Classification Of AD Progression Using SVM Model On Structure MRI Data

Objective:To utilize support vector machines(SVM) method on structure magnetic resonance imaging(MRI) data for distinguishing Alzheimer’s disease(AD) from cognitively normal, age-equivalent control subjects.Methods:A total of 543 participants(CN=139, EMCI=220, LMCI=108 and AD=76) from Alzheimer’s Disease Neuroimaging Initiative were included in this study. We got their272 MRI features for everyone, which including 49 subcortical volumes, 69 cortical volumes, 68 cortical thicknesses, 70 surface areas and 16 hippocampal subfields, as well as Mini Mental State Examination scores, age, gender, and years of education. The data was preprogressed with statistical methods such as one-way analyses of variance(ANOVA), nonparametric test and correlation analysis, which were applied for assessing differences in features among CN, EMCI, LMCI and AD groups. Then support vector machine prediction model would be built. Using the different features among groups based on SVM model, we could predict different stages of AD.Results:In 300 samples of the training set, the prediction accuracy on 276 indexes of training set was 78.07%, and test set result was 46.91%. The prediction accuracy on 54 indexes of training set was 100%, and test set result was 99.17%. The prediction accuracy on 8 indexes of training set was 100%, and test set result was 64.19%. Theprediction accuracy on 135 indexes of training set was 100%, and test set result was56.37%. The prediction accuracy on 45 indexes of training set was 100%, and test set result was 77.77%. The prediction accuracy on 25 indexes of training set was 100%, and test set result was 61.31%. In 400 samples of the training set, the prediction accuracy on276 indexes of training set was 88.50%, and test set result was 50.34%. The prediction accuracy on 54 indexes of training set was 100%, and test set result was 99.30%. In 488 samples of the training set, the prediction accuracy on 276 indexes of training set was78.07%, and test set result was 54.54%. The prediction accuracy on 54 indexes of training set was 100%, and test set result was 100%.Conclusion:It is by using the SVM-based prediction model combined with some meaningful biomarkers that accurate results are obtained. The best prediction accuracy of SVM model is achieved from the extracted features with ANOVA. According to the results, we get the indexes closely related to the disease and obtain a further evidence for the clinical and basic research. With increasing sample size, the conversion rate from MCI to AD can be predicted.