Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag Original papers 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 David E. Amrine a, ⁎ , Jiena G. McLellan a , Brad J. White a , Robert L. Larson a , David G. Renter b , Mike Sanderson b a Beef Cattle Institute, College of Veterinary Medicine, Kansas State University, Manhattan, KS 66506, United States b Center for Outcomes Research and Epidemiology, College of Veterinary Medicine, Kansas State University, Manhattan, KS 66506, United States ARTICLE INFO Keywords: Bovine respiratory disease Predictive models Risk classification ABSTRACT 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 metho- dology 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 al- gorithms 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 classifi- cation 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 (ac- curacy, 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. 1. Introduction Bovine respiratory disease (BRD) continues to be the most common cause of morbidity and mortality in weaned cattle (Woolums et al., 2013). The cumulative incidence of BRD in feeder cattle in the U.S. feedlots increased from 14.4% in 1999 to 16.2% in 2011 and costs associated with each case have almost doubled from US$12.59 per case in 1999 to US$23.60 in 2011 (USDA, 1999, 2013). Upon arrival to a feedlot, cattle are managed in groups and the decision to alter arrival management procedures such as administering antimicrobial meta- phylaxis is frequently based on the perceived risk of those cattle de- veloping BRD (Ives and Richeson, 2015). This perceived risk classifi- cation varies by organization and is based on many different quantitative and qualitative factors; however, misclassifications https://doi.org/10.1016/j.compag.2018.11.035 Received 9 April 2018; Received in revised form 28 August 2018; Accepted 26 November 2018 ⁎ Corresponding author. E-mail address: damrine@vet.k-state.edu (D.E. Amrine). Computers and Electronics in Agriculture 156 (2019) 439–446 Available online 10 December 2018 0168-1699/ © 2018 Elsevier B.V. All rights reserved. T