Machine learning model predicts carbapenem-resistant Klebsiella pneumoniae risk in ICU patients

BackgroundCarbapenem-resistant Klebsiella pneumoniae (CRKP) infections in intensive care units are associated with poor outcomes. The delay in obtaining culture-based susceptibility results often forces clinicians to choose between under-treatment and overtreatment with empirical antibiotics. A reliable early risk assessment using only standard clinical data could help address this challenge.MethodsThis single-center retrospective cohort study included 401 ICU patients with culture-confirmed K. pneumoniae infections (January 2022 to January 2025). Patients were randomly allocated to training (n = 281) and validation (n = 120) sets. Predictors extracted from electronic health records comprised demographics, severity scores (APACHE II, SOFA), comorbidities, invasive procedures, inflammatory markers, specimen type, history of multidrug-resistant (MDR) infection, and antibiotic exposure within the preceding 90 days. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression on the training set. The selected features were used to develop an XGBoost model, whose performance was compared against six other machine learning algorithms (logistic regression, random forest, etc.). Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC), calibration with Brier scores and calibration curves, and clinical utility with decision curve analysis. SHapley Additive exPlanations (SHAP) values were employed to interpret the model.ResultsCRKP isolates accounted for 15.7% (63/401) of cases. LASSO regression identified nine predictors: procalcitonin (PCT), specimen type, prior MDR infection, prior carbapenem exposure, history of stroke, APACHE II score, white blood cell count, age, and hemoglobin. In the independent validation set, the XGBoost model achieved an AUC of 0.852 (95% CI: 0.745–0.959), with a sensitivity of 0.737, specificity of 0.891, accuracy of 0.867, and an F1-score of 0.636. The model demonstrated good calibration (Brier score: 0.088) and provided a net clinical benefit across a wide range of risk thresholds. SHAP analysis highlighted PCT, specimen source (blood), and prior resistance-related exposures as the most influential predictors.ConclusionThe integration of LASSO feature selection with the XGBoost algorithm, utilizing only routine clinical data, generates a reliable early-warning model for CRKP infection risk prior to the availability of susceptibility reports. This tool shows promising discriminative ability and calibration, offering potential to guide empirical therapy and support antimicrobial stewardship. Future multicenter prospective studies are warranted to validate its generalizability and real-world clinical impact.