Show description

Then, CCA is applied to trans‐ form the remained WM/GM features to an updated common space. This transformingeliminating scheme is iteratively executed till the number of iterations exceeds a predefined threshold. In other words, the iterations are stopped, when the classification performance in the subsequent steps does not increase anymore. Robust Linear Discriminant Analysis (RLDA): In this study, we use the robust discriminant analysis (RLDA) [14] to classify PD from the normal subjects based on be the matrix containing the -dimensional samples the selected features.

Performance comparison of different methods in MCI classification. 82 Conclusion In this paper, we propose to fuse information contained in multiple HOFC networks for a better MCI classification. To this end, hierarchical clustering is utilized to generate multiple HOFC networks, each being located at one layer. With such a framework, features extracted from the network at each layer can be refined, and only the informative feature block is taken into account. Specifically, by combining the sequential forward selection and sparse regression, a novel feature fusion method is developed.

In such a way, we can eventually generate 4 HOFC networks from layer 1 to layer 4: HON1, HON2, HON3, and HON4, where the number of clusters equals 220, 190, 160, and 130, respectively. The averaged HOFC networks in layer 1 for MCI and NC subjects are shown in the left and middle of Fig. 2. The corresponding high-order Ensemble Hierarchical High-Order Functional Connectivity Networks 23 feature vectors (WLCC) Fea1 ∈ R220, Fea2 ∈ R190, Fea3 ∈ R160, and Fea4 ∈ R130 are extracted. Since our method is a feature fusion method, the correlation between features from different HOFC networks provides important information about redundancy.