shows the tendency for the determination accuracy rates to change with the number of selected features before converging. We found that the determination accuracy rates showed a consistent change with an increase in feature number. The accuracy rates increased initially and then reached convergence, that is, leveled off, for any given frequency band. Determination accuracy rates in all three frequency bands reached convergence with 100% accuracy. The speeds of convergence, that is, the number of iterations needed to obtain convergence, in the three frequency bands differed, with the slow-5 band showing the fastest convergence speed.
Graph of the convergence of the accuracy rate with the number of features in the pattern classification subsets.
In order to examine whether the high determination accuracy rates resulted from overfitting the classifier, we re-validated the SVM classifier with the selected subset data under the condition that everything was kept in correspondence with the above analysis, but the subject labels (1 and −1) were randomized. When we did this, we obtained accuracy rates of 56.4%, 38.46%, and 58.9% corresponding to the slow-5, slow-4, and whole-band, respectively. This indicates that the determination accuracy rates when the labels were randomized were approximately random, indicating that the 100% accuracy in the classification result shown in indeed reflects actual group differences in the rsFC patterns between the VaD patients and the controls.
shows the selected features for the VaD group and for the control group in the rsFC pattern classification at the point where the accuracy rates reached convergence for the slow-5, slow-4 and whole-band frequency bands. The slow-5 band reached convergence with the fewest number of features, nine, whereas the slow-4 reached convergence with the greatest number of features, fourteen. shows that fifteen, twenty-two, and nineteen brain regions were involved in the features that were selected for pattern classification in the slow-5, slow-4, and whole-band frequency bands, respectively. These brain regions are widely distributed in the frontal cortex, temporal lobe, parietal lobe, visual cortex, and subcortical regions. Three brain regions, the CAL.L, IPL.L, and LING.L, appeared among the selected features for all three frequency bands.
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.
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.
shows statistical comparisons between the VaD patient group and the control group for each feature selected in the rsFC pattern classifications. In the pattern classification the selected features are listed in the order of their feature weight from high to low. Both the feature weight in the pattern classification and the number of selected features differed between the three frequency bands. We found nine, fourteen, and twelve features for the slow-5, slow-4, and whole-band frequency bands, respectively. For the slow-5 frequency band, we found that five features decreased and four increased in the VaD group compared with the control group. The features that decreased were inter-regional connections between the IPL.L and the MTG.L, between the ORBmid.L and the CUN.L, between the OLF.L and the REC.L, between the CAL.L and the CAL.R, and between the CAL.R and the LING.R, whereas the features that increased were the inter-regional functional connections between the MFG.L and the PCG.R, between the ORBsupmed.L and the PUT.L, between the PCG.R and the PCL.R, and between the LING.L and the CAL.R. Four of these features were located in the left hemisphere, three features linked both hemispheres, but only two features were located in the right hemisphere. For the slow-4 frequency band, the SVM calculations selected fourteen features that were able to discriminate each individual VaD brain from the healthy ones. Five of these features were in the left hemisphere, three features linked both hemispheres, and six features were located in the right hemisphere. Specially, among the fourteen features, we detected four significantly increased features (inter-regional connections between the SFGdor.R and IFGoperc.R, between the IPL.L and TPOmid.L, between the SMG.R and TPOmid.L, and between the THA.L and HES.L) and three decreased features (inter-regional connections between the HIP.R and CAU.R, between the LING.L and LING.R, and between the CAU.L and PUT.L). For the whole–band frequency, we detected six features that decreased (inter-regional connections between the LING.L and the CAL.R, between the MTG.R and the SFGdor.R, between the FFG.L and the HES.L, between the AMYG.L and the IFGoperc.L, between the IOG.R and the PreCG.R, and between the CAL.L and the LING.L) and six features that increased (between the ORBinf.L and the SFGdor.R, between the HIP.R and the IOG.R, between the SFGmed.L and the MFG.R, between the AMYG.R and the IOG.R, between the LING.L and the TPOsup.R, and between the IPL.L and the TPOmid.L) in the VaD group compared with the control group. Four features were located in the left hemisphere, another four features linked the two hemispheres, and the remaining four features were located in the right hemisphere. In the three frequency bands, we detected several decreased inter-regional functional connections that were located in vision-related regions (in the slow-5: between the CAL.R and the LING.R, between the CAL.R and the LING.L, and between the CAL.L and the CAL.R; in the slow-4: between the LING.L and the LING.R and between the CAL.L and the LING.L; in the whole-band: between the CAL.R and the LING.L and between the CAL.L and the LING.L).
Statistical comparisons between the selected features of the VaD patients and the controls in the three frequency bands.