Construction and validation of a machine learning-based model for predicting the risk of carbapenem-resistant gram-negative bacteria infections in neurosurgical ICU patients
10.11816/cn.ni.2025-241773
- VernacularTitle:基于机器学习的神经外科ICU患者耐碳青霉烯革兰阴性菌感染风险预测模型构建及验证
- Author:
Xiaochao SONG
1
;
Meijuan JIN
;
Wei DING
;
Li YANG
;
Bo YANG
Author Information
1. 苏州大学附属第一医院感染管理处,江苏苏州 215006
- Publication Type:Journal Article
- Keywords:
Department of neurosurgery;
Carbapenem-resistant organism;
Hospital-acquired infection;
Random forest;
Prediction model;
Intensive care unit
- From:
Chinese Journal of Nosocomiology
2025;35(11):1690-1696
- CountryChina
- Language:Chinese
-
Abstract:
OBJECTIVE To investigate the current status and risk factors of carbapenem-resistant gram-negative or-ganisms(CRO)infections in neurosurgical ICU patients,and to construct and validate their prediction models.METHODS Clinical data and active screening microbiological results of 113 patients admitted to the Neurosurgical Intensive Care Unit(ICU)of The First Affiliated Hospital of Soochow University between Jul.2023 and Jan.2024 were retrospectively collected,and the patients were divided into a CRO-infected group(n=28)and a non-CRO-infected group(n=85).Predictive variables were screened using LASSO regression and logistic regression.Risk prediction models were constructed using random forest(RF)and logistic regression,the performance of the mod-el was evaluated by analyzing the area under the receiver operating characteristic(ROC)curve(AUC),calibration curves,and decision curves,and internal validation was performed using the bootstrap resampling method.RESULTS Among 113 neurosurgical ICU patients,28 cases developed CRO infections,with an infection rate of 24.78％.The highest infection rate was observed in lower respiratory tract infection,with 17 cases(15.04％).The predominant CRO pathogens were carbapenem-resistant Klebsiella pneumoniae(CRKP)and carbapenem-re-sistant Acinetobacter baumannii(CRAB),accounting for 50.00％and 42.86％of cases respectively.The AUC values for the RF prediction model and nomogram prediction modeling groups were 0.881 and 0.787 respectively,with Brier scores of 0.114 and 0.146,and threshold probabilities of net benefit ranging from 10％to 97％and 12％to 62％respectively.The RF prediction model exhibited superior discrimination,calibration,and clinical u-tility.The RF prediction model demonstrated that days of combined use of meropenem and vancomycin,GCS score,intestinal colonization,and hospitalization history were important predictors for CRO infections.CONCLUSION The prediction model for CRO infections in neurosurgical ICU patients established based on random forest algorithm has good predictive performance,and can be intervened with preventive and control measures for important predictive factors.
