Interpretable Machine Learning for Early Risk Stratification of Carbapenem Resistance Among ICU Patients with Sterile-Site Pseudomonas aeruginosa Isolates: Development and Internal Validation Using MIMIC-IV

OBJECTIVES: We developed and internally validated an interpretable machine learning model to stratify carbapenem-resistance probability among ICU patients with sterile-site Pseudomonas aeruginosa isolates, using clinical data from the first 24 hours of ICU admission. METHODS: This retrospective study used the MIMIC-IV database. Adults with P. aeruginosa isolated from sterile sites >48 hours post-ICU admission were included. Feature selection employed least absolute shrinkage and selection operator (LASSO), recursive feature elimination, and SHapley Additive exPlanations (SHAP) ranking, followed by bootstrap stability filtering (≥93% selection frequency). Six algorithms were compared using area under the receiver operating characteristic curve (AUROC), calibration, and decision curve analysis. RESULTS: Among 567 patients (resistant: 213 [37.6%]; susceptible: 354 [62.4%]), 15 predictors were identified. XGBoost achieved the numerically highest AUROC of 0.865 (95% confidence interval [CI]: 0.789-0.929), sensitivity 0.791, specificity 0.873, and negative predictive value (NPV) 0.873, though DeLong tests showed no significant differences among ensemble models. Cross-validation confirmed internal stability (mean AUROC: 0.855 ± 0.043). SHAP analysis identified age, respiratory rate, thrombocytopenia, and invasive line placement as key predictors. CONCLUSIONS: The model demonstrates promising internal performance for carbapenem-resistance risk stratification. However, the cohort is conditional on culture-positive patients, predictive values are prevalence-dependent, and no external validation was performed. Results should be interpreted as hypothesis-generating, not as evidence for safe exclusion of resistance in routine practice. External validation and prospective studies are essential next steps.