Summary:Human brain is a dynamic modular network, which can be decomposed into a series of modules, and its activities change with time. In the resting state, several brain networks, namely resting state networks (RSNs), will appear on the sub second time scale and communicate with each other. This paper attempts to explore the rapid remodeling of spontaneous brain modularization and its relationship with RSNs. Three independent resting state data sets of healthy subjects (n = 568) were used to explore the dynamic activity of modular brain network by EEG / MEG. This paper confirms the existence of RSNs, and some of them have splitting phenomena, especially the default mode network, visual, temporal and dorsal attention network. This paper also proves that the individual differences in psychological images are related to the time characteristics of specific modules, especially the visual network. In conclusion, the results of this paper show that large-scale electrophysiological networks have modular dynamic fingerprints in the resting state.
Spontaneous brain activity varies from sub second to year. These fluctuations involve a series of networks called resting networks (RSNs). Most studies believe that spontaneous brain activity can be explained by a series of spatiotemporal network models. It mainly combines dimension reduction algorithms (such as k-means clustering), principal component analysis, orthogonal connectivity factor decomposition or model-based methods (such as hidden Markov model). The characteristics derived from these rapid dynamic analysis can also be used as potential neural markers of some brain diseases and behavioral characteristics.
Some studies have shown that the human brain is a modular network, which can be divided into “modules” (also known as communities or clusters), that is, areas closely connected within the brain but not closely connected with other regions. Although several studies have explored the time dependence of brain modular networks, there are few studies on the rapid remodeling of spontaneous modular brain networks and its relationship with RSNs on the time scale related to rapid cognition. In order to accurately track network dynamics, EEG / MEG is used to describe those fast (sub millisecond) fluctuations that depend on modularization.
This paper assumes that the dynamic modular reorganization of the brain in the resting state is characterized by a continuous process of separation and merger within and between different RSNs. This paper uses a recently developed framework to detect “modular brain state”, which can accurately quantify “state” fluctuations over time. Different from other clustering algorithms, this framework can detect the rapid and instantaneous changes of brain modular structure, and is superior to other existing clustering algorithms in terms of temporal and spatial accuracy. The hypothesis was tested on EEG / MEG data sets of three independent resting healthy subjects (n = 568), and the source reconstructed to 68 regions of the whole cortex. EEG / MEG source connection technology (using power coupling and phase coupling), combined with sliding window method and algorithm for detecting modular state, is used to reconstruct dynamic brain network (Fig. 1). It is worth noting that the results of this paper show that most subjects have consistent network patterns, and some of them have splitting phenomena, such as default patterns, visual, temporal and dorsal attention networks. This paper speculates that the fast modular architecture for tracking ongoing neuronal activities provides new insights into the dynamics of large-scale electrophysiological network organization of human brain.
Figure 1 research framework (a) analyzes three data sets. (B) Based on the anatomical atlas, the template MRI was divided into regions of interest (ROI). Then, the weighted minimum norm estimation inverse solution (wmne) is used for data sets 1 and 2 and beamforming is used for data set 3 to reconstruct the regional time series of each subject. (C) The dynamic brain network is calculated by sliding window technique. (D) The first step based on modular algorithm is to divide each temporal network into communities. Then, the similarity between time module structures is evaluated. (E) The similarity matrix is divided into different communities, and each community represents the modular state of a specific spatial topology combined with different time windows. (F) Extract all modules of each subject (from different MSS). Only those that are strongly correlated with RSNs are retained (more than 80% of nodes overlap). (G) Calculate the average residence time and fraction occupancy of the main modules related to RSNs.
Three independent data sets were analyzed: 1) resting EEG data of 444 subjects provided by healthy brain network biobank; 2) Resting EEG data of 56 healthy subjects; 3) Resting state MEG data of 61 subjects provided by the human connectivity group program. For EEG data sets (data sets 1 and 2), weighted minimum norm estimation (wmne) and phase coupling are used. For MEG data set (data set 3), beamforming method, envelope coupling and correction of spatial leakage are used to reduce volume conduction effect. Then, a series of continuous fragments were obtained by sliding window technology to characterize the changes of each person’s functional brain network. Then, the algorithm based on modularization is applied to take the tensor of the dynamic network as the input, and generate modular states (MSS) that fluctuate with time. Each MS represents a specific modular topology. In short, the algorithm determines the modular structure with the same topology by quantifying the similarity between all time modules. Then, modules strongly associated with one or more RSNs are determined. The overall framework of the study is shown in Figure 1.
2.1 the first data set identifies 16 states
Figure 2 shows the 16 modules obtained from the alpha band of 444 subjects and reports the proportion of subjects with each module. Three derivative modules related to DMN were obtained: post-dmn included the posterior components of DMN (posterior cingulate gyrus, parahippocampal, precuneus and inferior parietal lobule regions), and ant-dmn included the anterior components of DMN (prefrontal region) and DMN (a large module integrating the posterior and anterior parts into the same module). In addition, three temporal regions are obtained: l-temp and r-temp represent left and right upper temporal regions and lower temporal regions respectively; Temp indicates the combination of left and right temporal modules. Two modules related to the visual network were also observed: ventral vis, including the ventral region of the visual network; Vis integrates ventral and dorsal visual areas. In general, the modules ranked from high to low in the percentage of subjects are: DMN (in 96% of subjects), post-dmn (in 88% of subjects), VIS visual network (in 86% of subjects), ant-dmn (in 86% of subjects), SMN body movement network (in 83% of subjects), Dan dorsal attention network (in 80% of subjects), l-temp (in 76% of subjects) Temp (present in 63% of subjects), San salience network (present in 63% of subjects), aud + vis – (a module combining auditory and visual network, present in 61% of subjects), DMN + FPN – (a module combining default mode and forehead network, present in 58% of subjects), Dan + vis – (a module combining Dan and vis, present in 53% of subjects), r-temp (present in 45% of subjects) FPN (present in 43% of subjects), aud + vis + Dan – (module combining AUD, vis and Dan networks; present in 35% of subjects) and ventral VIS (including tongue and spindle visual areas; present in 25% of subjects). The results are consistent between several thresholds of the functional connection matrix and within the beta band.
Figure 2 results of extracting alpha band in dataset 1: derivative modules related to RSNs and their corresponding percentage of subjects.
2.2 the second data set identifies 12 states
According to the second data set (Figure 3), 12 modules were obtained from 57 subjects in the alpha band. These modules are: post-dmn (present in 98% of subjects), VIS (present in 94% of subjects), Dan (present in 91% of subjects), DMN (present in 84% of subjects), l-temp (present in 82% of subjects), ant-dmn (present in 81% of subjects), SMN (present in 73% of subjects), aud + VIS (present in 68% of subjects), Dan + VIS (present in 60% of subjects) Temp (present in 45% of subjects), DMN + CCN (a module that combines DMN with cognitive control, present in 32% of subjects), and San (present in 22% of subjects). The results are also consistent between several thresholds of the functional connection matrix and in the beta band.
Figure 3 results of extracting alpha band in dataset 2: derivative modules related to RSNs and their corresponding percentage of subjects.
2.3 the third data set identifies 10 states
Figure 4 shows the results of alpha band in dataset 3. 61 subjects identified 10 modules: DMN (in 100% of subjects), post-dmn (in 95% of subjects), VIS (in 88% of subjects), l-temp (in 78% of subjects), San (in 72% of subjects), Dan (in 79% of subjects), SMN (in 62% of subjects) Aud + VIS (present in 49% of subjects), Dan + VIS (present in 43% of subjects), and CCN (present in 18% of subjects). The results are consistent between several thresholds of the functional connection matrix and within the beta band.
Figure 4 results of extracting alpha band in dataset 3: derivative modules related to RSNs and their corresponding percentage of subjects
In conclusion, the fluctuation module consistent with RSNs is obtained from the three data sets. Among subjects in all data sets, the default mode network was the most consistent network (the highest percentage among subjects). The results also show that some RSNs show various modular topologies over time, such as DMN, temporal region and visual network. In addition, the module combination of multiple RSNs is presented over time, reflecting the cross network interaction.
Figure 5 shows the steps to obtain the results of a specific subject, resulting in 8 MSS. The similarity between the 30 modules extracted from all MSS and RSNs templates was evaluated (Fig. 5C). Among the 30 modules, 18 modules meet the threshold condition that the node overlap is greater than 80%. These 18 modules are associated with 11 RSNs: DMN, San, SMN, VIS, LTEMP, Dan + vis, aud + vis, post-dmn, DMN + FPN, FPN, ant-dmn. By observing the similarity matrix in Figure 5C, it is shown that DMN, SMN, VIS, ant-dmn, aud + vis and LTEMP represent two or more modules, while San, Dan + vis and post-dmn are related to a single module. Figure 5D shows the dynamic fluctuations of the resulting associated modules, where each module is color coded according to its corresponding RSN.
Figure 5 Results for specific subjects. (A) Extract the dynamic fluctuation of 8 module states of specific subjects, (b) obtain the spatial representation of all modules from each Ms. (C) Similarity matrix between all MSS modules and RSNs templates (only RSNs with an overlap of more than 80% with one module are mentioned), * tag overlap of more than 80%. (D) Dynamic fluctuations of modules that meet the 80% overlap threshold and are related to the RSN template. These modules are color coded according to the corresponding RSNs (as shown on the right side of figure 5C).
2.4. Residence time and fraction occupancy
In order to quantify the time characteristics of each module, two indexes are calculated: residence time (DT), that is, the average number of continuous windows in a module; And score occupancy (FO), that is, the proportion of time spent in each module. For example, the modules obtained in Figure 5 are sorted by fo: DMN (fo = 56%), VIS (fo = 43%), Dan + VIS (fo = 43%), SMN (fo = 31%), San (fo = 19%), LTEMP (fo = 14%), aud + VIS (fo = 10%), ant-dmn (fo = 4%), postdmn (fo = 3%), DMN + FPN (fo = 3%), FPN (fo = 3%). According to DT, the modules are sorted as follows: DMN (DT = 16%), VIS (DT = 10%), Dan + VIS (DT = 7%), SMN (DT = 7%), aud + VIS (DT = 7%), San (DT = 6%), postdmn (DT = 6%), LTEMP (DT = 6%), FPN (DT = 5%), DMN + FPN (DT = 5%).
Figure 6 shows the FOS and average DTS of the modules obtained from each dataset. DMN has the highest fo and DT in all data sets. From the perspective of fo, the VIS network in dataset 2 is more important. From the perspective of DT, the SMN and San in dataset 1 are very stable modules.
In conclusion, the results obtained from all data sets confirm the importance (and stability) of DMN and its function as a functional core network in a resting state.
Figure 6 shows the score occupancy (FO) and average residence time (DT) of the derived modules obtained from (a) dataset1, (b) dataset2 and (c) dataset3, respectively. The horizontal dotted line appearing in each figure represents the mean plus two standard deviations. * mark important modules (mean > mean + two standard deviations).
2.5. Correlation between derivative module and psychological image
This paper attempts to explore whether there is a correlation between the derivative module and the psychological image experienced during the acquisition of resting state (measured by resting state questionnaire, RSQ). Only dataset 2 has this data. From RSQ, five main indicators (i.e. visual psychological image, inner language, somatosensory consciousness, inner music experience and psychological manipulation of numbers) are related to the time characteristics of each module. There was no significant correlation between any RSQ score and DT of the extracted module.
Figure 7 shows the positive correlation between visual psychological image and fo of vis, Dan and aud + vis obtained from alpha band. In the beta band, the results showed that the visual image was positively correlated with the Fo of AUD + vis and Dan.
In conclusion, these results show that the individual differences of visual images experienced in the acquisition process are positively correlated with the Fo of specific modules, mainly vis, Dan and vis + aud.
Figure 7 visual psychological image is significantly correlated with Fo of vis, AUD and Dan modules in alpha band (data set 2)
This paper shows how the rapid changes in the modular structure of large-scale electrophysiological networks shape spontaneous brain activity. Recently developed algorithms are used to extract repetitive modular brain states over time. The framework is applied to three independent EEG / MEG data sets, and it is confirmed that RSNs have experienced continuous modular changes, which exist in the process of separation and merging within and between resting networks.
DMN dynamically switches its modular topology, which proves that DMN can actually be decomposed into sub components, mainly the front and rear. Studies have also shown that dynamic state transitions lead to the inclusion of FPN regions in some brain states (results of dataset 1). Similarly, the temporal network can be divided into left, right and complete modules. The dynamic modular behavior of resting brain was confirmed by integrating the modules composed of different RSNs.
The results of this study show that there are dynamic changes of brain network during spontaneous activity. These brain regions are highly dynamic and change their modular arrangement over time. In order to test whether there is some hierarchy in the consistency of brain regions, this paper reports the region contained in each RSN and its proportion in% of all subjects. Once modules are associated with a specific RSN, they are identified as overlapping nodes to calculate their inclusion rate in all subjects and data sets.
DMN is the most consistent module among subjects because it has the highest incidence on subjects / data sets and has high fo values. It shows that DMN plays an important role in integrating information in spontaneous brain activity, because there are a large number of centers related to this network. In addition, the DMN configuration has the maximum stability over time (with the longest average residence time). Modules with short residence time are short-lived and show great variability in the time reflecting dynamic functional coordination. In this study, the derived transient modules depend on each subject and data set, and most of them belong to advanced cognitive network, attention network and sensory network. In all data sets, these short-lived modules with low residence time are the integration of multiple RSNs (Dan + vis, aud + vis…). It may be because during the rest period, individuals dynamically carry out a variety of thinking, and imagery and distraction are still the main activities. Therefore, the brain is a dynamic system with stable activity and transient functional changes alternating with each other.
In addition, this study emphasizes the significant existence of visual network, showing a high occupancy rate over a period of time (the results of data sets 1 and 2). This observation may be related to the dominant visual image activity of most subjects during resting acquisition. The individual differences of visual images experienced in the acquisition process are positively correlated with the Fo of VIS network (Fig. 7). The significant relationship between mental image and AUD, Dan and vis may indicate that the interaction between these networks forms a large module over time (the results observed in dataset 1, the emergence of Dan + vis + AUD, figure 2).
In the three data sets, nine common RSNs are displayed: DMN, post-dmn, VIS, SMN, San, LTEMP, VIS + AUD, Dan + vis. The results are based on the recorded signal technology, preprocessing steps, source reconstruction, adjacent matrix thresholding, EEG / MEG frequency band（ α Wave and β Wave), spectrum segmentation, functional connection measurement and correction or no correction of zero lag correlation.
The other modules are from each data set (Figures 2, 3 and 4). The percentage of subjects per derivative module also confirmed the difference between subjects. In the same data set, these individual differences are related to the variability of cognitive and behavioral functions. There are individual differences in time characteristics in specific modules (mainly vis, AUD and Dan), which is related to the self-report score of psychovisual image measured by resting state questionnaire.
Although it is found that the results are generally consistent, the results of the three data sets are not exactly the same. For example, the presence rates of DMN on data sets 1, 2 and 3 are 96%, 98% and 100% respectively, while the presence rates of VIS network are 86%, 94% and 88%. The differences in these results may be related to the differences in the data set itself, such as sample size, subject age and experimental conditions (i.e. closed / open eyes). Another reason for the difference in results is the use of different modes (MEG / EEG). Studies have confirmed that the difference of MEG / EEG is more obvious in the study of transient resting state functional connection mode.
A large number of EEG / MEG studies confirmed α to β Correlation of frequency range (8-30hz) in driving spontaneous large-scale neuronal interactions. This paper draws the following conclusions: 1) different modules fluctuate with time, 2) the default mode network is considered to be the most consistent, independent and stable module, and its modular topology changes dynamically, 3) it is observed that the module integrating several RSNs reflects the interaction between networks (dan-vis) over a period of time, and 4) the Fo of some specific modules is positively correlated with psychological image.
From the perspective of methodology, this paper uses the same method (from data processing to network construction) as the previous research dealing with the same data set. Wmne / PLV combination is used to reconstruct the dynamic network for EEG data set. For MEG data sets, the amplitude correlation (and orthogonalization) between beamforming and band limited power envelope is constructed. Appropriate window width is the key problem of reconstructing dynamic functional network. This depends on the frequency band of interest that affects the degree of freedom of the time series and the correlation measure used. In the EEG data set, select the minimum appropriate window length to provide 6 “cycles” in a given frequency band. In MEG data, select the method with sliding window size of 0.5s. In order to extract the fast transient module, a modular algorithm is applied to extract the main modular brain states over time. The algorithm used in this study regards the brain as a dynamic modular network for segmentation.
4. Materials and methods
4.1 EEG dataset
4.1.1 dataset 1 (hBN)
The resting EEG data of 444 healthy subjects were obtained from the healthy brain network (hBN) database. The age is between 5-21 years old.
18.104.22.168 data acquisition and preprocessing
In the sound shielding room, 128 channel EEG equipment is used to collect high-density EEG data at a sampling rate of 500Hz, and the band-pass frequency is 0.1-100hz. Record the reference at CZ (head vertex). Ensure that the impedance is below 40kohm. Collect 5-minute resting data (eye closed state). EEG data were preprocessed using automagic matlab toolbox. Multiple artifact suppression algorithm (Mara) is used to detect and suppress artifacts. Four 40s artifact free periods were selected for each subject.
4.1.2 dataset 2
A total of 59 healthy subjects (29 women). The average age was 34.7 years (SD = 9.1 years, range 18-55 years). Education ranges from 10 years of school education to a doctorate. No medication, no neurological or mental illness. After collecting EEG data, all subjects completed the rest state questionnaire (ResQ), which is composed of 62 items and consists of five main types of psychological activities: visual psychological image, inner language, somatosensory consciousness, inner music experience and mental manipulation of numbers.
22.214.171.124 data acquisition and preprocessing
10 minutes of resting data (eye closed state) were collected each time. The 64 channel biosemi activetwo system is used to collect EEG data, and the standard 10-20 lead system is used. Ensure that the impedance is kept below 20kohm. Pretreatment was consistent with previous studies. The bad channel signal (completely flat or contaminated by artifact) is restored to the adjacent electrode within 5cm radius by interpolation program. Keep the signal with voltage fluctuation between + 80 and – 80. Four 40s artifact free periods were selected for each subject.
4.1.3 building dynamic brain network
For two EEG data sets, the “EEG sourceconnectivity” method combined with the sliding window method is used to construct the dynamic brain network. “EEG sourceconnectivity” includes two main steps: 1) solving the inverse problem to estimate cortical sources and reconstruct their time dynamics, and 2) measuring the functional connection between reconstruction sequences.
The steps are as follows:
1. EEG and MRI templates (icbm152) were registered by using brainstorm to determine anatomical marker points.
2. Use openmeeg software to build a real head model.
3. Based on desikan killiani map, the cortical surface was divided into 68 regions.
4. The weighted minimum norm estimation algorithm is used to estimate the timing of the region.
5. Alpha wave 8-13hz and beta wave 13-30hz are used to filter the timing of the reconstructed region.
6. The phase locking value (PLV) is used to measure the functional connection between the reconstructed regional timing. To ensure that all dynamic network densities calculated across time are equal, a proportional (density based) threshold method is used to retain the first 15% of the connection values in each network.
4.2 MEG dataset (HCP)
MEG data in resting state of 61 healthy subjects (38 women) were obtained from HCP MEG2 database. The age ranges from 22 to 35.
4.2.2 MEG acquisition and preprocessing
MEG data were collected using a whole brain magens 3600 scanner. Collect three consecutive sessions, each for 6 minutes. Subjects were asked to relax, keep their eyes open and keep looking at the crosshairs projected on a dark background. Preprocessing includes removing artifact segmentation, determining the wrong recorded channel signal and regressing the artifact that appears as an independent component in ICA decomposition.
4.2.3 building dynamic brain network
The construction method is consistent with the previous method of processing the same data set. In order to solve the inverse problem, a linearly constrained minimum variance beamformer is applied. Calculate the sine shellsource models in advance, and calculate the data covariance in 1-30Hz and 30-48hz respectively. The data is beamformed onto a 6mm grid using a standardized lead field. Then, the source estimation is standardized by projecting the power of sensor noise. The source spatial data is filtered in alpha wave 1-13hz and beta wave 13-30hz. After obtaining the regional timing based on AAL template, the symmetrical orthogonalization program is used to eliminate the signal leakage. Amplitude envelope correlation measurement (AEC) is used to estimate the functional connection between regional time series. This method is simply summarized as 1) calculating the power envelope as signal amplitude using Hilbert change, and 2) measuring the linear amplitude correlation between the logarithms of ROI power envelope. Finally, the sliding window (length = 6S, step = 0.5s) is used to construct the dynamic connection matrix. Retain 15% of the strongest connections in each network for thresholding.
4.3 extracting modular brain state
Each modular state reflects the modular organization of a specific space. The algorithm steps are as follows:
1. Girvan Newman and Louvain algorithms are used to decompose each temporal network into modules. Each module algorithm repeats 200 iterations and generates 400 modular organizations (200 runs2 algorithms) for each network. The consumption algorithm proposed by Bassett is used to determine the final modular organization. The algorithm includes calculating the incidence matrix of n (n is the number of nodes), and calculating the times that two nodes in the 400 modular organizations are assigned to the same module. Then, the incidence matrix is compared with the zero model incidence matrix calculated from the random arrangement of the original partition. Only the significant values of the incidence matrix are retained. Finally, the threshold incidence matrix is clustered by using the Louvain algorithm which repeats 100 iterations.
2. The Z-score of Rand coefficient (value between 0 and 1) is used to evaluate the similarity between time modular structures. In this step, a moment similarity matrix of TT is obtained, and t is the number of time windows.
3. The similarity matrix is clustered by the consumption modularity method to obtain the module state (MS).
4.4 extract modules related to RNSs
This paper evaluates the similarity between each module (from MSS) and RSNs. Different masks or templates are formed, and each mask is integrated with one RSN, sub parts of RSN or several different RSNs. Then, the overall matching between each module and each template is calculated. If the overlap between modules and templates exceeds 80%, they are considered to be related.
For each module related to RSN template, two indicators are calculated.
1. Time fraction proportion (FO), which represents the total time of each module, is measured as a percentage. Therefore, the high FO value reflects the high time proportion of the module.
2. The average residence time (DT) is defined as the average number of consecutive windows that appear in a particular module. Therefore, a high DT value is considered a stable module and a low DT value is considered a transient module.
4.6 statistical test
The statistical relationship between cognitive phenotypes measured by the resting state questionnaire (RSQ) was evaluated to study whether the observed brain modules were related to subjective internal thoughts and feelings in the process of resting state acquisition. That is to calculate the Pearson correlation between the five main indicators in RSQ (visual psychological image, inner language, somatosensory perception, inner music experience and digital mental manipulation) and the Fo of the derived module. In order to solve the problem of multiple comparison (5 mental activities and 11 modules), Bonferroni was used to correct the p value to obtain an adjusted threshold of P < 0.0009.