"The transition period is defined as the time that spans approximately 3 weeks before to 3 weeks post-partum. During the transition period, dairy cattle experience tremendous physiological changes that occur to prepare for milk production. Cows that transition poorly undergo excessive metabolic stress, which increases the risk for disease. Metabolic stress is a physiological state composed of 3 components: inflammation, oxidative stress, and nutrient metabolism. Cows undergoing metabolic stress are typically identified by measuring biomarkers, monitoring feed intake, and reviewing health records. The biomarkers that are typically used to monitor metabolic stress are non-esterified fatty acids (NEFA) and beta-hydroxybutryrate (BHB), which are nutrient metabolism biomarkers. Although NEFA and BHB are effective for identifying cows undergoing metabolic stress, they are measured too close to calving to make proactive changes in management to reduce disease incidence in the current calving cohort. Therefore, our objective was to build predictive models for transition cow diseases using all 3 components of metabolic stress at dry-off, which would allow more time to intervene with proactive interventions. To address this objective, we designed a prospective cohort study carried out on 5 Michigan herds (N = 277 cows). We randomly selected cows to form 18 calving cohorts, which were defined as cows at the same stage of lactation that were expected to calve at approximately the same time. We followed the cows until 30 days post-partum to monitor for disease occurrence. Disease was defined as an aggregate outcome where a cow was defined as positive if she was diagnosed with 1) one or more clinical transition disease outcome (mastitis, metritis, retained placenta, ketosis, lameness, pneumonia, milk fever, displaced abomasum) and/or 2) one or more adverse health events (i.e. abortion, death of calf or cow). Our first specific aim was to build individual cow-level models, which are presented in Chapter 2. We built separate model sets for each component of metabolic stress, and a model set that included all 3 components. We hypothesized that the combined model would be better able to predict transition cow diseases. Our hypothesis was supported and the area under the curve calculated from receiver operator curve curve analyses was significantly higher for the combined model compared to the separate model sets (P < 0.05). Our second specific aim was to build cohort-level models, which we addressed in Chapters 3 and 4. In Chapter 3, we tested ways to aggregate individual-level biomarker and covariate data to the cohort-level. We tested 3 different methods that we referred to as the "central," "dispersion," and "count" methods. We also tested combining all 3 methods for prediction. We hypothesized that these methods would be effective at aggregating data to the cohort level. We found that the central method, which aggregated data either by calculating the mean (for continuous variables) or the median (for categorical variables) of values within a cohort, was the only method that produced viable models. Chapter 4 tested if individual level predictive probabilities generated from models in Chapter 2 can be aggregated to predict disease incidence at the cohort-level. We tested 3 different methods to aggregate predicted probabilities which we referred to as the "p-central," "p-dispersion," and "p-count" methods, and we tested the efficacy of combining all 3 methods. We hypothesized that the models would produce valid disease estimates. We found that the p-dispersion method was the only method that produced viable models, which involved calculating the standard deviation of predicted probabilities within a cohort."--Pages ii-iii.
Read
