Application Of Machine Learning Models For Prediction Of Advanced Schistosomiasis Prognosis

Objectives(1)In order to understand the current distribution and treatment status of advanced schistosomiasis patients in Hubei province,a prevalence survey was conducted,which could also provide references for improving provincial management level of advanced schistosomiasis treatment.(2)Taking the survey data in the first part as a sample,the prognosis of the advanced schistosomiasis patients of integral province was predicted based on machine learning algorithm models.The prediction performance of the machine learning algorithm models were compared by the area under the ROC curve(AUC),sensitivity,specificity and so on.(3)Taking the advanced schistosomiasis patients in Xiaonan as an example,the relationship between four indexes of liver fibrosis,B-mode ultrasonography examination indicators,liver function indices and the prognosis of advanced schistosomiasis were studied.Methods(1)The survey program was designed by Hubei Schistosomiasis Control Institute and field investigation was organized by professional physicians in all epidemic counties in Hubei province.According to the corresponding case inclusion and exclusion criteria,4136 eligible patients were incorporated in this study.The clinical,laboratory examination and epidemiological survey data of advanced schistosomiasis patients were collected for analysis.Clinical outcome and treatment costs were also investigated.(2)Based on the survey data of the first part,the patients were divided into two groups:good prognosis and poor prognosis.The occurrence of the event(death or deterioration)was coded as 1,and the event not occurred(clinically cured or improved)was coded as 0.Mortality in advanced schistosomiasis was mainly due to schistosomiasis and complications caused by schistosomiasis such as liver cancer,hepatic coma,hepatorenal syndrome(HRS)and upper gastrointestinal bleeding.Therefore,death in this study referred to all-cause death.The deterioration referred to the persistence of the main symptoms(e.g.no evidence of ascites reduction or the absence of surgical indications for splenomegaly type patients).70% of the patients(2896)were randomly assigned to the training group,while 30%(1240)were assigned to the testing group.The k NN,SVM,ANN,DT,LR,NBN,TAN,GBN,RF,and GBDT models were all constructed in R3.4.3 software(R Core Team R,2017).The Euclidean distance was calculated in k NN model.The SVM model applied the kernel function to map the low dimensional data to the high dimensional plane.The ANN model applied the standard feed forward back propagation(BP)network structure.DT model was based on the C4.5 algorithm.The three Bayesian network models incorporated the prior knowledge.RF and GBDT models integrated 5,000 decision trees for prediction.For all comparisons,P<0.05 was considered statistically significant with a two-sided test.The models' predictive performance comparison was based on a comprehensive analysis of the area under the ROC curve(AUC),accuracy,sensitivity,specificity and so on.(3)In the sixth part,104 cases of advanced schistosomiasis patients diagnosed in 2015 were collected from Xiaonan schistosomiasis specialized hospital.The patients were divided into two groups: good prognosis and poor prognosis.The relationship between four indexes of liver fibrosis,B-mode ultrasonography examination indicators,liver function indices and prognosis of advanced schistosomiasis were studied.Results(1)In this study,the advanced schistosomiasis patients were mainly in Jingzhou,Huangshi,Xiaogan and so on.The clinical types were mainly ascites and splenomegaly types.In demographic characteristics,age,gender,BMI,developmental and nutritional status differed between ascites and splenomegaly groups which has statistical significance.A number of indicators differed between the elderly group and the young adults group through the investigation of past medical history,symptoms and signs,laboratory tests,B-mode ultrosound and X-ray examination.Of clinical treatment,splenectomy,history of ascites,treatment means,treatment costs and treatment outcome differed between the two age groups which had statistical significance.(2)The study of k NN and SVM models have shown that they were both appropriate for predicting the prognosis of advanced schistosomiasis patients(AUC> 0.75).The k NN model was more sensitive and the SVM model was better in specificity.The study also found that,under different parameters,the prediction performance of the model gradually increased with the decreasing of k parameter.The comparison of ANN,DT and LR model have shown that each prediction model was proved to be effective and had its own advantages,the ANN model performed superior to LR and DT models in terms of AUC and sensitivity.Comparisons of the three Bayesian network models have shown that the predictive performance of the three models were applicable.NBN(AUC = 0.724)and TAN(AUC=0.737)were more suitable for prognosis prediction of advanced schistosomiasis patients,because the AUC values of NBN and TAN were higher than GBN(AUC = 0.658).However,of the models of NBN and TAN,although AUCs were approximate,TAN model may be more consistent with people's daily cognition and more interpretable because the interdependence of different variables were taken into account.For the ensemble learning model,both random forest and gradient boosting decision trees(parameter n.tree = 5000)yielded good results(AUC> 0.75).In the training group,the main indexes(AUC,sensitivity and specificity)of the RF model were better than the GBDT and DT models.In the testing group,the AUC of RF model was also superior to GBDT and DT model,while the sensitivity and specificity were closed to each other.However,it was noteworthy that the predictive performance of GBDT did not show superiority on DT model.This may be related to the adjustment of the parameters of the model or the characteristics of the dataset itself.The impact of the parameter adjustment process on the prediction performance of the ensemble learning model was also explored in this section.The choice of prognosis prediction model should be made after performance comparison,as well as combining with the actual needs of specific medical problems.(3)Of four liver fibrosis indices,hyaluronic acid(HA)and laminin(LN)could be used as clinical indicators for prognosis of advanced schistosomiasis.B-ultrasound results have shown that ascites could be used as a key prognostic indicator in advanced schistosomiasis patients.Of the liver function indices,AST / ALT could be applied as prognosis index.Conclusions(1)Investigating the epidemiological status of advanced schistosomiasis patients in Hubei Province was helpful to provide baseline information and basis for understanding the distribution of advanced schistosomiasis patients,improving the management level of treatment and formulating reasonable prevention and treatment strategy for advanced schistosomiasis patients.(2)In this study,we applied 9 kinds of machine learning models such as kNN,SVM,to predict the prognosis of advanced schistosomiasis and achieved good prediction performance.If data materials did not satisfy specific distribution requirements,parametric models(such as LR models)and semi-parametric models(such as Cox proportional hazards models)were not applicable.This study provided new methods which helped to compare the results between different methods.The k NN,SVM and ANN model's data requirements were lower than traditional prognostic prediction models.The DT model has shown a clear process of variable selection and was easy to understand.The three BN models were explicable when analyzing the interaction among many independent variables.The ensemble learning model overcame the shortcomings of the lack of generalization in a single decision tree.Moreover,the machine learning model was easy to adjust parameters and could thus generated prediction models with better predictive performance.(3)Some indices of liver fibrosis(hyaluronic acid and laminin),B-ultrasound examination(ascites)and liver function tests(AST/ALT)could be used as clinical indicators of advanced schistosomiasis prognosis.