A data-driven physiological profile classification for dairy cows using milk mid-infrared spectra and machine learning.

Abstract

Composite physiological status in dairy cattle reflects the integration of production, metabolic, and clinical information, yet scalable individual-level assessment remains limited. Milk mid-infrared (MIR) spectroscopy, routinely collected in dairy systems, provides a high-throughput opportunity to capture physiology-related signals. This study aimed to develop a data-driven classification framework using MIR-derived information and to evaluate its relevance with an independent score global individual de bien-être (SGI). A total of 27,780,928 records from 11 dairy cattle breeds were analyzed. Unsupervised hierarchical clustering was performed using 35 MIR-related traits in the subdata set (n = 5,478,789 records), and 3 distinct cow clusters were identified. Supervised machine learning models, including partial least squares discriminant analysis (PLS-DA) and random forest (RF), were then developed to predict clusters. The RF model outperformed PLS-DA, achieving an accuracy of 0.88 using MIR-related traits. Direct prediction from MIR spectra reproduced the MIR-derived cluster labels with comparable performance (accuracy = 0.89), suggesting that the intermediate trait-prediction step may be simplified. In the independent ScorWelCow interpretation data set (n = 1,870), Cluster 3 showed the highest mean SGI, supporting its interpretation as the most favorable MIR-derived composite physiological profile. These results suggested that MIR-derived information could contribute to a biologically meaningful, scalable composite physiological profiling framework at the individual cow level, although further validation with independent clinical or physiological reference data is required.