| Abstract |
Monitoring biomarkers in milk or blood is widely used to assess dairy cow health and to detect diseases such as (sub)clinical ketosis. Blood β-hydroxybutyrate (BHB) is a key indicator of metabolic imbalance, and its prediction from milk mid-infrared (FT-MIR) spectra has been extensively explored. However, most of the predictive models applied in spectrometry fail to achieve acceptable accuracy due to the highly skewed distribution of BHB concentrations (many blood samples have low concentrations while a few have high concentrations). Imbalanced distributions complicate both quantitative and class modeling. Thus, the objective of this work was to undertake discrete predictions of BHB (low versus high level) by implementing a specific weighting in a usual discrimination algorithm to overcome the class imbalance between healthy cows and those suffering of hyperketonemia. The threshold of 1.2 mmol/l was used to distinguish low from high BHB, i.e., absence or presence of hyperketonemia. The cleaned data set combined 4,221 records comprising milk FT-MIR spectra and blood BHB content, that were collected across 10 countries between 2013 and 2024, and in multiple breeds, and numerous research projects. Milk MIR spectra (selected wavenumber regions, second derivative Savitzky–Golay preprocessing) served as predictors. A replicated external herd validation strategy was implemented. The weighted model outperformed the standard model, achieving average intra-class error rates of around 18% on external test sets. It reached 82% accuracy, 83% sensitivity and 81% specificity, with these results comparable to or better than previously published discriminant models, despite using only raw MIR spectra. These findings demonstrate that the use of weighting within discriminant methods is a valid option for the class prediction of asymmetrically distributed health biomarkers. |
| Authors |
Lesnoff, M., Grelet, C., Wolf, V., Millot, L., Köck, A., Leblois, J., Bapst, B., Crowe, M.A., Larsen, T., Wathes, D.C., Ferris, C.P., Ingvartsen, K.L., Marchitelli, C., Becker, F., Hummel, J., Gelé, M., Soyeurt, H., Gengler, N., Dehareng, F. |
