We are first to compare the rsEEG mMSE features between the electrode sets corresponding to the main resting-state networks. A novel result here is also an identification of the sex/gender differences in the rsEEG complexity at the fine and the coarse scales. To the best of our knowledge, this work is the second one, following Dreszer et al. (2020), that quantified the rsEEG signal complexity using the mMSE algorithm, i.e. an extension of the multiscale SampEn (MSE) (Costa et al., 2005) to the multivariate timeseries (signals, e.g., EEG), proposed by Ahmed and Mandic (25), Ahmed et al. (2012) and Looney et al. (2018). Therefore, the key property of the mMSE, not present in the SampEn and MSE methods, is that it is designed for analysis of multivariate signals whereas the SampEn and MSE are applicable to univariate signals only. Moreover, just as the MSE, the algorithm, used here, examines the complexity at both fine-grained (short) and coarse-grained (long) timescales (M.U. Ahmed et al., 2012; Ahmed and Mandic, 2011; Looney et al., 2018). The shape of mMSE vectors in our study (Fig. 1) resembles a typical skewed inverted U-pattern (Costa et al., 2005, 2002) which is believed to result from the shortest scales representing only a random signal and the longest scales reflecting a more stable system, characterized by a reduced variance (Costa et al., 2005; McIntosh, 2018).
4.1. The highest resting-state EEG entropy at the scalp locations corresponding to the SMN and the DAN
We found the highest, relative to other networks, AUC values, representing a total entropy, in the electrode sets corresponding to the SMN and the DAN (Fig. 1, Table 2). A similar location of the largest complexity (a central area) has been identified by other authors (Fernández et al., 2012; Racz et al., 2019; Vecchio et al., 2021), however, the SMN also demonstrated relatively low entropy levels (Jia et al., 2017; Mediano et al., 2021; Nezafati et al., 2020). These inconsistent results may arise from different methodology, i.e. various brain signals (EEG,MEG or fMRI), resting-state conditions (eyes open or eyes closed) and algorithms used for data computation (e.g., Lempel–Ziv Complexity, ApEn, SampEn).
Spontaneous SMN fluctuations reflect activation of the motor system in the absence of any movement (Biswal et al., 1995) whereas the DAN is involved in the top–down attention control and the expectation of objects at a particular location or with certain features (Corbetta et al., 2000). The topography and strength of resting-state networks are associated with the history of network activation (Deco and Jirsa, 2012). Therefore, the highest degrees of entropy in the SMN and DAN may reflect their increased functional connectivity with other regions or a large repertoire of possible responses to stimuli resulting from a great experience in coding features/spatial locations of the objects and motor response to these stimuli. Since the SMN and the DAN are the “task-positive” networks, their complexity levels might be more influenced by online processing and the changes in cognitive demands than the intrinsic networks such as the DMN. When there are no external tasks or stimuli, the SMN and DAN are just prepared to respond to them and in this condition there is probably no need to communicate extensively with other regions. In this context, we would be rather inclined to accept the theory postulating that greater degrees of neural complexity represent less information exchanged across brain areas (Ghanbari et al., 2015). Otherwise, the brain signal entropy may not directly reflect the “activity” or amount of information processing which has been already suggested by other authors (e.g., Nezafati et al., 2020).
In the current study the SMN and the DAN demonstrated the greatest AvgEnt values corresponding to the complexity at the coarse scales (Fig. 1, Table 2) which might represent a greater capacity of these networks to process information globally, across distributed brain regions (Courtiol et al., 2016; McIntosh et al., 2014; Vakorin et al., 2011). The long-range connections in the brain play an important role in perception (Sporns et al., 2000) and information integration (Just, 2004) which makes them crucial for the main functions of the SMN and the DAN. Congruently with Omidvarnia et al. (2019) who found the highest positive correlations between the complexity at the coarse timescales and functional connectivity of the fMRI resting-state signal for the SMN and VN, the entropy at the scalp location corresponding to the SMN in our study might represent long-range interactions with other areas. The rsEEG signal from the electrode sets corresponding to the SMN also demonstrated the greatest complexity at the fine scales (Fig. 1, Table 2) suggesting the strongest, compared to other networks, short-distant functional couplings, and local synchronization across connected regions (McIntosh et al., 2014; Vakorin et al., 2011).
Inconsistently with the previous results (McDonough and Nashiro, 2014; Nezafati et al., 2020; Omidvarnia et al., 2019), the DMN in our study did not demonstrate the largest rsEEG complexity but, still, the AUC, MaxSlope and AvgEnt values for this network were among the highest, same as for the VAN (Fig. 1, Table 2). A relatively high entropy level in the DMN and the VAN across the timescales might be interpreted as representing greater information processing by both local and distributed neural assemblies. The DMN is considered a critical gateway for transferring information within local and across distributed networks (Bassett and Bullmore, 2006; Buckner et al., 2009) which may be reflected in high degrees of entropy at the fine and the coarse scales. However, in light of the evidence showing that the DMN complexity is one of the least associated with the functional couplings (Omidvarnia et al., 2019), the interpretation of entropy in the context of functional connectivity becomes less obvious than in the case of other networks. The DMN is “active” at rest (Buckner et al., 2008; Raichle et al., 2001) when its “information processing” relies on the spontaneous generation of images, voices, thoughts, and feelings that are stimulus-independent and resulted from mind wandering (Mason et al., 2007) and/or monitoring the external environment (Shulman et al., 1997). Therefore, it is even more natural to interpret a high complexity of this network in terms of extensive transitions between states and the brain’s tendency to wander (Honey et al., 2009; McIntosh et al., 2010).
4.2. The lowest resting-state EEG entropy at the scalp locations corresponding to the LN, VN and the FPN
The lowest total rsEEG entropies were observed for the channel sets linked to the LN, VN and the FPN (Fig. 1, Table 2). Similar results have been obtained by other authors in the case of VN (Jia et al., 2017; Mediano et al., 2021; Nezafati et al., 2020; Wang et al., 2014) or the LN (Mediano et al., 2021; Nezafati et al., 2020), however, the results suggesting the highest degrees of complexity in these two networks, compared to others, have been also reported (Jia et al., 2017; Smith et al., 2015). In the current study the VN showed the smallest total entropy and the complexity at both fine and coarse scales. Similarly to the VN, the LN demonstrated one of the lowest AUC and AvgEnt values. These outcomes might reflect less transitions between states (Faisal et al., 2008) and/or reduced (short- and long-distant) functional connections of these networks with other regions (McIntosh et al., 2014). Since the VN is mainly activated by visual stimuli, it was probably not very involved at rest when the participants are asked to keep their eyes fixated on one point in space and avoid any ocular movements. Similarly, the resting protocol favors a state of relaxation and calmness where the LN might be not particularly engaged. In this context the brain entropy would directly reflect its activity.
In contrast to other authors (McDonough and Nashiro, 2014; Mediano et al., 2021; Nezafati et al., 2020; Omidvarnia et al., 2019), we demonstrated that the FPN was characterized by one of the lowest entropies (Fig. 1, Table 2). The FPN is considered as an intrinsic network, recruited by executive control tasks (Dosenbach et al., 2006), and its complexity, similarly to the DMN, is rather expected to be among the highest at rest. A relatively low FPN entropy level, found here, may be explained referring to the method of obtaining the mMSE vectors. In this context a small complexity in the FPN might represent a lack of meaningful spontaneous interactions between the frontal and parietal regions forming this network. Indeed, in our previous study (Dreszer et al., 2020) where the mMSE method was applied to quantify rsEEG complexity, the parietal region was characterized by the greatest entropy whereas the frontal areas demonstrated one of the lowest entropies. Thus, we might assume that the frontal and the parietal regions, analyzed together as the FPN, will demonstrate different entropy levels than the same areas examined separately. To test this assumption, an additional mMSE analysis was performed on the current database but using the channel sets from our previous work (Dreszer et al., 2020). We found a similar complexity pattern in case of both former and current datasets revealing the highest degrees of total complexity and the entropy at the coarse scales separately for the frontal and parietal regions (see: Fig. A.1. and Fig. A.3., Appendix). Therefore, we might carefully conclude that the low FPN entropy in the present study results from different spatiotemporal patterns of spontaneous fluctuations in the anterior and posterior part of this network.
4.3. Resting-state EEG entropy range between the fine and the coarse timescales
In the current study the LN showed the greatest complexity range, i.e. among the highest entropy levels at the fine scales and the lowest at the coarse scales (Figs. 1 and 3, Table 2). A similar complexity pattern was reported by McDonough and Nashiro (2014) for the Cingulo-Opercular Network (CON) whose core hubs (the dorsal anterior cingulate cortex and the insula) are involved in emotional processing (Dosenbach et al., 2007). Such a mMSE profile of the LN, possibly reflecting the synchrony of neural assemblies at the fine scales and desynchrony at the coarse scales (Baptista and Kurths, 2008), may represent a process of emotional adaptation to the resting-state condition.
The DAN and the VN demonstrated relatively stable mMSE patterns across the timescales (had the smallest DiffEnt values, Fig. 3, Table 2) suggesting comparable amount and strength of local and global information processing. After reaching a relatively high complexity level both these networks showed a slow decrease of entropy (Fig. 1) which may reflect keeping balance between neural excitation and inhibition processes.
The DAN could be also considered as a separable, stable, internally more coherent, module in the brain (Hilger et al., 2020). The complexity of the task-positive networks such as the DAN and the VN, may change more in a task than at rest (McDonough and Nashiro, 2014). Therefore, a small differences in the entropy level at the fine and the coarse scales (low DiffEnt values) in the DAN and the VN may result from a relatively stable amount of visual stimulation and recruited cognitive resources during the resting state.
Interestingly, the DMN and the VAN in our study had almost overlapping mMSE vector shapes (Fig. 1) showing among the highest MaxSlope and AvgEnt values relative to other networks (Table 2). This effect is hard to explain in terms of sharing the same channels by these networks (the DMN and the VAN had only two common positions: T7 and T8). It is of note that we analyzed the rsEEG signal from the electrode set corresponding to the VAN, extracted by Yeo et al. (2011), and this network includes both the CON and the “classic” VAN which have demonstrated strong synchronization at rest (Dosenbach et al., 2007; Power et al., 2011) possibly reflected in the great degrees of rsEEG entropy.
The DMN is considered as a part of the task-negative system whereas the VAN/CON is identified as a task-positive network (Lee et al., 2012). In this context the DMN and the VAN/CON were supposed to indicate some inverse (anti-correlated) functionality resulting from the allocation of limited resources. As a support for this claim McDonough and Nashiro (2014) found the lowest complexity at the fine scales and the highest entropy at the coarse scales in the DMN whereas for the CON the inverse neural complexity profile was obtained. In our study similar mMSE patterns in the DMN and the VAN may result from the function that they both share, e.g. tonic alertness (Buckner et al., 2008; Sadaghiani and D’Esposito, 2015) which is highly involved during the resting state.
In our study the SMN and the LN had similar mMSE vector shape to the DMN/VAN (Fig. 1). While the VAN and the LN signals were acquired from very similar sets of electrodes (which may, to some extent, explain the similarity of their rsEEG complexity profiles), for the DMN or the SMN this was not the case. We might speculate then that the resemblance of mMSE patterns between particular channel sets (networks) might reflect their strong interactions or interdependencies. Conversely, different entropy patterns in given networks might represent weak connections between these systems. Therefore, in the present study, similar complexity profiles, especially in the DMN and the VAN and, to a lesser extent, in the SMN and the LN, might suggest the increased communication between these networks at rest.
4.4. Higher resting-state EEG complexity at the fine timescales and lower at the coarse scales in women compared to men
Although we did not find any significant s/g differences in the AUC values, corresponding to the total rsEEG complexity, women, relative to men, showed higher MaxSlope values (the fine scales) but lower AvgEnt values (the coarse scales) (Fig. 1). Since we are not aware of any work presenting the s/g differences in the mMSE values, we decided to quantify the complexity of the current data using exactly the same method and regions of interest as in our previous work (Dreszer et al., 2020). The detailed results of this re-analysis were shown in the Appendix in Fig. A. 1–3. Basically, we found a reproducible pattern of the s/g differences in the entropy level at the fine and the coarse timescales: for both datasets, women, compared to men, produced significantly higher MaxSlope values in the frontal areas and men showed greater AvgEnt values than women for all analyzed channel sets. Considering the AUC, the outcomes across both our studies were inconsistent: females demonstrated a greater total entropy level than males only for our previous dataset and we did not observe such s/g differences in the current study.
The way how the MaxSlope is calculated (the maximum change in the EEG signal complexity at the fine scales) allows to compare this parameter with the entropy, determined using other methods and measured at a single timescale. In this context, the s/g differences in the MaxSlope value are congruent with some previous findings (Fernández et al., 2012; Kumral et al., 2020; Pravitha et al., 2005; Wang, 2021). Higher degrees of entropy at the short scales and lower complexity at the long scales in women might reflect enhanced information processing in local neural assemblies but reduced large-scale interactions in the resting brain (McIntosh et al., 2014; Vakorin, 2011). The previous resting-state fMRI evidence suggesting a greater overall brain integration of specialized information at a global network level in males and higher segregation, i.e. specialized processing of the brain at a local level, in females supports the aforementioned interpretation of the brain complexity (Allen et al., 2011; Jalili, 2015; Satterthwaite et al., 2015; Tomasi and Volkow, 2012). Some authors, however, reported the opposite s/g differences in the short- and long-range functional connectivity (Foo et al., 2021; Zhang et al., 2016). There is also evidence on greater anatomical connectivity and higher both local and global efficiency of the cortical networks in women compared to men (Gong et al., 2009).
Greater neural complexity at the coarse scales as reflecting more distributed organization of the brain in men might result from optimal functioning of specialized and complex processes such as visuospatial imagery or orientation, which recruit long-distance connections. For example, the mental rotation tasks, where men often outperformed women (Gong et al., 2009; Hattemer et al., 2011; Hoppe et al., 2012), require visualization of the rotation of objects in space and, then, correctly matching them with exemplars which involve testing and comparisons before a decision is made. Furthermore, previous studies have revealed that men do not exhibit a higher local efficiency in visuospatial processing regions (Gong et al., 2009) suggesting that men’s superior performance in such tasks may also rely on the long-range interactions of these areas. The greater local information processing in females, on the other hand, might optimize functions that require synchronization across local networks such as those supporting verbal fluency in which women achieve better scores (Gauthier et al., 2009; Scheuringer et al., 2020; Schlosser et al., 1998). The s/g differences in the neural complexity at the fine and the coarse scales (and corresponding local and global network processing) might also reflect a predisposition of one of the sexes to develop certain disorders including autism which occur the 5–10 times more often in men than in women and is characterized by reduced local functional connections in the brain (Tommerdahl et al., 2008).
In the present study we did not find any significant interaction effects between the s/g factor and the channel sets corresponding to the resting-state networks, distinguished by Yeo et al. (2011). The s/g differences in the neural complexity at the short and the long scales, were observed in the whole brain, not in particular networks. However, when we analyzed the current data using the same regions of interest as in our previous study (Dreszer et al., 2020), it turned out that women, relative to men, showed higher MaxSlope values in the frontal areas which is, basically, in line with other findings (Ahmadi et al., 2013; Pravitha et al., 2005). In the current study a lack of specific scalp locations with significant s/g differences in the rsEEG entropy level might result from overlapping channel sets forming particular networks. For example, the frontal electrodes from our previous work (Dreszer et al., 2020), where women showed higher MaxSlope values than men, are here distributed among 4 networks (FPN, DMN, LN and VAN). Therefore, it is possible that the signals from these channels contributed significantly to the temporal dynamics of all aforementioned networks producing the cumulative effect observed at the whole brain level.
Interestingly, in our study women showed higher DiffEnt values (less stable mMSE profiles) than men in all analyzed networks (Fig. 4). These outcomes may suggest comparable levels of local and global information processing in male brain during the resting state and the advantage of short-distance over the long-distance interactions in females.
4.5. Limitations of the study and further directions
The present outcomes should be interpreted in light of several limitations. The most serious concern the validity and reliability of the analyzed variables. The particular aspects of current study burdened by the most important limitations are listed below.
4.5.1. Interpretation of entropy values at the fine and the coarse scales
Referring the brain signal complexity at the short and the long scales to local and global information processing respectively, has already met some criticism (Kosciessa et al., 2019; Omidvarnia et al., 2019) and needs further verification. In future studies it would be beneficial to take into account the anatomical data while interpreting the neural complexity results, especially information about the white matter microstructure since the white matter integrity has been already related to the network complexity at both fine and coarse scales (McDonough and Siegel, 2018). Furthermore, the coarse-graining procedure acts as a moving average low-pass filter that impedes separation of particular frequency bands from the signal. Therefore, the relationship between the mMSE and the frequency content of EEG signal is difficult to determine (Courtiol et al., 2016), although a certain progress has been made recently in this area (Kosciessa et al., 2019).
4.5.2. Reliability of the mMSE vectors and their features
In our study stability of the mMSE vectors has been checked using an internal consistency method. The obtained results suggest either a high or a very high reliability, both in the case of timescales and features of the mMSE vectors. To date, attempts to check the internal consistency of complexity indexes have been made in a few EEG studies (e.g., Tomasi and Volkow, 2012). However, it does not exhaust the need for in-depth research on the reliability of MSE measures. In future research the test-retest stability should be determined (comparable to some fMRI studies, e.g., Omidvarnia et al., 2019).
4.5.3. The necessity to replicate the study results
The present study is the second in the literature (after our previous work) where the mMSE algorithm was used to describe the dynamics of resting-state EEG activity at the short and the long timescales and is also the first attempt to use the multivariate method to describe the dynamics of the brain networks during the resting state. Hence, it would be strongly recommended to ensure to what extent obtained results could be replicated on a different sample. Being aware of this problem in this work we re-analyzed the current results with the use of electrode sets from our previous work (Dreszer et al., 2020) and demonstrated a substantial consistency in both our studies (Appendix File). In future research, however, it would be desired to replicate the results also from the brain network perspective.
4.5.4. Weaknesses of the methods: the mMSE algorithm and the proximity maps
4.5.4.1. Limitations of the mMSE algorithm
In the present study the same mMSE algorithm with the same parameters (EEG signal segmented into multivariate time series of length n = 1024, with m = 2, r = 0.15, p = 4, ε = 12, τk was set to 1 for k = 1,2,…,p, where m is the embedding coefficient, r is the similarity threshold, p is the number of channels in a given channel set, and ε is the timescale factor, τk is the time delay) as in our previous work (Dreszer et al., 2020) was used which allows for direct result comparisons. On the other hand, both these analyses share the same limitations of the mMSE algorithm which was already discussed elsewhere (Dreszer et al., 2020, Section 4.3.).
4.5.4.2. A lack of direct localization of EEG sources/problem of proximity maps
An important limitation of our analysis is the selection of channel sets based on the average electrode proximity maps provided by Giacometti et al. (2014) without a direct localization of EEG sources in our dataset. Unfortunately in the present study the MRI anatomical data useful for proper source localization were not available. A lack of EEG source locations did not allow to exact identification of the networks, therefore, the proximity parcellations can ONLY be treated as an approximation.
Since our sample comprised only young healthy adults, the models (the regions of interest from the standard brain atlases), used by Giacometti et al. (2014), might be a quite good approximation. However, in the case of more specific groups, these models could be hardly applied (a decreased effect size, DeYoung et al., 2022). In our opinion this method is only applicable when we may assume that the variability of anatomical locations of EEG sources will be comparable across participants, i.e. mainly in the case of a relatively homogenous sample, e.g. healthy young adults. In future research on the resting networks complexity, tested using EEG, it would be reasonable to perform the source localization.
4.5.4.3. Electrode sets
We should also mention about the electrode sets used in our study. They correspond to 7 main resting networks determined using the algorithm developed by Giacometti et al. (2014) and have the important disadvantage of including overlapping channels forming particular networks. It would be advisable to give up the overlapping electrodes and assign each channel to only one network. Otherwise, it would be recommended to resign from including to the analysis both LN and VAN channel sets differing in only one electrode. However, this change would cause a serious interference in the original parcellation which was created based on the localization and reconstruction of the EEG sources. In this case the confirmation of the procedure validity would be required.
Finally, our study was not free from limitations that also exist in almost every resting-state EEG study on the s/g differences (e.g., lack of controlling for the menstrual cycle, McEwen and Milner, 2017).