Analysis Of Risk Factors Of Sepsis In Chidren Based On Contribution Of Features

Sepsis in children have a high incidence rate,rapid progression,complex pathogenesis and diverse clinical manifestations.Due to the lack of gold diagnostic criteria,it is difficult to make early diagnosis.In the case that traditional data analysis is difficult to effectively analyze high-dimensional and complex linear data,machine learning can be used to identify risk factors of sepsis in children,so as to achieve early diagnosis.Aiming at the characteristics of high missing rate and complex correlation of medical data,a feature contribution evaluation method based on machine learning is proposed.The gradient lifting tree is used to construct the classification prediction model,and then the Shapley additive explanation method of tree model(Tree SHAP)is used to construct the interpreter model,and the contribution of each feature to the model prediction is analyzed to measure the importance of the feature.Due to the heterogeneity of sepsis,it is necessary to classify the subtypes of sepsis,so a subtype classification method based on the non-zero estimation feature of Tree SHAP method is proposed.In order to ensure the interpretability of medicine and the operability of indicators in clinic,the iterative feature selection driven by medical knowledge is adopted.After finding the important indicators in each subtype,the statistical method have limited effect on the analysis of risk factors.Tree SHAP method is used to analyze the performance of indicators in the prediction of sepsis,and Bayesian network is constructed to find the dependence between abnormal indicators.By using the method of risk factor identification and reasoning based on feature contribution,the children sepsis cases in Tongji Hospital were divided into 3 age groups according to the non-zero estimation feature of age.In each age group,12 important indicators were identified and screened,and the number of indicators decreased by 86%.However,the performance of the prediction model constructed by 12 indicators is similar to that of 87 indicators,which verifies the importance of these 12 indicators.There are differences in risk factors for different age groups,which has been recognized by doctors.