Evaluating the performance of supervised machine learning algorithms on predicting treatment outcomes of BRD
Bovine respiratory disease (BRD) causes 45-55% of all deaths in feedlots. Feedlot managers tend to identify BRD from visual observations with a sensitivity of about 62%. The purpose of this research is to compare the ability of various classifiers to predict with high precision unsuccessful treatment of BRD and to assess the economic viability of incorporating a fit model in a feedlot operation's culling process. To build the models, we use hyperparameter tuning, validation data, and a weighted-cost matrix. Models are fit using training examples from the same feedlot as the test data but coming from different animal cohorts spanning four years. We repeat the experiments for ten different feedlots ranging in size from 10,135 to 31,975 total head of cattle. In addition, while fitting the model, we explore various combinations of attributes to determine what information needs to be gathered and calculated to produce a successful model. We assess the classifiers' generalization ability by examining the confusion matrices produced when applying the fit models to unseen test data. To evaluate our models' economic viability in facilitating culling, we calculate the predicted revenue by applying the typical feedlot cost per animal for successful and unsuccessful treatment to the model predictions. Our results indicate that using machine learning algorithms to determine whether an animal should be treated results in a net increase in potential revenue for the evaluated feedlot.