Improving bovine health outcomes and management strategies using machine learning in cattle production systems

Abstract

The objective of this dissertation is to explore the use of machine learning to enhance health and management of cattle in beef and dairy systems. One chapter is dedicated to a review of the literature of machine learning application to the diagnosis and prognosis of cattle diseases in the beef and dairy production sectors. A total of four studies were designed and conducted to determine the predictive ability of machine learning algorithms to determine health outcomes or ideal management strategies in beef cattle in confined feeding operations at both group- and individual animal-levels. These four studies focused on bovine respiratory disease as the major health concern. The remaining study evaluated post-treatment intervals in stocker cattle treated with a novel antimicrobial for bovine respiratory disease. Two studies focused on the prediction of a group-level outcome using various machine learning algorithms. One study aimed to predict morbidity class (high = ≥ 15% or low = < 15%) of pens of feedlot cattle when given demographic data and feed delivery data from the first 15 days on feed. The algorithms tested included advanced perceptron, logistic regression, neural network, boosted decision tree, and random forest models. Model performance was poor as determined by area under the receiver operating characteristics curve (AUC) but showed promise for determining which pens of cattle would have low morbidity using specificity as a performance metric. Non-linear algorithms tended to perform better than logistic regression. The second study aimed to predict optimal metaphylaxis application strategy (yes or no) to groups of cattle arriving to the feedlot given demographic, origin, and external economic data. Algorithms tested were similar to the first study. High AUC values were achieved when models were given demographic data alone. Model performance based on AUC increased given external economic data but decreased given origin data. Again, non-linear models had higher predictive power than logistic regression. Two studies focused on the prediction of individual animal outcome at the time of first or second treatment for bovine respiratory disease in feedyard cattle. One study aimed to predict mortality or early culling (did not finish = yes or no) when given individual animal-, group-, and feedyard-level data on demographics. Additionally, the addition of weather data, the impact of a balancing strategy applied to the training data, and the development of individualized models on predictive model performance was tested. Similar algorithms to the previously mentioned studies were utilized. Models performed well with moderate AUC values. No improvement in AUC values were noted with the addition of weather data or the application of a balancing strategy. Creation of individualized models for each feedyard resulted in increased predictive performance for some feedyards but not others. The second study determined the economic effect of using a predictive model to cull sick animals at the time of diagnosis for first or second treatment of bovine respiratory disease. In this study, only a boosted decision tree model was investigated. It was found that minimizing the number of false positives (i.e. animals that the model has determined will die but would have actually lived) is the biggest driver of positive net return. Therefore, the specificity of the predictive model is the most important metric when the goal of using a predictive model to make culling decisions is profit. A final study was a prospective trial to investigate the effect of post-treatment interval on health outcomes following administration of pradofloxacin for the treatment of bovine respiratory disease in stocker cattle. Three-day, six-day, and nine-day post-treatment intervals were investigated in a multi-site trial. Post-treatment interval had no effect on first treatment success or case fatality. Additionally, there was no difference in number of days to death from initial treatment amongst the mortalities by post-treatment interval. In conclusion, the current dissertation has contributed to the body of literature surrounding the use of machine learning algorithms for health outcomes in cattle production systems. By investigating various algorithms, data types, and strategies for machine learning application to cattle health, this work will provide a foundation for future research. Additionally, this work provides a background on predictive model development and proper model training strategies that could be applied to other livestock production species.