Assessment of reproductive performance in dairy cows using explainable machine learning

Reproductive performance is a key determinant of productivity and economic sustainability in dairy farming and is influenced by a complex interaction of biological and management-related factors. This study aimed to evaluate reproductive performance in dairy cows and to identify the most influential risk factors at both global and individual animal levels using an explainable machine learning framework. An eXtreme Gradient Boosting (XGBoost) regression model was applied to evaluate reproductive performance based on days to first insemination. Model performance was assessed using standard regression metrics. Model interpretability was achieved through SHapley Additive exPlanations (SHAP), allowing both global feature importance assessment and local, animal-specific interpretations. SHAP analysis indicated that age had the greatest contribution to the model predictions, followed by mastitis, retained placenta, ovarian cysts, ketosis, and metritis. The direction and magnitude of the SHAP contributions varied across individual animals, highlighting heterogeneity in the model explanations. Model performance on the independent test dataset (RMSE $=$ 39.05 d, MAE $=$ 18.58 d, $R^2$ $=$ 0.01) indicated limited predictive generalizability. Nevertheless, SHAP analysis provided transparent global and local explanations of model behavior, demonstrating how explainable machine learning can be applied to investigate reproductive performance in dairy cows. These findings should be regarded as a methodological demonstration, and further studies using larger, more comprehensive datasets that include a broader range of variables are required before clinical decision-support applications can be considered.