Farm-Level Biosecurity and Antimicrobial Use in Pig Production: An Integrated Inferential and Explainable Machine Learning Analysis

Internal biosecurity may be associated with antimicrobial use (AMU) in pig production, but observational associations, predictions and explanations of fitted models answer different questions. We analysed 1110 farm–month–age-group units from 18 Hungarian pig farms; an original post hoc restriction retained 939 units from 13 farms. The continuous endpoint was the natural logarithm of one plus a defined AMU index formed by summing substance-specific active ingredient mass-to-recorded-weight ratios. The study-specific biosecurity instrument was used for within-farm monitoring and was not externally validated. We retained a 2 × 2 matrix of unfiltered/filtered and composite/separate-variable models, with categorical animal-group separation, age-group adjustment and farm-cluster uncertainty. In the unfiltered separate-variable model, recorded separation code 2 versus code 0 was associated with a higher log AMU index (β = 0.205; CR1-t 95% CI 0.013–0.397; p = 0.037), although the wild-cluster p-value was 0.116 and sensitivity estimates varied with adjustment. Code 3 binary contrasts were too sparse for precise effect quantification. Four ML families were evaluated using nested farm-grouped and leave-one-farm-out validation. For the inner-selected biosecurity-plus-age procedure, the unfiltered pooled R2 was −0.101 and −0.187, respectively. SHAP retained an exploratory role in explaining held-out tree model predictions. The results identify model-dependent observational patterns but do not establish a causal biosecurity effect or reliable prediction for new farms.