Cohort-level disease prediction using aggregate biomarker data measured at dry-off in transition dairy cattle: A proof-of-concept study.

During the transition from late gestation to early lactation, dairy cattle are at increased risk for disease. Herd-level monitoring for disease risk involves evaluating multiple factors, including food intake, cow density, and biomarkers of nutrient metabolism. Biomarkers that are measured include non-esterified fatty acids (NEFA) and beta-hydroxybutyrate (BHB), which are usually measured in a subset of the herd (i.e. cohort). If a certain proportion of cows in the cohort are above a specific threshold for a biomarker, the cohort is considered at high risk of disease. Few previous studies have investigated other methods to aggregate individual cow-level data to the cohort level. We designed a proof-of-concept study to determine if biomarker aggregation methods may be useful to predict cohort incidence of adverse health events including 1) clinical diseases: mastitis, metritis, retained placenta, ketosis, lameness, pneumonia, milk fever, displaced abomasum, 2) and abortion or death of the calf or the cow. The study design was a prospective cohort study that used cows (N = 277) from five Michigan commercial dairy herds. Multiple cohorts of cows (two to four cohorts per farm, 18 total) were enrolled that shared the same dry-off date. We tested three different methods (central, dispersion, and count) to aggregate individual cow data (i.e. biomarkers and covariates) measured at dry-off. The central method consisted of calculating the average value of each variable within a cohort, and the dispersion method involved taking the standard deviation or mean absolute deviation about the median of each variable within a cohort. The count method consisting of counting the number of cows above a specific threshold for each variable within a cohort. We used best subsets selection to select a bouquet of candidate models for each aggregation method and averaged the predictions over the model set. We built 4 sets of Poisson regression models: one for each aggregation method and a combined model that included all three methods. We evaluated the models based on goodness-of-fit, model calibration using scoring rules, and comparison of observed versus predicted counts. The central and the combined method produced models that had good fit and model calibration. These results indicate that it may be possible to use aggregate measures to predict cohort disease incidence as early as dry-off. The next step is to test biomarker aggregation methods in studies with larger sample sizes.