Research On Alzheimer's Disease Classification Algorithm Based On Regularized Logistic Regression

Alzheimer's disease(AD)is a degenerative neurological disease,once the disease is likely to be irreversible.The mild cognitive impairment(MCI)is early stage of Alzheimer's disease,and there is a high risk of conversion to AD.Computer-assisted diagnosis of AD is one of the most popular researches in AD diagnosis recently,and it is deeply loved by researchers as a very effective supplement to traditional diagnosis methods.Logistic Regression(LR)classifier is a powerful binary classifier and an important means for machine learning classification tasks.However,the neuroimaging data used in thesis is a very typical high-dimensional small sample data.LR cannot handle this kind of data well.This kind of data may cause the model to have over-fitting problems and poor classification results.This thesis uses magnetic resonance imaging data(MRI)to construct a regularized LR-based AD and its early stage diagnosis algorithm.The experimental effect is based on the classification accuracy as the judgment index.First,an AD diagnosis algorithm based on L2 regularized logistic regression is proposed.This algorithm can effectively diagnose AD and its early stage MCI.The LR model in this part uses the L2 regularization norm to regularize it.The L2 regularization LR model is smooth and convex,and the size of the feature can be limited to a small enough range for easy calculation.The regularization parameter is crossed by ten times Verification is selected,and independent component analysis is used to reduce the dimensionality of the preprocessed data.Finally,Newton's algorithm is used to find the optimal weight of the model.The experimental judgment results AD vs CN(cognitive normal),MCI vs CN,LMCI(late mild cognitive impairment)vs EMCI(early mild cognitive impairment)were 95.22%,81.22% and 74.35%,respectively.Secondly,on the basis of the first diagnosis algorithm,an early AD diagnosis algorithm based on L1/2+2 regularized logistic regression is further proposed.This algorithm classifies AD and its early stage MCI.This method uses L1/2 regularization,a typical Lq regularization method,and at the same time,uses the L2 regularization method to optimize the model.These two regularization norms are linearly combined and can be combined with L1/2 regularization.Make the feature sparse enough while keeping the unbiased oracle attribute and the advantages of the L2 regularization model penalty term as small as possible.After the data is preprocessed and input into the proposed algorithm model,it is possible to select sufficiently sparse features while ensuring the advantages of easy calculation.The experimental judgment results of AD vs CN,MCI vs CN,LMCI vs EMCI are 96.21%,84.23% and 75.26%,respectively.The experiment proved that this is an effective AD and its early stage diagnosis algorithm.