1. Neuroscience
Download icon

Frontal beta-theta network during REM sleep

  1. Sujith Vijayan  Is a corresponding author
  2. Kyle Q Lepage
  3. Nancy J Kopell
  4. Sydney S Cash
  1. Boston University, United States
  2. Harvard Medical School and Massachusetts General Hospital, United States
Research Article
  • Cited 7
  • Views 1,520
  • Annotations
Cite as: eLife 2017;6:e18894 doi: 10.7554/eLife.18894

Abstract

We lack detailed knowledge about the spatio-temporal physiological signatures of REM sleep, especially in humans. By analyzing intracranial electrode data from humans, we demonstrate for the first time that there are prominent beta (15–35 Hz) and theta (4–8 Hz) oscillations in both the anterior cingulate cortex (ACC) and the DLPFC during REM sleep. We further show that these theta and beta activities in the ACC and the DLPFC, two relatively distant but reciprocally connected regions, are coherent. These findings suggest that, counter to current prevailing thought, the DLPFC is active during REM sleep and likely interacting with other areas. Since the DLPFC and the ACC are implicated in memory and emotional regulation, and the ACC has motor areas and is thought to be important for error detection, the dialogue between these two areas could play a role in the regulation of emotions and in procedural motor and emotional memory consolidation.

https://doi.org/10.7554/eLife.18894.001

eLife digest

Over the course of a night we cycle through several different stages of sleep. During one of these stages, our eyes move rapidly from side to side behind our closed eyelids. This movement gives this stage its name: rapid eye movement sleep, or REM sleep for short. Most other muscles are paralyzed during REM sleep, possibly to prevent us from acting out the vivid dreams that also occur during this stage of sleep. But despite the distinctive properties of REM sleep, relatively little is known about about why we need it or how the brain generates it.

Vijayan et al. have now obtained new insights into the brain activity that underlies REM sleep by recording from the brains of human patients with epilepsy. The patients all had electrodes temporarily inserted into their brains to help neurologists identify the area of the brain that was responsible for their seizures. By recording from these electrodes overnight, Vijayan et al. were able to study the activity of individual brain regions while the patients slept.

Analysis of the recordings revealed rhythmic waves of neuronal activity in areas at the front of the brain during REM sleep. Two types of brain waves dominated: theta waves, which are relatively slow waves with a frequency of 4–8 cycles per second (Hertz), and beta waves, which are faster with a frequency of 15–35 Hertz. These theta and beta waves were especially pronounced in two subregions of the frontal lobe of the brain, called the dorsolateral prefrontal cortex (DLPFC) and the anterior cingulate cortex (ACC).

The discovery of prominent rhythmic activity in the DLPFC was unexpected. This is because previous studies had shown that this region, which is involved in decision-making and planning, was relatively inactive during REM sleep. Indeed it had been suggested that the limited activity of the DLPFC subregion might be responsible for the often bizarre and illogical nature of our dreams. Instead, Vijayan et al. showed that the ACC and the DLPFC coordinate their activity during REM sleep. The next challenge is to find out whether this dual activity helps support other roles that the two regions share in common, such as the strengthening of memories and the regulation of emotions.

https://doi.org/10.7554/eLife.18894.002

Introduction

A night’s sleep consists of periods of rapid eye movement (REM) sleep and periods of non rapid eye movement (NREM) sleep. We have a relatively detailed picture of the spatio-temporal pattern of activity that occurs during NREM sleep, and this has allowed detailed hypotheses concerning the circuits and mechanisms at play during NREM sleep. For example, the transfer of memories from the hippocampus to the neocortex during NREM sleep is thought to be mediated by the coordination of oscillatory activity seen in the local field potentials, namely by the interaction of hippocampal ripples (80–250 Hz) with neocortical spindles (11–16 Hz) and slow (<1 Hz) oscillations (Battaglia et al., 2004; Ji and Wilson, 2007; Lee and Wilson, 2002; Mölle et al., 2006; Siapas and Wilson, 1998; Sirota et al., 2003; Skaggs and McNaughton, 1996; Staresina et al., 2015; Wilson and McNaughton, 1994).

The spatio-temporal patterns of activity during REM sleep are not as well characterized. This may be due in part to the fact that the activity patterns seen during REM sleep do not easily lend themselves to neat description like those seen during NREM sleep. During REM sleep field potential recordings are not dominated by prominent oscillations; rather they are irregular and of low voltage, similar to recordings during awake behavior (Rasch and Born, 2013). During awake activity, the oscillations that emerge during cognitive tasks are thought to play a key role in aspects of cognition such as working memory and communication between brain areas (Başar et al., 2001). Thus a detailed understanding of the spatio-temporal characteristics of emergent oscillations during REM sleep may shed light on which areas are interacting during REM sleep, how dreams are generated, and how some of the putative functions of REM sleep, such as memory consolidation, are carried out. Little is known about oscillatory activity during REM sleep with the notable exceptions of theta oscillations in the rodent hippocampus (Buzsáki, 2002; Louie and Wilson, 2001; Montgomery et al., 2008), gamma oscillations in the neocortex (Castro et al., 2013; Llinás and Ribary, 1993; Steriade et al., 1996), and pontine-geniculate-occipital (PGO) waves (Calvo and Fernández-Guardiola, 1984; Datta, 1997; Datta and Hobson, 1994; Steriade et al., 1989).

We decided to focus on the oscillatory dynamics of the frontal cortices during REM sleep based on imaging studies and REM sleep’s putative functional roles. Imaging studies suggest that activity patterns in certain regions of the frontal cortices are markedly different during REM sleep than during awake activity, with some areas (e.g., the anterior cingulate (ACC)) being more active and some (e.g., the dorsolateral prefrontal cortex (DLPFC)) less active (Braun et al., 1997; Maquet et al., 1996). These differences in activity patterns suggest that important REM sleep processes might be taking place in the frontal cortices (Braun et al., 1997; Maquet et al., 1996; Muzur et al., 2002). At a functional level, REM sleep has been hypothesized to play a role in such processes as the consolidation of emotional and procedural motor memories (Fischer et al., 2002; Gilson et al., 2016; Nishida et al., 2009; Nitsche et al., 2010; Rasch and Born, 2013; Smith, 2001) and the regulation of emotions. Since the frontal cortices serve memory, emotional regulation (Buhle et al., 2014; Etkin et al., 2011, 2006), and motor functions, they make good candidate areas for the mediation of procedural motor memory and emotional memory consolidation as well as emotional regulation during REM sleep.

We show that the frontal cortices are dominated by bursts of beta activity and theta oscillations, in particular in the ACC and the DLPFC. The prominent oscillatory activity in the DLPFC is especially surprising since current thinking, based on imaging studies (Braun et al., 1997; Maquet et al., 2000; Muzur et al., 2002), is that the DLPFC is relatively quiet during REM sleep in comparison with awake periods, and it has been postulated that the relative quiescence of this ‘executive’ structure might be responsible for the bizarre nature of our dreams. We also show the novel result that beta bursts and theta oscillations during REM sleep are coherent between the DLPFC and the ACC, two relatively distant structures. This finding is especially intriguing since the ACC and the DLPFC are both implicated in memory and emotional processing (Buhle et al., 2014; Etkin et al., 2011, 2006; Goldman-Rakic, 1995), and the ACC is interconnected with motor areas (Bates and Goldman-Rakic, 1993; Dum and Strick, 2002, 1991; Hatanaka et al., 2003; Morecraft and Van Hoesen, 1993, 1992; Picard and Strick, 1996), is thought to have motor regions itself (Paus, 2001), and is thought to be important in error detection (Holroyd and Yeung, 2011; Paus, 2001). Therefore the rhythmic activity we observe in the frontal cortices may play a role in some of the hypothesized functions of REM sleep such as the consolidation of emotional and procedural motor memories and the regulation of emotions.

Results

We examined overnight sleep recordings of five patients with intractable epilepsy undergoing invasive recordings to localize the focus of their seizures. Patient information is in Table 1. We used electroencephalography (EEG), electromyography (EMG), and electrooculography (EOG) to score the sleep recordings in 30 s epochs as either N1, N2, N3 (NREM stages), REM, or Awake using standard criteria (Iber et al., 2007). According to these criteria, REM periods are identified by low EMG power (the lowest of a night’s sleep), irregular EEG of low voltage, and conjugate eye movements (Figure 1a,b). Identified REM periods were selected for further analysis.

Sleep recordings and electrode localization.

(a) EMG, delta band, and spindle band power, and the cross-covariance of the EOG leads during a portion of a night’s sleep from one subject (bin size 30 s). The dotted box indicates a period of REM sleep. During this REM period, EMG, delta, and spindle band powers are all relatively low and the sums of the cross-covariance functions over the 30 s bins are negative. The EOG montages are arranged such that rapid eye movements result in voltage deflections of the opposite polarity resulting in negative cross-covariance values during rapid eye movements. (b) Voltage traces of the EMG, scalp EEG, and EOG (purple trace for left eye, red trace for right eye) during a 30 s period from the REM episode demarcated by the dotted box in (a). Note that the EMG is of relatively low voltage, the EEG is of low voltage and irregular, and there are rapid eye movements in the EOG with the deflections in the right and left EOG of opposite polarity. (c) A coronal MRI image showing the locations of the electrode contacts (red dots) of a frontal electrode in one of the subjects. (d) A sagittal MRI image showing the location of the most medial contact in (c), which is located in the anterior cingulate.

https://doi.org/10.7554/eLife.18894.003
Table 1

Patient Information. Each row provides a patient’s demographic information and diagnostic information about their epilepsy.

https://doi.org/10.7554/eLife.18894.004

Subject

Gender

Age

Handedness

Diagnosis

Imaging

1

F

45

R

Bilateral temporal lobe epilepsy

MRI-normal

PET-hypometabolism right side

2

F

55

R

Bilateral temporal lobe epilepsy

MRI-mild T2 hyperintensity and volume loss (right greater than left) in the mesial temporal lobes

3

F

45

R

Multifocal

temporal parietal-occipital epilepsy

MRI- multifocal FlAIR abnormalities - second lymphoma?

4

F

42

Ambidextrous

Multifocal epilepsy with involvement of bilateral temporal lobes

MRI- nonspecific T2 hyperintensities

PET-without clear lesion

5

M

53

Left

Focal epilepsy etiology and localization unknown

MRI- normal

PET- normal

All patients were monitored using intracranial depth electrodes whose distal-most contacts targeted the medial temporal and frontal regions (Figure 1c,d). Many of the frontal electrodes showed a marked change in the spectral content during REM sleep onset. In particular, REM sleep onset resulted in the appearance of prominent bursts of beta activity. Figure 2a depicts the spectrogram from an exemplar frontal electrode. The activity shown here and in other figures (and the activity included in all analyses) was detected by bipolar derivations from two neighboring contacts; thus this physiological activity is relatively local in nature. As REM sleep begins there is a change in the spectral content; beta bursts appear and then occur throughout the REM sleep episode. To get an impression of the global occurrence of the beta bursts we identified periods during which there was a prominent beta activity on a given electrode (three standard deviations above the median power and at least two cycles in length). This analysis suggested that the beta bursts indeed tended to occur in the frontal electrodes after onset and throughout the REM sleep episode, but not in temporal lobe electrode leads (Figure 2b). Furthermore, the beta bursts seemed to occur in close proximity to one another across the frontal electrodes.

Oscillatory activity across scales.

(a) Spectrogram shows the activity pattern detected from a bipolar derivation of an electrode located in the frontal cortices during REM sleep. The black line indicates the beginning of the REM sleep episode. (b) Pattern of beta activity across all the electrode contacts of one subject. There are two frontal depth electrodes (Frontal 1 and Frontal 2), two temporal depth electrodes (Temporal 1 and Temporal 2), and one parietal depth electrode (Parietal). Each depth electrode has eight contacts. Each row represents activity from one bipolar derivation. Blue dots indicate periods in which beta power in a given contact was 3.0 standard deviations above the median. The black line indicates the beginning of the REM sleep episode.

https://doi.org/10.7554/eLife.18894.005

To determine where exactly the beta bursts were prominent within the frontal cortices, and to determine the prominent spectral features of the frontal electrodes, we focused on specific regions within the frontal cortices. We first examined the activity of contacts in the DLPFC (See Methods for electrode localization details), an area that imaging studies suggest is relatively inactive during REM sleep relative to awake periods (Braun et al., 1997; Maquet et al., 1996; Muzur et al., 2002). Surprisingly, the DLPFC contacts showed bursts of oscillatory activity throughout each period of REM sleep, with activity occurring most prevalently in the beta and theta bands (Figure 3a,b). All of the DLPFC electrode contacts (15/15) showed significant theta activity (t test, p<0.05/15, Bonferroni corrected for number of electrodes (15), median 5.2 Hz, Figure 3c,d), while ~87% (13/15) of the electrodes showed significant beta activity (median 20.9 Hz, Figure 3c,d). The frequency range of the beta activity could be clearly distinguished from the frequency range of the spindling activity that occurs during NREM sleep (Figure 3—figure supplement 1).

Figure 3 with 2 supplements see all
Oscillatory activity in the DLPFC.

(a) Spectrogram shows the activity pattern of an electrode in the DLPFC during a period of REM sleep. The lower panel shows the voltage trace during the period demarcated by the dotted lines in the spectrogram. (b) Average power spectrum across all subjects (green trace) of electrodes located in the DLPFC. The red trace is the average of the fits of the power spectra to the model a*fb and the black traces are ± the standard error of the mean. (c) Whisker plot of the peak frequency in the beta band (top) and peak frequency in the theta band (bottom) for electrodes showing significant power in those bands. (d) Bar graph showing the fraction of total electrodes in the DLPFC with significant beta or theta power.

https://doi.org/10.7554/eLife.18894.006

We then examined electrode contacts in the ACC, which is thought to be relatively active during REM sleep. They also showed significant activity in the beta and theta bands (Figure 4a,b). All of the ACC contacts (22/22) showed significant beta activity (p<0.05/22, Bonferroni corrected for number of electrodes (22), median 19.7 Hz, Figure 4c,d) and ~86% (19/22) showed significant theta activity (median 4.6 Hz, Figure 4c,d). Significant activity in the beta and theta bands was seen not only in the DLPFC and ACC areas of the frontal cortices, but also in the inferior frontal gyrus (Figure 4—figure supplement 1).

Figure 4 with 1 supplement see all
Oscillatory activity in the ACC.

Same as Figure 3a–d, but for electrode contacts in the ACC.

https://doi.org/10.7554/eLife.18894.009

We next explored whether theta and beta activity is as prevalent outside the frontal cortices. Theta activity was prevalent in the medial temporal regions, but not beta activity. For example, in the medial temporal gyrus we found ~58% (11/19) of the electrode contacts had prominent theta activity ((p<0.05/19, Bonferroni corrected for number of electrodes (19), median 6.1 Hz, Figure 5a–c), but only ~16% (3/19) had prominent beta activity (median 17.9 Hz, Figure 5a–c).

Oscillatory Activity Outside the Frontal Cortices.

(a–c) Same as Figure 3b–d, but for electrode contacts in the middle temporal gyrus.

https://doi.org/10.7554/eLife.18894.011

We then examined whether the DLPFC and the ACC interacted, since bursts of oscillatory activity appeared to occur in concert between the two areas (Figure 6a). We found that a large number (23/31, ~74%) of ACC-DLPFC electrode pairs showed a significant coherence in the beta band (p<0.05, Bonferroni corrected for number of electrode pairs (31 and 34 for beta and theta respectively), Figure 6b,c) and 26/34 (~76%) showed a significant coherence in the theta band (Figure 6b,c). We found the same results whether we employed the confidence levels provided by Chronux or a Monte Carlo simulation method (see Methods for more details).

Simultaneous oscillatory activity in the DLPFC and ACC.

(a) Spectrograms showing the simultaneous activity patterns of an electrode in the ACC (left) and an electrode in the DLPFC (right) during a period of REM sleep. Lower panels show the voltage traces during the periods demarcated by the dotted lines in the spectrograms. (b) Coherence during REM sleep between the DLPFC electrode depicted above and the two ACC contacts on the ipsilateral side of the DLPFC contact. The coherence plot on the left is for the ACC contact depicted above. (c) Bar graph of the fraction of total ACC-DLPFC electrode pairs showing significant theta or beta coherence.

https://doi.org/10.7554/eLife.18894.012

We then attempted to determine if either the beta or the theta activity occurs first in either the DLPFC or the ACC, on average. We looked at the cross-covariance in theta power and in beta power between all electrode pairs that showed significant coherence. Both theta and beta cross-covariance functions showed a central peak. The theta peak occurred at approximately 22 ms (Figure 7a), suggesting that theta activity on average tended to occur first in the DLPFC. The peak of the beta cross-covariance function occurred very close to zero but was slightly positive (2 ms, Figure 7b), suggesting that the beta activity occurred nearly simultaneously in the DLPFC and ACC but that on average it occurred very slightly earlier in the DLPFC. To further examine the difference in the timing of the occurrence of beta and theta activity, we quantified the asymmetry in the cross-covariance function in the −100 to 100 ms window. In particular, we subtracted the area under the cross-covariance curve from −100 to 0 ms from that of 0 to 100 ms. A positive value would suggest that DLPFC activity preceded ACC activity and a negative value the opposite. We found that 69% (18/26, Figure 7c) of electrode pairs had a positive value with respect to theta, while 65% (15/23, Figure 7d) of electrode pairs had a positive value with respect to beta, suggesting that although on average theta activity and beta activity tended to occur first in the DLPFC, there was a fair amount of variability in timing among the electrodes.

ACC and DLPC oscillatory power relationships.

(a) Cross-covariance of theta power across ACC-DLPFC electrode pairs (solid black trace). The dotted red lines correspond to the standard error of the mean and the dotted black line indicates the zero lag time point. (b) Same as (a) but for beta power. (c) Histogram of the difference between the area under the theta power cross-covariance function from −100 ms to 0 ms and the area under the function from 0 ms to 100 ms. (d) Same as (c) but for the beta power cross-covariance function.

https://doi.org/10.7554/eLife.18894.013

Discussion

Here we present the novel finding of a theta and beta network in the frontal cortices of humans during REM sleep. The network includes the DLPFC, which prior studies had suggested was relatively quiescent during REM sleep, and the ACC. The bursts of theta and beta activity in the DLPFC and the ACC are coherent between these structures. We suggest that interactions between these two areas may play an important role in the function served by REM sleep.

DLPFC activity during REM sleep

Imaging studies, in particular Positron Emission Tomography (PET) studies, suggest that the DLPFC is relatively quiescent during REM sleep in comparison with quiet awake periods (Braun et al., 1997; Maquet et al., 1996; Muzur et al., 2002). The relative silence of the DLPFC, a structure thought to be important for executive functions, during REM sleep has been postulated to underlie the often bizarre and illogical nature of our dreams (Muzur et al., 2002). Our data suggest that the DLPFC is actually active during REM sleep, since it displays significant theta and beta oscillations (Figure 3a,b,d). Furthermore, the DLPFC might be communicating with other structures, since theta and beta bands in the DLPFC are coherent with those in the ACC (Figure 6a–c). The DLPFC might communicate directly with the ACC or it might be driven by a third structure that coordinates the activity of the DLPFC with other areas. A discrepancy in brain activity levels assessed by PET studies and electrophysiological techniques has also occurred with respect to sleep spindles (Dang-Vu et al., 2010). PET studies suggest that the thalamus is relatively quiet during sleep spindles (Hofle et al., 1997), while a large body of animal studies shows that spindles are generated precisely by thalamocortical interactions (McCormick and Bal, 1997; Steriade et al., 1985, 1987). One possibility is that during REM sleep the periods outside the oscillatory events may show less overall neural activity than is seen during awake periods, so that PET measurements, which average over relatively large periods of time (on the order of minutes), indicate that the DLPFC is less active during REM sleep than during quiet awake periods. Another possibility is that the oscillatory activity observed in the DLPFC during REM sleep may manifest under relatively inhibited conditions; the connectivity patterns between the DLPFC and the ACC suggest that the DLPFC may receive more inhibition from the ACC under the physiological conditions of REM sleep in comparison with awake conditions (Medalla and Barbas, 2012).

Functional implications

What functional role sleep plays has yet to be resolved. There is a large body of evidence suggesting that NREM sleep may play an important role in memory consolidation (Rasch and Born, 2013), but whether or not REM sleep also plays a role in memory consolidation remains unclear (Dudai et al., 2015; Rasch and Born, 2013). Some animal and human studies suggest that REM sleep plays a role in procedural and emotional memory consolidation (Fischer et al., 2002; Fu et al., 2007; Karni et al., 1994; Louie and Wilson, 2001; Nishida et al., 2009; Nitsche et al., 2010; Popa et al., 2010; Smith, 2001; Wagner et al., 2001), but other studies, especially human studies, do not corroborate these findings and suggest that REM sleep may be important for other cognitive functions (Cai et al., 2009; Walker et al., 2002b) such as emotional regulation (Baran et al., 2012) while stage 2 NREM sleep may be more important for the consolidation of procedural memories (Mantua et al., 2016; Tamaki et al., 2008; Walker et al., 2002a). Whether these differential findings arise because of the different natures of the tasks, differences across species, or simply because REM sleep plays little or no role in memory consolidation remains to be determined. Below we discuss our findings in light of previously known facts about the structures and oscillations at play during REM sleep. Our discussion is driven by the perspective that REM oscillatory activity may play an important role in memory, and we put forward hypotheses for how the rhythmic activity that we have observed during REM sleep could serve a role in memory consolidation, but we note that rhythmic activity during REM sleep could serve some other function or possibly no function at all.

We observed rhythmic activity in the beta and theta bands that was coherent between the ACC and the DLPFC, two relatively distant structures that are reciprocally connected and both implicated in memory (Goldman-Rakic, 1995). It is possible that this activity could be important in the procedural motor memory consolidation since the ACC is interconnected with motor areas (Bates and Goldman-Rakic, 1993; Dum and Strick, 2002, 1991; Hatanaka et al., 2003; Morecraft and Van Hoesen, 1993, 1992; Picard and Strick, 1996), is thought to have motor regions itself (Paus, 2001), and is thought to be important in error detection (Holroyd and Yeung, 2011; Paus, 2001). In humans, beta activity has been shown to follow PGO waves during REM sleep in the subthalamic nucleus of the basal ganglia, so the beta activity observed in our work could result from the transmission of the PGO-related beta activity from the basal ganglia (Amzica and Steriade, 1996; Fernández-Mendoza et al., 2009); pontine wave density has been correlated with sleep-related improvements in memory tasks and the expression of plasticity-related genes (Datta, 2000; Datta et al., 2008). This possibility is in line with our findings that the observed beta activity tends to manifest at approximately the same time in both the ACC and the DLPFC, suggesting that it may originate from a subcortical source. Such a scenario leads to the intriguing possibility that PGO waves may trigger an evaluation of a motor plan by a dialogue occurring via the beta band channel (Figure 8).

Schematic of Hypothesized Circuit Involved in REM Sleep Memory Consolidation.

PGO waves originating from the PPN may trigger beta activity in the basal ganglia. This beta activity in the basal ganglia may then be transmitted to cortical areas including the ACC and DLPFC. The triggering of the beta activity by the PGO waves may begin a dialogue between the basal ganglia, ACC, and DLPC (as well as other areas) through the beta band in which a motor plan is evaluated, ultimately leading to motor memory consolidation. Theta activity may be a channel through which the hippocampus provides contextual and spatial information about a given memory to the ACC and other parts of the frontal cortex. (PPN = Pedunculopontine tegmental nucleus and Hipp = Hippocampus).

https://doi.org/10.7554/eLife.18894.014

With respect to theta, experimental studies suggest that the hippocampus provides contextual and spatial information to the cingulate and prefrontal cortex via the theta band (Jones and Wilson, 2005; Remondes and Wilson, 2013) and disrupting hippocampal theta activity during REM sleep results in contextual memory deficits (Boyce et al., 2016). Therefore, the theta band may be a channel through which spatial content, e.g., the location at which a procedural motor task was acquired, is provided to the ACC.

With respect to emotional memories the theta band may be a channel through which the amygdala provides the emotional content or valence of a particular memory to the ACC since both human and animal studies suggest that prefrontal theta and the interaction of prefrontal theta activity with both the hippocampus and the amygdala may be important for the consolidation of emotional memories during REM sleep (Fu et al., 2007; Nishida et al., 2009; Popa et al., 2010; Rasch and Born, 2013; Wagner et al., 2001). Furthermore, emotional information from the limbic system conveyed via the theta band may bias the processing and consolidation of procedural memories with high emotional content (Popa et al., 2010). Note that many different functions have been attributed to the ACC (Paus, 2001), so its activation during REM sleep does not necessarily mean that it serves the function of emotional or procedural memory consolidation during REM sleep.

Conclusion

We generally lack detailed knowledge about the physiological characteristics of REM sleep, especially in humans. Recent intracranial studies in humans have begun to provide us with more detailed information about the physiological processes underlying REM sleep (Andrillon et al., 2015; Cantero et al., 2003). These studies have primarily focused on REM-related processes (both at the unit and LFP level) in the medial temporal lobe. We build on this knowledge base by examining the spatio-temporal characteristics of rhythmic activity in the frontal cortices. Here we have made significant inroads toward better understanding the physiological processes underlying REM sleep through the discovery of a frontal beta-theta network and we hypothesize that this network might mediate memory consolidation during REM sleep.

Materials and methods

Subjects and recordings

Data were collected from five patients (four females and one male) between the ages of 42 and 55 years (mean 48.0 ± 5.6 (SD) years) undergoing invasive monitoring using intracranial depth electrodes for intractable epilepsy (see Table 1). Each depth electrode had 6 to 8 contacts. Subject 1 had 10 depth electrodes (two anterior temporal (bilateral), two posterior temporal (bilateral), and six frontal [bilateral]), subject 2 had eight depth electrodes (two anterior temporal (bilateral), two posterior temporal (bilateral), four frontal [bilateral]), subject 3 had five depth electrodes (one anterior temporal, one posterior temporal, two frontal, and one parietal), subject 4 had 10 depth electrodes (two anterior temporal (bilateral), two posterior temporal (bilateral), and six frontal [bilateral]), and subject 5 had 10 depth electrodes (two anterior temporal (bilateral), two posterior temporal (bilateral), and six frontal [bilateral]). The placement of the electrodes was determined strictly by clinical criteria. The patients gave informed consent. The research protocol was approved by the Partners Human Research Committee.

The contacts on the depth electrodes were spaced by 10 mm and had an impedance of ~100 Ω. The electrophysiological signals were sampled at 500 Hz (four subjects) or at 512 Hz (one subject).

Electrode localization

The locations of the depth electrode contacts were determined using a preoperative high-resolution T1-weighted magnetic resonance imaging (MRI) scan and a postoperative computerized tomography (CT) scan. The preoperative MRI scan was used to generate a 3D rendering of the brain using FreeSurfer (Dale et al., 1999; Fischl et al., 1999); this rendering was placed in a 3D coordinate system (RAS). The postoperative CT scan was then coregistered with the 3D rendering, thus providing a RAS coordinate for each electrode (Dykstra et al., 2012). FreeSurfer automatically generated subcortical and cortical parcellations with anatomical labels using the Desikan-Killiany Atlas (Desikan et al., 2006; Fischl et al., 2002). Using these parcellations and the RAS coordinates, we determined the anatomical location of each electrode contact. We considered contacts labeled as being in rostral middle frontal by FreeSurfer according to the Desikan-Killiany Atlas (Desikan et al., 2006) to be in the DLPFC. We considered the inferior frontal gyrus to consist of contacts labeled as parstriangularis, parsopercularis, and parsorbitalis. Contacts labeled as middle temporal gyrus were categorized as such. The locations of the contacts were also determined by a trained neurologist. In the case of the midline contacts (in relation to this paper, the anterior cingulate contacts) there was some discrepancy between the locations of the contacts determined by the neurologist and by the FreeSurfer labeling scheme. The locations determined by the neurologist were used. We employed bipolar derivations using two abutting contacts so that we could be sure the activity being analyzed was relatively local in nature. Only electrode derivations for which both contacts had labels within the same defined region and such that at least one of the contacts was in the gray matter were used for subsequent analyses. We did not use bipolar derivations with contacts in two different regions since in these cases it was not possible to determine from which region the activity originated. In addition, contacts that showed epileptic activity were eliminated from subsequent analyses (36 bipolar derivations were eliminated by this criterion). After eliminating bipolar derivations based on these criteria, we were left with 15 bipolar derivations in the DLPFC, 22 in the ACC, 12 in the inferior frontal gyrus, and 19 in the medial temporal gyrus. In other defined regions, there were relatively few bipolar derivations that met the criteria for inclusion in our analyses, and not all subjects had bipolar derivations in these regions.

Sleep staging

Scalp electroencephalography (EEG), electromyography (EMG), and electrooculography (EOG) were used to score sleep. Sleep was first scored using custom written software (https://github.com/svijayan9/Sleep-Viewer) that incorporated the criteria specified by the American Association of Sleep Medicine for sleep scoring (Iber et al., 2007). Those periods specified as REM sleep were subsequently examined by eye and verified as REM sleep epochs by experienced sleep scorers (SSC). We also verified via video recordings that the patients were asleep during identified REM sleep epochs. The subjects spent 73.18% ± 6.82% (SEM) of the time in NREM sleep, 16.23% ± 2.85% of the time in REM sleep, and 10.59% ± 1.78% of the time awake after sleep onset (WASO), with a total sleep time of 378.13 ± 22.67 min. We used conservative measures in identifying REM periods for the purposes of further analysis. We chose relatively long (>5 min) REM periods that were not interrupted by non-REM sleep stages. Furthermore, we selected the beginning and the end of each REM period (first and last 10 s) such that there were no other stages or transition stages present. Thus, our results were not influenced by the spectral content from any other sleep stages.

Data analysis

Those electrode contacts identified to be in the epileptic focus or identified as having interictal activity by the epileptologists were excluded from further analysis. In addition, nights during which seizures occurred were excluded from further analysis. All analyses were done using a bipolar montage, specifically by subtracting adjacent contacts.

Identification of contacts with significant theta and beta power

REM episodes were divided into 10 s windows. For each electrode, a power spectrum estimate, Se,j(f), was calculated for each of the ten-second windows using the Chronux (Bokil et al., 2010) toolbox for Matlab (MathWorks). Here, e is the electrode number (e{ 1,,n }) and j specifies the trial number (j {1,…, J}). For each 10 s window, the power spectrum estimate was fit as a function of frequency, f, to the model afb, using a robust least-squares regression to obtain Se,jr(f), the robust fit for a given electrode and trial (see Figure 3—figure supplement 1). For each frequency band of interest (e.g., f  [4, 8] for theta), the mean values S¯e,j and S¯e,jrwere calculated for each 10 s window for both the spectrum estimates Se,j(f) and the fitted curves Se,jr(f). For each electrode, a t-test was used to determine whether the set of mean values of the spectrum estimates for a given band was significantly different (p<0.05, Bonferroni corrected for number of electrodes) from the set of mean values of the fitted curves. We used this method rather than a comparison with awake activity since during awake conditions the calculated power spectrum would depend on the cognitive state and the task being performed. See Figure 3—figure supplement 2 for power spectrum comparisons between NREM sleep and REM sleep, which demonstrate that the spectral content is very different between the two states and show the clear distinction between spindling activity during NREM sleep and beta activity during REM sleep. Note that all subjects had at least 21 min of REM sleep, resulting in over 125 10 s windows per electrode in each subject. Only electrode contacts showing significant power in the band of interest were used for subsequent analysis. However, this resulted in the elimination of very few electrodes. In the DLPFC no electrodes were eliminated from the theta band analyses (0/15) and two electrodes were eliminated from the beta band analyses (2/15). In the ACC no electrodes were eliminated from the beta band analyses (0/22) and three electrodes were eliminated from the theta band analyses (3/22).

Coherence estimates

We estimated coherence values using the Chronux toolbox. Coherence was examined for 34 ACC-DLPFC electrode pairs in the theta band (3 ACC electrodes were eliminated because theta power was not significant) and 31 ACC-DLPFC electrode pairs in the beta band (2 DLPFC electrodes were eliminated because beta power was not significant). We estimated the coherence during REM sleep by first dividing all REM sessions of a subject into 2 s windows. All 2 s intervals of data were included in a single coherence calculation for that subject using the coherence formula employed by Chronux (see equations 7.77 and 7.80 in Mitra and Bokil, 2008). The time-bandwidth was set such that the coherence estimate for the center band (e.g., 6 Hz for theta) integrated the entire band of interest. We assessed significance using two methods: by using the confidence level provided by Chronux and by doing Monte Carlo simulations (2000 shuffles). The confidence level provided by Chronux is based on the distribution of the statistic (i.e., the coherence measure) given the null hypothesis of zero coherence (Bokil et al., 2010; Brillinger, 2001; Jarvis and Mitra, 2001). The distribution of the estimator is given by the following formula.

p(|C|2)=1m1(1|C|2)m2

In the formula m is the degrees of freedom, which is equal to the number of 2 s intervals multiplied by the number of data tapers used in the calculation. See Brillinger (2001) for details regarding the asymptotic convergence of the coherence estimator under the hypothesis of zero coherence. For the Monte Carlo simulations, a shuffle was done by permuting the 2 s windows of one of the electrode pairs and then calculating the coherence. Using the distribution of coherences from the shuffles we determined the probability of getting the coherence value we got using our data. Both methods identified the same set of DLPFC-ACC electrode derivation pairs as having significant (p<0.05, Bonferroni corrected for number of electrode pairs) beta or theta coherence.

Power Cross-covariance estimates

The cross-covariance of the power was calculated by first filtering the raw signal in the band of interest. The instantaneous amplitude was then calculated using the Hilbert transform of the filtered signal. Using these calculated instantaneous amplitudes the cross-covariance of the power (amplitude) of an ACC-DLPFC electrode pair was calculated for each two-second REM sleep window. For a given electrode the cross-covariance functions for all two-second windows were averaged to obtain an average cross-covariance function.

Only those REM windows in which both electrodes had power in the top quartile in the band of interest with respect to the REM episode were used for the analysis. We also restricted our analysis to those electrode pairs that showed a significant coherence in the band of interest.

Finally the average cross-covariance functions for all electrodes were averaged to obtain an overall average cross-covariance function. The max lag reported is simply the time lag at the maximum value of the overall average cross-covariance function. We used 23 ACC-DLPFC bipolar electrode pair derivations for the beta band cross-covariance analysis and 26 ACC-DLPFC bipolar electrode pair derivations for the theta band cross-covariance analysis (we omitted eight electrode pairs from the theta and beta analyses).

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
    Gamma, alpha, delta, and theta oscillations govern cognitive processes
    1. E Başar
    2. C Başar-Eroglu
    3. S Karakaş
    4. M Schürmann
    (2001)
    International Journal of Psychophysiology : Official Journal of the International Organization of Psychophysiology 39:241–248.
    https://doi.org/10.1016/s0167-8760(00)00145-8
  7. 7
  8. 8
  9. 9
  10. 10
    Time Series: Data Analysis and Theory
    1. DR Brillinger
    (2001)
    SIAM, Time Series: Data Analysis and Theory.
  11. 11
  12. 12
  13. 13
  14. 14
    Phasic activity of the basolateral amygdala, cingulate gyrus, and hippocampus during REM sleep in the cat
    1. JM Calvo
    2. A Fernández-Guardiola
    (1984)
    Sleep 7:202–210.
  15. 15
    Sleep-dependent theta oscillations in the human hippocampus and neocortex
    1. JL Cantero
    2. M Atienza
    3. R Stickgold
    4. MJ Kahana
    5. JR Madsen
    6. B Kocsis
    (2003)
    The Journal of Neuroscience : The Official Journal of the Society for Neuroscience 23:10897–10903.
  16. 16
  17. 17
  18. 18
    Functional neuroimaging insights into the physiology of human sleep
    1. TT Dang-Vu
    2. M Schabus
    3. M Desseilles
    4. V Sterpenich
    5. M Bonjean
    6. P Maquet
    (2010)
    Sleep 33:1589–1603.
  19. 19
    Neuronal activity in the caudolateral peribrachial pons: relationship to PGO waves and rapid eye movements
    1. S Datta
    2. JA Hobson
    (1994)
    Journal of Neurophysiology 71:95–109.
  20. 20
    Cellular basis of pontine ponto-geniculo-occipital wave generation and modulation
    1. S Datta
    (1997)
    Cellular and Molecular Neurobiology 17:341–365.
  21. 21
    Avoidance task training potentiates phasic pontine-wave density in the rat: A mechanism for sleep-dependent plasticity
    1. S Datta
    (2000)
    The Journal of Neuroscience : The Official Journal of the Society for Neuroscience 20:8607–8613.
  22. 22
  23. 23
  24. 24
  25. 25
    The origin of corticospinal projections from the premotor areas in the frontal lobe
    1. RP Dum
    2. PL Strick
    (1991)
    The Journal of Neuroscience : The Official Journal of the Society for Neuroscience 11:667–689.
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
    Evidence of subthalamic PGO-like waves during REM sleep in humans: a deep brain polysomnographic study
    1. J Fernández-Mendoza
    2. B Lozano
    3. F Seijo
    4. E Santamarta-Liébana
    5. MJ Ramos-Platón
    6. A Vela-Bueno
    7. F Fernández-González
    (2009)
    Sleep 32:1117–1126.
  31. 31
  32. 32
  33. 33
    Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
    1. B Fischl
    2. DH Salat
    3. E Busa
    4. M Albert
    5. M Dieterich
    6. C Haselgrove
    7. A van der Kouwe
    8. R Killiany
    9. D Kennedy
    10. S Klaveness
    11. A Montillo
    12. N Makris
    13. B Rosen
    14. AM Dale
    (2002)
    Neuron 33:341–355.
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
    Regional cerebral blood flow changes as a function of delta and spindle activity during slow wave sleep in humans
    1. N Hofle
    2. T Paus
    3. D Reutens
    4. P Fiset
    5. J Gotman
    6. AC Evans
    7. BE Jones
    (1997)
    The Journal of Neuroscience : The Official Journal of the Society for Neuroscience 17:4800–4808.
  39. 39
  40. 40
    The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specifications
    1. C Iber
    2. S Ancoli-Israel
    3. AL Chesson
    4. SF Quan
    (2007)
    Westchester, IL: American Association of Sleep Medicine.
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
    Observed brain dynamics
    1. P Mitra
    2. H Bokil
    (2008)
    Oxford University Press.
  54. 54
  55. 55
  56. 56
  57. 57
    The prefrontal cortex in sleep
    1. A Muzur
    2. EF Pace-Schott
    3. JA Hobson
    (2002)
    Trends in cognitive sciences 6:475–481.
  58. 58
  59. 59
  60. 60
  61. 61
  62. 62
  63. 63
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
  69. 69
  70. 70
  71. 71
    Abolition of spindle oscillations in thalamic neurons disconnected from nucleus reticularis thalami
    1. M Steriade
    2. M Deschênes
    3. L Domich
    4. C Mulle
    (1985)
    Journal of Neurophysiology 54:1473–1497.
  72. 72
    The deafferented reticular thalamic nucleus generates spindle rhythmicity
    1. M Steriade
    2. L Domich
    3. G Oakson
    4. M Deschênes
    (1987)
    Journal of Neurophysiology 57:260–273.
  73. 73
    Phasic activation of lateral geniculate and perigeniculate thalamic neurons during sleep with ponto-geniculo-occipital waves
    1. M Steriade
    2. D Paré
    3. D Bouhassira
    4. M Deschênes
    5. G Oakson
    (1989)
    The Journal of neuroscience : the official journal of the Society for Neuroscience 9:2215–2229.
  74. 74
    Synchronization of fast (30-40 Hz) spontaneous cortical rhythms during brain activation
    1. M Steriade
    2. F Amzica
    3. D Contreras
    (1996)
    The Journal of neuroscience : the official journal of the Society for Neuroscience 16:392–417.
  75. 75
    Fast sleep spindle (13-15 hz) activity correlates with sleep-dependent improvement in visuomotor performance
    1. M Tamaki
    2. T Matsuoka
    3. H Nittono
    4. T Hori
    5. T Masako
    6. T Hori
    (2008)
    Sleep 31:204–211.
  76. 76
  77. 77
    Practice with sleep makes perfect: sleep-dependent motor skill learning
    1. MP Walker
    2. T Brakefield
    3. A Morgan
    4. JA Hobson
    5. R Stickgold
    (2002)
    Neuron, 35, 12123620.
  78. 78
  79. 79

Decision letter

  1. Takeo Watanabe
    Reviewing Editor; Brown

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Frontal Βeta-Theta Network during REM Sleep" for consideration by eLife. Your article has been reviewed by three peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by Sabine Kastner as the Senior Editor.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

This manuscript presents a study in which intracranial recordings of REM sleep were conducted in humans (epileptic patients). Based on neuroimaging studies, the authors have thought of the ACC and DLPFC as regions of specific interest, which may be inversely active to each other (ACC high, DLPFC low) during REM. However, the results of the present study suggest that both areas are active, with significant beta and theta activity that occurs coherently between the regions. The authors suggest that the results indicate a neurological basis for a proposed function of sleep in memory consolidation.

All reviewers acknowledge the importance of spatio-temporal analyses of spontaneous oscillations during REM sleep in human. However, a multitude of analyses and statistical procedures are unclear.

Essential revisions:

1) All reviewers have agreed that significantly better clarification of analyses and control analyses are necessary.

For example:

“With regards to the data analysis, it's not clear how conclusions such as "showed significant theta activity" are drawn. It seems that frequency analyses were done in either 10- or 30-sec bins. Either way, a bin may not be 100% REM. Presumably the 'edges' on the sleep stage would have some other sleep (and frequencies). And, what make it 'significant'? A t-test is used (as indicated by the reported statistic) but this is not in the Data Analysis section”.

“It was not clear how many out of all channels were analyzed and where these channels were placed (scalp locations). "Frontal cortices" usually indicates a huge extent and thus ambiguous. How specific could the ACC and DLPFC be in comparison to other frontal regions (superior frontal cortex, orbitofrontal cortex etc.)? Is it the case that the ACC and DLPFC were more active compared to other frontal cortices? Or is it the case that the ACC and DLPFC were as active as other frontal cortices but more active than other non-frontal cortices? It is stated that Bonferroni correction was made based on the number of electrodes. However, the number of electrodes is not reported”.

“The authors limited their analysis to a potentially quite restricted subsample of the data. First, data from frontal electrodes were reduced to electrodes in the ACC and DLPFC, because beta and theta activity was most prominent in these regions. From these electrodes, only those were used for coherence analysis that showed a significant power increase in the frequency bands of interest. From these electrode pairs, only those that showed significant coherence were used for the subsequent power cross-covariance analysis, which was moreover only calculated for time windows with power in the top 25% within each REM episode. All of this data-driven sub-selection potentially bias the results towards positive findings. As an example, using only electrode pairs with high coherence for power cross-covariance can be problematic since these two measures are not independent (Lachaux, Rodriguez et al. 1999, Srinath and Ray 2014). Furthermore, it is not clear how much data (sleep epochs, electrodes, time windows, electrode pairs) were discarded in absolute numbers in each step (only the ratios of electrodes used based on power and coherence are given in Figures 36). Please report the number of recorded electrodes per subject including their location. Please give an idea on how many data were finally used at each analysis step. Coherence was tested for statistically significant differences using Monte Carlo simulations as well as the "confidence level provided by Chronux" (subsection “Coherence Estimates”). What is this confidence level based on? Apparently, both methods yielded "the same results" (Paragraph six Results section). Please elaborate, are the results indeed fully identical? How did the authors deal with the influence of the number of trials for the estimation of coherence (e.g. Maris, Schoffelen, Fries 2007). Was the number of observations equalized across REM sleep periods? In the subsection “Coherence Estimates”, the calculation of coherence reads as if the coherence between two electrodes was calculated for each window and then was averaged. However, since calculating coherence for a single trial (here: 2-s time window) will always result in a coherence of 1 (as well as their average), I assume that the cross-spectra were calculated for each time window and then used to calculate coherence (not "average coherence") across time windows. Please clarify the procedure”.

Please state in significantly clearer way how you conducted the analyses, what was compared, or how the statistical significance was obtained. Bin length, definition of regions, data selection (if any), procedures of a coherence analysis, cross-covariance analysis, and data visualization (Figures 26 each show spectral data from only one bipolar derivation or electrode pair. How representative are these data? Is there a way to present data aggregated from more than one electrode / subject?), etc. are unclear. In addition, please clarify whether you compared REM sleep vs. NREM sleep or Wake stage.

2) All reviewers agreed that discussion about REM sleep and memory consolidation are too far-reaching based on the present data, particularly because it is not shown that the obtained effects are specific to REM sleep.

“The function of REM in procedural memory consolidation is questionable as the result is not found in many studies. Recent studies seem to suggest that REMs role in emotional memory consolidation is limited (e.g., see Payne et al., 2015) and perhaps reflects emotional processing without consolidation (see Baran et al., 2012). Keeping in mind that the present data contain no cognitive measures, it cannot be stated "These findings, in conjunction with previous work on REM sleep, suggest a system for the consolidation of memories during REM sleep." It seems that much of the discussion is focused on just this point – the REM role in consolidation – which I see no clear support for in the present data. Sure they support some cognitive role of REM but is it necessarily emotional or procedural memory consolidation? That link is absent”.

“The authors speculate that activation of the ACC during REM sleep relates to consolidation motor learning. However, there are many functions in which the ACC is suggested to play roles. Mere activation of the ACC during REM sleep may leave other possibilities than its involvement in consolidation of motor learning”.

Please revise the relevant statements so that the manuscript would not go too far beyond the current data.

https://doi.org/10.7554/eLife.18894.015

Author response

Essential revisions:

1) All reviewers have agreed that significantly better clarification of analyses and control analyses are necessary.

We have substantially modified the Methods and Results sections to make our approaches to analyses clearer. We have added much more detail on electrode selection, electrode localization, statistical measures for frequency, coherence and cross-covariance estimates, sleep staging, determination of statistical significance, and more. The changes to the manuscript are detailed below in response to specific reviewer queries.

For example:

“With regards to the data analysis, it's not clear how conclusions such as "showed significant theta activity" are drawn. It seems that frequency analyses were done in either 10- or 30-sec bins. Either way, a bin may not be 100% REM. Presumably the 'edges' on the sleep stage would have some other sleep (and frequencies). And, what make it 'significant'? A t-test is used (as indicated by the reported statistic) but this is not in the Data Analysis section”.

As stated in the section entitled Determination of Significant Theta and Β Activity, power spectra were calculated in 10-second bins. For each calculated power spectrum a 1/f fit was determined using a robust least-squares regression (see Figure 3—figure supplement 1). We then calculated the mean power for each power spectrum and each fit in the band of interest. We then used a t-test to determine if the set of means of the power spectra differed significantly from the set of means of the fits in the bands of interest. An electrode is said to show significant theta activity if the means of the power spectra in the theta band were significantly different from the means of the fits in the theta band as assessed by a t-test (P<0.05, Bonferroni corrected for number of electrodes).

We used conservative measures in identifying REM periods. We chose relatively long REM periods that were not interrupted by non-REM sleep stages. Furthermore, we selected the beginning and the end of each REM period (first and last 10 seconds) such that there were no other stages or transition stages present. Thus our results were not influenced by spectral content from any other sleep stages.

To address the reviewer's concerns, we have added the relevant details to the section of the Methods entitled Contacts with Significant Theta and Β Power (see the section entitled Determination of Significant Theta and Β Activity for details). To further address the reviewer’s concerns we have added details to the section of the Methods entitled Sleep Staging. The relevant text is pasted below.

We used conservative measures in identifying REM periods for the purposes of further analysis. We chose relatively long (>5 minutes) REM periods that were not interrupted by non-REM sleep stages. Furthermore, we selected the beginning and the end of each REM period (first and last 10 seconds) such that there were no other stages or transition stages present. Thus our results were not influenced by spectral content from any other sleep stages.”

“It was not clear how many out of all channels were analyzed and where these channels were placed (scalp locations). "Frontal cortices" usually indicates a huge extent and thus ambiguous. How specific could the ACC and DLPFC be in comparison to other frontal regions (superior frontal cortex, orbitofrontal cortex etc.)? Is it the case that the ACC and DLPFC were more active compared to other frontal cortices? Or is it the case that the ACC and DLPFC were as active as other frontal cortices but more active than other non-frontal cortices? It is stated that Bonferroni correction was made based on the number of electrodes. However, the number of electrodes is not reported”.

Identification of Electrode Localization

Subjects were implanted with depth electrodes. Each depth electrode had 6 to 8 contacts. Subject 1 had 10 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), and 6 frontal (bilateral)), subject 2 had 8 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), 4 frontal (bilateral)), subject 3 had 5 depth electrodes (1 anterior temporal, 1 posterior temporal, 2 frontal, and 1 parietal), subject 4 had 10 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), and 6 frontal (bilateral)), subject 5 had 10 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), and 6 frontal (bilateral)).

Electrode locations were determined by co-registering the postoperative CT scan with the preoperative MRI scan (see Figure 1C, D in the manuscript for an example) using techniques previously developed in the lab (Dykstra et al., 2012). In brief, FreeSurfer (Dale et al., 1999; Fischl et al., 1999) was used to create a 3-D rendering of the brain using the MRI scan and place the rendering in a 3-D coordinate system, thus providing a 3-D coordinate for each electrode contact. FreeSurfer was also used to automatically parcellate the cortical and subcortical structures and provide anatomical labels using the Deskian-Killany atlas; from Figure 1 in Deskian et al. (2006). The 3-D coordinate assigned to each electrode contact was used to assign a unique anatomical label to each contact.

We employed bipolar derivations using two adjacent contacts so that we could be sure the activity being analyzed was relatively local in nature. Only electrode derivations for which both contacts had labels within the same defined region and such that at least one of the contacts was in the gray matter were used for subsequent analyses. We did not use bipolar derivations with contacts in two different regions since in these cases it was not possible to determine from which region the activity originated. In addition, contacts that showed epileptic activity were eliminated from analyses (36 bipolar derivations were eliminated by this criterion). After eliminating bipolar derivations based on these criteria, we were left with 15 bipolar derivations in the DLPFC, 22 in the ACC, 12 in the inferior frontal gyrus, and 19 in the medial temporal gyrus. In other defined regions there were relatively few bipolar derivations that met the criteria for inclusion in our analyses; furthermore, not all of the subjects had bipolar derivations in these regions. For instance, other frontal areas (e.g. orbitofrontal cortex) had very few bipolar electrode derivations that satisfied our criteria and each of these areas had electrodes in only 2 or fewer out of 5 subjects.

In this study, scalp electrodes were used only for sleep staging. All other analyses utilized depth electrodes, which do permit anatomical localization using the methods outlined above.

To address the reviewer's concerns, a paragraph has been added to the methods section entitled Subjects and Recordings. The relevant text is pasted below.

“Each depth electrode had 6 to 8 contacts. Subject 1 had 10 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), and 6 frontal (bilateral)), subject 2 had 8 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), 4 frontal (bilateral)), subject 3 had 5 depth electrodes (1 anterior temporal, 1 posterior temporal, 2 frontal, and 1 parietal), subject 4 had 10 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), and 6 frontal (bilateral)), and subject 5 had 10 depth electrodes (2 anterior temporal (bilateral), 2 posterior temporal (bilateral), and 6 frontal (bilateral)).”

To address the reviewer's concerns, a paragraph has also been added to the methods section entitled Electrode Localization. The relevant text is pasted below.

“We employed bipolar derivations using two abutting contacts so that we could be sure the activity being analyzed was relatively local in nature. Only electrode derivations for which both contacts had labels within the same defined region and such that at least one of the contacts was in the gray matter were used for subsequent analyses. We did not use bipolar derivations with contacts in two different regions since in these cases it was not possible to determine from which region the activity originated. In addition, contacts that showed epileptic activity were eliminated from subsequent analyses (36 bipolar derivations were eliminated by this criterion). After eliminating bipolar derivations based on these criteria, we were left with 15 bipolar derivations in the DLPFC, 22 in the ACC, 12 in the inferior frontal gyrus, and 19 in the medial temporal gyrus. In other defined regions there were relatively few bipolar derivations that met the criteria for inclusion in our analyses, and not all subjects had bipolar derivations in these regions.”

Determination of Significant Theta and Beta Activity

For each region of interest (e.g., the DLPFC) all bipolar electrode derivations in the region were tested for significant theta and beta activity. Identified REM periods were divided into 10-second windows (i.e., the bin size was 10s). The power spectrum was calculated for each 10-second window (Figure 3—figure supplement 1) and a 1/f fit was calculated for each calculated power spectrum using a robust least-squares regression (Figure 3—figure supplement 1). To determine if there was significant power in a given band (e.g., theta) we took the mean power within the given band for each spectrum estimate and each fit, for each 10 -second bin. This gave a set of paired values corresponding to the mean powers (for the given band) for the spectrum estimates and for the fitted curves of the various 10- second bins. For each bipolar derivation, a t-test (P<.05, Bonferroni corrected for number of bipolar derivations) was used to determine whether there was significant power in the band of interest.

We used this method rather than a comparison with awake activity since during awake conditions the calculated power spectrum would depend on the cognitive state and the task being performed. We have included in the figure supplements power spectrum comparisons between NREM sleep and REM sleep (Figure 3—figure supplement 2) to demonstrate that the spectral content is very different between the two states and to show the clear distinction between spindling activity (11-16 Hz) during NREM sleep and beta activity during REM sleep.

We found there to be significant beta and theta activity in the ACC and the DLPFC (see Figures 3 and 4 in the manuscript). The presence of rhythmic activity in the DLPFC is surprising since imaging studies suggest that the DLPFC is relatively inactive during REM sleep; the relative inactivity of the DLPFC, a structure thought to be important for executive functions, has been argued to be the cause of the illogical nature of our dreams. The high-frequency activity generated in the DLPFC presumably requires a fair amount of metabolic energy. Furthermore, the ACC and the DLPFC, two relatively distant structures, were coherent in the theta and beta bands, suggesting that the DLPFC may in fact be interacting with other structures during REM sleep. There was also significant theta and beta activity in the inferior frontal gyrus (see Figure 4 supplement in manuscript). There were not enough bipolar derivations to determine the nature of the rhythmic activity present in other frontal structures; note that the placement of the electrodes was determined strictly by clinical criteria. The frontal structures that were examined exhibited more high frequency activity (i.e., beta) than did temporal areas (see Figures 2 and 5 in the manuscript). The difference in power between the various frontal structures cannot be determined using our data since power depends on the geometric arrangement of the neuronal process relative to a given electrode as well as the proximity to the electrode.

For subsequent analyses we used only electrodes that showed significant power in the band relevant to the analysis. However, this resulted in the elimination of very few electrodes. In the DLPFC no electrodes were eliminated from the theta band analyses (0/15) and 2 electrodes were eliminated from the beta band analyses (2/15). In the ACC no electrodes were eliminated from the beta band analyses (0/22) and 3 electrodes were eliminated from the theta band analyses (3/22).

We have now included the number of electrodes used for each of the relevant analyses in the manuscript.

To further address the reviewer's concerns, the following two paragraphs have been added to the section of the Methods entitled Identification of Contacts with Significant Theta and Beta Power. The relevant text is pasted below.

“We used this method rather than a comparison with awake activity since during awake conditions the calculated power spectrum would depend on the cognitive state and the task being performed. See Figure 3—figure supplement 2 for power spectrum comparisons between NREM sleep and REM sleep, which demonstrate that the spectral content is very different between the two states and show the clear distinction between spindling activity during NREM sleep and beta activity during REM sleep.”

“The authors limited their analysis to a potentially quite restricted subsample of the data. First, data from frontal electrodes were reduced to electrodes in the ACC and DLPFC, because beta and theta activity was most prominent in these regions. From these electrodes, only those were used for coherence analysis that showed a significant power increase in the frequency bands of interest. From these electrode pairs, only those that showed significant coherence were used for the subsequent power cross-covariance analysis, which was moreover only calculated for time windows with power in the top 25% within each REM episode. All of this data-driven sub-selection potentially bias the results towards positive findings. As an example, using only electrode pairs with high coherence for power cross-covariance can be problematic since these two measures are not independent (Lachaux, Rodriguez et al. 1999, Srinath and Ray 2014). Furthermore, it is not clear how much data (sleep epochs, electrodes, time windows, electrode pairs) were discarded in absolute numbers in each step (only the ratios of electrodes used based on power and coherence are given in Figures 36). Please report the number of recorded electrodes per subject including their location. Please give an idea on how many data were finally used at each analysis step.

We have now added details about the selection of data to be included in the analyses. The selection of the particular frontal areas used in the analyses was based on anatomical criteria and the requirement that each area has a sufficient number of electrodes across all subjects. As stated, since we used bipolar derivations using neighboring electrodes, we required that both electrodes be from the same anatomically defined region and that at least one of the electrode contacts be in the grey matter. These selection criteria resulted in 15 bipolar derivations in the DLPFC, 22 in the ACC, and 12 in the inferior frontal gyrus. These selection criteria allowed us to have confidence that the activity of a given derivation was local in nature and that it originated from the region of interest; if two contacts had been from two different regions it would not have been possible to determine the regional origin of the activity.

The reviewer correctly points out that we eliminated electrodes that did not show significant power in the theta or beta bands for subsequent analyses. However, this resulted in the elimination of relatively few electrodes. In the DLPFC no electrodes were eliminated from the theta band analyses (0/15) and 2 electrodes were eliminated from the beta band analyses (2/15). In the ACC no electrodes were eliminated from the beta band analyses (0/22) and 3 electrodes were eliminated from the theta band analyses (3/22). Therefore, for the coherence analysis we examined 34 ACC-DLPFC electrode pairs in the theta band (3 ACC electrodes eliminated) and 31 ACC-DLPFC electrode pairs in the beta band (2 DLPFC electrodes eliminated).

Cross-Covariance Calculations

Since the coherence measure indicated that there was a relationship between the time series of the electrode pairs, we were interested in determining whether on average a rhythm originated in one region or the other when there was ongoing rhythmic activity in both bipolar derivations. In other words, the primary motivation was to determine the time lag, not the amplitude, of the relationship. Since we were interested in electrode pairs whose time series had a temporal relationship, we restricted our analysis to pairs of electrodes whose coherence measure was significant. Also, to get a better estimate of the lag we restricted the analysis to periods when the power was high and employed the cross-covariance method as described in (Adhikari et al., 2010). We used 23 ACC-DLPFC bipolar electrode pair derivations for the beta band cross-covariance analysis and 26 ACC-DLPFC bipolar electrode pair derivations for the theta band cross-covariance analysis (we omitted 8 electrode pairs from the theta and beta analyses).

We calculated cross-covariance by first filtering the raw signal in the band of interest during the REM episodes and five minutes before and after the REM episodes to avoid edge effects. The instantaneous amplitude was then calculated using the Hilbert transform of the filtered signal; following this step we used the data from only the REM period. Using the calculated instantaneous amplitude the cross-covariance of the power (amplitude) of an ACC-DLPFC electrode pair was calculated for each two- second REM sleep window. For a given electrode the cross-covariance functions for all two-second windows were averaged to obtain an average cross-covariance function (Adhikari et al., 2010).

To address the reviewer’s concerns we have added details about the number of electrodes used in each step. We have also added the following sentence to the section entitled Power Cross-Covariance Estimates (in the Materials and,methods section).

We used 23 ACC-DLPFC bipolar electrode pair derivations for the beta band cross- covariance analysis and 26 ACC-DLPFC bipolar electrode pair derivations for the theta band cross-covariance analysis (we omitted 8 electrode pairs from the theta and beta analyses).

Activity Across Subjects

The examples of spectral data that we included in the manuscript are representative. For each example spectrogram from a single subject, we also included the average power spectrum (averaged across all subjects). Note that these average power spectra show significant power in the same bands in which the example spectrogram shows power. For example, the example spectrogram in Figure 3a in the manuscript shows activity in the theta and beta bands, and the average power spectrum in Figure 3b in the manuscript also shows power in the theta and beta bands. To give a better sense of the distribution of the power across the subjects and to address the reviewer’s concerns we have now added lines showing +/- the standard error for the power spectrum across the subjects. In Figure 3b we have added traces showing +/- the standard error for the power spectrum across the subjects. The included standard error traces for the ACC, DLPFC, and Middle Temporal Gyrus indicate that the observed beta and theta activity is consistent across the subjects. Note that in the instance of the inferior frontal gyrus, for which we provide data in the figure supplements, the standard error traces indicate that theta and beta activities are variable across the subjects.

Coherence was tested for statistically significant differences using Monte Carlo simulations as well as the "confidence level provided by Chronux" (subsection “Coherence Estimates”). What is this confidence level based on? Apparently, both methods yielded "the same results" (Paragraph six Results section). Please elaborate, are the results indeed fully identical?

Coherence Calculations II

We determined significance using two different methods. One method was to use the confidence level provided by Chronux. The confidence level is based on the distribution of the statistic (i.e., the coherence measure) given the null hypothesis of zero coherence (Bokil et al., 2010; Brillinger, 2001; Jarvis and Mitra, 2001). The distribution of the estimator is given by the following formula.

In the formula m is the degrees of freedom, which is equal to the number of 2-second intervals multiplied by the number of data tapers used in the calculation. See Brillinger (2001) for details regarding the asymptotic convergence of the coherence estimator under the hypothesis of zero coherence. The other method was via Monte Carlo simulations (2,000 shuffles). A shuffle was done by permuting the 2-second windows of one of the electrode pairs and then calculating the coherence. Using the distribution of coherences from the shuffles we determined the probability of getting the coherence value we got using our data. Both of these methods gave identical results in the sense that they identified the same set of electrodes as having significant beta activity and the same set of electrodes as having significant theta activity.

To address the reviewer's concerns and make these details clearer, we have amended the text in the Materials and methods section entitled Coherence Estimates. The relevant text is pasted below.

We assessed significance using two methods: by using the confidence level provided by Chronux and by doing Monte Carlo simulations (2,000 shuffles). The confidence level provided by Chronux is based on distribution of the statistic (i.e., the coherence measure) given the null hypothesis of zero coherence (Bokil et al., 2010; Brillinger, 2001; Jarvis and Mitra, 2001). The distribution of the estimator is given by the following formula.

In the formula m is the degrees of freedom, which is equal to the number of 2-second intervals multiplied by the number of data tapers used in the calculation. See Brillinger (2001) for details regarding the asymptotic convergence of the coherence estimator under the hypothesis of zero coherence. For the Monte Carlo simulations, a shuffle was done by permuting the 2-second windows of one of the electrode pairs and then calculating the coherence. Using the distribution of coherences from the shuffles we determined the probability of getting the coherence value we got using our data. Both methods identified the same set of DLPFC-ACC electrode derivation pairs as having significant (p <.05, Bonferroni corrected for number of electrode pairs) beta or theta coherence.”

How did the authors deal with the influence of the number of trials for the estimation of coherence (e.g. Maris, Schoffelen, Fries 2007). Was the number of observations equalized across REM sleep periods?

Note that we did not calculate any coherence differences as in (Maris et al., 2007). For a given electrode pair in a given subject we divided all the REM episodes into 2-second bins and then made a single coherence calculation. Because we were interested not in differences but rather in the significance relative to a null hypothesis we did not need to adjust for the number of observations (the number of observations is taken into account by the confidence level). We have clarified this point by amendments to the text, which are detailed in the responses to 1c-i and 1c-ii.

In the subsection “Coherence Estimates”, the calculation of coherence reads as if the coherence between two electrodes was calculated for each window and then was averaged. However, since calculating coherence for a single trial (here: 2-s time window) will always result in a coherence of 1 (as well as their average), I assume that the cross-spectra were calculated for each time window and then used to calculate coherence (not "average coherence") across time windows. Please clarify the procedure”.

As the reviewer points out our description of the coherence calculation needs clarification.

Coherence Calculations I

Coherence calculations were made using the Chronux toolbox. Coherence was examined for 34 ACC-DLPFC bipolar electrode derivations in the theta band (3 ACC electrodes were eliminated because theta power was not significant) and 31 ACC- DLPFC electrode pairs in the beta band (2 DLPFC electrodes were eliminated because beta power was not significant). For each subject REM episodes were divided into 2- second windows. All of the 2-second windows were used to make a single coherence calculation for that subject using the coherence formula employed by Chronux (see equations 7.77 and 7.80 in Mitra and Bokil (2008)). For each frequency band, the time- bandwidth was set such that the coherence estimates for the center band (e.g., 6 Hz for theta) integrated the entire band of interest.

To address the reviewer's concerns, we have modified the Methods section entitled Coherence Estimates. The relevant text is pasted below.

“Coherence was examined for 34 ACC-DLPFC electrode pairs in the theta band (3 ACC electrodes were eliminated because theta power was not significant) and 31 ACC- DLPFC electrode pairs in the beta band (2 DLPFC electrodes were eliminated because beta power was not significant). We estimated the coherence during REM sleep by first dividing all REM sessions of a subject into 2-second windows. All 2-second intervals of data were included in a single coherence calculation for that subject using the coherence formula employed by Chronux (see equations 7.77 and 7.80 in Mitra and Bokil (2008)). The time-bandwidth was set such that the coherence estimate for the center band (e.g., 6 Hz for theta) integrated the entire band of interest.“

Please state in significantly clearer way how you conducted the analyses, what was compared, or how the statistical significance was obtained. Bin length, definition of regions, data selection (if any), procedures of a coherence analysis, cross-covariance analysis, and data visualization (Figures 26 each show spectral data from only one bipolar derivation or electrode pair. How representative are these data? Is there a way to present data aggregated from more than one electrode / subject?), etc. are unclear. In addition, please clarify whether you compared REM sleep vs. NREM sleep or Wake stage.

We have now included bin lengths and data selection criteria in each of the relevant sections. Details about the definition of regions can be found in the response titled Identification of Electrode Localization. Details about coherence analyses can be found in the responses sections titled Coherence Calculations I and Coherence Calculations II. Details about cross-covariance analyses can be found in the response section titled Cross-Covariance Calculations.

The data are representative across the subjects; we have modified some of the figures to give a better sense of the consistency across the subjects. Please see the section entitled Activity Across Subjects for more details. We have also provided details about the comparisons that were made and how significances were obtained; these details are in multiple sections, but most notably in the response section titled Determination of Significant Theta and Β Activity.

2) All reviewers agreed that discussion about REM sleep and memory consolidation are too far-reaching based on the present data, particularly because it is not shown that the obtained effects are specific to REM sleep.

Please revise the relevant statements so that the manuscript would not go too far beyond the current data.

We have substantially modified and reduced our discussion of REM sleep and memory. We have made it clear that our claims are hypotheses and we have added the necessary caveats. Since the functional role of REM sleep remains largely a mystery, we feel that some hypotheses about the putative functional role of the rhythmic activity we observed are warranted in the discussion to motivate future studies. Of course, we are open to further modifications if the changes we have made are unsatisfactory. We have rewritten this section and pasted it below in response to one of the reviewer comments.

“The function of REM in procedural memory consolidation is questionable as the result is not found in many studies. Recent studies seem to suggest that REMs role in emotional memory consolidation is limited (e.g., see Payne et al., 2015) and perhaps reflects emotional processing without consolidation (see Baran et al., 2012). Keeping in mind that the present data contain no cognitive measures, it cannot be stated "These findings, in conjunction with previous work on REM sleep, suggest a system for the consolidation of memories during REM sleep." It seems that much of the discussion is focused on just this point – the REM role in consolidation – which I see no clear support for in the present data. Sure they support some cognitive role of REM but is it necessarily emotional or procedural memory consolidation? That link is absent”.

To address the reviewer’s concerns we have substantially modified the Discussion section and any claims about the functional role of REM sleep in the manuscript. The new Discussion section is pasted below.

What functional role sleep plays has yet to be resolved. There is a large body of evidence suggesting that NREM sleep may play an important role in memory consolidation (Rasch and Born, 2013), but whether or not REM sleep also plays a role in memory consolidation remains unclear (Dudai et al., 2015; Rasch and Born, 2013).[…] Furthermore, emotional information from the limbic system conveyed via the theta band may bias the processing and consolidation of procedural memories with high emotional content (Popa et al., 2010). Note that many different functions have been attributed to the ACC (Paus, 2001), so its activation during REM sleep does not necessarily mean that it serves the function of emotional or procedural memory consolidation during REM sleep.”

“The authors speculate that activation of the ACC during REM sleep relates to consolidation motor learning. However, there are many functions in which the ACC is suggested to play roles. Mere activation of the ACC during REM sleep may leave other possibilities than its involvement in consolidation of motor learning”.

As the reviewer points out activation of the ACC during REM sleep could serve some other role than the consolidation of motor learning.

To address the reviewer's concerns and make these details clearer, we have added the following to the text in the Discussion section entitled Functional Implications. The relevant text is pasted below.

“Note that many different functions have been attributed to the ACC (Paus, 2001), so its activation during REM sleep does not necessarily mean that it serves the function of emotional or procedural memory consolidation during REM sleep.”

https://doi.org/10.7554/eLife.18894.016

Article and author information

Author details

  1. Sujith Vijayan

    1. Department of Mathematics and Statistics, Boston University, Boston, United States
    2. Department of Neurology, Harvard Medical School and Massachusetts General Hospital, Boston, United States
    Contribution
    SV, Conceptualization, Formal analysis, Investigation, Methodology, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    svijayan9@gmail.com
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon 0000-0001-6385-3922
  2. Kyle Q Lepage

    Department of Mathematics and Statistics, Boston University, Boston, United States
    Contribution
    KQL, Formal analysis
    Competing interests
    The authors declare that no competing interests exist.
  3. Nancy J Kopell

    Department of Mathematics and Statistics, Boston University, Boston, United States
    Contribution
    NJK, Funding acquisition, Writing—review and editing
    Competing interests
    The authors declare that no competing interests exist.
  4. Sydney S Cash

    Department of Neurology, Harvard Medical School and Massachusetts General Hospital, Boston, United States
    Contribution
    SSC, Funding acquisition, Writing—review and editing
    Competing interests
    The authors declare that no competing interests exist.

Funding

National Science Foundation (DMS-1042134)

  • Sujith Vijayan
  • Kyle Q Lepage
  • Nancy J Kopell

Office of Naval Research (N00014-13-1-0672)

  • Sydney S Cash

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

SV, KL, and NK acknowledge support from the National Science Foundation (NSF) Grant DMS-1042134. SSC received support for this work from ONR N00014-13-1-0672 and an MGH ECOR Research Scholars award. The authors thank Anna Vijayan, Giovanni Piantoni, Jung H Lee, Benjamin Pittman-Polletta, Michelle M McCarthy, Jean-Baptiste Eichenlaub, Salva Ardid, Jason Sherfey, and David Stanley for their helpful comments on this work. The authors would also like to thank Brandon Lockyer for helpful input on sleep staging.

Ethics

Human subjects: All patients gave informed consent. The research protocol was approved by the Partners Human Research Committee. IRB number 2007P000165.

Reviewing Editor

  1. Takeo Watanabe, Brown

Publication history

  1. Received: June 16, 2016
  2. Accepted: December 26, 2016
  3. Version of Record published: January 25, 2017 (version 1)

Copyright

© 2017, Vijayan et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,520
    Page views
  • 279
    Downloads
  • 7
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, Scopus, PubMed Central.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Jimmie M Gmaz et al.
    Research Article
    1. Computational and Systems Biology
    2. Neuroscience
    Christopher M Kim, Carson C Chow
    Research Article