The Prediction Study Of Mild Cognitive Impairment Progresses Using MRI

Mild cognitive impairment（MCI）, commonly suffered from a high risk of conversion from MCI to Alzheimer’s diasease（AD）, is a transitional stage between the healthy aging and dementia and lack of scientific clinical diagnosis method. In recent year, the continuous magnetic resonance imaging provides the necessary basis to delineate the abnormalities of brain morphometric and network architectures. However, MCI remains challenging for early detecion due to the mild symptoms of cognitive impairment, various etiologies and pathologies, and high rates of reversion back to normal. In order to understand better the impact of differernt disease progress and brain networks on MCI classification, the article analyses as follows:（1） Prediction of conversion from MCI to AD using brain morphometric and thickness network measures. Firstly, according to the different time periods before diagnosis of probale AD, the MCI converters（MCIc） were subdivided into four groups: 6 months（MCIc_m6）、12 months（MCIc_m12）、18 months（MCIc₁8） and mixed conversion time（MCIc_mixed）. Then cortical thickness、cortical volume and cortical surface area were extracted, and the network for each individual was obtained from cortical thickness and nodal properties were then extracted. Finally, repeated sparse linear regression with stability selection（repeated SS_LR） and minimal-redundancy-maximalrelevance（mRMR） were used for feature selection, and the classifications were implemented using support vector machine（SVM） classifier based on a nested cross validation.（2） Prediction of early MCI（EMCI） and late MCI（LMCI） using the combination of different frequency band functional network measures. Firstly, the resting-state functional magnetic resonance imaging（rs-fMRI） of each subject was pre-processed and subdivided into three frequency bands（full-band, slow-4 and slow-5）. The time series of each region were extracted and used for function network contraction, then the graph theory was used to extracte diagnostic features and to analyze the differences of topological properties between two groups and three bands. Finally, SS_LR and mRMR were also used for feature selection, and the SVM classifiers with the nested cross validation were employed for classification.The classification results of MCIc and MCI non-conversters（MCInc） showed that the better classification performances were observed in short-term prediction compared with long-term prediction. Moreover, the repeated SS_LR achieved better results compared with mRMR in the short-term and mixed conversion time classification. Observing the selected features, we found that more consistent features were included in the classifer models using repeated SS_LR, and the features of brain morphometric and thickness network shared smaller overlap compared with mRMR. Overall, homogenization of MCIc sub-groups can promote the classification, and the combination of brain morphometric and thickness measures can provide complementary information for the prediction of MCI conversion.Observing the results of EMCI and LMCI, we found that lower classification performances were obtained using single frequency band measures and the combination of three frequency bands meaures could significantly improve the classification results. Furthermore, SS_LR obtained more stable and better classification performances compared with mRMR. Observing the features which were included in the classifier models, we also found that the features were selected predominately in slow-4 band. Using single features of slow-4 band achieved encouraging but unfortunately unstable results, in contrast, using the features of full-band and slow-5 did not obtain favourable results. Thus, the different low-frequency bands show different properties and physiological functions and the features in the slow-4 band may be more stable and robust than those in the full-band and slow-5 band, and efficient combination of the features from different low-frequency bands can promote the classification.