Risk prediction models for carbapenem-resistant Acinetobacter baumannii infection in ICU patients established based on 5 types of machine learning algorithms
10.11816/cn.ni.2025-250868
- VernacularTitle:基于5种机器学习算法构建ICU患者耐碳青霉烯类鲍曼不动杆菌感染风险预测模型
- Author:
Chen JIA
1
;
Yan GAO
;
Xili XIE
;
Feng ZHAO
;
Haiming QING
;
Lu WANG
Author Information
1. 北部战区总医院疾病预防控制科,辽宁沈阳 110016
- Publication Type:Journal Article
- Keywords:
Machine learning;
Intensive care unit;
Carbapenem-resistant Acinetobacter baumannii;
Predic-tion model;
Extreme gradient boosting model
- From:
Chinese Journal of Nosocomiology
2025;35(17):2586-2591
- CountryChina
- Language:Chinese
-
Abstract:
OBJECTIVE To establish the an optimal prediction model for carbapenem-resistant Acinetobacter bau-mannii(CRAB)infection in ICU patients based on machine learning(ML)so as to help clinicians to diagnose and make decisions.METHODS The clinical data were collected from the patients who were hospitalized in ICUs of a three-A hospital from Jan.1,2017 to Dec.31,2024 and were randomly divided into the training set and the test set in a 7∶3 ratio.The characteristic variables were selected by means of LASSO regression analysis in combina-tion with multivariate logistic regression analysis.Five types of M L classification models were integrated,the opti-mal model was analyzed and identified.The performance of the prediction model for CRAB infection in the ICU patients was evaluated with the use of sensitivity,specificity,accuracy,areas under receiver operating characteris-tic curves(AUCs),calibration curves,Hosmer-Lemeshow test and decision curve analysis(DCA).The outputs of the ML models were interpreted by Shapley additive explanations(SHAP)and permutation importance.RESULTS A total of 2 904 patients were enrolled in the study,695(23.93％)of whom had CRAB infection.The AUC of XGBoost model was highest in the training set and the test set,respectively(0.994 and 0.907).The result of Hosmer-Lemeshow test showed that the calibration curves of the XGBoost model indicated that the predicated risk was highly con-sistent with the observed risk(x2=7.323 and 4.609,P=0.513 and 0.764,respectively).The DCA curves showed that the XGBoost model performed best within the whole range of threshold,with the highest net profit.The length of ICU stay,use of tigecycline,central venous catheterization,use of carbapenems and use of ventilator were determined as the major predictive factors by means of SHAP.CONCLUSIONS The XGBoost model is established and interpreted by SHAP.It provides bases for screening of the ICU patients at high risk of CRAB infection.
