Prediction of the growth in grower-finisher pigs using biologically constrained machine learning under small-sample data conditions

Accurate prediction of pig growth is essential for feed planning and market decisions in precision pig farming, but farm data are often small and fragmented. To address this challenge, a biologically constrained machine learning framework is proposed to predict the time required for grower-finisher pigs to reach the target weight of 100 kg under small sample conditions. By applying biologically constrained modifications to machine learning models (SC-GAM and Monotone XGBoost), the constrained models outperformed unconstrained baselines, which achieved higher accuracy (RMSE ≤ 4.73 d, ACC±7d ≥ 92%), greater robustness, and improved biological realism. An evaluation system was developed that combines traditional accuracy indicators with biologically grounded metrics. Practical applicability was examined via an on-farm shadow test on an independent batch. The models delivered reliable predictions that support finishing scheduling and feed-related management decisions. These findings highlight the potential of biologically constrained models to improve operational efficiency and reduce resource wastage in commercial pig farming.