SLEEP AND THE FUNCTIONAL CONNECTOME

INTRODUCTION

A major goal of recent efforts to map the human functional connectome (Biswal et al., 2010) is to chart circuits that underlie the brain’s functions. These efforts have been facilitated by the use of functional MRI (fMRI), which allows the study of large-scale circuits with full-brain coverage, provides a spatial resolution of a few millimeters, and provides a temporal resolution of a few seconds. Many of the recent fMRI-based studies performed in the context of the functional connectome have relied on the notion that human subjects not involved with any particular task or behavior display a remarkable amount of brain activity that involves most brain regions (Biswal et al., 2010). Although the function of this activity is still incompletely understood, it appears to reflect important aspects of the brain’s circuitry. Thus fMRI during rest may provide information that is complementary to information from studies performed during specific tasks or behaviors.

Interestingly, much of this spontaneous fMRI activity continues during conditions (e.g., anesthesia and sleep) when interaction with the environment and voluntary mentation are reduced (Boveroux et al., 2010; Fukunaga et al., 2006; Vincent et al., 2007). Understanding the nature of the spontaneous fMRI activity is complicated by the fact that both sleep and spontaneous fMRI activity are far from well understood. Nevertheless, because cortical circuits that differ from wakefulness to sleep and between substates of sleep have been identified with both PET and fMRI, the study of fMRI activity during sleep may allow the identification of circuits that support the unique functions and behaviors that are associated with sleep and thus improve our understanding of sleep. These circuits can be divided into those that regulate sleep in a bottom-up manner versus those that regulate sleep in a top-down manner. The former could be said to regulate sleep behavior in general and require nuclei in the brainstem/hypothalamus and basal forebrain that project to the cortex, whereas the latter may be more important for the use-dependent functions of sleep that originate in the cortex (Krueger et al., 2008). fMRI may not be optimal to map circuits that involve subcortical nuclei, but it may be well-suited to measure cortical circuits associated with use-dependent activity during wakefulness.

Use-dependent theories of sleep function as well as other theoretical and methodological expositions on the functions of sleep are often limited to mechanistic functions. This means that they do not necessarily relate brain activity during sleep to a specific feature of cognition that changes during a subsequent period of wakefulness and that is evolutionarily adaptive. Recent work suggests that sleep indeed has a number of adaptive functions that support waking activity in this way (e.g., memory consolidation; (Walker and Stickgold, 2004), emotion regulation (Stickgold, 2002; Walker, 2009)). Thus the study of sleep may elucidate the circuits and processes that support waking function and complement information obtained from fMRI during waking conditions. In this review, we will discuss examples of this for memory, arousal, and consciousness after providing a brief background on sleep and on studying it with fMRI. Although other methods can be used to measure connectivity during sleep (e.g., magnetoencephalography), we will focus on fMRI to provide a more concise and clearly circumscribed review.

BRAIN CIRCUITS THAT SUPPORT SLEEP

The identification of different brain states is accomplished with polysomnography, which is the combined use electroencephalography (EEG), electro-oculography, and electromyography. The different brain states are classified as wakefulness, REM sleep, and non-REM sleep, and non-REM sleep is further divided according to depth of sleep (i.e., stages N1–N3 (AASM, 2007) or stages S1–S4 (Rechtschaffen and Kales, 1968)). Non-REM sleep is also divided in light sleep (including N1 (or S1)), N2 (or S2), and deep sleep, also called slow wave sleep (SWS) and includes N3 (and S3–S4) (see Table). Sleep onset is thought to result from an increased sleep drive mediated by homeostatic and circadian factors. The homeostatic factors are reflected in ATP depletion and adenosine accumulation, which eventually lead to reduced acetylcholine release, for example, by the main wake-promoting system know as the ascending reticular activating system (ARAS) (Moruzzi and Magoun, 1949). Cholinergic activity in this system is also responsible for REM sleep. Animal studies have shown that sleep control may also be effectuated by inhibitory inputs to wake-promoting systems such as the ARAS by the ventro-lateral preoptic area (McGinty and Szymusiak, 2000; Saper et al., 2001).

Another important region involved in the modulatory control of brain activity during sleep is the thalamus, which serves as the gatekeeper for sensory input to the cortex and which has extensive, bidirectional connections with the cortex through thalamocortical loops. The thalamus contains nonspecific nuclei that relay signals from the ARAS to the entire neocortex (see Figure). The thalamus is thus well-suited to mediate neocortical arousal throughout the sleep-wake cycle. In fact, many studies have demonstrated thalamic involvement in the electrophysiological manifestations of sleep including K-complexes, spindles, and slow oscillations (see review by (Llinas and Steriade, 2006)).

During sleep, the brain remains metabolically and electrically active. In part, this activity serves to regulate sleep. This includes circadian cycling from wake to sleep and ultradian cycling between rapid eye movement (REM) and non-REM sleep. Importantly, we know of several brain mechanisms that explain the relatively rapid transitions (seconds to minutes) that occur between these stages. However, when considering these mechanisms, it is important to bear in mind that other unknown repetitive patterns may exist during sleep beyond the traditional REM/non-REM patterns.

A primary mechanism that supports sleep regulation is neuromodulation by regions such as the hypothalamus (e.g., the ventro-lateral preoptic nucleus), pons, and basal forebrain. These brain regions are involved in modulating excitability and activity in downstream regions through neurotransmitters such as GABA, acetylcholine, hypocretin/orexin, norepinephrine, and serotonin. Such neurotransmitters can be released by modulatory neurons and remotely alter the configuration of and activity in downstream micro- and macro-circuits (Marder, 2012). The various basal forebrain and brainstem nuclei that have been identified to play a role in the regulation of sleep interact in a complicated fashion and generally involve multiple neurotransmitters (Saper et al., 2005). Distinct mechanisms have been proposed that facilitate the transitions between wake and non-REM sleep and between non-REM and REM sleep (McCarley and Hobson, 1975; Saper et al., 2010; Saper et al., 2005). Generally, the transitions between wakefulness and sleep are controlled via hypothalamic modulation of brainstem regions, whereas ultradian cycles are controlled via interaction of brainstem reticular regions.

In terms of the mnemonic functions of sleep, some have suggested that different sleep stages support different memory systems (Plihal and Born, 1997; Smith, 2001), whereas others have suggested that the cycling itself is important so that different sleep stages support different memory processes (Giuditta et al., 1995). For example, non-REM might weaken superfluous memories through global synaptic downscaling (Tononi and Cirelli, 2006), and REM sleep may strengthen the remaining memories through ponto-geniculo-occipital (PGO) wave-based reactivation of allocortical and neocortical regions (Mavanji and Datta, 2003). In addition, sleep features other than PGO- waves and slow waves may be relevant for memory consolidation (e.g., sleep spindles and hippocampal sharp-wave ripples).

EEG-FMRI OF SLEEP

The gold standard for defining sleep is a set of behavioral characteristics that include an increased arousal threshold (Campbell and Tobler, 1984; Moruzzi, 1969; Piéron, 1913). However, in practice, the EEG has become a widely accepted surrogate for the behavioral definitions because of the strong correlation between EEG slow waves and arousal threshold (e.g., (Blake and Gerard, 1937)). Therefore, the study of network activity specific to sleep is approached with the combined use of electroencephalography (EEG) scalp recordings and neuroimaging methods such as Positron Emission Tomography (PET) and fMRI (for review see (Duyn, 2012)). Conjoint analysis of EEG and positron emission tomography (PET) or fMRI overcomes some of the drawbacks that each method has when used in isolation. For example, many sleep-specific neuro-electrical phenomena such as sleep, K-complexes, and slow waves can be detected easily and with temporal precision (ms) with EEG. Currently, this is not possible with fMRI alone. On the other hand, a drawback of EEG is its poor (cm) spatial resolution and the fact that it only captures a small fraction of the broad repertoire of electrical activity generated in the brain. For example, EEG is relatively sensitive to superficial sources (close to the scalp) of synchronized electrical activity in large neuronal ensembles with isotropically organized dendrites but much less so for other types of sources. FMRI, on the other hand, has better spatial resolution (mm) and can detect deep sources, but it has a poor temporal resolution (seconds) and only indirectly detects neuronal activity through blood flow and BOLD effects. Like EEG, fMRI does not detect all types of activity and is likely not equally sensitive in all brain regions (Logothetis, 2008). In summary, combined EEG-fMRI may allow the characterization of sleep-specific brain activity with better sensitivity and precision than each modality alone.

Conjoint analysis of EEG-fMRI data can proceed in various ways (Duyn, 2012). In order to study differences in brain activity between the different stages of sleep, EEG can be used simply to detect or stage sleep, after which, fMRI can be analyzed for correlational structure or fluctuation level in a stage-specific manner (e.g., (Horovitz et al., 2009; Horovitz et al., 2008; Spoormaker et al., 2010)). Alternatively, fMRI correlates of specific EEG features can be analyzed. For this purpose, power in specific EEG frequency bands or specific microarchitecture features such as K-complexes or spindles can be extracted and regressed against the fMRI signal (e.g., (Schabus et al., 2007)).

In general, it has been found that the correlation between EEG and fMRI signals is generally low (<0.5), and this may not be surprising given the two modalities measure different aspects of neuronal activity. Interpretation of EEG-fMRI correlations should be done with caution and should take into consideration the fact that the underlying signals originate from the simultaneous activity of large neuronal ensembles and have inadequate sensitivity and temporal and spatial resolution to capture neuronal spiking in local circuits. In particular, when applied to the study of sleep, connectivity measures derived from EEG-fMRI may emphasize the relatively slow modulatory effects of neuronal activity and excitability that are prevalent in sleep (Olbrich et al., 2011). Another limitation that applies to functional connectivity studies in general is the risk of false negatives, i.e. the absence of signal correlation does not negate the presence of a network connection because it may simply be caused by down regulation of activity.

FMRI CHANGES WITH SLEEP

Although the application of fMRI to sleep is a relatively novel research area that is still in its infancy, initial studies are providing a wealth of information about sleep networks and their various functions that support waking behavior. A brief overview of these findings will be presented in this section, whereas specific examples of network functions that undergo changes during sleep, namely memory, consciousness, and arousal, are discussed in following sections (see Figure).

Most fMRI studies have focused on non-REM sleep and a general finding has been a preservation of fMRI signal fluctuations and correlations across the brain (e.g. (Horovitz et al., 2008)). Earlier PET work found a reduction of CBF with the deepening of sleep in most brain areas (Braun et al., 1997; Maquet, 2000), with the exception of unimodal sensory areas where rCBF was preserved. The reticular formation and its dorsal and ventral projections have reduced rCBF during sleep compared to wake (Braun et al., 1997). Within these projections, the hypothalamus and the heteromodal cortices have a strong functional connectivity during sleep and an absence of this connectivity during wakefulness (Kaufmann et al., 2006). These apparent discrepancies between modalities may result from the fact that rCBF (measured with PET) and its fluctuation (measured with fMRI) may provide divergent information. Nevertheless, the continuous presence of the fMRI signal and its correlations during sleep supports the existence of brain activity that does not require attention, alertness, or voluntary mentation.

Studies that have more specifically examined the effect of sleep depth (judged from EEG features such as the level of theta activity (4–8Hz) or slow-wave activity (SWA) (< 4.5Hz)) found a preservation of corticocortical connectivity during light sleep (Horovitz et al., 2008; Larson-Prior et al., 2009; Spoormaker et al., 2010) but a decrease in such connectivity during deep sleep (Horovitz et al., 2009; Larson-Prior et al., 2011; Samann et al., 2011; Spoormaker et al., 2011; Spoormaker et al., 2012; Wu et al., 2012), an observation that is consistent with EEG studies of connectivity and cortical excitability (Boly et al., 2012a; De Gennaro et al., 2001; Massimini et al., 2010). In particular, neuroimaging studies have reported reductions in corticocortical connectivity with sleep in fronto-parietal networks (Braun et al., 1997; Hofle et al., 1997; Horovitz et al., 2009; Kaufmann et al., 2006; Larson-Prior et al., 2011; Samann et al., 2011; Spoormaker et al., 2012). The results from PET studies indicate that these networks exhibit a reduction in CBF (Braun et al., 1997; Hofle et al., 1997), thus they may support waking functions such as attention, memory, and consciousness (see following sections). An interesting, but poorly understood, finding in this regard has been a decrease in the anti-correlation during light sleep between the default mode network (DMN) on one hand and the attention (DAN) and executive control (ECN) networks on the other (Larson-Prior et al., 2011; Samann et al., 2011).

During REM sleep, PET has shown rCBF increases in brainstem, thalamus, and extra-striate occipital cortex, an observation that is consistent with a network that supports the PGO waves that are characteristic of this sleep stage (Braun et al., 1998). In addition, rCBF increases have been observed in hippocampal and amygdalar regions. These increases may play a role in memory consolidation and/or emotion regulation (Braun et al., 1998; Hong et al., 2009). The few fMRI studies that have been performed so far have been largely consistent with this. Specifically, fMRI studies have shown a correlation between level and onset of REM with signal in the thalamus, occipital cortex, and pons (Miyauchi et al., 2009; Wehrle et al., 2005). Exploiting the superior temporal resolution of fMRI relative to PET, instances of rapid eye movements were found to correlate with signal in the amygdala, hippocampus, cingulate, and sensory cortices; this is consistent with the PET results and further points to their possible involvement in memory consolidation (Wehrle et al., 2007).

Sleep-specific changes in fMRI activity have also been studied by correlating it with sleep-specific EEG features. In fact, most EEG patterns used to define sleep stage (e.g., power in alpha, delta, and sub-delta bands, K-complexes, spindles, see Table) have been studied in combination with fMRI (Andrade et al., 2011; Bergmann et al., 2012; Dang-Vu et al., 2005; Hofle et al., 1997; Jahnke et al., 2012; Larson-Prior et al., 2011; Laufs, 2008; Laufs et al., 2006; Olbrich et al., 2009; Picchioni et al., 2011; Picchioni et al., 2009; Schabus et al., 2007; Schabus et al., 2012) and, prior to that, with PET (e.g. (Dang-Vu et al., 2005; Hofle et al., 1997).

Quiet waking is characterized by alpha activity. An increase in alpha activity has been interpreted as a disconnection from the external world (Hobson and Pace-Schott, 2002) and correlates with nodes of the DMN (Larson-Prior et al., 2011). Decreases in alpha could indicate falling asleep (N1), when accompanied by an increase in theta activity, or becoming more alert and engaged with the external world, when accompanied by faster EEG activity (beta and gamma) (see Table). In the descent to sleep, the correlation between alpha activity and the DAN decreases (Larson-Prior et al., 2011). Over longer time periods, alpha activity also correlates with the anterior insula; this could be related to the subject trying to avoid falling asleep and remain awake, or it could be a reflection of intermittent transitions between wake and N1 (Larson-Prior et al., 2011).

The EEG landmarks of N2 are spindles and K-complexes (Table). PET studies show that spindle incidence negatively correlates with thalamic CBF (Dang-Vu et al., 2010; Hofle et al., 1997). fMRI studies provide better temporal discrimination and detect two networks: one involving paralimbic structures (anterior cigulate cortex and insula) and neocortical areas (superior temporal gyrus) and another involving motor networks (Schabus et al., 2007). K-complexes correlate with BOLD signal levels in thalamus, brainstem, and sensorimotor areas and anti-correlate with signal in anterior insula (Jahnke et al., 2012). These two networks are thought to support the dual function of K-complexes in the promotion of sleep and arousal (Colrain, 2005).

N3 is characterized by increased SWA. PET studies show a global decreases in rCBF with increase in SWA (Dang-Vu et al., 2005; Dang-Vu et al., 2008; Hofle et al., 1997). This is consistent with the findings from the comparison of sleep stages reported above. It has been suggested (Dang-Vu et al., 2010) that these changes may be driven by the intermittent neuronal hyperpolarizations characteristic of the so called “down” portion of the “up–down” state cycling that is characteristic of N3 (Steriade and Timofeev, 2003) the better temporal discrimination of fMRI, this phasic activity produced increases in the BOLD signal in a variety of areas, including medial frontal gyrus, parahippocampal gyrus, precuneus, posterior cingulate cortex (PCC), ponto-mesencephalic tegmentum, and cerebellum (Dang-Vu et al., 2008). The correlated activity in the pons includes nucleus coeruleus, an area originally associated with wake but seen active in synchrony with slow waves in animal models (Steriade, 2005). It should be noted that these phasic changes occur against a background of reduced baseline activity in N3 compared to wakefulness and therefore may be of a different nature than phasic changes that occur during wakefulness.

SLEEP AND MEMORY

SLEEP AND AROUSAL

SLEEP AND CONSCIOUSNESS

Consciousness can be thought of as having a level and a content, a viewpoint which can be paralleled with the distinct concepts of arousal and awareness (Laureys, 2005; Zeman, 2001). Both decrease during the progressive stages of non-REM sleep and may be even further reduced in disorders of consciousness such as coma (Laureys, 2005; Zeman, 2001). Arousal level (i.e., wakefulness) and awareness increase or decrease together with few exceptions. During REM the content of consciousness (awareness) is high but the level of consciousness (arousal level) is low. Patients in a vegetative state can be awake and can have their eyes open but their content of consciousness is low (Laureys, 2005).

The natural variations in awareness and arousal that occur during sleep have been studied to investigate the major networks that support these processes (see review by (Heine et al., 2012)). In the following subsections, we will first describe these networks and then describe how they change during sleep.

SUMMARY AND OUTLOOK

Sleep and the functional connectome are research areas with considerable overlap. Neuroimaging studies of sleep based on EEG-PET and EEG-fMRI are revealing the brain networks that support sleep, as well networks that may support the roles and processes attributed to sleep. For example, phenomena such as arousal and consciousness are substantially modulated during sleep, and one would expect this modulation to be reflected in altered network activity. Thus, characterizing brain activity changes across the sleep-wake cycle may provide a fruitful basis for the mapping of networks underlying these phenomena. It should be realized, however, that the neuromodulatory changes during sleep may not only alter network configurations but also may change the level and location of network input. This complicates the interpretation of fMRI correlations in terms of connectivity in cortical networks.

A number of recent studies have started to explore the presence of transient network activity underlying spontaneous fMRI signal fluctuations (Liu and Duyn, 2013; Smith et al., 2012; Tagliazucchi et al., 2013) based on the notion that much of the brain’s activity may not obey the stationarity of network properties assumed by conventional analysis method such as independent component analysis and seed-based correlation. Such novel analysis approaches may be particularly useful for the study of sleep, when many transient brain interactions occur (e.g., those related to EEG features such as K-complexes, spindles, and slow waves). A particular example is the intriguing possibility of studying the hippocampal-cortical interactions that have been suggested to underlie memory consolidation, which is one of the functions attributed to sleep. In this and other networks, the study of transient network activity based on fMRI data alone may provide information complementary to that obtained from correlating EEG features with fMRI signals (Liu and Duyn, 2013).

This review has discussed several examples in which roles attributed to sleep are reflected in network connectivity. It should be realized that some of these roles are not exclusively supported by sleep. For example, some evidence exists that memory consolidation, can occur during quiet waking or rest (Carr et al., 2011). In addition, there is some preliminary evidence that neocortico-hippocampal connectivity during resting-state fMRI has relevance to subsequent memory retrieval (Tambini et al., 2010). This is an intriguing area that is still relatively unexplored.

It is also important to consider that it is difficult to draw causal conclusions from many of these connectivity studies because of their reliance on correlational measures. Improving the research in this area would necessitate the use of more sophisticated statistical techniques such as dynamic causal modeling and effective connectivity analysis (for review see (Duyn, 2012)).