Table 1.
Demographic characteristics and primary neuropsychological information of the subjects in the present study.
Table 2.
List of brain regions extracted from the Automated Anatomical Labeling (AAL) template (Tzourio-Mazoyer et al., 2002) and their abbreviations as used in this study.
Figure 1.
The preprocessing procedure used in the construction of the sample dataset based on resting state fMRI (rsfMRI) data in three frequency bands.
The rsfMRI data were normalized to the MNI standard space using the EPI template and filtered using 0.01∼0.027 Hz (slow-5), 0.027∼0.073 Hz (slow-4), and 0.01∼0.073 Hz (whole-band) frequencies. Brain regions were defined according to the AAL atlas, and the time series was extracted for each region. The whole brain functional network was constructed for each frequency band by taking each cortical region as a node and the inter-regional Pearson’s correlation coefficient as the edge for each subject. All the independent elements of the individual functional connectivity matrix (90×90) were arranged into a 1-by-4005 row vector. We assembled all the row vectors for all 39 subjects into a 39-by-4005 matrix and normalized the data using their z-scores. The normalized matrix and the subject labels (patients were labeled by 1s and healthy subjects as -1s) constituted the sample dataset for further SVM analysis. In total, we obtained three sample datasets corresponding to the three different frequency bands.
Figure 2.
The flow chart of the rsFC pattern classification obtained by applying MVPA to the sample dataset in the slow-5 frequency band.
The sample dataset was a 39-by-4005 matrix in this study and was comprised of 4005 features, that is, inter-regional functional connections. Each row and each column of the sample dataset represented a subject and a feature, respectively. Recursive feature elimination (RFE) was applied to the sample dataset for feature selection and to calculate the rank of the feature weight (a row vector). The subsets of the sample dataset were produced according to the feature weighting rank. The classifier was used to separate the VaD brains from those of the controls in each subset of the sample dataset by using a leave one-out-subject-cross-validation (LOOCV). The accuracy rate was used to compare the discriminative information in each subset of the sample dataset.
Figure 3.
Graph of the convergence of the accuracy rate with the number of features in the pattern classification subsets.
Triangles in red, green, and blue show how the accuracy rates changed with the number of features in the slow-5, slow-4, and whole-band frequency bands, respectively.
Figure 4.
The selected features, that is, the inter-regional functional connections in the resting state functional connectivity pattern classification for the VaD group and the control group in the three frequency bands.
Upper panel: The features selected in the VaD patients. The node size is proportional to the frequency occurrence of the brain region in the selected features and the line thickness to the mean value of the feature. The red (green) lines represent features that were increased (decreased) in the VaD group compared with the controls. Lower panel: Same as the upper panel but for the control group. For both panels: slow-5∶0.01∼0.27 Hz (low frequency); slow-4∶0.027∼0.073 Hz (high frequency); whole-band: 0.01∼0.073 Hz.
Table 3.
Brain regions involved in the features selected from the functional connectivity pattern for differentiating the brain states of the VD patients from those of the controls in the three frequency bands.
Table 4.
Statistical comparisons between the selected features of the VaD patients and the controls in the three frequency bands.