Evaluation of three classification models to predict risk class of cattle cohorts developing bovine respiratory disease within the first 14 days on feed using on-arrival and/or pre-arrival information

Bovine respiratory disease (BRD) remains the leading cause of morbidity and mortality in feedlot cattle. At feedlot arrival, classification of cattle groups into high- or low-risk based on their expected level of BRD is common, highly variable, and based on many subjective criteria. An accurate objective classification methodology would provide a tool to more efficiently allocate resources and promote judicious use of antimicrobial therapy. The objective of this research was to evaluate the diagnostic performance of three classification algorithms to classify cattle into risk classes based on the expected BRD morbidity in the first 14 days on feed (DOF) and to evaluate if data collected at the sale barn would provide information useful to increase classification performance. Data from 141 lots representing 618 purchase groups and 35,027 animals were used to predict the BRD risk class of cattle groups on arrival at the first management location (lot) following purchase. Sale barn, lot-level, and weather variables at each location were used to determine the combination of data most beneficial to diagnostic performance. Three classification algorithms were evaluated for their diagnostic performance (accuracy, sensitivity, specificity) in classifying cattle groups into risk classes based on three BRD morbidity cutoffs (2%, 4%, 6%) within the first 14 DOF. Bootstrapping methods were applied to estimate confidence intervals around the diagnostic performance point estimates. The predictive performance of individual algorithms varied by different cutoffs in BRD morbidity within the first 14 DOF and the predictors provided to the algorithms. The median morbidity within the first 14 DOF was 2.1% and using a 2% cutoff to classify cattle groups into high- or low-risk, using only lot level information provided the highest accuracy and specificity and was as good as the same model trained with additional lot and sale barn information with respect to sensitivity. At the 4% cutoff, the lot level dataset also provided the highest accuracy and sensitivity and the same level of specificity as using the full dataset. With a limited dataset, using cutoffs in BRD morbidity within the first 14 DOF of 2% and 4%, we found collecting sale barn data did not provide any additional benefit over collecting only on-arrival data with respect to classifying lots of cattle into high- or low-risk. A 6% cutoff was not useful due to the highly imbalanced dataset that is created with respect to our outcome of interest.