Risk prediction of carbapenem-resistant Pseudomonas aeruginosa infection in children

Objective This study aimed to investigate the infection distribution, drug resistance, and risk factors associated with carbapenem-resistant Pseudomonas aeruginosa (PA) in children, and to construct a risk prediction model. Methods Two retrospective cohorts were established based on the Children’s Hospital Affiliated to Shandong University. The derivation cohort included 1397 children with Pseudomonas aeruginosa infection hospitalized from January 2020 to December 2024, divided into the carbapenem-resistant PA (CRPA) group and the carbapenem-sensitive PA (CSPA) group based on antimicrobial susceptibility testing. LASSO regression was used for feature selection, and an XGBoost model was constructed to predict CRPA infection, with internal validation performed using a randomly split test set. Subsequently, an independent temporal validation cohort of 431 children (January 2025 to March 2026) was used to evaluate model generalizability. SHapley Additive exPlanations (SHAP) was employed to interpret variable importance. Results Among the 1397 PA isolates, 1177 were CSPA and 220 were CRPA. Children in the CRPA group were younger, had longer hospital stays, and were more prevalent in the NICU and PICU than those in the CSPA group. Among CRPA isolates, 88.2% met multidrug-resistant criteria and 46.4% met extensively drug-resistant criteria, while maintaining high sensitivity to amikacin (95%). The XGBoost model, incorporating 10 key variables, achieved an AUC of 0.848 (95% CI, 0.783–0.912) on the internal test set and 0.715 (95% CI, 0.622–0.807) on the temporal validation cohort. SHAP interpretability analysis showed that length of hospital stay, CD8 + T cells, ventilator use, and pre-infection carbapenem exposure were important risk factors for CRPA infection. Conclusion CRPA infection in children is closely associated with prolonged hospitalization, immune dysregulation, invasive procedures, and prior carbapenem exposure. The XGBoost prediction model demonstrated good discrimination in internal validation, and its performance remained stable on temporal validation, suggesting potential utility in identifying high-risk children for early clinical intervention.