Predicting Subclinical Ketosis in Dairy Cows Using Machine Learning Techniques.

Area of Science:

• Veterinary Medicine
• Animal Science
• Machine Learning in Agriculture

Background:

• Subclinical ketosis diagnosis in dairy cows via blood ketone bodies is difficult and expensive.
• There is a need for accessible tools to monitor subclinical ketosis risk using milk performance data.

Purpose of the Study:

• To design a scoring system for selecting optimal machine learning models to identify dairy cows at risk of subclinical ketosis.
• To select the best-performing models and validate them on unseen data.

Main Methods:

• Developed two machine learning pipelines (regression and classification) incorporating feature selection, outlier detection, data scaling, and oversampling.
• Fit and evaluated various linear and non-linear models using training and testing datasets.
• Assessed model suitability using three blood β-hydroxybutyrate (bBHB) concentration thresholds (1.0, 1.2, and 1.4 mmol/L).

Main Results:

• For bBHB thresholds of 1.2 and 1.4 mmol/L, logistic regression was the best model, utilizing milk fat-to-protein ratio, acetone, milk β-hydroxybutyrate, lactose, lactation number, and days in milk. Cross-validation and external validation showed good sensitivity and specificity.
• For the 1.0 mmol/L bBHB threshold, Support Vector Classification (SVC) performed best in cross-validation but had lower performance on the testing dataset.
• Regression models demonstrated poor data fitness (R² < 0.4).

Conclusions:

• Machine learning classification models utilizing test-day milk records offer a viable tool for monitoring subclinical ketosis incidence in dairy herds.
• The developed scoring system and selected models provide a practical approach to risk assessment, complementing traditional diagnostic methods.