Simulated poultry antibiotic residue data test a machine-learning pipeline for Salmonella resistance analysis

Antimicrobial resistance in foodborne pathogens such as Salmonella enterica is widely recognised as a One Health challenge, but researchers face a stubborn data problem: no linked datasets currently pair poultry antibiotic residue measurements with phenotypic resistance profiles in human Salmonella isolates. Without such data, integrative analytical pipelines cannot be tested on real-world inputs. A team led by Yara Al-Hallak, publishing in BMC Veterinary Research, has now proposed a way around this impasse — not by collecting new field data, but by building a fully synthetic environment in which the true answer is known by construction.

The study is explicitly a methodological proof-of-concept. The authors generated a synthetic dataset of 2,500 observations representing hypothetical human cases, created by explicit rules with ranges informed by field observations and CLSI/EUCAST criteria. The dataset encoded four antibiotics used in both veterinary and human medicine, four anonymized synthetic regions, poultry residue levels ranging from 0.02 to 5.0 mg/kg, and partially overlapping minimum inhibitory concentration (MIC) ranges. Crucially, the resistance label depended on a residue threshold, and the MIC value was generated from that label, so the researchers could evaluate whether their pipeline could recover a predefined relationship.

Three models were compared on an 80/20 train-test split with cross-validation, calibration analysis and SHAP interpretability: a categorical baseline Random Forest, a Logistic Regression model, and an optimized Random Forest. The optimized Random Forest reached 79.6% accuracy and an AUC-ROC of 0.860 on the held-out set, outperforming the categorical baseline, which scored an AUC of 0.471, and Logistic Regression, which reached 0.818. Residue level and MIC dominated the feature importance rankings, a result the authors attribute directly to the data-generating design rather than to any discovered biological mechanism.

Robustness was assessed across 14 scenarios, each consisting of 30 independent synthetic datasets. These scenarios varied the encoded effect size — including a null effect of none — as well as MIC overlap and distribution, class imbalance, and the form and antibiotic-specificity of the relationship between residues and resistance. Each scenario was run both with and without the MIC variable included as a model feature, a deliberate choice that proved analytically important.

The scenario analysis clarified what the models were actually learning. When no association was encoded, AUC-ROC sat at 0.50, equivalent to chance, and performance rose in step with the encoded effect size. Because the MIC had been generated from the resistance label, the MIC variable alone achieved an AUC-ROC of approximately 0.76 under the base design. In contrast, when the MIC was removed, the model isolated the residue signal alone, with AUC-ROC ranging from 0.50 under the null to 0.92 under a near-deterministic encoded effect.

The authors are careful about what these numbers do and do not mean. 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. They explicitly state that the results do not demonstrate a biological association between poultry antimicrobial residues and resistance in human Salmonella isolates, and that any practical application remains conditional on validation with independent biological data.

The background motivating the work is itself contested territory. According to the paper, poultry antibiotic residues arise mainly from improper administration or non-compliance with withdrawal periods, and their contribution to resistance selection remains debated in the scientific literature. The absence of linked residue-resistance datasets has limited the ability to test integrative pipelines that could one day inform such debates, which is precisely the gap the simulation framework is designed to bridge.

Methodologically, the study illustrates a broader principle for computational research: when the ground truth is known, a pipeline’s successes and failures can be diagnosed precisely. The near-perfect recovery of a strong encoded effect, the collapse to chance performance under the null, and the disentangling of the MIC variable’s circular contribution all show that the framework behaves as intended. Whether the same behaviour carries over to messier real-world data, where residue thresholds, MIC distributions and confounding factors are not designed but observed, remains an open question the authors themselves flag.

For now, the contribution is a validated analytical scaffold rather than a finding about food safety. The framework offers researchers a reproducible testbed for building and stress-testing integrative models before scarce and expensive linked biological datasets become available. Until such validation with independent data is performed, the pipeline’s strong simulated performance should be read as a statement about the method, not about the relationship between residues on poultry and resistance in human Salmonella infections.

Subject of Research: A simulation-based machine-learning framework for integrating poultry antibiotic residue data with Salmonella resistance profiles

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

Article References: Al-Hallak, Y. (2026). A Simulation-based framework for integrating poultry antibiotic residue data with Salmonella resistance profiles: a methodological proof-of-concept. BMC Veterinary Research. https://doi.org/10.1186/s12917-026-06050-1

Image Credits: AI Generated

DOI: 10.1186/s12917-026-06050-1

Keywords: antimicrobial resistance, Salmonella, poultry, antibiotic residues, machine learning, Random Forest, simulation study, MIC, One Health, proof-of-concept, SHAP, BMC Veterinary Research