Classifier Of Kernel Logistic Neural Network And Its Application In The Diagnosis Of Hemodialysis

Classifier of kernel logistic neural network,on the one hand,combines the capacity of dealing with nonlinear characteristics and the probability description ability of kernel logistic regression,which ensure the generalization performance of the model.On the other hand,the innovation and improvement of the network structure make the model structure more optimized.We use the Restricted Boltzmann Machine(RBM)to pre-train the model in order to get the initial parameters of the network.Respectively,the thesis discusses the binary classification model and multi-class classification model of kernel logistic neural network based on RBM.Then the model is applied to the hemodialysis diagnosis,establishing the hemodialysis evaluation model.The main contents of this thesis are listed as follows:Firstly,a binary classification model of kernel logistic neural network based on RBM is proposed in this thesis.The model which is the combination of kernel logistic regression and artificial neural network has the advantages of independent learning ability and dealing with linear inseparable problem.The initial value of parameters to be identified can be got by unsupervised learning of RBM,which reduces the effects of the randomness of initial parameters.The model adopts variable learning rate with scaling factor to identify parameters,and the convergence speed of the model can be improved by dynamic regulating the learning speed.Numerical simulation results shows that the proposed binary classification model of UCI data sets and the measured medical data sets both can get better classification results compared to the other classification algorithms.Secondly,multi-classification model of kernel logistic neural network is proposed based on the binary classification model in this thesis to deal with multi-class classification problems.The model uses the “one-versus-all”(OVA)multi-classification algorithm to divide the model into more than one binary classifier.A maximum strategy which chooses class label with the biggest output probability as the final category is applied to the output.Numerical simulation using cross validation method,takes the average of simulation results as the final results.The simulation results of the UCI data sets and the measured medical data sets show that the multi-classification model can obtain more accurate classification results than other classification method.Finally,in allusion to the key attribute selection problem of hemodialysis,the thesis employs multi-classification model of kernel logistic neural network to classify and predict the uremia data.T-S fuzzy neural network is used to set up a systematic hemodialysis evaluation model based on the selected nine key indicators.And the parameters of the evaluation model are optimized used optimization algorithm.Then the effectiveness of the evaluation model is verified by comparing the hemodialysis data with the other ill data and health data and analyzing the survival curve of the model.