Spontaneous brain activity and EEG microstates. A novel EEG/fMRI analysis approach to explore resting-state networks

Abstract

The brain is active even in the absence of explicit input or output as demonstrated from electrophysiological as well as imaging studies. Using a combined approach we measured spontaneous fluctuations in the blood oxygen level dependent (BOLD) signal along with electroencephalography (EEG) in eleven healthy subjects during relaxed wakefulness (eyes closed). In contrast to other studies which used the EEG frequency information to guide the functional MRI (fMRI) analysis, we opted for transient EEG events, which identify and quantify brain electric microstates as time epochs with quasi-stable field topography. We then used this microstate information as regressors for the BOLD fluctuations. Single trial EEGs were segmented with a specific module of the LORETA (low resolution electromagnetic tomography) software package in which microstates are represented as normalized vectors constituted by scalp electric potentials, i.e., the related 3-dimensional distribution of cortical current density in the brain. Using the occurrence and the duration of each microstate, we modeled the hemodynamic response function (HRF) which revealed BOLD activation in all subjects. The BOLD activation patterns resembled well known resting-state networks (RSNs) such as the default mode network. Furthermore we “cross validated” the data performing a BOLD independent component analysis (ICA) and computing the correlation between each ICs and the EEG microstates across all subjects. This study shows for the first time that the information contained within EEG microstates on a millisecond timescale is able to elicit BOLD activation patterns consistent with well known RSNs, opening new avenues for multimodal imaging data processing.

Introduction

In almost two decades of functional magnetic resonance imaging (fMRI) studies the dominant method to obtain information on brain function has been measuring changes of brain activity in response to a task. At present, however, we are experiencing a “turn of the tide” in our analytical approach; a new avenue of neuroimaging research on spontaneous fluctuations of brain activity is emerging. These spontaneous fluctuations can be defined as “changes in brain activity not externally induced or voluntarily generated by the subject which occur during relaxed wakefulness” (Raichle and Snyder, 2007, Laufs, 2008). As highlighted by Fox and Raichle (2007) in their comprehensive review, there are good reasons to explore the resting fluctuation of brain activity. This is because the brain at rest is responsible for most of the ongoing cerebral energy consumption and task related increases in neuronal metabolism are usually less than 5% (Raichle and Mintun, 2006).

In the fMRI field, the spontaneous modulation of the blood oxygen level dependent (BOLD) signal, which cannot be attributed to the experimental paradigm, has been viewed for many years as “noise” and is usually minimized through averaging. Several groups observed by means of functional connectivity analyses that spontaneous BOLD fluctuations appear not to be just random noise, but are specifically organized in functionally relevant resting-state networks (RSNs) (Biswal et al., 1995, Lowe et al., 1998, Cordes, 2001, Raichle et al., 2001, Greicius et al., 2003; Fransson, 2005; Fox et al. 2006). Even so, it is still a matter of discussion to what extent the observed fluctuations in the RSNs of the resting fMRI signal are a direct consequence of neuronal activity or whether they are low-frequency artifacts due to other processes (physiological and scanner related artifacts). A growing body of evidence suggests that despite being able to detect about a dozen spatially consistent RSNs across subjects (Fox and Raichle, 2007), fMRI alone, as an indirect measure of neural activity, is not well suited to assess the functional significance of these RSNs. A further disadvantage is that the BOLD signal changes are delayed and temporally blurred with respect to the underlying neural events.

On the contrary, electroencephalography (EEG) is a direct measure of the activation status of large cooperating neuronal assemblies. In the electrophysiology field, “spontaneous activity,” has been observed already from the earliest recordings by Berger in the late twenties. The posterior 8–12 Hz oscillations were named “alpha rhythm” and it is the most prominent EEG rhythm during the awake resting state with closed eyes (Berger, 1929).

There is good evidence that spontaneous electrophysiological activity is coherently expressed in larger neuronal populations (Arieli et al., 1996, Tsodyks et al., 1999, Salek-Haddadi et al., 2003). Resting EEG activity appears to be macroscopically organized across the brain and fluctuates coherently in specific brain circuits (Laufs, 2008). This might be related to fluctuations of specific mental activities corroborating the concept of a “default mode” of brain function as proposed by imaging studies (Raichle et al., 2001). EEG is able to provide information on the subjects' vigilance and to some degree the state of mind, endowed in the frequency domain and electric field potential topography. Despite its richness of direct information about the neuronal activity, EEG is deficient with regard to the spatial localization of sources and space resolution.

The obvious solution to combine the two measurements was at first motivated by the clinical interest in mapping changes in neural activity associated with epileptic discharges onto MRI images of brain anatomy (Ives et al., 1993). Borrowing a metaphor from Laufs, epilepsy researchers adopting simultaneous multimodal experiments were forced to “marry the blind and the lame,” in order to explore neural processes which cannot be monitored or recalibrated by neurophysiological behavior such as resting brain activity (Laufs, 2008, Laufs et al., 2008). Recently the technical aspects of recording electrophysiological data simultaneously with fMRI have rapidly progressed thanks to dedicated EEG hardware and artifacts removal procedures (Laufs et al., 2008) so that it has become possible to exploit both signals to extract the best temporal and spatial resolution.

Most multimodal studies on spontaneous brain activity use regressors for the model derived from convolving the power time courses of the frequency bands of interest with a canonical hemodynamic response function. As expected, the first EEG/fMRI investigations of this kind were concerned with the BOLD correlates of alpha rhythm, which traditionally desynchronize with engagement in attention demanding tasks on the one hand and with sleepiness on the other. The majority of those experiments, in line with electrophysiological animal studies, identified thalamic BOLD activity to be positively and occipital–parietal areas to be inversely correlated with alpha oscillations (Goldman et al., 2002, Moosmann, 2003, Laufs et al., 2003a, Laufs et al., 2003b, De Munck et al., 2007). By identifying inverse relationships, those studies identified brain regions that increase their activity in the absence of marked alpha activity. The failure to identify an average cortical BOLD signal pattern which is positively correlated with alpha power (except for the thalamic activation) may be explained by the fact that fMRI group analysis fails to detect non-uniform brain activity during periods of prominent alpha oscillations (Laufs et al., 2006) suggesting that for a more comprehensive assessment of neural oscillations broader EEG spectral properties should be incorporated into the analysis procedure. Notably, performing the same type of analysis as for alpha power with frequency oscillations in the “beta” band (13–30 Hz), led to positive correlation of β-2 power (17–23 Hz) with the default mode network (Laufs et al., 2003b). Through a sophisticated approach, which combines data-driven fMRI analysis such as independent components analysis (ICA) to identify RSNs and which correlates the representative time courses to EEG bands between 1 and 50 Hz averaged across the entire scalp, Mantini et al. (2007) demonstrated that each of these networks could be generally associated with more than one specific electrical oscillation frequency band. Moreover, showing that different RSNs might exhibit positive as well negative correlation with α and β-oscillations, the work of Mantini et al. (2007) contradicts in part previous findings about the positive correlation of β-2 power band with the default mode network. So far, it appears that similar EEG frequency bands recorded at the scalp can be associated with different fMRI-generated spatial maps, while BOLD fluctuations of a distinct single network may correspond to diverse EEG patterns (Laufs et al., 2008). One potential reason for this observation may lie in the temporo-spatial properties of BOLD and surface EEG signals which are generated differently. Indeed, the EEG has been traditionally decomposed into a series of fixed broad spectral bands (delta, theta, alpha, beta, gamma) based more on history and discovery than on a theoretical framework. This way to look at the data, although computationally convenient, may obscure the fact that the sources of each of these characteristic oscillations may or may not be unique (Szava et al., 1994).

In the present study we start from the assumption that a complex spectral signature is more likely to characterize the hemodynamic fluctuations at rest (Mantini et al., 2007), and that isolating specific frequency bands might result in an over-simplification of data space. Pioneering work by Lehmann (1987) introduced a novel method to investigate the time domain of the EEG, the so-called “EEG segmentation” which focuses on transient EEG events and identifies and quantifies brain electric microstates as time epochs with quasi-stable field topography. A brain microstate is defined as a functional/physiological state of the brain during which specific neural processes occur. It is characterized uniquely by a fixed spatial distribution of active neuronal generators with time varying intensity. Brain electrical activity is modeled as being composed of a time sequence of non-overlapping microstates represented as normalized vectors constituted by scalp electric potentials with variable duration (Pascual-Marqui et al., 1995). In addition, we suggest that EEG segmentation techniques might prove particularly well suited to explore complex neuronal networks, identifying direct neuronal activity and associating it with a specific source distribution at a given time point. In the present study, we modeled the hemodynamic response function (HRF) using the microstates as regressors for the BOLD fluctuations during relaxed wakefulness.