2.3. fMRI data analysis

Full details of data analysis were reported in our previous study.8 In summary, BOLD signals were first bandpass filtered between 0.04 and 0.07 Hz.12 We then parcellated the brain into 90 regions using automatic anatomical labeling.13 The first principal component of voxel time series was used to represent each brain region.14 Next, the Hilbert transform was applied to estimate the instantaneous phase of the first principal component in each region. Subsequently, we estimated a time‐varying phase difference matrix by subtracting the phase angle between pairs of regions, resulting in a 90 × 90 × 285 adjacency matrix for each fMRI run. Note that for each run, the first 10 TRs and the last 5 TRs from 300 TRs were excluded because fMRI noise was seen in the EEG. We binarized these matrices using a threshold of pi/6.8, 15 Tensor decomposition was applied to the series of adjacency matrices for each run to try to reduce the number of spurious network connections.8