Construction and validation of machine learning-based dynamic early warning model for mortality risk in trauma-induced hypothermia patients
10.19745/j.1003-8868.2025041
- VernacularTitle:基于机器学习的创伤低体温患者死亡风险动态预警模型构建及验证
- Author:
Yi-jing FU
1
;
Jing YUAN
;
Guan-jun LIU
;
Qing-yan XIE
;
Jia-meng XU
;
Wei CHEN
;
Guang ZHANG
Author Information
1. 天津理工大学天津市先进机电系统设计与智能控制重点实验室,天津 300384;天津理工大学机电工程国家级实验教学示范中心,天津 300384;军事科学院系统工程研究院,天津 300161
- Publication Type:Journal Article
- Keywords:
machine learning;
hypothermia;
mortality risk;
real-time dynamic early warning;
trauma
- From:
Chinese Medical Equipment Journal
2025;46(3):9-14
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a dynamic early warning model based on machine learning methods and validate its predi-ctive efficacy so as to achieve precise assessment and early warning of mortality risk in patients with traumatic hypothermia.Methods Firstly,a total of 480 patients who met inclusion criteria were retrospectively selected from the eICU database and randomly divided into training and test sets at an 8∶2 ratio.Secondly,physiological parameters were extracted from these patients,and five machine learning algorithms including XGBoost,AdaBoost,LightGBM,logistic regression(LR)and random forest(RF)were employed respectively to develop dynamic mortality risk warning models for traumatic hypothermia patients,utilizing a 1-hour observation window.Thirdly,receiver operating characteristic curves(ROC)were plotted using the test set data and the effects of different warning windows on the model performance were analyzed by calculating the AUC.Finally,the interpretability of the models was analyzed using the SHapley Additive exPlanations(SHAP)algorithm to elucidate the contribution of each feature to predictive performance.Results The optimal warning window for the dynamic warning model constructed using the eICU database was 12 hours,and in case of 12-hour warning window the logistic regression model achieved the highest AUC of 0.935 and showed optimal predictive performance.The results of the interpretability analysis by the SHAP algorithm showed that body temperature was the feature that had the greatest impact on the model results,and its reduction was positively correlated with the increased risk of death.Conclusion The machine learning-based dynamic warning model for mortality risk in traumatic hypothermia patients enables real-time dynamic risk assessment,providing robust support for clinicians to identify the patient's condition changes at an early stage and references for the adjustment of clinical treatment programs.[Chinese Medical Equipment Journal,2025,46(3):9-14]
