A Simulation-based framework for integrating poultry antibiotic residue data with Salmonella resistance profiles: a methodological proof-of-concept

Background Antimicrobial resistance (AMR) in foodborne pathogens such as Salmonella enterica is a One Health challenge. Poultry antibiotic residues arise mainly from improper administration or non-compliance with withdrawal periods, and their contribution to resistance selection remains debated. Linked datasets pairing poultry residues with phenotypic resistance in human Salmonella isolates are unavailable, limiting the testing of integrative analytical pipelines. Methods A synthetic dataset of 2,500 observations representing hypothetical human cases was generated by explicit rules, with ranges informed by field observations and CLSI/EUCAST criteria. It encoded four antibiotics used in veterinary and human medicine, four anonymized synthetic regions, poultry residue levels (0.02–5.0 mg/kg), and partially overlapping MIC ranges. The resistance label depended on a residue threshold, and the MIC was generated from the label, so that recovery of a predefined relationship could be evaluated. A categorical baseline Random Forest, Logistic Regression, and an optimized Random Forest were compared on an 80/20 split with cross-validation, calibration, and SHAP. Robustness was assessed in 14 scenarios of 30 independent datasets each (varying effect size, including none; MIC overlap and distribution; class imbalance; and the form and antibiotic-specificity of the relationship), with and without the MIC variable. Results The optimized Random Forest reached 79.6% accuracy and AUC-ROC 0.860 on the held-out set, outperforming the categorical baseline (AUC 0.471) and Logistic Regression (AUC 0.818). Residue and MIC dominated feature importance, reflecting the data-generating design. Across scenarios, AUC-ROC was 0.50 when no association was encoded and rose with the encoded effect size; MIC alone gave AUC-ROC ≈ 0.76 under the base design because it was generated from the label, whereas the model without MIC isolated the residue signal (AUC-ROC 0.50 under the null to 0.92 under a near-deterministic effect). Conclusions The findings demonstrate the performance of a machine-learning pipeline under predefined simulation assumptions: performance was at chance when no association was encoded, followed the encoded effect size, and was partly attributable by construction to the label-conditional MIC. The results do not demonstrate a biological association between poultry antimicrobial residues and resistance in human Salmonella isolates; any practical application remains conditional on validation with independent biological data.