Leveraging Machine Learning to Predict Subsequent Farrowing Success: A Decision Support Tool for Sow Management

Abstract Background Optimizing sow herd productivity and profitability is dependent on accurately predicting the likelihood of farrowing in subsequent cycles. Reliable predictions support timely culling decisions, optimize resource allocation, and enhance overall herd productivity. Integrating machine learning algorithms into reproductive decision-making offers a promising approach to improve breeding herd efficiency. Therefore, the objective of this study was to identify predictors of farrowing success using various machine learning algorithms to assist breeding decisions at the farm level. Methods Experimental data was sourced from six lactation trials (2021 and 2022) conducted on a commercial sow farm in the Carthage System (Carthage, IL, USA). Animals were housed in the same lactation rooms, had the same genetics, and were not challenged with Porcine Reproductive and Respiratory Syndrome virus (PRRSV) or Porcine Epidemic Diarrhea virus (PEDV). The final dataset consisted of 4,300 observations, including reproductive performance, environmental temperature and humidity at farrowing, daily lactation feed intake, and sow and litter characteristics. The dataset was divided into 80% for training, which was evaluated using three repeats of 10-fold cross-validation. The remaining 20% of the dataset was held out for final model evaluation. Five machine learning algorithms logistic regression (LR), support vector machines (SVM), random forest (RF), extreme gradient boosting (XGBoost), and neural networks (NN) were trained after random down-sampling and Bayesian optimization of hyperparameter tunning. Each model's performance was evaluated based on accuracy, recall, specificity, precision, and F1-score, and the best model was selected based on both accuracyCG1.1KEK[1.2 and recall. Feature importance and model interpretability were further evaluated using Gini Index, partial dependence and SHAP (Shapley Additive exPlanations) plots. DSESG[2.1KEK[2.2 Results XGBoost achieved the highest predictive performance with an accuracy of 92.1%, outperforming SVM (86.1%), RF (91.2%), ENET (89.1%), and NN (86.1%). Across all models, farrowing season, environmental temperature, sow weight at entry to farrow, lactation ADFI (total and within the first 3 days), wean-to-estrus interval, and stillbirth rate were identified as the most influential predictors of farrowing outcome, in order of importance. The significance of farrowing season and temperature highlights the influence of environmental conditions on reproductive performance. SHAP analysis revealed nonlinear associations between lactation feed intake and weight change on the probability of subsequent farrowing, underscoring the relevance of early lactation feeding and dynamic weight monitoring for subsequent farrowing success. Conclusion This study demonstrates that machine learning models provide swine veterinarians and producers with practical insights for precision sow management. By proactively identifying at-risk sows, these models enable targeted interventions, including adjusting wean-to-service intervals, optimizing lactation feed intake, and refining breeding strategies. Ultimately, this serves as a valuable resource for improving sow longevity and maximizing overall herd productivity and profitability.