Diagnostic milk biomarkers for predicting the metabolic health status of dairy cattle during early lactation

Abstract

ABSTRACT: Data on metabolic profiles of blood sampled at d 3, 6, 9, and 21 in lactation from 117 lactations (99 cows) were used for unsupervised k-means clustering. Blood metabolic parameters included β-hydroxybutyrate (BHB), nonesterified fatty acids, glucose, insulin-like growth factor-1 (IGF-1) and insulin. Clustering relied on the average and range of the 5 blood parameters of all 4 sampling days. The clusters were labeled as imbalanced (n = 42) and balanced (n = 72) metabolic status based on the values of the blood parameters. Various random forest models were built to predict the metabolic cluster of cows during early lactation from the milk composition. All the models were evaluated using a leave-group-out cross-validation, meaning data from a single cow were always present in either train or test data to avoid any data leakage. Features were either milk fatty acids (MFA) determined by gas chromatography (MFA [GC]) or features that could be determined during a routine dairy herd improvement (DHI) analysis, such as concentration of fat, protein, lactose, fat/protein ratio, urea, and somatic cell count (determined and reported routinely in DHI registrations), either or not in combination with MFA and BHB determined by mid-infrared (MIR), denoted as MFA [MIR] and BHB [MIR], respectively, which are routinely analyzed but not routinely reported in DHI registrations yet. Models solely based on fat, protein, lactose, fat/protein ratio, urea and somatic cell count (i.e., DHI model) were characterized by the lowest predictive performance [area under the receiver operating characteristic curve (AUCROC) = 0.69]. The combination of the features of the DHI model with BHB [MIR] and MFA [MIR] powerfully increased the predictive performance (AUCROC = 0.81). The model based on the detailed MFA profile determined by GC analysis did not outperform (AUCROC = 0.81) the model using the DHI-features in combination with BHB [MIR] and MFA [MIR]. Predictions solely based on samples at d 3 were characterized by lower performance (AUCROC DHI + BHB [MIR] + MFA [MIR] model at d 3: 0.75; AUCROC MFA [GC] model at d 3: 0.73). High predictive performance was found using samples from d 9 and 21. To conclude, overall, the DHI + BHB [MIR] + MFA [MIR] model allowed to predict metabolic status during early lactation. Accordingly, these parameters show potential for routine prediction of metabolic status.

Keywords

• metabolic status
• milk composition
• predictive modeling
• dairy cattle