Cohort-level disease prediction by extrapolation of individual-level predictions in transition dairy cattle

Dairy cattle experience metabolic stress during the transition from late gestation to early lactation resulting in higher risk for several economically important diseases (e.g. mastitis, metritis, and ketosis). Metabolic stress is described as a physiological state composed of 3 processes: nutrient metabolism, oxidative stress, and inflammation. Current strategies for monitoring transition cow nutrient metabolism include assessment of plasma non-esterified fatty acids and beta-hydroxybutyrate concentrations around the time of calving. Although this method is effective at identifying cows with higher disease risk, there is often not enough time to implement intervention strategies to prevent health disorders from occurring around the time of calving. Previously, we published predictive models for early lactation diseases at the individual cow level at dry-off. However, it is unknown if predictive probabilities from individual-level models can be aggregated to the cohort level to predict cohort-level incidence. Therefore, our objective was to test different data aggregation methods using previously published models that represented the 3 components of metabolic stress (nutrient metabolism, oxidative stress, and inflammation). We included 277 cows from five Michigan dairy herds for this prospective cohort study. On each farm, two to four calving cohorts were formed, totaling 18 cohorts. We measured biomarker data at dry-off and followed the cows until 30 days post-parturition for cohort disease incidence, which was defined as the number of cows: 1) having one or more clinical transition disease outcome, and/or 2) having an adverse health event (abortion or death of calf or cow) within each cohort. We tested three different aggregation methods that we refer to as the p-central, p-dispersion, and p-count methods. For the p-central method, we calculated the averaged predicted probability within each cohort. For the p-dispersion method, we calculated the standard deviation of the predicted probabilities within a cohort. For the p-count method, we counted the number of cows above a specified threshold of predicted probability within each cohort. We built four sets of models: one for each aggregation method and one that included all three aggregation methods (p-combined method). We found that the p-dispersion method was the only method that produced viable predictive models. However, these models tended to overestimate incidence in cohorts with low observed counts and underestimate risk in cohorts with high observed counts.