Machine learning-based detection of individual cow global health using MIR-predicted traits and big data

Global health in dairy cows reflects health status and production traits, indicating overall farm performance. Traditional monitoring methods are labor-intensive, underscoring the need for data-driven approaches. Milk mid-infrared (MIR) spectra effectively predict production traits and health-related biomarkers in dairy cows. This study aimed to: (1) identify individual-level global health cluster using unsupervised hierarchical clustering with 35 MIR-predicted traits; (2) develop a predictive model for global health classification with four supervised machine learning algorithms and 35 MIR-predicted traits; and (3) assess the feasibility of directly predicting global health classification from MIR using two supervised algorithms. A total of 27,765,481 records from 11 cattle breeds, each with 35 MIR-predicted traits, were processed. For unsupervised clustering, a balanced subset of 5,845,345 records was used. The global health cluster was defined by calculating the mean values of the 35 MIR-predicted traits within each cluster from the full dataset. For the predictive model, the dataset was split (7:3) for calibration and validation. Four supervised learning algorithms—partial least squares discriminant analysis (PLS-DA), support vector machine, neural network, and random forest (RF)—were used to develop prediction models. Five clusters were identified, with the fourth and fifth clusters considered the global health cluster. In validation, the prediction accuracy of the five clusters using 35 MIR-predicted traits ranged from 0.70 (PLS-DA) to 0.83 (RF). Directly predicting the five clusters from MIR achieved 0.75 accuracy with PLS-DA and 0.87 accuracy with RF in validation. Our findings suggest that MIR-predicted traits can be used to assess individual-level global health in dairy cows and demonstrate the potential for directly predicting the five clusters from MIR. This newly defined global health cluster may provide opportunities for genomic evaluation of global health in cattle; however, further validation in practical applications is necessary.