Machine learning models based on facial images alone have moderate ability to predict negative outcomes in feedyard cattle treated for bovine respiratory disease

Abstract Objective To evaluate the use of facial images and deep learning models to predict subsequent did-not-finish (DNF) outcomes in cattle at the time of bovine respiratory disease (BRD) treatment. Methods A dataset of 1,271 facial images was obtained from feedyard cattle undergoing treatment for clinically identified BRD and classified based on 60-day outcome as finish (n = 891) or DNF (380). Five image classification models (MobileNetV2, ResNet18, ResNet50, SE-ResNeXt50, and ViT-l16-r224) were trained and evaluated using a stratified train-test split. Model performance was assessed using receiver operating characteristic area under the curve (AUC), precision-recall AUC, and threshold-dependent metrics. Results Across models, receiver operating characteristic AUC values ranged from 0.488 to 0.593 and precision-recall AUC from 0.296 to 0.380, indicating moderate discriminative ability. Sensitivity and specificity varied substantially across models and decision thresholds, reflecting trade-offs between identifying DNF animals and minimizing false positive classifications. Conclusions Models with higher specificity may offer greater practical value for feedyard decision-making due to the economic impact of incorrect culling decisions. The modest predictive performance suggests that facial images alone may not capture sufficient information to reliably predict cattle outcomes following BRD treatment. Clinical Relevance Facial images alone had moderate ability to predict DNF outcomes in feedyard cattle treated for BRD. Incorporating facial imaging with clinical signs, health history, or other animal-level information may contribute as part of a multimodal decision-support system for BRD outcome prediction.