The Application Of Neural Network Model Based On Principal Component Analysis In Predicting The Cognitive Status Of The Elderly

With the rapid aging of the China population,the problem of Alzheimer's disease(AD)becomes more and more serious,which will become a severe public health problem.The etiology of the disease is not clear and the pathogenesis is very complex.There is no effective treatment for this disease in clinical practice.Therefore,early detection and early diagnosis are especially important.At present,there are many researches about AD.The developed countries attach great importance to the basic research of AD,and have established follow-up cohort of AD.In our country,same studies have also been conducted in the relevant population.There are several significant features of such follow-up data:(1)Data is longitudinal,which means the research objects accept multiple measurements at different time points.The relationship among the measure values is interrelated and complex;(2)Data variables are various and there are many complicated factors affect the disease.(3)The progress of the disease is not linear but has a long incubation period.As the follow-up data of Alzheimer's disease has several features,such as non-linear feature,unbalanced type,complex relationship,and diversification of influencing factors.So it is not easy using statistical methods to accurately predict the cognitive status of the elderly.At present,the linear mixed model is used to deal with the longitudinal data and establish prediction model in most cases.In this model,the data should obey the normal distribution or a specific distribution,the intercept and slope should obey the multivariate normal distribution,and the progress hypothesis of the disease should be linear generally.However,the follow-up data of Alzheimer's disease is difficult to meet the above assumptions fully in practice,In particular,due to the non-linear characteristics of data,making the traditional linear mixed model has some limitations.Therefore,lack of appropriate statistical analysis methods has become an obstacle in predicting nonlinear disease progression.In recent years,Artificial Neural Network(ANN)has been widely used in medical diagnosis,prognosis,survival analysis,clinical decision support and other fields,showing a strong ability of prediction.The input variables of the neural network model are critical to the accuracy of the prediction,and the Alzheimer's follow-up data is a large sample of multivariate.When the follow-up data are analyzed and studied,the information among the variables is highly correlated and overlapping,which makes the statistical analysis of the difficulty and complexity greatly increased.If you blindly reduce the number of variables,you may lose a lot of important information.The principal component analysis can be used to find a few unrelated variables instead of the original variables.The purpose of simplifying the original variables and reducing the number of input nodes neural network model can be achieved,and the speed and precision of network training are improved.Based on the actual data of 4642 cases of the National Alzheimer's Coordinating Center(NACC),this paper studies BP and RBF neural networks based on the principal component analysis to predict the cognitive status of the elderly,comparing the predictive capabilities of these two models.Because this study is a dynamic study on the cognitive status of the elderly,the MMSE difference of the five-year follow-up of the respondents was used as the indicator of cognitive function.The cognitive status of the elderly was predicted to be related to the associated risk factors Value of the nonlinear problem.In this study,the principal component analysis was used to descend the dimension of the follow-up data of Alzheimer's disease.In the premise that the original data information can be preserved as much as possible,the cumulative contribution rate of the top 8 principal components can reach 80.15%.The 8 principal components were used as the input variables of BP and RBF neural network prediction model for the network training of cognitive status of the elderly.On the bases of principal component analysis results,the BP neural network model is constructed to determine the number of BP neural networks and the number of neurons in each layer.The Sigmiod function was selected as the activation function of the hidden layer and the output layer.The follow-up data of 3000 respondents were selected as training samples,and the BP neural network was trained and studied.The results show that,the Levenberg-Marquardt learning algorithm was chosen to be the optimal algorithm for the network after comparing various kinds of BP neural network algorithms.When the number of nodes in the hidden layer is 12,the BP neural network has a good approximation to the function,and the training speed is the fastest.Then,the follow-up data of 1642 respondents were simulated by the trained network.Finally,the BP neural network prediction result was obtained with the maximum relative error of 2.0471 and the average relative error of 0.4410.At the same time,the RBF neural network model is constructed.RBF neural network is selected as a three-layer network,the number of nodes in each layer is determined,the transfer function of hidden layer is Gaussian function,and the transfer function of output layer is linear function.The principal component analysis results are taken as input variables,and the network training and simulation are carried out with the same data set of BP neural network.In the process of network training,the Spread value is adjusted continuously,and we drew the conclusion that when the Spread value is 1.0,the training time is the shortest,while the prediction accuracy is the best.It was confirmed RBF neural network has better approximation ability on data.Then,the follow-up data were simulated by the trained network,it was confirmed that RBF neural network prediction result of the maximum relative error is 1.2572 and the average relative error is 0.3364.The conclusions of this study are as follows: 1.The principal component analysis method can reduce the number of variables involved in data modeling largely,the training speed of the neural network is greatly improved,and the training accuracy is guaranteed;2.It was confirmed that BP neural network and RBF neural network both have strong ability to deal with nonlinear problems,by adjusting the structure and parameters of BP neural network and RBF neural network,the two models can get the ideal prediction results;3.Compared with BP neural network,the RBF neural network has faster convergence,smaller prediction error and better stability;4.On the basis of RBF neural network model,We can predict cognitive status of the elderly over longer periods of time.This study is of great significance for predicting the cognitive status of the elderly,taking early targeted prevention and treatment,and the timely preparation of the family,improving the quality of healthcare and reduce the burden of the family.