Study On Partial Least Squares Variable Screening Method For Chinese Dosage Effectiveness Data

The dose-efficacy study of traditional Chinese medicine(TCM)is a key research direction in TCM clinical practice,which reflects the thought of TCM syndrome differentiation and treatment.However,the dose-effect relationship of TCM is still in the exploratory stage,so we mainly study it based on data mining.Partial least squares(PLS)is widely used in data analysis of TCM,but the variable selection of linear and nonlinear models is based on the selection of components.The variables were recombined through the principal components,and then the first n components were selected according to the weight.There was no direct variable screening for independent variables.In other words,the interaction of variables and the interaction of predictive variables are not involved.Therefore,in the analysis of the dose-effect relationship of TCM prescriptions,it is necessary to use the data processing method to process the data,and then apply it to the model analysis of the dose-effect relationship.We mainly carry out a series of research work based on the optimization of PLS,the specific content is as follows:(1)Combining with the characteristics of path analysis,we analyze the relationship between variables from the perspective of covariance matrix and correlation matrix,and put forward a PLS variable screening method based on path analysis.By calculating the direct and indirect path coefficients,the direct effect of each variable on the dependent variable and the indirect effect of other independent variables on the dependent variable can be obtained.Then screen out and eliminate some redundant variables that are of little importance,but because of collinearity with important variables,the overall importance of the redundant variables is very high.In order to verify the effectiveness of this method,two sets of TCM dose-effectiveness datasets and four sets of UCI dataset were used to test its performance.The experimental results show that the selection of independent variables by path analysis can improve the fitting regression of experimental data by PLS method.(2)According to the different compatibility laws of the constituent drugs in TCM prescriptions,a comprehensive weight analysis method combining the analytic hierarchy process(AHP)and deep Belief network(DBN)is proposed.This method ranks the characteristics of the dose-effect data of traditional Chinese medicine and filters out that it has no influence on the dependent variable or Irrelevant variables with little impact.Firstly,AHP was used to introduce the compatibility rules of TCM prescriptions to identify and distinguish different diseases treated by the same prescriptions.Then the interlayer conduction characteristics of the DBN are used to extract the features from the independent variables and obtain the weight matrix.Datasets of Maxingshigan decoction for asthma and fever,and Dachengqi decoction for pancreatitis and obstruction were used to verify the effectiveness of the method.The results show that the comprehensive weight analysis method combining the AHP and DBN can sort and screen the characteristics of TCM data in order of importance,and the model has better fitting degree and higher prediction accuracy.(3)Using PYTHON programming language and development tools,we designed and developed a data analysis auxiliary system for the dose-effect data of Chinese medicine.