Research On High-dimensional Feature Analysis Methods Of Alzheimer's Disease Based On ELM And FMR

The utilization of information technology to analyze brain images from functional Magnetic Resonance Imaging(fMRI)is one of the important methods to study brain science.Effective fMRI brain image analysis methods are an important foundation for medical image information management,neurological information management,and health management,which can promote brain science research.The fMRI brain image data has a high feature dimension.In research,feature screening is usually required in order to reduce the time and space complexity of the algorithm,avoid dimensionality catastrophe,and obtain more desirable analysis results.The common feature selection methods often have incomplete theoretical foundation or partiality of feature to law fitting.To address these problems,this paper proposes to incorporate all features into a unified framework based on the extreme learning machine(ELM)model,and conducts related research such as data classification,feature importance identification,and analysis.First,a Parallel Extreme Learning Machine(P-ELM)method is proposed for the fMRI brain image-based assisted diagnosis of Alzheimers’ diease(AD).Based on ELM,this research includes all brain functional connectivity into the training of the classification model to avoid the problems of inaccurate and incomplete feature selection.The concept of average accuracy convergence of ELM models,and the training and screening methods of optimal ELM classifiers are proposed to improve the stability of classifier accuracy.In addition,the P-ELM method has the advantage of low time complexity and achieved a diagnostic accuracy of 96.85%and 95.05% for AD and Mild Cognitive Impairment(MCI),respectively.Second,to address the problem of unreliable prior information often used in feature screening and the problem that features do not fit the labeling law well,this paper proposes the idea of pruning artificial neural networks to achieve key feature screening and designs the Key Features Screen on Extreme Learning Machine(KFS-ELM)method.The method uses all features and data category labels for model training,fitting the pattern between the features and the category labels.The research work focuses on dealing with the overfitting problem of classifiers brought by pruning,avoiding the regularity problem of a single ELM classifier fitting only some features,and also achieving the identification of feature importance.The experimental results of using the KFS-ELM method to screen key features of Cognitively Normal(CN)subjects and AD subjects show that the method can effectively screen out key features and assign weights to the importance of key features.The higher the weight value of key features,the higher their importance for AD diagnosis.Features with a weight value of 0 have a negative impact on the accuracy of AD diagnosis.The accuracy rate of AD diagnosis using key features reached 99.20%.Compared to using all features for AD diagnosis,the number of features used was reduced by about 77% and the accuracy rate was increased by3.87%.Further,to address the problems of imprecision of various methods to identify the importance of features,or that increasing the accuracy will bring higher space complexity and time complexity,the Extreme Learning Machine Feature Lens(ELM-FL)method is proposed.By analyzing the ELM model,it is found that the part most closely related to the feature importance is the output weight.In this paper,we propose the idea of projecting the output weights onto the space of input data to identify the feature importance,and implement the ELM-FL algorithm based on the Moore-Penrose generalized inverse of the input weight matrix.Experimental results of importance identification of AD subjects using the ELM-FL method show that the method can label the importance weights of all features.The higher the weight of a feature,the higher its importance for classification.The lower the weights of the features,the lower their importance to the classification.It is even found that when the features are ranked from highest to lowest weight,the smaller the weight of the features ranked after 1000,the greater the negative impact on the classification accuracy.By comparing the importance of the features identified by the KFS-ELM method and the ELM-FL method,it was found that the results of the two methods have a high degree of similarity and can verify each other’s validity.In addition,the ELM-FL method has the ability to identify the feature importance of a single subject,which has positive implications for the study of disease specificity and the diagnosis of a single subject.Finally,a management information system was constructed with the optimal ELM classifier and ELM-FL as the core method to analyze the important characteristic changes in the development of AD-related diseases.The feature images of two adjacent stages were calculated in the order of four stages,Control Normal(CN),Early Mild Cognitive Impairment(EMCI),Late Mild Cognitive Impairment(LMCI)and AD.The analysis of the features images and features found some patterns in the progress of AD-related disorders.Changes in functional connectivity in most of the whole brain were not obviously related to AD and its developmental process.Among the relevant functional connectivity,those that could express disease-common patterns were ranked high in terms of their weight size and showed a more concentrated state.The functional connectivity that could express subject-specific patterns were ranked relatively low and showed a wide range of distribution.In the progression from CN to AD,there are different regions and forms of the brain affected by each stage of the disease,and there are both changes in the Region of Interest(ROI)itself and changes in functional connectivity between ROIs.The changes in functional connectivity do not occur in one direction during the process.At each stage there is a rise of some functional connectivity and a decline of some others.In general the functional connectivity show a decrease from the CN to EMCI stage,a greater decrease from EMCI to LMCI stage,and an increasing trend from LMCI to AD stage.The experimental results show that the management information system designed in this paper can effectively extract,organize and apply the information related to Alzheimer’s disease in fMRI brain images.The management system and the method used can be applied to neuroinformation management,health management,and assisted decision making in medical diagnosis.