Random forests prediction of blood metabolic clusters of dairy cows: comparing three types of milk biomarkers

Milk biomarkers as predictors of physiological imbalances and subclinical or clinical diseases in dairy cows may be used in management and selection. Cows were categorised into three clusters differentiating the metabolic status and defined by clustering of plasma hormone Insulin-like Growth Factor-1 (IGF-1) and three plasma metabolites: Glucose, beta-hydroxybutyrate (BHB), and non-esterified fatty acids (NEFA). We aimed to compare prediction of cluster category by use of milk metabolites and enzymes, selected milk MIR spectra wavenumbers, and milk IgG glycan profiles. Data between calving and 50 days in milk (DIM) were obtained from 235 Holstein Friesian cows in six research herds: AU (Denmark), UCD (Ireland) AFBI (UK), CRA-W (Belgium), FBN (Germany), CREA (Italy). Some diets were designed to challenge the cow and provoke production diseases but there were too few incidences to establish predictions. Instead, k-means clustering was performed over scaled residuals from linear mixed effects models for each of IGF-1, Glucose, log10(BHB) and log10(NEFA) measured around 14 and 35 DIM. Biomarker predictors were: Averages from 2 milkings/week of milk metabolites (free glucose, glucose-6-phosphate, BHB, isocitrate, uric acid, urea) and enzymes (LDH, NAGase) from same week as blood sampling; Closest MIR spectra sample within a limit of +/-3 days from blood sampling date; IgG glycans sampled 14 and 35 days post-calving used for corresponding period. Random forests (RF) predictions by one biomarker-type and parity (1, 2 and 3+) were evaluated by leave-one-cow-out (internal) cross–validation. Results are shown in Table 1 as proportion of correctly classified observations (overall accuracy, i.e. diagonal of confusion matrix) for each of the three biomarkers and for each period (DIM14 and DIM35), and sensitivity, specificity and balanced accuracy (average of these) for each cluster. We obtained predictions of metabolic clusters from milk biomarkers of fair accuracy. Sample sizes may constraint both estimation of clusters, number of classes, and RF training. External validation in separate herds is warranted. Still, we see potential in using milk biomarkers to support management of dairy cows. Milk metabolites/enzymes resulted in slightly better accuracy of RF predictions but other methods should be investigated too. Practical matters may also influence choice of biomarkers.