| ObjectivesIn medical research,mixed linear models are often used for effect estimation of actual data with features of Hierarchical Structure,but it is less efficient for further exploration of complex(high-order)interactions between patient subgroups and treatment modalities.As a kind of prediction model in data mining method,decision tree model has strong exploration performance,can analyze multiple predictive variables at the same time,and can automatically detect the potential relationship between predictors,which is of great significance for medical efficacy evaluation.However,for the Hierarchical Structure Data,the traditional CART regression tree model is inefficient,and the model construction accuracy is low and the bias is large.Therefore,improving the fitting effect and predictive performance of the regression tree model is an important prerequisite for applying it to the Medical Hierarchical Structure Data.The Random Effect-Expectation Maximization Regression Tree(RE-EM)is based on the idea of the regression tree model and the liner mixed effects model,and separates the random effects from the Hierarchical Structure Data to improve the accuracy and predictive performance of regression tree model.The purpose of this study is to introduce and validate the principle of the Random Effect-Expectation Maximization Regression Tree model and to compare it horizontally with the traditional CART regression tree model.At the same time,based on the real data of the diagnosis and treatment of chronic hepatitis B patients,the empirical analysis and simulation experiments were carried out to further evaluate the fitting effect and prediction performance of RE-EM regression tree model under different data structures.Provide methodological support and model selection for more accurate medical efficacy evaluation and regression tree model construction,and provide methodological advice for patient diagnosis and intervention.MethodsIn the first part,the random effect and residual covariance structure are set,and the regression tree model simulation data of three terminal nodes is generated.The RE-EM regression tree model and the CART regression tree model were respectively fitted to evaluate the accuracy and bias of the two regression tree models,and the applicability of the RE-EM regression tree model to the fitting Hierarchical Structure Data was verified.The second part is based on the antiviral treatment information of patients with chronic hepatitis B in the clinical HIS data,and fits the general linear model,liner mixed effects model,CART regression tree model and RE-EM regression tree model for evaluating antiviral effect.In the third part,based on the results of four models of chronic hepatitis B antiviral efficacy analysis,the fitting effect and predictive performance of RE-EM regression tree model under different parameters(sample size,time point,residual correlation,underlying model)were explored.For example,the sample size of the research object is set to 50,100,200,500,1000,corresponding to the time points 10,20,50,100.The predictive performance evaluation is divided into two parts:(1)Predicting new observations of the research object,using the top 70% of the observations as the training set,respectively fitting the four models,and the remaining 30% as the test set to evaluate model prediction performance;(2)Predict the new research object,with 70% of the research objects as the training set,respectively fit the four models,and the remaining 30% as the test set to evaluate the model prediction performance.ResultsThe first part of the results show that the RE-EM regression tree model outperforms the CART regression tree model in terms of model construction accuracy and fit bias.Under different data structures,the REEM regression tree model can accurately construct the assumed regression tree model,while the CART regression tree model is not ideal for the fitting of system structure data,the MSE value is large and the assumed regression tree model cannot be accurately constructed.The second part of the results show that the liner mixed effects model performs better than the general linear model in fitting fit and bias.In the analysis of the efficacy of chronic hepatitis B patients based on the general linear model,the effect of time-dependent ALT levels on the quantitative HBV DNA detection was statistically significant,white the liner mixed effects model considering the random effect and the residual covariance structure has no such relationship.In the analysis of the regression tree model,the MSE of the RE-EM regression tree model is the smallest,0.8048,which is lower than the general linear model,the liner mixed effects model and the CART regression tree model.The third part of the results show that in the model fitting effect evaluation,when the linear simulation data does not contain random effects,the linear model fitting effect is better than the regression tree model,while for the nonlinear data,the regression tree model has a good fitting effect.It is better than the linear model,and the fitting effect of the RE-EM regression tree model is similar to the CART regression tree model.When the simulation data is Hierarchical Structure Data,the liner mixed effects model is similar to the RE-EM regression tree model,which is better than the general linear model and the CART regression tree model.In combination with various situations,the RE-EM regression tree model has a better fitting effect and is superior to the CART regression tree model.In the performance evaluation of model prediction new observations,for linear simulation data with features of Hierarchical Structure,the performance of linear model prediction new observation is better than regression tree model,and the prediction performance of RE-EM regression tree model is better than CART regression tree model.For nonlinear simulation data with hierarchical structural features,RE-EM regression tree model and liner mixed effects model predict the best performance of new observations,which is better than CART regression tree model.The general linear model has the worst fitting effect.In the performance evaluation of the model prediction new object,when the linear simulation data does not have the hierarchical structural features,the linear model predicts the performance of the new object better than the regression tree model.When the nonlinear simulation data does not have hierarchical structural features,the regression tree model predicts that the performance of the new object is better than the linear model,which is consistent with the foregoing.For system structure data,the RE-EM regression tree model predicts that the performance of new objects is always optimal.Combining various situations,the RE-EM regression tree model predicts better performance of new objects.ConclusionsFor the Hierarchical Structure Data,the RE-EM regression tree model can effectively identify the potential connections between the predictors and improve the fitting effect of the model,which reflects the applicability and feasibility of the RE-EM regression model in the Hierarchical Structure Data.It can be seen from the modeling process of the regression tree model that compared with the linear model,the tree model is composed of the root node to the terminal node,which is similar to the human decision form,and the result is intuitive and concise and has strong explanatory.In this study,the effectiveness of the RE-EM regression tree model for analysis of medical Hierarchical Structure Data was verified by simulation experiments and evaluation of anti-viral efficacy of chronic hepatitis B.The fitting effect and prediction performance of the RE-EM regression tree model for the Hierarchical Structure Data are better than the CART regression tree model.For linear Hierarchical Structure Data,the RE-EM regression tree model predicts that the performance of new objects is close to or even better than that of the liner mixed effects model. |
PDF Full Text Request