Establishment of children's septic shock prediction model with"Grid search+XGBoost"algorithm
10.11855/j.issn.0577-7402.2020.12.10
- VernacularTitle:"网格搜索+XGBoost"算法建立儿童脓毒性休克预测模型
- Author:
Jun GONG
1
;
Xiao-Gang ZHONG
;
Jun-Tao TAN
;
Yun-Yu LIU
;
Qing-Mao RAO
;
Tian-Yu XIANG
;
Hui-Lai WANG
Author Information
1. 重庆医科大学医学数据研究院,重庆 400016
- Keywords:
XGBoost algorithm;
logistic regression analysis;
children septic shock;
sepsis;
prediction model
- From:
Medical Journal of Chinese People's Liberation Army
2020;45(12):1270-1276
- CountryChina
- Language:Chinese
-
Abstract:
Objective To screen the risk factors of children septic shock and establish the children septic shock predictive model,so as to provide an early warning for the occurrence of disease.Methods One thousand five hundred and fifty eight cases of sepsis children,aged less than 14 and admitted to the 7 affiliated medical institutions of Chongqing Medical University from January 1,2015 to August 31,2019,were selected in present study.According to whether septic shock occurred during hospitalization,they were divided into septic shock group(study group,287 cases)and simple sepsis group(control group,1271 cases).The independent risk factors were selected using univariate analysis+logistic regression analysis,grid search algorithm was used to search for the optimal parameters of XGBoost algorithm,the prediction model was constructed by XGBoost algorithm.Results A total of 80 indicators were collected,14 indicators with a missing rate of>30%were excluded,and 66 indicators were finally included.Forty one differential indicators with statistical significance were screened by using univariate analysis,and 10 independent risk factors were screened by using logistic regression analysis,including:increased urine microalbumin,more common leukocytes,urine protein positive,low calcium ion,high lactate dehydrogenase,high uric acid,low albumin,high myoglobin,high creatine kinase isoenzyme MB,and high procalcitonin.The grid search showed the best model performs when the XGBoost algorithm parameter max_depth=6 and eta=0.1.The tested prediction model had an area under the curve(AUC)of 0.757,a sensitivity of 0.727,and a specificity of 0.768.The performance of the model was improved compared to that of previous studies.Conclusion The clinical prediction model constructed by"grid search+XGBoost"algorithm has a good prediction effect on septic shock in children,and can be used to predict children septic shock in Chongqing.
