Early Diagnosis Of Alzheimer’s Disease Based On Machine Learning And Neuroimaging

Objective:In order to build a kind of can be applied to Alzheimer’s Disease（Alzheimer’s diseases,AD）and mild cognitive decline in early classification problems into the problem of machine learning algorithm,and verify the prediction accuracy in neuroimaging,by considering the characteristics of the indicators to find the model has higher prediction efficiency,improve the model generalization of force on the selected data set,the auxiliary tool for clinical early diagnosis of AD.Methods:1.In this study,543 subjects were selected from the public database of Alzheimer’s Disease Neuroimaging Initiative（ADNI）.According to the course of Disease,they were divided into four groups.There were 139 cases of normal cognitive（NC）groups,22cases of early mild cognitive impairment（EMCI）group,108 cases of late mild cognitive impairment（LMCI）and 76 cases of AD.The four groups were combined in pairs to obtain six dichonomic groups,including NC-EMCI,NC-LMCI,NC-AD,EMCI-LMCI,EMCI-AD and LMCI-AD group.2.Collect Structural magnetic resonance imaging（s MRI）data of these subjects as well as their gender,age and education level,as well as their MMSE score statistical data.The s MRI image is normalized into 272 feature index results through the pretreatment of Freesurfer software.They included 68 Cortical thickness（TA）,70 surface area（SA）,69 Cortical volumes（CV）,49 subcortical volumes（SV）and 16 hippocampal subfields（HS）.This study focused on 272 characteristic data sets and 276 data sets after adding gender,age,education level and Mini Mental State Examination（MMSE）scores.3.Two feature selection algorithms, L₁ norm-regularized support vector machine（SVM）and L₁ norm-regularized logistic regression（LR）,were used for feature selection.The optimal number of feature subsets was obtained for each group,and then the lesion location of the group was determined by exploring the feature with the highest correlation after feature selection.4.The best selection was put into four commonly used machine learning classification models（LR,back propagation neural network（BPNN）,random forest and SVM）for the classification of AD course.By comparing the performance of different classification algorithms,the optimal classification prediction model is obtained.The data set was divided into a test set and a sample set,and the accuracy of the four algorithms was compared by using the ten-bend cross validation method.The sensitivity,specificity,and Area under receiver operating characteristic curve（AUC）were used to evaluate the accuracy.The AUC value of each group was obtained by comparing the classification results based on SVM and RF classification models before feature selection,traditional statistical feature selection method, L₁-LR feature selection method and L₁-SVM feature selection method before feature selection.Results:1. L₁-SVM feature selection model was used for feature selection.Among the 276features,121 features were the best in the NC-EMCI group,while 82,22,113,39,and 53features were selected into each classification model in order for the other five groups.MMSE score showed significant changes in all stages of the disease course of AD,and age became an important factor in the identification of NC-LMCI group,EMCI-AD group,and LMCI-AD group.2. L₁-LR feature selection model was used for feature selection.Among the 276features,42,11,56,21 and 20 items of the six groups were selected into the classification model.MMSE score ranked first in the NC-AD group and LMCI-AD group.Ranked 7th in NC-AD group in education level.In the four groups of NC-LMCI,and LMCI-AD,the age ranked 5th and 3rd,respectively.Gender ranked the fourth in the NC-AD group.3.The combination of L₁-SVM feature selection algorithm and BP neural network algorithm has a better prediction effect on AD course transformation,but there are slight differences among different classification groups.In the process from NC to AD,the accuracy of BPNN is up to 98.90%.The accuracy of BPNN in differentiating NC-LMCI,EMCI-LMCI and LMCI-AD groups was 95.04%,93.01%and 92.41%,respectively.However,in the process of identifying NC-EMCI and EMCI-AD groups,SVM performed best,with the accuracy of 85.4%and 97.63%,respectively.Under the selection of 276 features using L₁-LR,the accuracy of SVM model in the NC-LMCI and EMCI-AD groups was 92.26%and 96.95%higher than that in the other groups,respectively.The accuracy of RF classification model in the NC-AD group and the EMCI-LMCI group was higher than other models,with 97.71%and 86.74%respectively.4.After L₁-SVM feature selection,the prediction accuracy of EMCI-AD group among 276 items was the highest（97.63%）,and the AUC value was 0.99.After L₁-LR feature selection,among 276 data sets,it was found that the accuracy rate of EMCI-AD group was the highest,which was 96.95%,and the AUC value was 0.99.After₁L-SVM feature selection,the prediction accuracy of the LR classification model among 276 data sets,the accuracy,specificity and AUC of EMCI-AD group were 95.25%,98.89%and0.99,respectively.After₁L-LR feature selection,the accuracy of EMCI-AD group was the highest（96.29%）,specificity was 98.75%,AUC value was 0.99 in 276 items of data sets.After L₁-SVM feature selection,the BPNN classification model has the highest prediction accuracy of NC-AD group is 98.90%,specificity（100%）,sensitivity（98.75%）and AUC（1.00）,respectively in 276 data sets.After L₁-LR feature selection,NC-AD group also had the highest accuracy（97.64%）among 276 items.After L₁-SVM feature selection,the RF classification model has the highest prediction accuracy of 96.77%,specificity（98.57%）and AUC（0.99）,respectively in NC-AD group among 276 items.After L₁-LR feature selection,in 276 data sets,NC-AD group had the highest prediction accuracy（97.71%）,specificity（98.75%）and AUC（0.99）.Conclusion:1.The 276 feature data set with the addition of four clinical demographic data has a better classification effect than the 272 feature data set with only s MRI.Increasing clinical demographic data can improve the prediction rate of the classification model.2.Most of the best features in the pylonized groups were distributed in the limbic system and temporal lobe.3.Under the L₁-LR feature selection,the SVM classification model has a better classification prediction effect on EMCI and AD patients with similar disease course and the most difficult to distinguish.4.After L₁-SVM feature selection,BPNN classification model has better prediction performance for NC and AD patients at these two stages.5.In the identification of NC-AD transformation group,the two feature selection methods used in this study have high prediction effects under SVM and RF classification models,but the₁L-SVM feature selection method is superior to the₁L-LR method.