Machine Learning for Predicting Somatic Cell Count in Holstein Cattle

Somatic cell count (SCC) is a key indicator of udder health and milk quality in dairy cattle, but predicting it from routinely recorded data is difficult because SCC is highly skewed and depends on production, reproductive, and management factors that interact with one another. Seven modeling approaches were evaluated for predicting natural log-transformed SCC (LSCC) in Holstein cows using data from a commercial dairy herd. The approaches included Multiple Regression (MR) and Generalized Linear Model (GLM) as linear benchmark methods, a Support Vector Machine (SVM) using a Radial Basis Function kernel, and four nonlinear methods, including Bayesian Neural Network (BNN), Random Forest (RF), Chi-squared Automatic Interaction Detection (CHAID), and Classification and Regression Tree (C&R Tree). After multivariate outlier removal, 9,915 records were retained and stratified into five lactation groups, and models were developed using on-farm data. Nonlinear models generally showed higher testing correlations and lower relative errors than MR, GLM, and SVM. The best-performing models achieved testing correlations of approximately 0.62 to 0.70 across the first four lactation groups, whereas predictive performance declined in cows with more than four lactations. BNN and RF showed broadly comparable predictive performance across lactation groups. Average protein, milk, and fat yields were among the most influential predictors of LSCC in both models, with average milk yield becoming the most important predictor in cows with more than four lactations. Overall, nonlinear models, particularly BNN and RF, provided better prediction of LSCC from routinely recorded data than the conventional approaches, although prediction accuracy decreased in later lactations.