Risk Factors And Prediction Model For Intravenous Immunoglobulin Resistant Kawasaki Disease In Shenzhen Children’s Hospital

Objective:The present study was designed to identify risk factors for intravenous immunoglobulin resistant Kawasaki disease and build a predictive model for intravenous immunoglobulin resistant Kawasaki disease in Shenzhen.Methods:(1)Clinical data collection:Patients hospitalized in Shenzhen Children’s Hospital from January 2014 to December 2018 with main discharge diagnosis of Kawasaki disease were enrolled into the study.Clinical data were logged into the Redcap program,comprising demographic information,laboratory examination before IVIG treatment(blood routine test data containing both before and48 hours after IVIG treatment),treatment,management,and echocardiogram results.(2)Processing the primary data: After establishing the KD database on the Redcap program,the primary data was exported.While the IVIGRKD cases were picked up according to diagnostic criteria from AHA,cleaning the data,missing data imputation,and standardize the variables.(3)Feature selection: Python was used for machine learning.70% cases were randomly selected as the training set and 30% cases as the verification set.Based on Support Vector Machine classifier modeling,there are three different methods to feature selection,including the classical difference test,Least Absolute Shrinkage and Selection Operator(Lasso),Recursive feature elimination(RFE),which were compared for the best AUC.(4)Building the prediction models:After the above steps,the method of optimal feature selection is obtained,and then the model is built with three commonly used classifiers suitable for the study with relatively small sample size,namely Logistic regression,SVM and Random Forest.(5)Verifying the prediction models: 30% of the cases were randomly selected as the validation set,and the AUC value,sensitivity,specificity,accuracy and 95%confidence interval of the three models were calculated for comparison.(6)Evaluatingreliability of the prediction models: The reliability of the new model was evaluated by comparing the indexes of sensitivity,specificity,area under ROC curve and 95%confidence interval with the existing IVIGRKD scoring systems.(7)Data analysis was conducted using R Project for Statistical Computing(R version 3.6.0).All data were presented as count with percentage for categorical variables and mean±standard deviation(SD)for continuous variables.Variables with missing value≥10% were deleted.Variables with missing value<10%,category variables were filled with the largest proportion and continuous variables were filled with Bayesian linear regression.The training set was used to calculate the mean and variance to standardize the variables.T test was used for quantitative data conforming to normal distribution,and Kruskal-wallis H test was used for qualitative data failing to conform to normal distribution.Chi-square test was used for qualitative data,P<0.05 suggesting that the difference was statistically significant.The prediction model was visualized by drawing a nomogram based on the Lasso-Logistic regression model.Results: Total of 833 subjects were enrolled into this study,including 107 cases of IVIG resistant(12.8%)and 726 cases of IVIG responsive(87.1%).Based on SVM classifier,to build models with three feature selection methods,Lasso-SVM showed an AUC value of 0.878 which is greater than that of Test-SVM(AUC=0.872)and RFE-SVM(AUC=0.860).It is believed that Lasso is more suitable for feature selection in this study than the classical difference test methods and RFE.13 IVIGR risk factors were selected using Lasso,including age,mean platelet volume(MPV),and platelet to lymphocyte ratio(PLR)in the blood routine test before IVIG treatment,Neutrophil(NEU),platelet count(PLT),hemoglobin(HB)and red blood cell count(RBC)in the blood routine test after IVIG treatment 48 hours,aspertate aminotransferase(AST),serum sodium concentration and total protein(TP),albumin(ALB),presence or absence of Pyuria,Z value of the anterior descending branch of the left coronary artery(LADZ).The AUC=0.885(95%CI 0.8505~0.9485),the sensitivity=0.888(95%CI 0.6384 ~0.8837),the specificity=0.685(95%CI 0.7027~0.9706),and the accuracy=0.886(95%CI 0.8536~0.9476)of the Lasso-Logistic regression model were constructed by Bootstrap method for 1000 times.And the results can beconverted into a visual nomogram,which is easy to understand.The AUC value of the Lasso-Logistic regression model was significantly higher than that of the existing scoring system.Conclusion:(1)We built up the prediction model of IVIGRKD in Shenzhen Children’s hospital and found three new risk factors including MPV and PLR before IVIG treatment and LADZ compared with the previous IVIGRKD scoring system.(2)The new model of predicting IVIGRKD was based on Lasso-Logistic regression and translated into a visual nomograph,which is suitable for clinical application.(3)The AUC value of the new prediction model is higher than the existing scoring system,which can provide a basis for the prediction of IVIGRKD in shenzhen.