248 Machine learning models to predict real-time mortality outcomes in commercial wean-to-finish pig barns

Abstract The adoption of precision livestock farming is steadily increasing in the United States swine industry due to the availability of innovative sensors, microphones, and cameras for automatic surveillance of pig barns. Wean-to-market mortality in pigs is of utmost economic importance to farmers, especially as mortality rates have increased over time. Machine learning models provide an interesting opportunity to combine data from multiple sources into one model to predict real-time daily mortality outcomes in growing pigs. In this study, we evaluated the ability of three machine learning frameworks [elastic net logistic regression (ELNR), support vector machine (SVM), and gradient-boosted decision trees (i.e., XGBoost; XGB)] to forecast daily sporadic and episodic mortality outcomes. These models were trained and tested on 5,364 daily observations collected from October 2020 to December 2023 at six commercial wean-to-finish farms owned and operated by The Maschhoffs, LLC (Carlyle, Illinois, USA). Features used in the analysis were derived from manually recorded antibiotic treatment data, SoundTalks® cough sensors (SoundTalks NV, Leuven, Belgium), and data from ventilation controller systems that recorded water disappearance and climatic conditions. Across all models, episodic mortality predictions were more accurate than those for sporadic mortality (cross-validation F0.5 scores = 0.30 to 0.51 and 0.16 to 0.24 for episodic and sporadic mortality, respectively). The most complex model, XGBoost, reported the highest F0.5 scores during cross-validation and holdout testing for mortality episode prediction (0.51 and 0.40, respectively), considerably outperforming a naïve baseline classifier that predicted outcomes based on majority class at different pig ages. In addition, XGBoost functioned as a conservative classifier for mortality episodes, reporting false positive rates of 0.08 and 0.14 during cross-validation and holdout testing, respectively. In this model, age, water disappearance, and SoundTalks® respiratory health status were the most important predictors of episodic mortality (0.26, 0.17, and 0.08 relative accuracy gain, respectively). Receiver operating characteristic and precision-recall curves were used to evaluate the overall classification ability of each model based on unseen data from subsequent pig cohorts. Each model framework achieved a moderately high area under the receiver operating characteristic curve (AUC = 0.77 to 0.78), indicating moderate classifier accuracy above baseline expectation for a non-informative classifier (baseline AUC = 0.50). The area under the precision-recall curve for each model framework was also moderately high relative to the naïve baseline classifier (AUCPR = 0.40, 0.34, and 0.46 for ELNR, SVM, and XGB, respectively; baseline AUCPR = 0.25). Findings from this analysis support data-driven decision-making to reduce mortality losses in commercial wean-to-finish pig barns through targeted preventative interventions. However, the development of more robust mortality prediction models will require larger, more diverse datasets, which could further enhance predictive accuracy and model applicability across different production settings.