Milk Spectra Meta-Clustering Identifies Dairy Cow Health Groups

Clustering techniques have proven valuable in identifying distinct groups within biological data, and Fourier-transform infrared (FTIR) spectroscopy, particularly mid-infrared (MIR) analysis of milk, is increasingly used to predict various milk traits and health conditions in individual dairy cows. This study explored the direct application of clustering to milk MIR spectral data to uncover latent groups of cows linked to specific milk traits or health disorders, with the goal of enabling early prevention or monitoring of at-risk animals. Researchers utilized a large dataset comprising over 400,000 individual milk MIR records from thousands of commercial farms. They combined several advanced data processing steps: spectral filtering to select informative wavenumbers, two dimensionality-reduction methods (Principal Component Analysis and an autoencoder), and two clustering algorithms (k-means and spectral clustering). This yielded eight different clustering approaches. The key innovation was regrouping the assigned clusters into "meta-clusters" that consolidated the most similar findings across all eight approaches. The analysis revealed five distinct meta-clusters of early-lactation dairy cows, each significantly associated with specific milk traits. These meta-clusters appeared to represent a gradient of negative energy balance (NEB) severity—severe, moderate, and possibly mild—with the remaining two clusters likely indicating cows recovering from NEB. Notably, despite the diversity of clustering methods employed, they largely converged on these same five meta-clusters, demonstrating the robustness of the findings. Crucially, the computationally efficient PCA-based k-means approach using the full spectrum was able to recapture the insights found by more complex methods.