关键词: 过敏性紫癜, 肾损害, 机器学习, XGBoost, 预测

Abstract: Ｏｂｊｅｃｔｉｖｅ：To investigate risk factors of renal damage in children with HenochSchnlein purpura (HSP),and to establish a prediction model for children with purpura nephritis.Ｍｅｔｈｏｄｏｌｏｇｙ：533 HSP patients from January 2016 to June 2018 were selected to obtain demographic characteristics,clinical symptoms and laboratory indicators. After feature selection of 31 indicators,Logistic regression and XGBoost algorithm were used for classification prediction,and 50% crossvalidation was used to test the accuracy of algorithm. Finally,we compare the performance of two models.Ｒｅｓｕｌｔｓ：The accuracy of XGBoost prediction model was 0818,higher than that of Logistic regression,which was 0727. The precision rate,recall rate and Fscore of XGBoost model were 0830,0930 and 0877,respectively,and the area under ROC curve was 088,each of which was higher than Logistic regression. After single factor test and ranking of important characteristics of XGBoost model,the variables in the top 10 were:antistreptococcal hemolysin ‘O’ (ASO),urinary NAG,urinary RB protein,serum IgA,age,purpura recurrence,skin purpura site,abdominal symptoms,24hour urinary protein quantification,and neutrophil percentage.Ｃｏｎｃｌｕｓｉｏｎ：XGBoost model can be used to predict occurrence of renal damage in HenochSchnlein purpura. Compared with the traditional Logistic regression algorithm,the prediction accuracy of XGBoost is higher.