Development and validation of an interpretable machine learning-based model for predicting carbapenem-resistant

Background: Carbapenem-resistant Methods: We conducted a retrospective cohort study including 2,195 postoperative ICU patients. Clinically available demographic, treatment-related, and laboratory variables were used to develop eight machine learning models. Feature selection was performed using Boruta, and model interpretability was enhanced using Shapley Additive Explanations (SHAP) analysis. Model performance was evaluated in an independent test set using the area under the receiver operating characteristic curve (AUC), with sensitivity analyses performed using reduced feature sets. Results: Among 2,195 postoperative ICU patients, 694 (31.6%) developed CRAB infection. Patients with CRAB infection had significantly longer ICU stays, greater exposure to invasive procedures, higher antimicrobial use, and worse laboratory profiles than non-infected patients. Using 19 features selected by the Boruta algorithm, all eight machine learning models achieved good discrimination in the test set (AUC > 0.83). Gradient Boosting demonstrated the best overall performance, with an AUC of 0.867 (95% CI: 0.836-0.892), good calibration, and the highest net clinical benefit. SHAP analysis identified duration of mechanical ventilation, central venous catheterization, ICU length of stay (LOS), and carbapenem exposure as the most influential predictors. Sensitivity analyses showed that models using only the top 10 or top 5 SHAP-ranked features achieved performance comparable to the full model, supporting the feasibility of feature reduction for clinical application. Conclusions: This study provides an interpretable and clinically applicable framework for early risk assessment of CRAB infection in postoperative ICU patients, supporting targeted prevention strategies and more rational antimicrobial stewardship.