XGBoost-based early warning framework for farm-level PED risk using meteorological and herd management variables

Porcine epidemic diarrhea (PED) imposes substantial economic losses on the global swine industry, with outbreaks driven by the complex interplay of meteorological conditions and herd management practices. A farm-level risk prediction framework was developed by integrating meteorological and reproductive performance variables from commercial pig farms in southern China. A triangulated variable selection strategy-combining hierarchical clustering analysis, variance inflation factor filtering, and principal-component analysis loading evaluation-reduced 22 meteorological candidate predictors to 14 consensus variables. Class imbalance was addressed through inverse-proportion class weighting. Among five evaluated classifiers, XGBoost achieved the most balanced predictive performance. Feature importance analysis identified seasonal periodicity and reproductive performance indicators, particularly pre-weaning mortality rate, as primary risk determinants. These findings support the incorporation of multimodal farm-level data into PED surveillance systems and provide a quantitative basis for precision biosecurity management.