Changes in Brain Functional Network Connectivity in Adult Moyamoya Diseases

Abstract

Moyamoya disease (MMD) is a cerebrovascular disease that is characterized by progressive stenosis or occlusion of the internal carotid arteries and its main branches, which leads to the formation of abnormal small collateral vessels. However, little is known about how these special vascular structures affect cortical network connectivity and brain function. By applying EEG analysis and graphic network analyses undergoing EEG recording of subjects with eyes-closed (EC) and eyes-open (EO) resting states, and working memory (WM) tasks, we examined the brain network features of hemorrhagic (H_MMD) and ischemic MMD (I_MMD) brains. For the first time, we observed that I_MMD had the much lower alpha-blocking rate during EO state than healthy controls while H_MMD exhibited the relatively low EEG activity rate across all the behavior states. Further, I_MMD showed strong network connections in the alpha-wave band in frontal and parietal regions during EO and WM states. EEG frequency and network topological maps during both resting and WM states indicated that the left frontal lobe and left parietal lobe in H_MMD patients and the right parietal lobe and temporal lobe in I_MMD patients have clear differences compared with controls, which provides a new insight to understand distinct electrophysiological features of MMD. However, due to the small sample size of recruited patient subjects, the result conclusion may be limited.

Supplementary information

The online version contains supplementary material available at (10.1007/s11571-021-09666-1)

Introduction

Moyamoya disease (MMD) is a cerebrovascular disease that is characterized by progressive stenosis or occlusion of the terminal portion of the internal carotid arteries and its main branches, leading to the formation of abnormal collateral vessels (Suzuki and Takaku 1969); (Kuroda and Houkin 2008); (Scott and Smith 2009). Clinical phenotypes of MMD are referred to as the first ischaemic and haemorrhagic strokes (Scott and Smith 2009); (Kim and Jeon 2014), and different phenotypes may result from different pathogeneses. Patients may experience severe neurological or cognitive impairment if not treated in a timely manner (Kuroda and Houkin 2008). Surgical revascularization is considered the only effective treatment to prevent recurrent stroke.

Electroencephalogram (EEG) is widely used to study healthy and diseased brain dynamics because of its noninvasiveness and low cost (Niedermeyer 2003). EEG studies in MMD were first reported in children. The observed typical features include posterior slow (P slow), centrotemporal slow (CT slow), sleep spindle depression and “rebuild-up” phenomenon after the end of hyperventilation (Kodama et al. 1979). However, the abnormal EEG pattern reported in paediatric patients is not well studied in adult MMD patients (Cho et al. 2014). The previous clinical EEG studies on MMD mainly focused on abnormalities in the EEG waveform shape and patterns (Frechette et al. 2015) while no deep investigation on MMD brain EEG network and power spectrum analysis. The EEG mean frequency (MF) is a measurement of the averaged brain activity rate, which is reportedly positively correlated with the mean regional cerebral blood flow (rCBF) in the cortex of animals (Baldy-Moulinier and Ingvar 1968) and human brains (Ingvar et al. 1976); (Tolonen et al. 1981). To evaluate the metabolic rate of MMD brains, we examined EEG mean activity rate in MMD patients.

Moreover, vascular structure abnormalities often lead to changes in brain functional connectivity (FC) (Honey et al. 2007); (Bullmore and Sporns 2009). Hence, we also examined the MMD brain network connectivity property (Stam et al. 2007); (Rubinov and Sporns 2010); (xxxx) and quantified the economic cost for connectivity and the efficiency for information transmission. (Achard and Bullmore 2007). The deviated small-world networks have been extensively reported in patients with mental illness, such as depression (Zhang et al. 2011), schizophrenia (Liu et al. 2008), and Alzheimer’s disease (Stam et al. 2008) while little is known in MMD brains. The investigated EEG characteristics in MMD brain networks may provide a new insight to understand distinct electrophysiological features of MMD.

Materials and methodology

Participants

Forty adult patients with MMD were enrolled consecutively for this experimental study from March 2017 to August 2018. We excluded 16 subjects who did not complete the experiment well, leaving 24 MMD subjects (11 with and 13 without a history of haemorrhage) for the EEG analyses. We set the criteria for rejection: (1) if the subjects have continuous unusual physical movement during task or couldn’t completed the whole experiments due to the physical discomfort; (2) if the subjects randomly pressed the keyboard regardless of right or wrong. The inclusion criteria were as follows: (1) right-handed, Chinese and aged between 18 and 68 years; (2) diagnosis confirmed by digital subtraction angiography; grades III and IV of Suzuki’s classification (Suzuki and Takaku 1969); (3) physically capable of undergoing cognitive testing; (4) no evidence of recent or remote infarct and haematoma larger than 8 mm in maximum dimension on structural images (Su et al. 2013); (Karzmark et al. 2011); (2014); (Lei et al. 2014); (5) no brain surgery before enrolment; (6) absence of other cerebrovascular or severe systemic diseases; and (7) absence of any situation that could compromise cognition, such as drug use and diseases. After screening by MR angiography, 21 healthy adult subjects without cerebrovascular or mental diseases were included in this study as controls with same age distribution as patients.

To ensure that all subjects were able to carry out the experiment in this study, we evaluated their cognitive status using the Mini-Mental State Examination (MMSE). This study was approved by the Institutional Review Board at Huashan Hospital of Fudan University and was conducted in accordance with the Helsinki Declaration. This project was approved by the Ethics Committee of Fudan University, and all subjects signed the informed consent form before the experiment.

Experimental procedure and data preprocessing

EEG recording was started with a 5-min eyes-closed, a 5-min eyes-open resting state, then 30 trials of a delayed-response working memory task (Fig. 1). The actiCHamp 64-channel EEG system from Brain Products was used, and the sampling rate was 1000 Hz. The impedance of all electrodes was below 10 KΩ. Subjects were asked to sit in a dim and soundproof room. All data were analysed offline and preprocessed using MATLAB R2017b software plug-in EEGLAB 14.0.0 (Delorme and Makeig 2004). EEG data were first referred to as the average reference. Then, 0.5 Hz high-pass filtering and 100 Hz low-pass filtering were performed using a finite impulse response (FIR) filter. The 50 Hz line noise was removed by a notch filter. Eye-blink and cardiac artefacts were removed with independent component analysis (Mognon et al. 2011). For some channels (especially Tp7 and Tp8) contaminated by muscle artefacts, channel interpolations were performed to ensure data quality. Preprocessing was carried out using EEGLAB scripts. Since some patients were unable to maintain a complete eyes-closed or eyes-open state for the whole 5 min, 90 s continuous clean data were selected for subsequent analysis.

Fig. 1.

Experimental procedure in this study

Power spectrum analysis

Mean frequency

The MF was calculated as follows, where P(f) was defined as a power spectrum (Chotas et al. 1979):

$$\overline{f}=\frac{{\sum }_{f=1}^{100}(P(f)\times f)}{{\sum }_{f=1}^{100}(P(f))}$$ 1

Notably, the MF was also used to detect the outliers. If the MF of a subject was twice the standard deviation higher or lower than the average value, it was removed as an outlier in all subsequent analyses. Thus, one subject was removed from all three groups as an outlier.

Response sparseness

The response sparseness of brain EEG activities was calculated by the following formula, where $\stackrel{-}{f_i}$ is the MF of each recording site in a set of N channels (N = 64), and S represents sparseness (Yu et al. 2014):

$$S=\raisebox{1ex}{$\left(1-{\frac{\left[{\sum }_{i=1}^N\left(\frac{{\overline{f}}_i}{N}\right)\right]}{{\sum }_{i=1}^N\frac{{\overline{f}}_i^2}{N}}}^2\right)$}\!\left/ \!\raisebox{-1ex}{$\left(1-\frac{1}{N}\right)$}\right.$$ 2

The value of S near 0 represented a dense code, and the value near 100% represented a highly sparse code (Olshausen and Field 1996); (Vinje and Gallant 2000). A high S value indicated that only a few units had high activity rates, while all others had low rates and followed exponential distributions. This finding implies high efficiency in employing unit activity in representing useful information with relatively low cost.

Alpha blocking ratio

The MATLAB function ‘pwelch’ was used to calculate the power spectrum. Alpha waves appear when relaxing and closing the eyes. When the subjects opened their eyes or if the subject performed a mental task or focused, the alpha wave pattern disappeared, and the beta pattern appeared. This alpha blocking phenomenon is well recognized (Kaiser 2005). The alpha blocking ratio is a sign of the existence of a higher level of cortical arousal in the subject. The alpha blocking ratio is calculated as (Wan et al. 2018):

$$Ratio(x)=\frac{(log10(P_{EC\left(\alpha \right)})-log10(P_{EC\_base(\alpha )}))-(log10(P_{x\left(\alpha \right)})-log10(P_{x\_base(\alpha )}))}{(log10(P_{EC\left(\alpha \right)})-log10(P_{EC\_base(\alpha )}))}\ast 100\%$$ 3

where P_EC(α) represents the alpha power (8–13 Hz) during the eyes-closed state, and $P_{x\left(\alpha \right)}$ represents the alpha power during the eyes-open or working memory state. Subtracting the respective baselines was necessary to obtain an accurate alpha blocking ratio. Thus, we defined the average power of 8 Hz and 13 Hz as the power baseline. For each state, the product of the power baseline and the number of the band sample were subtracted.

Graph-based network analysis

Functional connectivity

To construct the EEG network, we defined electrodes as the nodes, and the correlation between any paired electrodes was defined as the edge. Here, the phase lag index (PLI) was used to measure the FC strength (FCS), which quantified the consistency of phase lags between two signals and was less sensitive to volume conduction (Stam et al. 2007). PLI ranges from 0 to 1, and a higher PLI value indicates a larger FCS. In this study, the weighed network was considered, and the PLI value was used as edge weights. Since different frequency bands were related to different neurophysiological nature, all frequency bands of full, δ (1–4 Hz), θ (4–7 Hz), α (8–13 Hz) and β (14–30 Hz) bands were analysed separately.

To construct the distance matrix, we defined the PLI as the edge weight, and the 1/PLI as the edge length. In other word, we first calculated the 64*64 functional connectivity (FC) matrices based on the PLI, then we calculate the reciprocal of FC matrices to obtain the adjacency matrices, and thirdly the shortest path was obtained using Dijkstra’s algorithm based on adjacency matrices.

The average strategy was used to improve the signal-to-noise ratio. Therefore, the 90-s clean data were divided into 9 epochs with 10 s, and a 64*64 PLI matrix of each epoch was calculated. The final PLI matrix for each subject was obtained by averaging among the 9 epochs. Then we average all the subjects and average the 64*64 functional connectivity matrices for each channel and obtained the final 64*1 functional connectivity matrices. And the 64*1 functional connectivity matrices was mapped to the topography map by using the matlab function “topoplot.m”. All the following network analyses were based on the average connectivity matrix (Zeng et al. 2017). To compare the abnormal FC topological pattern among the three groups (control, H_MMD, I_MMD), we retained the top 128 PLI values (average degree = 2) to construct the connectivity map. Finally, the graph similarities between the individual and averaged FC patterns were used to assess the representativeness of averaged FC patterns in each group (supplementary materials) (Onnela et al. 2002); (Yu et al. 2016). We have also applied the average strategy for calculating the Mean Frequency (MF) and Response Sparseness (S) for each subject and then do the average to have the final results.

Small-world characteristics

Functional segregation was calculated using the weighted clustering coefficient (C):

$$C_i=\frac{\sum _{j\ne i}\sum _{\begin{array}{c}k\ne i\hfill \\ k\ne j\hfill \end{array}}{w}_{\mathit{ij}}{w}_{\mathit{ik}}{w}_{\mathit{jk}}}{\sum _{j\ne i}\sum _{\begin{array}{c}k\ne i\hfill \\ k\ne j\hfill \end{array}}{w}_{\mathit{ij}}{w}_{\mathit{ik}}}$$ 4

where w_ij was the edge weight (measured as PLI) between nodes i and j, and the clustering coefficient of the whole network was the mean clustering coefficient over all nodes (N is the total electrode number):

$$C=\frac{1}{N}\sum _{i=1}^N{C}_i$$ 5

Functional integration was calculated using the characterized path length (L):

$$L=\frac{1}{E}=\frac{1}{\left(1/(N(N-1))\right)\sum _{i=1}^N\sum _{\begin{array}{c}j=1\hfill \\ j\ne i\hfill \end{array}}^N{d}_{\mathit{ij}}^{-1}}$$ 6

where d_ij was the shortest path between node i and node j. Here we defined the reciprocal of functional connectivity matrices as the adjacency matrices. Then, the shortest path between two paired nodes was calculated by using Dijkstra’s algorithm based on the adjacency matrices.

According to the definition, C and L largely depended on the edge weights and network size (Stam et al. 2008). To avoid these effects, we considered the normalized clustering coefficient (C/C_r) and normalized path length (L/L_r). The original PLI matrix was shuffled, and 100 random networks were generated. Then, the average weighted clustering coefficients (C_r) and path length (L_r) were calculated.

Small-worldness (σ) was used to integrate the ability of network functional segregation and integration, which was calculated as the normalized clustering coefficient (C/C_r) divided by the normalized path length (L/L_r):

$$\sigma =\gamma /\lambda =(C/Cr)/(L/Lr)$$ 7

In the original model, the small-world network has a higher C and a similar but slightly higher L than the random network did, which also acts as γ = C/C_r > 1, λ = L/L_r ≈ 1, σ = γ/λ > 1 (xxxx); (Humphries et al. 2006).

Cost-efficiency Index (CEI)

The CEI evaluates the balance between network efficiency and cost and is calculated as efficiency minus cost. For the ranges of cost thresholds, the CEI is positive and shows a trend of rising and then falling.

The network cost is used to characterize the density of the network (Achard and Bullmore 2007), which can be calculated as

$$\text{cost}=\frac{1}{N(N-1)}\sum _i{K}_i$$ 8

where K_i is the degree of node i and is defined as the number of nodes connected to node i in the graph. N is the total node number, which equals 64 in this study. Network efficiency (E) is calculated as

$$E=\frac{1}{N(N-1)}\sum _{i\ne j}\frac{1}{d_{\mathit{ij}}}$$ 9

Here, we considered the cost ranges from 6.635% to 50.79% (the average degree ranges from 4 to 32).

Statistical analysis

All statistical analyses were performed using the MATLAB Statistics and Machine Learning Toolbox. The one-way ANOVA was used to evaluate the differences among groups, and a post hoc test was used to check subgroup differences. The P value was set at 0.05. Localization of electrodes on the brain cortex was performed in reference to an open-source brain atlas (Koessler et al. 2009).

Results

General demographics

According to the medical history of stroke, patients were assigned to either a haemorrhagic group (11 patients, H_MMD) or an ischaemic group (13 patients, I_MMD). The H_MMD, I_MMD, and control groups did not significantly differ in age, gender, or education level (Table1). Although the MMSE scores of both the H_MMD and I_MMD groups appeared lower than those of the control group, the difference was non significant (Table 1).

Table 1.

Demographic features of the three groups

Index	HMMD (n = 11)	IMMD (n = 13)	NC (n = 21)	F value (p value)
Age (years)	40.91 ± 9.58	41.23 ± 14.60	43.38 ± 9.06	0.33(0.858)
Male (%)	4(36.4%)	5(38.5%)	12(57.1%)	2.06(0.357)
Education (years)	8.84 ± 2.60	8.72 ± 2.45	8.33 ± 2.39	0.41(0.815)
MMSE	24.92 ± 1.44	24.82 ± 1.17	25.16 ± 2.39	2.31(0.316)
WM accuracy	0.773 ± 0.217	0.759 ± 0.150	0.827 ± 0.146	0.67(0.717)

MMSE Mini-Mental State Examination

Mean frequency

In the eyes-closed state, the MF difference among the three groups was significant (F = 7.00, p = 0.030). Although no significant difference was found between the two subgroups, H_MMD patients exhibited a trend of lower MF (over 30%) than did controls in the left frontal lobe (FC5) and left parietal lobe (C5). Next, in the eyes-open and working memory states, the differences among the three groups were non significant (p = 0.062 in eyes-open state, p = 0.318 in working memory state). In detail, in the eyes-open state, I_MMD patients exhibited a trend of lower MF than did controls, mainly in the right parietal lobe (CP4), while H_MMD patients exhibited lower MF than did controls, mainly in the left frontal lobe (F7, FC5, AF7 and AF3), right temporal lobe (FT10, TP10), left parietal lobe (C3, CP5, C5 and CP3) and right parietal lobe (CP6 and CP4). Additionally, in detail, in the working memory state, I_MMD patients exhibited a trend of higher MF than did controls, mainly in the left frontal lobe (AF3) and temporal lobe (TP8) (Fig. 2a and Table 2).

Fig. 2.

Comparison of mean frequency and sparseness among the three groups. a Topology contrast of the mean frequency distribution under different states among MMD patients and controls. In the right panel (the fourth, fifth column), the ratio (defined as $100\%\ast ({\overline{f}}_{\mathit{MMD}}-{\overline{f}}_{\mathit{Control}})/{\overline{f}}_{\mathit{MMD}}$) was calculated to evaluate the relative mean frequency distribution between MMD patients and controls. b Dynamic and static mean frequency with non-overlap 10 s epoch among the three groups. c Dynamic and static sparseness contrast with non-overlap 10 s epoch among the three groups

Table 2.

Mean frequency among three groups in different states

Mean frequency		IMMD vs control	HMMD vs control	IMMD	HMMD	Control	F value (p value)
EC	> 30%	–	–	11.63 ± 0.69	12.04 ± 0.73	13.98 ± 0.49	7.00(0.030)
	< −30%	–	FC5; C5
EO	> 30%	–	–	14.56 ± 1.08	13.27 ± 1.36	17.22 ± 0.87	5.56(0.062)
	< −30%	CP4	F7; FC5; FT10; C3; CP5; CP6; TP10; P7; AF7; AF3; C5; CP3; CP4
WM	> 30%	AF3; TP8	–	17.62 ± 1.39	14.48 ± 1.62	15.94 ± 1.31	2.29(0.318)
	< −30%	–	–

EC Eyes-closed state; EO Eyes-open state; WM Working memory state

Considering the state transition, the control group exhibited a pattern of increased MF from the eyes-closed to eyes-open state and then a decrease from the eyes-open to working memory state. However, the I_MMD and H_MMD groups presented a pattern of increased MF from the eyes-closed to eyes-open and working memory states (Fig. 2b).

Response sparseness

In resting states (both eyes-closed and eyes-open), the I_MMD group exhibited a trend toward the lowest sparseness among all three groups (p = 0.259 in eyes-closed state; p = 0.641 in eyes-open state), while in the working memory state, the H_MMD group exhibited the lowest sparseness (p = 0.048, Supplementary Material, Table S1).

Considering the state transition, all three groups exhibited a pattern of increased sparseness increase from the eyes-closed to eyes-open and working memory states. However, the H_MMD group presented with an obvious deviation from the other two groups when converting from the eyes-open to working memory state (F = 3.60, p = 0.166, Fig. 2c).

Alpha blocking ratio

Within-group analysis implied that all three groups exhibited an obvious alpha blocking phenomenon during state transition (Fig. 3a–c). However, visual inspection indicates that the I_MMD and H_MMD groups presented with a trend of fewer alpha blocking phenomena from the eyes-closed to eyes-open and working memory states.

Fig. 3.

The alpha blocking effect under different states in MMD patients and controls. Power spectra of the three groups are shown in A ~ C. A comparison of the alpha blocking ratio among the three groups is shown in D

Between-group analysis indicated that the alpha blocking ratio of the three groups exhibited a trend of control (77.13% ± 5.44%) > H_MMD (73.67% ± 10.20%) > I_MMD (42.42% ± 11.85%) from the eyes-closed to eyes-open state (p = 0.105). The trend changed to control (85.94% ± 5.53%) > I_MMD (77.85% ± 7.22%) > H_MMD (63.82% ± 11.96%) from the eyes-closed to working memory state (p = 0.192) (Fig. 3d).

Graph-based network analysis

Functional connectivity

The differences in the mean PLI among the three groups in different states of all frequency bands were examined (Supplementary Material, Fig S3). The results indicated that only in the δ (p = 0.039) and α bands (p = 0.044) of the eyes-open state did the difference among the three groups reach statistical significance. Referring to the previous results of alpha blocking analysis, we subjected the α band to further FC contrast analysis.

Figure 4a indicates that the three groups exhibited significant differences only in the eyes-open state (F = 6.23; p = 0.043). Further analysis showed that the I_MMD group exhibited a trend of higher PLI (over 30%) than did controls, mainly in the left frontal lobe (AF3 and C1), middle frontal lobe (FCz and Cz), left parietal lobe (C3, CP1, CP3, P1, P3), middle parietal lobe (CPz), right parietal lobe (CP2 and CP4) and left occipital lobe (O1). The difference between the H_MMD and control groups was not obvious (less than 30%).

Fig. 4.

Comparison of functional connectivity topology among the three groups in the α band. a Comparison of the PLI distribution under different states among the three groups. The ratio (defined as 100%*(PLI_MMD-PLI_Control)/PLI_Control) was calculated to evaluate the relative PLI distribution of MMD patients and controls and is presented by the colour bar. b The PLI of each electrode (average subjects) in all three groups. c Functional connectivity pattern of the three groups with an averaged network degree of 2. d Graph similarity between individuals and the average of the functional connectivity map in each group

Considering the state transition, all three groups exhibited a pattern of PLI decrease from the eyes-closed to eyes-open and working memory states (Fig. 4b and S4). Differences in ΔPLI among the three groups reached statistical significance from the eyes-closed to eyes-open state (p = 0.0166) and the eyes-closed to working memory state (p = 0.0408, Table S3). Further analysis indicated that only I_MMD patients exhibited a trend of less ΔPLI decrease (less than 50%) than did controls in the right parietal lobe (C4 and CP4) from the eyes-closed to eyes-open state.

Figure 4c and Table S4 shows the FC topological map based on electrodes. In the eyes-closed state, the FC of the controls was mainly concentrated on the posterior area, and long-range connections were found from the posterior occipital lobe to the frontal lobe, while the key electrodes were located in the occipital lobe (Oz, O2). In contrast, the I_MMD group exhibited less FC in the right posterior area and key electrodes located in the left parietal lobe (CP1, P1 and P3). Interestingly, the H_MMD group exhibited less FC in the bilateral frontal lobe and key electrodes located in the right parietal lobe (P4 and PO7). When the state was transferred to the eyes-open state, the controls exhibited a diffused pattern of the whole brain. Both the I_MMD and H_MMD groups retained some long-range connections from the posterior occipital lobe to the frontal lobe. When the state was transferred to the working memory state, the controls presented with long-range connections from the posterior occipital lobe to the frontal lobe again. However, these long-range connections disappeared in both the I_MMD and H_MMD groups.

Figure 4d demonstrates the topological similarity (for the method details, please see the supplementary materials) between the average FC patterns and the individual FC patterns. The higher similarity indicates that we could use the average results (Fig. 4c) to represent the results in group level. Thus, we found the differences of FC patterns between the MMD groups and healthy control group are obvious (Fig. 4d).

Small-world characteristics

The differences in small-world related parameters are shown in Fig S5. Significant differences among the three groups were observed only in full frequency bands of normalized path length ($\mathrm{\lambda }$) and small-world characteristics ($\mathrm{\sigma }$) in the eyes-open state. Subgroup analysis showed that the I_MMD group exhibited significantly higher $\mathrm{\lambda }$ (1.016 vs 1.006, p = 0.010) and lower $\mathrm{\sigma }$ (1.002 vs 1.012, p = 0.015) than did the control group.

Cost-efficiency index

Figure 5 indicates significant CEI differences among the three groups in a certain range of network cost of the α band in the eyes-closed state and the θ and α bands in the eyes-open state. Further subgroup analysis in these ranges implied that the I_MMD group exhibited the lowest CEI in the α band of the resting state (p < 0.05).

Fig. 5.

Comparison of cost-efficiency balance among the three groups under different states in the θ a and α b bands. The X-axis represents the cost (also called density) of the unweighted network, i.e., a cost of 10% indicates that the top 10% of the PLI weights construct the network. The Y-axis represents the cost-efficiency index (CEI). The solid line indicates the average CEI of different subjects. The shaded area indicates the standard error. The black asterisk represents a significant difference between the two groups (the alpha level was set to 0.05)

Discussion

For the first time, we applied both power spectrum and graph-based network analyses to explore the neurophysiological nature of adult MMD and its different clinical phenotypes. By assessing the metrics under both resting states and working memory tasks, we identified several key EEG features that can aid in representing distinct electrophysiological features underneath H_MMD and I_MMD brains.

Abnormal power spectrum of MMD

Previous studies indicated a positive correlation between the MF of EEG and the brain metabolic rate (Ingvar et al. 1976); (Tolonen et al. 1981); (1969); (Ingvar 1971). In the present study, the MF of MMD patients was significantly lower than that of controls under the eyes-closed state. This result is reasonable because MMD brains are typically ischaemic and hypoxic due to stenosis and occlusion of small blood vessels. These conditions result in a low metabolic rate and cannot support a high level of brain activity.

When healthy controls switched to a focused working memory state from eyes-open resting, the EEG MF decreased slightly because unrelated brain regions may be suppressed, while only the involved brain regions are engaged with a certain activity level for memory tasks.

However, both I_MMD and H_MMD patients presented an increased brain activity MF when switching from resting to the working memory state. This finding suggests that MMD patients may have to recruit more brain regions and cost more to accomplish certain tasks. In addition, the present study provided different global MF maps of moyamoya phenotypes in both resting and working memory states. Although the regions with either higher or lower MF may not be the exact locations of the lesions (Nagata 1989), it could provide insight into the intrinsic distinct pathophysiological processes of H_MMD and I_MMD. The diverse MF dynamics between haemorrhagic and ischaemic brains may suggest that it could serve as a potential biomarker for preclinical differential diagnosis.

Substantial evidence indicates that cortical networks may employ a sparse coding algorithm in coding the outward signals or perform cognitive computations. For example, visual V1 neurons exhibit a sparse response to the natural images (Vinje and Gallant 2000); (Simoncelli and Olshausen 2001), and olfactory bulb neurons exhibit sparse and selective coding in response to odour (Jortner et al. 2007). The sparse coding principle indicates that neuronal populations encode rich information with fewer neural activities. A neural system with high sparseness indicates high efficiency (Yu et al. 2014); (Olshausen and Field 1996); (Vinje and Gallant 2000) Based on our results, both MMD patients and healthy controls presented increased efficiency when brain behaviour switched from resting to focused cognitive tasks, while healthy controls reached higher sparseness levels. Among the three groups, H_MMD patients have the lowest brain response sparseness in the working memory task, suggesting that bleeding inside certain brain regions may have damaged cognition in the past.

Alpha blocking with visual stimuli (eyes-open) is a classical response (Niedermeyer 1997). In our study, we observed a decreased alpha blockade rate in both H_MMD and I_MMD under the eyes-open and working memory states, suggesting that patients may be less alert and conscious during the waking state. Moreover, the particular lowest alpha blocking rate in I_MMD patients under the eyes-open state may serve as a clinical phenotype for discriminating ischaemic patients from haemorrhagic patients. This result was consistent with a previous study of cerebral ischaemia, which also found unsuppressed alpha activity in patients in the eyes-open state (1984).

Abnormal graph-based network of MMD

We compared both FCS and topological mapping between the MMD and healthy control groups in the present study. A previous study demonstrated that FCS is reduced in healthy controls when switching from the eyes-closed to eyes-open state or working memory state in the α band, which suggests that the brain becomes desynchronized when processing the incoming external stimulus (Pfurtscheller 2001). In our study, MMD patients exhibited a similar decrease in FCS, indicating the highly difficult desynchronization of brain correlations during active cognitive processes.

Further topological analysis showed that in the eyes-closed state, healthy controls maintained strong long-range connections from posterior regions to the frontal lobe. This pattern may represent the functional brain networks of cognitive control, such as frontoparietal, cingulo-opercular, and salience networks (Marek and Dosenbach 2018). However, this topological pattern changed when mapping the FCS of the MMD groups, indicating deteriorated functional networks. This result is consistent with our previous fMRI network studies in the resting state (Lei et al. 2018). In addition, the topological patterns of H_MMD and I_MMD changed in different manners when the state transferred from the eyes-closed to eyes-open and working memory states, respectively, implying their different pathophysiological processes.

Graph theoretical analysis provided intuitive evidence of network segregation and integration characteristics of the brain (Rubinov and Sporns 2010). In the full frequency band of the eyes-open state, the three groups exhibited significant differences in global integration. Moreover, a detailed calculation indicated the most randomized topology in the ischaemic phenotype. This result represents the advancement of a previous graph theoretical study of the moyamoya structural network, evaluating more behaviour states and frequency bands (Kazumata et al. 2016). Furthermore, we considered lagged phase-based connectivity to efficiently avoid volume conduction and assess weighted network characteristics.

Neural interaction has been proven efficient at a relatively low cost and can be measured by a parameter of CEI (Achard and Bullmore 2007); (Bullmore and Sporns 2012). In the present study, the three groups exhibited significantly different CEI in certain ranges of the α and $\theta$ bands of the resting state, while ischaemic phenotype exhibited the lowest CEI in the α band of the resting state. This result implied that I_MMD contained more low-frequency components and fewer high-frequency components than did the other two groups because of different pathophysiological durations.

Limitations

Primarily, the statistical power of this study may be affected by the small sample size. MMD is a rare disease, it is not easy to recruit a lot of subjects for the experiments. Moreover, due to the declined effect of MMD on the cognitive functions of patients, 16 subjects did not complete the experiment well, leaving 24 MMD subjects for the data analysis. This small size may limit the result conclusion in this paper. Since EEG has low spatial resolution, we combined EEG and fMRI techniques in another recent study of the same cohort. Finally, cerebral blood perfusion is very complicated in MMD. Thus, further studies with large sample sizes and haemodynamic evaluations (Xie et al. 2020) as well as mechanism study with animals (de Oliveira et al. 2020) are needed.

Conclusion

In sum, we applied power spectrum and graph-based theory analysis to examine the electroencephalographic interaction characteristics of adult MMD. This study helps not only in understanding the pathophysiological nature of MMD and its different phenotypes, but also in laying the foundation for further differential diagnosis and treatment.

Author contributions

YY, YG and YM supervised the research, YL, GZ and YY designed the research, GZ, YL, YL, WZ, JS, XQ, LC and XZ performed the research, and GZ, YL and YY wrote the paper. All authors reviewed the manuscript.

Data availability

EEG data collected in patients in this paper are owned by Huashan Hospital, Fudan University. If you have any questions about the authenticity of the data, you can contact the corresponding author by email.

Compliance with ethical standards

Conflict of interest

The authors declare no conflict of interests.

Ethical approval

This study was approved and supervised by the Ethics Committee of Huashan Hospital.

Informed consent

All participants signed written informed consent.