1. Introduction
Complex network analysis applied to neuroscience research has shed light on how the human brain efficiently transfers information using limited physical connections [
1,
2]. Modular processing of information occurs through segregated, specialized, and functionally modularized brain networks, and this highly efficient clustering can be observed in the visual system. For example, even though color and motion are temporally asymmetrical and autonomous modules, such that color is processed before motion [
3], during the perceptual processing of a scene, information from both sensory modalities is employed simultaneously through the integration of specialized processors. Thus, cross-modal activations of specialized processors may play a role in broadcasting information to a cognitive network endorsed by hubs [
4] (a set of mainly central nodes engaged in information flow and also known as the connective core).
The existence of a cognitive network that integrates information during high-level cognitive processing is predicted by the global workspace theory [
5]. This theory postulates that the information processed in primary sensory areas is transmitted to a global neuronal workspace [
6] that mobilizes excitatory neurons with long-range axons capable of interconnecting sensory-level and high-level processing areas [
7]. The cognitive network is a locus within this global workspace and consists of interconnected brain sites [
6,
7]. Therefore, the global workspace establishes a framework for network cooperation, competition and state transitions for processing segregated information during effortful cognitive tasks [
8]. For that reason, the cognitive network topology is posited as a dynamic state with decreased segregation of specialized sub-networks in exchange for increased brain connectivity.
The transition between segregated modular processing of specialized sub-networks to integrative processing of the cognitive networks creates a temporary metastable state of global activity [
5]. Thus, determining dynamic changes in the topology of neuronal networks induced by different cognitive states enables the estimation of task effects on speed and specialization of information processing [
8]. A previous study investigated the reconfiguration of network topology during the transition from resting-state to a visually-cued finger-tapping task [
8]. One of the major findings of this study was that network parameters of the global topology were conserved in both states. This result suggests that, within a small-world topology, functional networks are able to support different, task-specific connectivity patterns without drastic changes in global topology. Another important finding was that motor tasks, compared to a resting-state, gave rise to higher frequency bands connecting pivotal nodes on the parietal and frontal regions. These frequencies also presented the lowest level of synchronicity of coupled oscillators in the limit between ordered and disordered behavior. The authors suggest that higher frequency networks support rapid and adaptive reconfiguration in changing environments [
8].
Characterizing the effective connectivity between nodes in the cognitive network may offer insight on the gating dynamics of functional networks and workspaces. For example, nodes that are not task-specific could mark functional areas engaged in cooperative dynamics. Indeed, if information is transmitted to pivotal hubs, the networks they make up become potential components of the cognitive network. That is, the engagement of pivotal hubs could indicate that information is being integrated in a cognitive network.
3. Discussion
In the present study, we analyzed the connective core [
4] of eleven subjects during effective attention [
11] and working memory [
12] tasks, as well as during resting-state [
13]. The connective core, also called rich-club, is a substructure composed by neural hub regions that are densely interconnected and promote efficient communication and functional integration. van den Heuvel & Sporns [
4] also reported a structural imbalance between incoming and outgoing projections of some brain areas, calling these regions “net receivers” and “net emitters”. This property can also affect the effective connectivity aspects, suggesting that these hubs with different behaviors can play a potential role in information flow as neural communication “sources” and “sinks”.
We proposed a similar quantified classification into three hub types: divergent hubs, convergent hubs, and neutral hubs. However, in addition to considering an imbalance to define sources (divergent hubs) and sinks (convergent hubs) nodes, we also introduced a range in which the incoming pathways are not representatively greater than the outgoing pathways or vice versa (neutral hubs). Furthermore, we evaluated the hubs’ contributions, dynamic behaviors and roles in the network. Our data reveal that divergent hubs are more effective in characterizing the network; when they are removed, there is a significant loss of effective connectivity of the whole network (see
Figure 1 and
Table 2). We observed that the average percent variation in the characteristic path length when a single divergent hub is removed from the network is generally greater than when any other hub type is removed, across behavioral paradigms. This effect was observed across all behavioral paradigms and supports a previous study by Kaiser et al. [
14], who evaluated the robustness of brain networks when removing their nodes and edges randomly or selectively. Similar to what our data reveal, the researchers concluded that, if structures with many connections (e.g., an isolated hub) or the connective core are removed, the functional effects on the brain network should be substantial.
After understanding the contribution of the connective core and the different hubs to the network effectiveness, we analyzed the connective core’s composition. In doing so, we observed that the connective core is composed mostly of divergent hubs (see
Table 3). Furthermore, we performed a tracking of the connective core across behavioral states. We observed that the connective core is composed of an invariant substructure, which is preserved between behavioral states, as well as a variant substructure that is characteristic of each paradigm (see
Table 4). In addition to the divergent hubs making up most of the connective core, they also account for most of the substructure preserved between behavioral states. When looking at the variant substructure during the transition between different paradigms, certain patterns are easily noticeable: (a) half of the divergent hubs remain divergent, while the rest cease to be hubs, suggesting that divergent hubs play an integrative role; (b) neutral hubs become divergent or cease to be hubs; and (c) most convergent hubs cease to be hubs, implicating that they may be paradigm-specific hubs. These effects were described by Sporns [
15], who defined central nodes as a mechanism of convergence and divergence of information flow, ensuring integrated processing. In another study, Meyer & Damasio [
16] described a structural convergence zone that is responsible for the specialized process of sensory input and a structural divergence zone that integrates segregated information to build a response to sensory stimuli.
We also investigated whether the increased contribution by divergent hubs to the connective core, described above, was caused by their greater volume relative to other hubs. To accomplish that, we assessed the individual net entropies of each hub type, as described in
Section 2.1. Importantly, we observed that divergent hubs contribute significantly to the effective connectivity of the global network across all behavioral states—an effect that was independent of their volume and is described in
Table 1. Moreover, we observed that convergent hubs have a greater contribution during resting-state than in cognitive tasks. However, despite this larger contribution by convergent hubs, we also see that the contribution from divergent hubs is still representative during resting-state. This result can be associated to an electrophysiological signature of the default mode network (DMN); van den Heuvel & Pol [
17] described the DMN as a spontaneous functional activation of anatomically distant brain regions during resting-state while the brain is waiting for a stimulus. It is well-established that the brain networks work with a balance between functional integration and segregation [
18]; thus, these results suggest that the large contribution of convergent hubs would be associated with convergence zones of these anatomically distant processing modules, and that the contribution of divergent hubs characterizes this balance between functional segregation and integration.
When we investigated the dynamic behavior of different hub types, some patterns were present. We observed a “swap” relationship between neutral and divergent hubs across all paradigms, as shown in
Table 5. This finding would be a consolidation of the tracking of the variant substructure over time described above, where we noted that neutral hubs have a preference to turn into divergent hubs or cease to be hubs. This preference may indicate that neutral hubs play a supporting role to divergent hubs in integrating the different sub-networks that can emerge during some behavioral states. However, we can also see that, during cognitive tasks, the “swap” between convergent and divergent hubs becomes greater than during resting-state. Analyzing the percent variation in relation to resting-state, we observed an increase of approximately 29% for
and of more than 100% for
. Fair et al. [
18] described that an integrative process can manage different control networks, offering a possible explanation for the increase of the observed “swap” effect during cognitive tasks.
EEG microstate topographies are thought to be electrophysiological correlates of these episodes of coherent activity proposed by the global workspace model [
5,
19,
20,
21,
22]. Researchers who have tried to establish a microstate with resting-state fMRI (rsfMRI) signal have frequently observed no correlation between discrete EEG frequency bands and corresponding hemodynamic states [
23,
24,
25]. On the other hand, microstate time courses, rather than discrete frequencies, have shown correlation with BOLD activation of distinct networks [
24]. The large-scale networks described in the global workspace model have to be both stable, through the duration of a specific cognitive process, and flexible, to rapidly transition between tasks. Thus, these networks must be able to change into different patterns of connectivity at a sub-second time scale that is consistent with that of EEG microstates.
A correlation analysis of resting-state fMRI and EEG microstates to determine if a single hemodynamic network is related to one or more microstates relies on the assumption that during this “rest” period no conscious cognitive process is occurring in the brain being analyzed. However, given that participants are awake, there is no way of monitoring the amount of cognitive processes that they may engage despite the participant not being explicitly instructed to do so (i.e., to complete a specific task). When these sub-second dynamics within a period of rest are taken into account, the EEG time course correlation with distinct networks of BOLD activation makes sense [
26,
27,
28].
Given these observed patterns, we performed an analysis of complexity of the microstates’ sequences in selected windows of EEG data, as described in
Section 4.6. In doing so, we observed that in windows with common high quantities of different hub types, the microstates’ sequences complexity are higher than in windows with only one high quantity of a given hub type (see
Section 2.4). Therefore, the presence of high quantities of different hub types in the same windows can be a signature of the transition between cognitive networks in a global neuronal workspace [
5]; this effect was also described by Sporns [
15], who asserted that the study of the connective core offers a potential substrate for understanding theories about the global neuronal workspace, high cognition, consciousness, and how disturbances on hub regions can affect integrative processes.
5. Conclusions
In this study, we analyzed the connective core network across three behavioral conditions: resting-state, Stroop and 2-back paradigms. We observed that divergent hubs are more effective in characterizing the network, independent of the fact that these hubs are the most abundant type in the the connective core. In addition, we noticed that the connective core is composed of both an invariant substructure that is preserved between behavioral states, and a variant substructure that is characteristic of each cognitive task and the resting-state. The constructs proposed here can be used to analyze transitions in brain states associated with task components and/or resting-states, playing a relevant role in the balance between functional segregation and integration. Moreover, we observed that the effectiveness of convergent hubs during resting-state may be associated with anatomically distant modules described in the default mode network. Furthermore, the prevalence of a specific hub type supports the existence of well-established functional networks, while the transitions observed characterize an increased expression of different hub types.
This description may bring light to the role of functional brain networks across different contexts, from cognitive processes in healthy subjects to disordered neural events. Thus, studying the connective core offers a potential substrate for understanding theories about the global neuronal workspace, high cognition, consciousness, and how disturbances on hub regions can affect integrative processes.