Machine Learning Applications for Chute-Side Decision-Making and Targeted Thoracic Ultrasound Prognostic Modeling in Feedyard Cattle

Abstract

Bovine respiratory disease (BRD) remains the leading cause of morbidity and mortality in feedyard cattle and continues to result in substantial economic losses, reduced animal performance, and animal welfare concerns in beef production systems. Although cattle diagnosed with BRD are commonly treated based on clinical assessment, subjective chute-side prognostic tools capable of improving treatment and management decisions remain limited. Machine learning and targeted thoracic point-of-care ultrasound (TT-POCUS) have emerged as promising approaches for integrating animal-level clinical and imaging data into practical decision-support systems. Therefore, the objectives of this thesis were to evaluate current machine learning applications relevant to chute-side beef cattle management and to assess the prognostic utility of TT-POCUS-based predictive models for bovine respiratory disease outcomes. The first chapter consists of a scoping review conducted to evaluate how machine learning is applied to support chute-side decision-making in beef cattle production. A structured literature search identifies 43 full-text studies for assessment, of which 26 meet the inclusion criteria. Included studies evaluate applications in disease detection and prognosis, treatment decision support, physiologic monitoring, imaging analysis, and nutritional prediction. Collectively, these findings demonstrate the broad versatility of machine learning for integrating complex animal-level data into actionable management tools while highlighting important knowledge gaps in external validation, implementation, and practical field adoption. The second chapter evaluates the repeatability of previously developed TT-POCUS prognostic logistic regression models for predicting first-treatment failure (FTF) and did-not-finish (DNF) outcomes in feedyard cattle at initial BRD treatment. A prospective observational study enrolls 603 cattle, and previously developed model structures are assessed in an independent population. Initial deployment identifies multicollinearity among predictors, limiting direct reproducibility of the previously developed model. Revised models retain meaningful prognostic performance, with DNF prediction achieving 39% sensitivity and 98% specificity, and FTF prediction achieving 55% sensitivity and 88% specificity, supporting the continued evaluation of the prognostic relevance of TT-POCUS-derived pulmonary variables. The third chapter evaluates and compares six supervised machine learning algorithms incorporating TT-POCUS-derived variables for predicting FTF and DNF outcomes in feedyard cattle diagnosed with BRD. A total of 1,422 first-treatment cattle and 150 chronic cattle are included for 60-day outcome assessment. Predictive performance varies across algorithms and outcomes, with logistic regression demonstrating the highest discrimination for FTF (AUC = 0.71) and first-treatment DNF (AUC = 0.80), while the perceptron model demonstrates the highest discrimination for chronic-treatment DNF (AUC = 0.84). Collectively, these studies demonstrate the growing potential for machine learning to support practical chute-side decision-making in beef production systems while emphasizing the importance of biologically relevant and operationally feasible inputs. The scoping review identified broad opportunities for machine learning integration across cattle health and production but highlighted persistent limitations in external validation and commercial implementation. The repeatability and model comparison studies demonstrated that TT-POCUS derived pulmonary biomarkers provide prognostic information that can improve risk stratification at the time of BRD treatment, a stage where clinical decision-making often relies heavily on subjective assessment alone. Importantly, findings from this thesis suggest that when using structured TT-POCUS and animal-level datasets, simpler linear modeling approaches can effectively capture clinically relevant prognostic information while maintaining flexibility for integration of additional predictive variables in future model development. These findings emphasize that model selection should align with available data structure, biological context, and intended field application rather than model complexity alone. Earlier identification of cattle at elevated risk for treatment failure or non-recovery may improve treatment selection, labor allocation, culling strategies, and antimicrobial stewardship while reducing unnecessary interventions and improving animal welfare outcomes. Collectively, this work contributes to the advancement of precision livestock medicine by supporting development of more data-driven tools for improving cattle health management and production efficiency in commercial feedyard systems.