Grad Student Competition
Feedyard cattle facial images at the time of bovine respiratory disease treatment to predict negative outcomes
| Date/Time: |
9/12/2025
09:45
|
| Author: |
Jordana A Zimmermann |
| Clinic: |
Kansas State University |
| City, State, ZIP: |
Manhattan, KS 66503 |
J.A.R. Zimmermann, M.V., MS.
1
;
L.F.B.B. Feitoza, DVM, PAS, Ph.D.
2
;
B.J. White, DVM, MS
2
;
E.M. Bortoluzzi, M.V., Ph.D.
1
;
1Department of Anatomy and Physiology, Beef Cattle Institute, Kansas State University, Manhattan, KS, 66506
2Beef Cattle Institute, Kansas State University, Manhattan, KS, 66506
3Beef Cattle Institute, Kansas State University, Manhattan, KS, 66506
4Department of Anatomy and Physiology, Beef Cattle Institute, Kansas State University, Manhattan, KS, 66506
Introduction:
Machine learning (ML) models have shown promise in predicting Bovine Respiratory Disease (BRD) outcomes in feedyard cattle. Facial imaging combined with ML to predict post-BRD treatment outcomes in feedyard remains unexplored. This study aimed to evaluate the feasibility of using facial images taken at time of BRD treatment to predict negative feedyard cattle outcomes (Recovered or Did-not-finish (DNF)).
Materials and methods:
A total of 923 facial images were collected in a cross-sectional observational study of commercial feedyard cattle at time of BRD treatment (July–December 2023). Cattle outcomes were determined 60 days post-enrollment as Recovered or DNF. Cattle mortalities or culling removals were considered DNF. Predictive models were created using a data subset with known outcomes, then image classification was evaluated using Microsoft Azure Machine Learning Studio’s automated ML function, with model selection based on performance metrics.
Results:
Images were removed due to poor quality (n=278), resulting in 645 labeled images available for testing and training the model. The true prevalence of DNF was 28%. Sensitivity, specificity, positive predictive value, and negative predictive value were 7%, 96%, 50%, and 70%, respectively.
Significance:
Sensitivity was relatively low; yet, when the model predicted DNF, 50% of those cattle did not finish, which is an improvement in the classification prediction relative to baseline DNF risk of 28%.
This study highlights the potential of facial imaging-based ML models for short-time period prediction of feedyard cattle outcomes. The ML model is time-efficient, and further refinement is needed to improve overall predictive performance.
