One of the keys to improving feed conversion rates in Precision Livestock Farming (PLF) is the early identification of growth impediments. However, the swine farming data collected by Electronic Feeding Station (EFS) are often disorganized and lack effective labeling. Data from healthy pigs are frequently intermixed with that from sick pigs, leading to label leakage and survivor bias in models, particularly when age is included as a feature. To address these known issues, this study breaks away from traditional modeling methods. First, we clean and classify the time-series data from electronic feeding stations, using age-cohort baselines as one of the criteria for determining high and low productivity, thereby avoiding problems such as label leakage. Next, we construct a high-dimensional feature matrix that captures dynamic derivatives such as feeding acceleration and weight gain acceleration, which together serve as behavioral feature fingerprints. To test the system, we optimized the mixed-model algorithm and evaluated the model based on behavioral deviations among individual pigs after removing all absolute age labels. Our results indicate that the full-feature model achieved an ROC-AUC of 0.778 and an F1-score of 0.4137 at the optimal threshold. Interestingly, SHAP attribution analysis revealed that “intake peer deviation,” “Cumulative Intake and Lifetime Avg Intake,” and “feeding acceleration” served as precursors to low productivity and growth retardation in this dataset, with these factors proving more significant than absolute feed intake or age. Our ablation experiments confirmed that a model based solely on behavioral features (excluding age labels) maintained an ROC-AUC of 0.773, successfully decoupling pig growth performance from growth stage. Our model can detect changes in feeding dynamic signatures at an average of 12.3 days, thereby providing insights for pig growth assessment, health monitoring, or more informed culling decisions.
Wang et al. (Wed,) studied this question.