured milk composition and behavioral data were used to monitor 206 Holstein cows from two commercial dairy farms. Milk yield, fat, protein, lactose, fat-to-protein ratio (FPR), somatic cell count (SCC), rumination time (RT), and body temperature were among the parameters that were noted. Cows were categorized as clinically healthy (n = 45), subclinical ketosis (n = 91), subclinical mastitis (n = 28), or clinical mastitis (n = 42) based on clinical examination, blood β-hydroxybutyrate (BHB) concentration, and milk indicators. Random Forest achieved the highest classification accuracy (0.857), followed by Gradient Boosting and Logistic Regression (0.833), while Decision Tree and Multilayer Perceptron reached 0.810. Compared to clinically healthy cows (4.45 ± 0.54%; 477.0 ± 36.0 min/day), subclinical ketosis cows had a greater milk fat content (5.21 ± 0.72%) and a shorter RT (336.9 ± 94.2 min/day). In comparison to clinically healthy cows (64.0 × 10(3) cells/mL; 4.63 ± 0.16%), cows with clinical mastitis showed significantly greater SCC (416.8 × 10(3) cells/mL) and lower lactose levels (4.56 ± 0.24%). These results demonstrate that integrating sensor-derived milk and behavioral data with ML algorithms enables early, non-invasive detection of health disorders, supporting proactive herd management.