Integrated prediction of ceftriaxone resistance in major Salmonella serotypes using MALDI-TOF MS data with balanced datasets and machine learning.

OBJECTIVES

Salmonella causes gastroenteritis and invasive infections worldwide, and rising ceftriaxone resistance complicates empirical therapy. Conventional culture-based susceptibility testing remains reliable but time-consuming. MALDI-TOF MS, widely adopted for rapid identification, offers potential for antimicrobial-resistance prediction when combined with machine learning.

METHODS

We analyzed 632 clinical Salmonella isolates, 536 from Zhejiang University Children's Hospital and 96 from Wanbei Coal-Electricity Group General Hospital. Serovar-stratified MALDI-TOF MS spectra were preprocessed and modeled with six classifiers (XGBoost, Random Forest, Logistic Regression, Naïve Bayes, Linear SVM, and RBF SVM) across five data-balancing strategies (none, SMOTE, ADASYN, undersampling, and cost-sensitive learning). Model performance was assessed by ROC-AUC, PR-AUC, sensitivity, F1 score, calibration, and decision-curve analysis. SHAP-based feature selection enhanced interpretability and reduced model complexity.

RESULTS

Ceftriaxone resistance varied markedly by serovar, highest in S. enteritidis (46.9%) and C1 (30.9%). Multivariate analyses and visualization confirmed distinct resistance-associated spectral signatures. Serovar-specific pipelines were identified; models for C1, E1, and S. typhimurium achieved validation ROC-AUCs of 0.883 to 0.924. SHAP revealed mainly serovar-specific discriminative peaks that provided transparent sample-level explanations. Decision-curve analysis demonstrated consistent clinical net benefit compared with treat-all or treat-none strategies across a range of decision thresholds.

CONCLUSIONS

Integrating MALDI-TOF MS, serovar stratification, data-balancing strategies, and machine learning yields interpretable, serovar-specific models for rapid ceftriaxone-resistance prediction in Salmonella. Implemented as a user-friendly web application, these models could shorten time to targeted therapy and support improved antibiotic stewardship and reduce unnecessary broad-spectrum antibiotic use across clinical settings worldwide.