Predicting postpartum diseases in Holstein cows using milk spectra and machine learning-Retrospective assessment from diagnosis date.

Abstract

Dairy cows commonly experience health disorders in the early-lactation period. Although Fourier-transform infrared (FTIR) spectroscopy offers a noninvasive and cost-effective method for analyzing milk composition, its potential in predicting subsequent early-lactation diseases has yet to be adequately explored. This study aimed to uncover the ability of milk FTIR spectra to predict postpartum diseases in 1,162 Holstein cows from a commercial dairy farm in Cayuga County, NY. We collected proportional milk samples daily on cows in the early-lactation pen and stored milk at 4°C until analysis via FTIR spectroscopy. Cows were monitored through 30 DIM and classified as healthy (n = 825; no adverse health events) or diseased (n = 311; diagnosis of clinical ketosis, metritis, displaced abomasum, or mastitis, or any combination of these). We developed predictive models for 8 distinct time periods preceding the diagnosis date (>10 d, 10-8 d, 7-6 d, 5-4 d, 3 d, 2 d, 1 d, and 0 d), using regression, machine learning, and deep-learning methods applied to milk FTIR spectral data. Model performance was evaluated through a repeated down-sampled double cross-validation framework and permutation tests. Our results showed that progressive changes in spectral regions related to the absorbance peaks of fat, protein, and lactose are correlated with disease progression, leading to an increase in average area under the receiver operating characteristic curve (AUROC) from 0.50 (>10 d before diagnosis) to 0.72 (1 d prior) and 0.76 (the day of diagnosis) across all model types. Partial least squares-discriminant analysis (PLS-DA) models using milk FTIR spectra achieved an average AUROC of 0.71 from 7 d before diagnosis, outperforming models based on cow-level features (0.62) or combined with spectra-predicted milk major components (0.67). Among spectral models, PLS-DA reached the highest average AUROC (0.74), followed by long short-term memory (0.72), and surpassed ridge regression (0.71) and random forest (0.69). These findings highlight the effectiveness of using milk FTIR spectra to predict upcoming health conditions in early-lactation Holstein dairy cows, although broader evaluation is necessary to assess generalizability and on-farm utility.