A machine learning framework for precision prediction of lactation performance in large dairy herds: Integrating dietary, environmental, and health risk factors

Precision management of large dairy herds requires accurate prediction of daily milk yield (DMY) and early identification of health risks. This study presents a machine learning framework addressing two interlinked objectives: (1) high-accuracy DMY prediction, and (2) early disease detection, using a 5-year dataset from 19,798 Hostein cows. Our novel dual modeling pipeline integrates advanced ML models, including Random Forest (RF), Gradient Boosting Regression (GBR), and Extreme Gradient Boosting (XGB), to predict DMY using dynamic environmental, dietary, and physiological inputs. RF emerged as the top model for DMY prediction (R² = 0.77, RMSE = 5.72 kg/d), capturing nonlinear interactions among lactation stage, parity, and nutrient intake. Shapley Additive exPlanations (SHAP) analysis identified lactation days, parity, and dietary ether extract (EE) as key drivers; a 1 % EE increase boosted DMY by 2.1 kg/d in multiparous cows, and neutral detergent fiber (NDF) > 35 % reduced DMY by 4.2 kg/d due to rumen fill limitations. For disease detection, an RF classifier integrated SMOTE (synthetic minority over-sampling technique) achieved robust performance (AUC = 0.93, sensitivity = 0.80), enabling early identification of mastitis and ketosis via yield deviations. Real-time temperature-humidity index (THI) alerts (>72) reduced yield losses by 4.8 kg/d. Practical applications include dietary optimization (16.8 % crude protein, 5.8 % EE during peak lactation) and automated health alerts. This study advances precision dairy farming by providing a scalable, interpretable framework that bridges ML innovation with actionable herd management strategies, validated across diverse lactation stages and environmental conditions. Future work will integrate genomic and IoT-enabled sensor data to enhance predictive accuracy and adaptability.