Abstract
Fourier-transform mid-infrared (FT-MIR) spectroscopy is a powerful tool to predict fine milk composition but also indirect traits such as cow health at a large scale. Previously, an unsupervised learning approach applied to primiparous Holstein cows identified a cluster associated with negative energy balance and elevated concentrations of blood BHB and milk acetone, which are recognized indicators of metabolic disruptions. This study extended the methodology to multiparous cows, representing a broader variability. A hierarchical clustering pipeline was applied to over 1,712,417 records from primiparous and multiparous Holsteins, using 27 FT-MIR-based predictions, SCS, and milk yield. The approach was repeated 5 times for 4 different datasets: first, second, and third lactations, and a merged dataset for a total of 20 iterations. Across those iterations, the metabolic cluster was consistently found, and 2 partial least square discriminant analysis prediction models were estimated for each iteration using either the predictions or directly the spectra. The correlations between these predictions of all the models were between 0.82 and 1, suggesting that the models are similar. The cluster previously associated with udder health conditions was not consistently found across repetitions. To ensure transferability and facilitate the deployment of the results, a final model based on FT-MIR was estimated, presenting good performance with an accuracy in validation of 0.92. Those results prove that the metabolic cluster is extendable to other lactations and predictable directly from milk FT-MIR spectra.