Predicting Methicillin Resistance in Staphylococcus aureus from Antibiotic Co-Resistance Profiles: A Machine Learning Approach Using XGBoost

Background Methicillin-resistant Staphylococcus aureus (MRSA) is a leading cause of hospital-acquired infection worldwide. Confirming MRSA status conventionally requires dedicated molecular or phenotypic testing that may be unavailable in resource-limited settings. We investigated whether co-resistance phenotypes to seven commonly tested antibiotics could predict methicillin resistance without genomic data. Methods We analysed 229 S. aureus isolates from the BV-BRC database (formerly PATRIC) with laboratory-confirmed resistance profiles for at least three antibiotics. Two classifiers were evaluated: XGBoost and Logistic Regression, using 5-fold stratified cross-validation. Bootstrap confidence intervals (n=1000 resamples) were computed for all AUC estimates. An ablation study assessed the contribution of penicillin resistance—near-universal in this species—to model performance. Results XGBoost achieved an AUC-ROC of 0.680 (95% CI: 0.606–0.745) and average precision of 0.716, outperforming Logistic Regression (AUC-ROC 0.628, 95% CI: 0.562–0.701). Removing penicillin from the feature set did not impair XGBoost performance (AUC 0.685), confirming that model discrimination did not rely on a near-ubiquitous feature. SHAP analysis identified ciprofloxacin resistance as the strongest predictor, consistent with co-selection on SCCmec-carrying clones. MRSA isolates showed markedly higher ciprofloxacin co-resistance rates (70.0% vs 11.8% in MSSA; Δ=+58.2 pp). Conclusions Phenotypic co-resistance profiles carry meaningful predictive signal for MRSA status. Performance is insufficient for standalone diagnosis but may support rapid triage in low-resource settings. Ciprofloxacin and clindamycin resistance are the most informative phenotypic markers for MRSA inference, with implications for antibiotic stewardship and surveillance prioritisation.