Research On Influencing Factors And Early Recognition Model Of IVIGR Kawasaki Disease Based On Machine Learning

Objective Our study uses hospital clinical medical records data to explore the clinical epidemiological characteristics of KD in Northwestern China with children with KD hospitalized in a tertiary general hospital.Machine learning is used to explore the influencing factors of IVIG-insensitive KD and construct IVIG suitable for the Northwestern region The insensitive KD prediction model is designed to identify children who are not sensitive to IVIG at an early stage,so as to guide clinicians to formulate reasonable preventive measures and treatment plans to prevent adverse events.Methods The case data of children with KD who were hospitalized in the Pediatric Department of the Second Hospital of Lanzhou University from February2015 to July 2020 were collected,including: demographic characteristics,clinical manifestations,laboratory examinations,cardiac color Doppler follow-up results,and summary of KD children Epidemiological characteristics.Divide children with KD into IVIG-sensitive group and IVIG-insensitive group,use Lasso-Logistic regression and GBDT decision tree machine learning methods to explore the influencing factors of IVIG-insensitive KD and build a predictive model.Use the area under the ROC curve and the accuracy,Sensitivity,specificity,and SHAP value to evaluate the performance of each model,and finally compare with the previous mature domestic and foreign prediction models(including Egami,Kobayashi,Formosa evaluation,and Yang score)to select the best prediction model.Results(1)A total of 603 children were included in the study,including 400 males and 203 females,with a male-to-female ratio of 1.97:1;the median age of onset was 25 months(range15～40 months),mainly between 0 and 5 years old((Accounting for 90.55%),the largest number of people in the age group of 1 to 3years(accounting for 48.75%);in terms of seasonal distribution: spring(160 cases,accounting for 26.53%),summer(190 cases,accounting for 31.51%),autumn(116cases,19.24%)and winter(137 cases,22.72%).(2)From 2015 to 2020,the number of cases of children with KD increased year by year,with an average annual growth rate of 136.00%;the composition ratio of CAL showed a downward trend(the average annual growth rate was-0.47%);the incidence of IVIGR It also shows a downward trend year by year(the average annual growth rate is-1.29%).(3)The most prone site for KD complicated with CAL is the left coronary artery,with a frequency of 224(42.83%),followed by the anterior descending artery(frequency of 190,36.23%),the right coronary artery(frequency of 80,15.30%),and the left coronary artery.Circumflex branch(29 frequency,5.54%).The types of concurrent CAL are mainly coronary artery dilation and small coronary aneurysms.(4)Follow-up observation was conducted on 325 children with KD who did not develop CAL at the time of admission.Among the 186 children with complete cardiac color Doppler ultrasound,24(12.90%)had complicated CAL.The median time to CAL was 3 weeks(1～6 weeks),the shortest occurrence time is 1 week,and the longest is 10 weeks.Follow-up observation of 278 patients with CAL who had developed CAL at the time of admission.49 children with dilated,small and medium-sized coronary aneurysms had not reached the end of follow-up(18 cases)or were lost to follow-up(31 cases)because of the end date of the study.Excluded,the remaining dilated,small and medium-sized coronary aneurysms all had outcomes.Among them,the median prognosis of coronary artery dilation was 3 weeks(6-10weeks),the fastest prognosis was only 1 week,and the slowest prognosis was 33Weeks;the median time to prognosis for small coronary aneurysms is 7 months(4-13months),the fastest prognosis is only 2 months,and the slowest prognosis is 27months;the median coronary aneurysm prognosis time The median was 37.5 months(32.5 to 44 months),the fastest outcome was 28 months,and the slowest outcome was59 months.By the end of the study,none of the 7 cases of large coronary artery tumors had an outcome.Thrombosis occurred in 4 cases.(5)The correlation analysis of KD with CAL and IVIG sensitivity found that the composition ratio of CAL in the IVIG-insensitive group was higher than that in the IVIG-sensitive group,and the difference was statistically significant(χ2=17.03,P<0.01).All types of CAL were statistically significant between the IVIG-insensitive group and the IVIG-sensitive group(P less than 0.01).The incidence of each type of CAL in the IVIG-insensitive group was higher than that of the IVIG-sensitive group.(6)Lasso-Logistic regression model: when λ=0.0155,Lasso regression obtains the optimal solution,2 variables are screened out,12 variables are screened out,and the Lasso screened variables are included in the Logistics regression to construct the Lasso-Logistics regression model It was found that long fever time,high levels of CRP,serum P,TBIL,LDL;low levels of PLT,serum Na,ALB and CHO are independent risk factors for IVIG insensitivity.The AUC of the model is 0.75,the sensitivity is 65.48%,the specificity is 75.66%,and the accuracy is 41.13%.(7)GBDT model: Use 80% of the data set as the test set to build the GBDT model,and use 20% of the data set as the verification set to adjust the hyperparameters in GBDT learning.When the depth of the hyperparameter tree is 5,the GBDT model performs best Based on the best parameters,the GBDT model is constructed using all the data sets.Its AUC is 0.87(95CI: 0.85～0.90),the sensitivity is72.62%,the specificity is 89.04%,and the accuracy is 61.65%.The SHAP value is used to evaluate each feature value pair.The degree of contribution of the model found that the most important features in the GBDT decision tree model were TBIL,ALB,CRP,fever time,and serum Na.Conclusions KD is mainly concentrated in boys between the ages of 1 and 3,and the incidence is high in spring and summer.From 2015 to 2020,the number of cases of KD has been increasing year by year,and the composition ratio of CAL and the incidence of IVIGR have shown a downward trend.KD complicated by CAL mainly occurs in the left coronary Arteries,mainly coronary artery dilation and small coronary aneurysms;children with KD should be followed up strictly after discharge to prevent the occurrence of adverse events;children with IVIG-insensitive KD are more likely to develop CAL;the GBDT model in this region is The optimal recognition model of IVIG-insensitive KD concurrent CAL.