Construction and validation of prediction models for delayed encephalopathy after acute carbon monoxide poisoning based on machine learning
10.3760/cma.j.cn114656-20250610-00440
- VernacularTitle:基于机器学习的一氧化碳中毒迟发性脑病的预测模型的建立与验证
- Author:
Yanwu YU
1
;
Yan ZHANG
;
Ding YUAN
;
Huihui HAO
;
Fang YANG
;
Hongyi YAN
;
Pin JIANG
;
Mengnan GUO
;
Zhigao XU
;
Changhua SUN
;
Gaiqin YAN
;
Lu CHE
;
Jianjun GUO
;
Jihong CHEN
;
Yan LI
;
Yanxia GAO
Author Information
1. 郑州大学第一附属医院郑东院区急救中心，郑州 450000
- Keywords:
Acute carbon monoxide poisoning;
Delayed encephalopathy after acute carbon monoxide poisoning;
Machine learning;
Prediction model
- From:
Chinese Journal of Emergency Medicine
2025;34(10):1403-1409
- CountryChina
- Language:Chinese
-
Abstract:
Objective:s To investigate the risk factors for delayed encephalopathy after acute carbon monoxide poisoning (DEACMP) in patients with acute carbon monoxide poisoning (ACOP) and to develop predictive models based on machine learning algorithms.Methods:Patients with ACOP hospitalized at the First Affiliated Hospital of Zhengzhou University from August 2019 to October 2024 were included, with the occurrence of DEACMP as the outcome measure. The dataset was randomly divided into training and validation sets at a ratio of 7:3. Lasso regression was used to select features influencing the outcome in training sets. Nine machine learning models—including Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM)—were constructed. Receiver operating characteristic (ROC) curves were plotted and the area under the curve (AUC) calculated for each model. Calibration curves were used to assess accuracy, and decision curve analysis (DCA) was applied to evaluate clinical utility. The SHapley Additive exPlanations (SHAP) method was employed to visualize and interpret the best-performing model.Results:A total of 264 ACOP patients were included, of whom 54 (20.5%) developed DEACMP. Lasso regression identified eight key feature variables. Based on these factors, predictive models were constructed, showing good AUC stability across the nine machine learning models in both training (0.92–0.99) and validation sets (0.85–0.91). The RF model performed best, with an AUC of 0.99 in the training set and 0.90 in the validation set; its calibration curve and DCA curve also demonstrated excellent performance. SHAP analysis of the RF model revealed the importance ranking of factors from highest to lowest as follows: Glasgow Coma Scale (GCS) score, duration of coma, age, history of coronary heart disease, CK-MB level, monocyte count, diastolic blood pressure (DBP), and drinking history.Conclusions:The RF model exhibited the highest predictive performance for DEACMP occurrence in ACOP patients. The influencing factors, ranked in order of importance from highest to lowest, are as follows: GCS score, duration of coma, age, history of coronary heart disease, CK-MB level, monocyte count, DBP, and drinking history.
