Prediction and classification of metritis and mastitis in Holstein cows using transition milk spectra under different modeling strategies

Metritis and mastitis are common early-lactation diseases of dairy cows that reduce milk production. Early prediction enables timely intervention and management, yet no studies have investigated the ability of milk Fourier-transform infrared (FTIR) spectroscopy for predicting the onset and development of metritis or mastitis within the first 2 wk postpartum. Our study aimed to assess the potential of milk FTIR spectra for early detection of postpartum metritis and clinical mastitis and to describe their spectral variations as lactation advances and diseases progress. Holstein cows (n = 1,103) from a commercial dairy farm in Cayuga County, New York, were monitored through 14 DIM and classified as healthy (n = 784; no adverse health events) or as diagnosed with metritis (n = 57; diagnosis of metritis but not ketosis, displaced abomasum, or clinical mastitis within 14 DIM) or clinical mastitis (n = 72; diagnosis of clinical mastitis but not ketosis, displaced abomasum, or metritis within 14 DIM). We constructed models for predicting and classifying postpartum metritis and mastitis using pooled, multiblock, and single-day partial least squares discriminant analysis (PLS-DA) strategies, assessed with repeated leave-one-out cross-validation and permutation tests. Across all modeling strategies, metritis was more distinguishable than mastitis, a pattern that corresponded with increasing fat and decreasing protein and lactose absorbance in transition milk from cows developing metritis. In the pooled strategy, models using spectra from DIM 1 to 7 achieved average area under the receiver operating characteristic curve of 79.4% for identifying metritis from healthy cows and 79.0% for distinguishing metritis from mastitis, whereas mastitis prediction reached only 60.7%. The multiblock and single-day PLS-DA models showed similarly strong performance for metritis (up to 79.2%) but failed to detect mastitis reliably. Furthermore, the added value of FTIR spectra for metritis prediction appeared contingent on sufficient sample size, as demonstrated by down-sampling experiments in the pooled strategy (with the down-sampled ratios of 80%, 60%, 40%, 20%, 10%, 5%), where models with spectral data outperformed those without only at or above 40% sampling. We conclude that transition milk FTIR spectra within the first 7 DIM showed disease-related signatures that may support early identification, although performance varied with sample size and modeling strategy, and multiherd validation is required to confirm generality and practical value.