Can we get information on dairy cow chronic stress biomarkers from milk mid-infrared spectra?

In recent years, animal welfare considerations have increased for citizens and public authorities. Within the several dimensions of welfare, chronic stress is particularly difficult to measure objectively, while in its chronic form, it may cause metabolic, inflammatory, and infectious diseases, fertility problems, or lower milk production. Therefore, the goal of this work was to evaluate the potential of milk mid-infrared (MIR) spectra for predicting at the individual cow level the content of 2 chronic stress biomarkers: hair cortisol and blood fructosamine. A dataset was specifically created by organizing a dedicated large-scale sampling. Sampling was performed in commercial dairy farms by several regional milk recording organizations in Austria, Belgium, France, Germany, and Luxembourg for a total of 1,412 individual cow records. Each of the 1,412 cows was sampled 1 time for milk, hair, and blood. Milk MIR analysis was conducted locally on a total of 30 standardized spectrometers. After removing the spectral and reference missing values and cleaning data, the final dataset comprised 1,071 hair cortisol values and 940 blood fructosamine values. Qualitative models were tested to discriminate low versus high contents of hair cortisol and blood fructosamine for each individual dairy cow from the milk MIR spectra through the use of partial least squares discriminant analysis. Four quantitative modeling strategies were also tested to predict the exact quantitative value of the 2 biomarkers. All models were evaluated in an external herd validation process by iteratively excluding 33% of herds (i.e., 26 herds out of 78). The best model discriminating high hair cortisol values for individual dairy cows showed a global accuracy of 71%, with a specificity of 74% and a sensitivity of 61%, and this model included MIR spectra, with milk yield, parity, square of parity, DIM, square of DIM, and breed as predictors. The best discrimination of high fructosamine values demonstrated a global accuracy of 73%, with a specificity of 75% and a sensitivity of 67% and was based on the combination of 2 models. When testing the 4 quantitative methodologies, the best R² were 0.13 for hair cortisol and 0.2 for blood fructosamine in external herd validation, showing a poor capacity of milk MIR spectra to predict the exact quantitative value of the 2 biomarkers. While the quantitative models did not supply satisfying results, the qualitative models provided accuracies of 71% and 73% in a robust external herd validation, respectively. Therefore, the accuracy obtained, even if not very high, can be considered very positive given the difficulty of predicting these totally indirect biomarkers of chronic stress. While there are currently no possibilities to get objective information on chronic stress in a routine frame, the prediction of both hair cortisol and blood fructosamine content, at the individual cow level, may facilitate large-scale chronic stress assessment and improvement by providing objective, cheap, and quantitative information.