Beyond spindles: interactions between sleep spindles and boundary frequencies during cued reactivation of motor memory representations

Abstract

There is now ample evidence that sleep spindles play a critical role in the consolidation of newly acquired motor sequences. Previous studies have also revealed that the interplay between different types of sleep oscillations (e.g. spindles, slow waves, sharp-wave ripples) promotes the consolidation process of declarative memories. Yet the functional contribution of this type of frequency-specific interactions to motor memory consolidation remains unknown. Thus, this study sought to investigate whether spindle oscillations are associated with low- or high-frequency activity at the regional (local) and interregional (connectivity) levels. Using an olfactory-targeted memory reactivation paradigm paired to a motor sequence learning task, we compared the effect of cuing (Cond) to no-cuing (NoCond) on frequency interactions during sleep spindles. Time–frequency decomposition analyses revealed that cuing induced significant differential and localized changes in delta (1–4 Hz) and theta (4–8 Hz) frequencies before, during, and after spindles, as well as changes in high-beta (20–30 Hz) during the spindle oscillation. Finally, coherence analyses yielded significant increases in connectivity during sleep spindles in both theta and sigma (11–17 Hz) bands in the cued group only. These results support the notion that the synchrony between spindle and associated low- or high-frequency rhythmic activity is an integral part of the memory reactivation process. Furthermore, they highlight the importance of not only measuring spindles’ characteristics, but to investigate such oscillations in both time and frequency domains when assessing memory consolidation-related changes.

Statement of Significance

Sleep spindles are thought to take part in the consolidation of newly learned memories. Also, declarative memory studies suggest that spindles interact with other sleep oscillations to promote consolidation. Yet, the functional contribution of frequency-specific interactions to motor memory consolidation remains unknown. This study provides the first evidence suggesting that the interaction between spindles and sleep-associated frequency bands may be linked to the reactivation of motor memory traces. Additionally, we report that interactions are found locally and between cortical regions. Thus, our results stress the importance of not only measuring spindles’ characteristics, but to investigate spindles within a broader time and frequency context when looking at their role in memory consolidation.

Introduction

The seemingly effortless manner in which we successfully produce complex tasks, such as walking, driving or typing on a keyboard, involve prior training of simple, stereotyped and sequential repetitions of movements—a form of procedural memory called motor sequence learning (MSL) [1, 2]. A large body of work has shown that memory traces associated with this type of learning are transformed and strengthened over time into an enduring and more accessible state. This memory consolidation process unfolds through a complex series of interactions between multiple cerebral structures [3–8], and relies upon mechanisms occurring during sleep (non-rapid eye movement [NREM] sleep, in particular), depending on the cognitive processes needed to perform the learned motor sequence task [9–11].

Numerous studies have investigated the nature of the changes in electrophysiological events that are specific to certain stages of sleep (e.g. sleep spindles, slow waves, ponto-geniculo-occipital waves) and that are associated with procedural memory consolidation [12–14]. Researchers have consistently reported that certain characteristics of sleep spindles (e.g. amplitude, density, duration) are modulated following MSL, and that sleep spindles correlate with gains in performance following training on a MSL task [13, 15–19]. Spindles are synchronous thalamocortical events specific to NREM sleep, which oscillates in the sigma band (~11–17 Hz), particularly abundant in NREM2 compared with slow wave sleep (SWS; stage 3 and 4 sleep, also known as NREM3). In support to this spindle/memory consolidation functional link, a recent targeted memory reactivation (TMR) study from our group, in which participants were exposed during NREM2 sleep to an olfactory stimulus previously associated to a MSL task, has confirmed the active involvement of sleep spindles in motor sequence memory consolidation [20]. More specifically, we found that parietal spindle events were significantly involved in the consolidation of a novel motor sequence memory trace.

However, there is evidence to suggest that lasting neuroplastic changes involved in memory consolidation are not only dependent upon sleep spindle activity, but upon brain functional dynamics involving several other types of oscillations as well [14, 21, 22]. In an influential model, called the active system consolidation model, sleep spindles are described as a single element in a complex homeostatic process that interact with other types of oscillations such as slow oscillations (SO) and sharp-wave ripples (SWR) during the consolidation process [23–28]. First, previous studies showed that non-procedural learning produces an increase in sleep spindle activity during SO up-states – corresponding to the depolarization phase of the participating neurons [29]. Still, it is not clear whether these learning-related changes in sleep spindles, time-locked to slow wave activity (SWA), are also manifest during other forms of memory, such as MSL consolidation. Second, several studies have also revealed evidence of temporal coupling between hippocampal SWR and spindles, and have proposed that such synchronous oscillations may constitute the underlying neuronal mechanism involved in information transfer and memory consolidation [30–32]. More recently, neocortical recordings in rodents [33] have revealed the presence of gamma activity synchronized to spindle oscillations. It has been hypothesized that this high-frequency activity may be involved in the reactivation of memory traces during sleep. There is also limited evidence of similar synchronization between spindles and higher frequency activity in humans [32]. Yet no studies have looked at inter-frequency dynamics, for example beta time-locked to spindles, in relation to memory consolidation. Thus, while previous human studies investigating sleep spindles mainly focused on the sigma band, it remains unclear whether task-related changes outside the spindle frequency range are involved in motor memory consolidation, and whether these boundary frequencies are coordinated with spindles. Finally, other studies looking at changes in cerebral connectivity, time-locked to spindle activity in animals [34] and humans [35–39] have described the cortico-thalamic and cortico-cortical interactions between different regions during sleep spindle events. MSL-related topographic connectivity changes occurring during spindles, however, remain to be described in humans.

Accordingly, considering that sleep spindles are part of a larger information transfer process observable in the frequency and time domains, it is thus critical to investigate the dynamical changes in related oscillations through spindles and their associated connectivity. Here, we examined the changes across frequency bands, from delta to beta, time-locked to spindle events, and observed spatial connectivity in relation to NREM2 parietal sleep spindles in an olfactory TMR paradigm design. This paradigm allowed us to investigate task-dependent changes in cortical activity associated with the reactivation of a newly acquired memory trace. Participants were exposed during NREM2 sleep to a rose-like odor, which was (or not) previously associated to performing the MSL task (Figure 1). The following morning, participants were then retested on the same task in order to evaluate the offline performance gains between both sessions, as an index of behavioral MSL consolidation. We hypothesized that cuing and reactivation of the motor memory trace would induce local spectral power increases in the delta band, time-locked to sleep spindles, as well as increases in sigma during spindles. Additionally, we expected that connectivity would significantly increase during sleep spindles in the sigma band in the cued group only.

Figure 1.

Experimental design. (A) Overview of the experimental design. Subjects participated in a TMR protocol, which consisted of the association (or not) of an olfactory stimulus during training on a motor sequence task. During subsequent sleep in the second half of the night, subjects were then reexposed to the associated olfactory cue. The effects of the TMR manipulation were then assessed by comparing the subjects’ performance between the training and retest sessions. (B) Experimental groups. Subjects were randomly assigned to one of two groups: The “Cond” group was first exposed to the odor during the evening MSL training session, and reexposed to the same stimulus during NREM2 sleep stage. By contrast, the “NoCond” group was not exposed to the odor while training, but received olfactory stimulation during NREM2 sleep. All groups were retested on the MSL task the next morning. (C) Sleep periods. The “stim” period comprised segments of sleep during which the odor was represented to the participants. By contrast, the “pre-stim” period consisted of NREM2 sleep episodes that occurred before the onset of the olfactory cuing and that were length-matched to the stim periods.

Materials and Methods

Participants

Participants in this study constituted a subset (N = 49; see subsection Experimental session for details) of the groups of subjects from our previous TMR experiment [20], in which we investigated the role of sleep stages and spindles in sequence motor learning consolidation. A third group, which was stimulated during rapid eye movement (REM) sleep in our previous work, was not included in this study for lack of spindles in this sleep stage.

Pre-selection and screening session

Prior to a screening night at our sleep laboratory, participants were selected following a series of strict inclusion criteria. Eligible participants had to be right-handed, between 20 and 35 years of age and free of any history of neurological, psychological, psychiatric, or sleep disorders. Subjects with previous formal training in playing a musical instrument, or any training as a professional typist were excluded in order to control for preexisting experience in tasks requiring highly coordinated finger movements. Obese individuals (body mass index > 30) and, those using nicotine regularly or users of recreational drugs were also excluded. Furthermore, individuals who worked night shifts, were engaged in trans-meridian trips in the 3 months prior to the study, or reported taking three or more servings of caffeinated beverages per day, were not included in the study. All eligible participants had to have a score lower than 10 on the Beck Anxiety Inventory [40] and the short version of the Beck Depression Inventory [41]. Sleep quality was also assessed with the Pittsburgh Sleep Quality Index questionnaire [42]. Participants who met the initial eligibility criteria underwent overnight polysomnographic (PSG) screening in the sleep laboratory according to the American Academy of Sleep Medicine guidelines [43], as described in our first study [20]. Upon completion of the screening session, participants were required to wear an actigraph (Actiwatch 2, Phillips Respironics) to the left wrist in order to monitor their sleep/wake cycles prior to the experimental session.

Experimental session

From the selection process, a total of 58 participants were retained to complete this study. From this group, nine were discarded from the analyses for the following reasons: two were excluded due to poor sleep efficiency (SE < 75%), one as a result of MSL performance identified as outlier and six were discarded due to poor-quality electroencephalography (EEG) recordings. Hence, 49 participants were included for subsequent analyses: 25 subjects were assigned to the Cond group (experimental group conditioned to the olfactory stimulus and cued during NREM2 sleep; mean age: 24.8 ± 5.0 years, 11 females) and 24 to the NoCond group (experimental group not previously conditioned to the olfactory stimulus but exposed to the odor during NREM2 sleep; mean age: 24.9 ± 4.0 years, 11 females). Due to the use of a different spindle detection algorithm (see section EEG pre-processing/Sleep spindle detection and selection), the current subset was slightly different than the one used in our previous study [20]. This detection tool allowed for the analyses of several participants (N = 9) that were initially rejected, while others (N = 3) could not be analyzed due to poor recordings quality (Cond: N = +4; NoCond: N = +5, −3). In view of these change within the subset, overnight MSL changes, sleep architecture difference and sleep spindles characteristics between groups and conditions were re-analyzed.

Overall experimental design and procedure

About a week following the screening process, participants meeting all the inclusion criteria were invited again to the sleep laboratory for the experimental session. After installation of the EEG electrodes and the olfactory delivery apparatus, subjects were assigned randomly to one of the two experimental groups (Cond, NoCond; see Figure 1B).

Prior to training on the MSL task, participants were asked to complete the Standford Sleepiness Scale questionnaire [44] to measure their subjective levels of alertness/sleepiness. Around 10:30 pm, subjects were then trained on the motor task during which they were exposed, or not, to a rose-like odor through a nasal cannula. Following completion of the MSL task, participants were instructed to get into bed and prepare for sleep. Overnight PSG recordings started from lights out to lights on for a total maximum duration of about 8 hours. After 4 hours of recordings, participants were then exposed to the rose-like olfactory stimulus during NREM2 sleep periods for a maximum of 60 min. When stimulation was completed, participants were allowed to complete their 8-hour night of sleep. Upon waking, a period of 2 hours preceded the retest session to ensure dissipation of sleep inertia.

Experimental groups

Two groups took part in this study (Figure 1B). The Cond group was exposed to the rose-like odor during the training session, and reexposed for a maximum of 60 min during NREM2 sleep comprised in the second half of the night. By contrast, the NoCond group was not exposed to the odor during training on the MSL task, but was exposed to the same olfactory stimulus during NREM2 sleep akin to the Cond group. Having not been exposed to the odor during training, the NoCond group thus allowed us to control for potential unknown effects of post-training and sleep-related olfactory stimulation. Furthermore, it served as a control condition to test for the effect of cuing on behavior (MSL task), sleep patterns and EEG data.

MSL: finger sequence task

The finger tapping task used for this study was an adapted version of the five-item MSL paradigm [45], as described in detail in Laventure et al. [20]. Briefly, subjects were first trained to explicitly remember an eight-item sequence of finger movements (2-4-1-3-4-2-3-1, where 1 stands for the index finger and 4 for the little finger). Following a verification procedure, subjects were asked to complete the formal training session, composed of 24 blocks of practice interspersed of 30 s of rest. Unknown to the subjects, each practice block consisted of 80 key presses. Prior to the beginning of the training session, participants were explicitly asked to try to reproduce the sequence “as fast and accurately as possible”. The next morning, they were asked to complete eight blocks of the same finger sequence motor task in a retest session.

Performance assessment and analyses

Based on previous work in our lab [20, 46], a global performance index (GPI) was used to assess performance on the MSL task. This index takes into consideration the requirements given to each subject regarding speed and accuracy (see MSL: finger sequence task section), and has the additional advantage of controlling for individual strategy. Outlier performance was identified by testing the learning curve characteristics from the training session of each individual using the generalized extreme studentized deviate [47] as described in our previous study [20].

Due to minor changes in the composition of the experimental groups—addition and withdrawal of some subjects, overnight changes in MSL performance were re-analyzed. The effect of the evening training session on motor abilities was measured using a mixed design ANOVA for repeated measures with blocks (n = 24) as the within-subjects factor and groups as the between-subjects factor. This allowed us to test for a learning effect during the training session (main effect of block), and for differences in learning rate between groups (block × group interaction) and differences between groups in terms of overall skill on the MSL task throughout the session (main effect of group).

Information about the end of training was examined using another mixed repeated-measures ANOVA with the average GPI from the last four blocks of this MSL session as repeated within-subjects factor and groups as between-subjects factor. This analysis assessed whether participants reached asymptotic performance at the end of training (main effect of block), and provided information about the learning rate (block × group interaction) and level of MSL performance (main effect of group) between groups. Finally, motor memory consolidation was determined using a repeated-measures ANOVA performed with sessions and blocks (i.e. last four blocks of training and first four blocks of retest) as repeated within-subjects factors and groups as the between-subjects factor. The use of four blocks instead of only one (i.e. last block of training and first of retest) was preferred to avoid fatigue effects at training, warm-up effects at retest blocks and to help reduce block-to-block performance variability [48, 49].

Olfactory stimulus

A solution of phenyl ethyl alcohol (PEA—concentration: 6.31 × 10^–3 [% v/v]) and heavy mineral oil (solvent—USP/FCC) was used as the odorant source. Delivery of the olfactory stimulus was carried out using an ON/OFF block design procedure as described in our previous work [20]. For the MSL training session, the ON blocks comprised blocks of practice, while the OFF blocks corresponded to the periods of 30 s of rest in-between. During NREM2 sleep exposure, the odor was delivered on a 30-s ON/ 30-s OFF block design for a maximum of 60 min.

PSG recording

Sleep recordings were acquired using a 16-channel, V-Amp 16 system (Brainamp, Brain Products GmbH, Gilching, Germany) from 10 scalp derivations (F3, Fz, F4, C3, Cz, C4, P3, Pz, P4, Oz) referenced to linked mastoids (A1, A2). They were recorded continuously (at <5 kOhm) during the whole night using Recorder software (Brain Products), and were visually inspected online for quality. Signals were digitized at 250 samples per second and bandpass filtered (high pass filter = 0.3 Hz, low pass filter = 70 Hz). PSG measurements included EEG, electro-oculogram (EOG), bipolar submental electromyogram (EMG) electrodes as well as a nasal airflow thermistor (Braebon, Ottawa, Canada) to monitor respiratory effort.

For all PSG recordings, including online scoring and stimulation periods, sleep stages were visually identified in 30-s epochs displaying EEG (high pass filter = 0.3 Hz, low pass filter = 35 Hz) from central and occipital derivations (C3, C4, and Oz) referenced to average mastoids (A1 and A2), EOG (high pass filter = 0.3 Hz, low pass filter = 35 Hz) from the lateral outer canthus of each eye, and bipolar sub-mental EMG (high pass filter of 10 Hz). Periods of cortical arousal or movement during sleep were identified using an automated detector when movement continuously exceeded 100 μV for more than 100 ms (detection was visually verified).

EEG pre-processing

Sleep architecture

Sleep stages were scored by a Registered Polysomnographic Technologist (RPSGT) expert according to standard criteria [50] using 30 s epochs. Sleep architecture analyses were conducted on periods defined by the onset and offset of stimulation (see Figure 1C).

Experimental sleep periods

Exposure to an olfactory stimulus during the second half of the night defined two different sleep periods called pre-stim and stim. The stim period refers to the exposure period during NREM2 sleep, while the pre-stim sleep period applies to the NREM2 epochs that occurred before the onset of the first stimulation. The stim and pre-stim periods were length-matched so that the pre-stim period could serve as a reliable within-group baseline for the subsequent analyses.

Sleep spindle detection and selection

Following our previous analysis of this dataset demonstrating that only spindles originating from the parietal region showed cue-related changes [20], spindles were automatically detected from the Pz derivation during NREM2 using an improved version (https://github.com/Sinergia-BMZ/swa-matlab) of a previously published algorithm [51] and rated as a reliable detection technique [52]. Detected events lasting between 0.3 and 2 s and occurring within the 11–17 Hz frequency range were identified as sleep spindles [39]. The toolbox was adapted to provide additional information related to spindle frequency including time markers for each oscillation within the events. Sleep spindles were then categorized by site using a previously described technique [20]. This method uses the detected onset of a given event to determine the site at which the spindle occurred first, as it is known that a single event can appear on multiple electrodes. Afterward, each sleep spindle within the pre-stim and stim sleep periods was extracted as a 5-s epoch (arbitrary window of 2 s before and 3 s after onset) using the EEGLab Matlab toolbox – version 13.6.5 [53, 54]. Finally, upon visual inspection, epochs containing artifacts, false-positive detections and multiple spindle events within the 5-s epoch were excluded from the analyses. This last step allowed a thorough inspection of the signal quality at each site.

EEG analyses

Sleep spindle characteristics

Following findings demonstrating the implication of parietal NREM2 sleep spindles in MSL consolidation [20], analyses of spindle characteristics (peak amplitude, duration, frequency, and density) were carried out on Pz NREM2 events. One-way ANOVAs were used to assess the pre-stim versus stim differences (percentage) in spindle characteristics (within-subjects factor) between groups (between-subjects factor).

Event-related signal perturbation processing and analyses

Time–frequency analyses were used to assess the temporal and spectral activity present before, during, and after a spindle. These event-related signal perturbation (ERSP) analyses permitted to assess the mean changes in the power spectrum over baseline. Each pre-selected sleep spindle epoch was transformed into a time–frequency matrix using an ERSP function from the EEGLab Matlab toolbox. The ERSP transformation was conducted from 3 to 36 Hz (3 Hz being the lower limit for this analysis) in the frequency domain and from −1000 to 2000 ms in the time domain on all 10 electrodes (F3, Fz, F4, C3, Cz, C4, P3, Pz, P4, Oz). The function used three-cycle Morlet wavelets at the lowest frequency (with Hanning-tapered window applied [n = 200]). ERSP matrices were normalized using a period preceding the analyzed event from −1440 to −1000 ms.

Average spindle. 

A time–frequency image of an average spindle was generated using ERSP data from all selected Pz sleep spindles comprised in both conditions and groups. Data from each pixel were average across epochs.

Comparing changes between conditions and groups. 

The number of spindles varied between conditions (within-subject) and participants. Hence, a bootstrap procedure [55] was performed to generate a set of 300 samples for both conditions in each subject. A verification step was performed to ensure that there was no difference between groups before the stimulation. Using the bootstrapped samples from the Pz derivation, unpaired t-tests with false discovery rate (FDR; p < 0.05) correction were performed comparing pre-stim ERSPs between groups. The stim condition was also tested between groups.

Then, with the same set of bootstrapped samples, contrasts between the stim and pre-stim sleep spindles (independent variable) were created using Wilcoxon–Mann–Whitney comparisons in each subject, and for each pixel of the time–frequency maps (dependent variable). This resulted in a single contrast map per derivation and subject.

Single subject contrast maps were then brought to a second level of analysis to test group differences in time–frequency changes using linear regressions. This set of analyses compared the difference in changes between groups (independent variable) at each point of a time–frequency matrix (dependent variable), ultimately generating a single image per derivation. Since samples drawn from smaller pools of events in the preceding step generated higher variance contrasts, regression weights were estimated to control for two attributes: first, the ratio of spindles from each subject compared with the total number of spindles from both groups, and second, the ratio of the number of spindles between pre-stim and stim sleep periods for each subject:

$$W=\frac{(S_{\text{pre-matched}}^i+S_{\text{during}}^i)}{{{\displaystyle \sum }}^{\text{}}(S_{\text{pre-matched}}+ {{\displaystyle \sum }}^{\text{}}(S_{\text{during}})}$$

where $S$ represents the quantity of spindle events and $i$ refers to the subject iteration. On each derivation, linear regressions assessed differences of stimulation-induced time-dependent spectral changes between groups. These regression maps, generated using plot functions provided by EEGlab, are only composed of significant regression coefficients betas (p < 0.001) corrected with a false discovery rate (FDR) method (see Figure 2C).

Figure 2.

Stimulation-dependent sleep spindle time–frequency decomposition. ERSP analyses were conducted on Pz sleep spindle to investigate stimulation-dependent changes in signal around and during the events. (A) Average spindle. Event-related spectral perturbation average of all spindles included in the analyses. Each event was aligned using the algorithm-detected spindle onset (point zero on the x-axis). The resulting time–frequency image ranged from −1000 to 2000 ms in reference to the onset. Frequencies (y-axis) ranged from 3 to 36 Hz and an overlay allowed for the identification of predefined frequency band used in the analyses—delta (δ), theta (θ), sigma (σ), and high-beta (hi-β). The unit of each pixel is defined in decibel (dB). (B) Group comparison at Pz with respect to changes between pre-stim and stim periods. Linear regression tested for differences between groups by comparing condition (stim vs pre-stim) contrast map. Positive (red) significant betas (p < 0.001, FDR corrected; significance thresholds are indicated by a dotted line on the color bar) identify higher ERSP changes in the Cond compared with the NoCond group while negative (blue) values indicate the contrary. (C) Topographical representation of significant differences between groups. Using maps resulting from linear regressions on the ten derivations, topographical representations of only the significant betas between groups were generated (non-green area represent a significant value). Results were fragmented and averaged into three time segments (−1000 to 0 ms, 0–1000 ms, 1000–2000 ms) and three predefined frequency bands (delta–theta, sigma, and high-beta). These topographic maps use the same units as the Pz time–frequency regression map shown in B. Hence, only significant results are represented with a color spectrum ranging from red-positive (Cond > NoCond) to blue-negative (NoCond > Cond). Significance thresholds are indicated by a dotted line on the color bar.

Changes in topographical representations. 

The regression analyses yielded one map per derivation (i.e. 10 maps). Topographical representations of the significant betas were estimated in specific periods of time (−1000 to 0 ms, 0–1000 ms, 1000−2000 ms) and frequency band (delta–theta: 3–8 Hz; sigma: 11–16 Hz; high-beta: 20–30 Hz). Since only frequencies from 3 to 4 Hz were available from the delta band (due to the restricted size of the epoch analyzed), the latter was concatenated to theta frequencies, thus forming a wider delta–theta band. These topographical representations allow for a better topographical localization and lateralization assessment of the difference in changes occurring during parietal spindles between groups.

Coherence analyses

Changes in connectivity between cortical regions during sleep spindles were investigated using coherence analyses. Thus, changes in phase and amplitude synchronization between electrode pairs that occurred between pre-stim and stim sleep spindles were tested in both groups. The same set of sleep spindle epochs as in ERSP analyses was used for the following connectivity tests. Each analyzed epoch covered a 2000-ms period, starting at the spindle onset.

Coherence and imaginary coherence. 

Coherency is a standard method to determine the spectral connectivity between two signals. It is a complex valued function of frequency. A set of coherencies between electrode pairs e and e′ with index n = 1 … N and a scalar function of frequency, f, is defined by:

$$C_{\text{e},e\prime }(f_n)=\frac{S_{\text{e},e\prime }[n]}{\sqrt{S_{\text{e},\text{e}}[n]}\sqrt{S_{e\prime ,e\prime }[n]}}$$

with frequency $f_n=n{f}_0$ and $f_0=1/T$  

where $S_{\text{e},\text{e}}[n]$ is the Fourier transform of a time domain signal of electrode e and $S_{e\prime ,e\prime }[n]$ is the transform of another time domain signal of electrode e′. The coherency can be split into real and imaginary part and different quantities can be analyzed further.

The most important are the coherence ${\text{Coh}}_{\text{e},e\prime }$ defined as the modulus of the coherency and the imaginary part of the coherency $i{\text{Coh}}_{\text{e},e\prime }$. iCoh accounts for the portion of coherence between two electrodes sensitive to interactions with non-vanishing degree of time lag [56].

Z-stat and average. 

Coherence was defined on selected frequency bands (delta: [1–4]Hz; theta: [4–8] Hz; sigma: [11–16] Hz; high-beta: [20–30] Hz) and Z-transformed:

$$i\text{Coh}(f_n)\to Z(f_n)=\text{tan}\hspace{0.17em}{h}^{-1}(i\text{Coh}(f_n))-\frac{1}{df-2}$$

where $df=2P_{S,c}$ is the degree of freedom for the coherence estimate. The last term in $Z$ corrects for the bias: $Z(f_n)$ is approximately normally distributed with mean $\text{tan}\hspace{0.17em}{h}^{-1}(i{\text{Coh}}_{\text{true}}(f_n))$ and variance $1/(df-2)$. Frequency bands were then defined (e.g. $B_i=[F_i,F_{i+1}]$) and finally, the coherence was computed in each band:

$${Z^{\prime }}_i=\frac{1}{\text{card}(B_i)}{\displaystyle \sum }_{f_n\in B_i}Z(f_n)$$

for each subject $S$, and for each pair of electrode $\text{e}$ and $e\prime$.

Non-parametric statistical test of significance. 

First, we considered the average over each subject, in both conditions,

$$z_{i,pre-stim}=\frac{1}{n_s}{\displaystyle \sum }_s{Z^{\prime }}_{i,s,pre-stim},z_{i,stim}=\frac{1}{n_s}{\displaystyle \sum }_s{Z^{\prime }}_{i,s,stim}$$

The contrast $\text{δ}{z}_i=\left|z_{i,stim}-z_{i,pre\text{-}stim}\right|$ was then tested for significant difference from zero with a non-parametric approach based on spatial clustering of the pairwise link. The null hypothesis was drawn from a random reallocation “$r$” of the two conditions (random permutation pre-stim and stim for each subject—there are $2^{n_s}$ such reallocations defining fate stim and fate stim′). In each case, $\text{δ}{z}_{i,r}=\left|z_{i,sti{m}^{\prime }}-z_{i,pre\text{-}sti{m}^{\prime }}\right|$ was computed and a pair of electrodes was selected for which $\text{δ}{z}_{i,r}\ge 0.5\underset{\text{pairs}}{\text{max}}\text{δ}{z}_{i,r}$. Then, the selected pair was clustered (2nd order of neighborhood in the electrode montage) and for each cluster, the contrasts were summed and all the values were accumulated along the random resampling. This constitutes the null hypothesis distribution from which a unique defined threshold $z_i^{}$ was computed (p < 0.00001). The significant links were defined as the one for which $z_i\ge z_i^{}$separately in each frequency band labeled with $i$. For each significant pair, the direction of the effect (increase or decrease) was computed through the group mean square coherence difference between pre-stim and stim, which is defined as the amplitude of the complex-valued expression.

Ethics statement

This study was revised and approved by an institutional ethics committee (“Comité mixte d’éthique de la recherche du Regroupement Neuroimagerie/Québec”; ID: CMER-RNQ 09-10-026). Upon their arrival at the sleep laboratory for the screening night, all participants were asked to provide written consent.

Results

MSL consolidation

As previously conducted in our initial study [20], a mixed repeated-measures ANOVA conducted on GPI scores of the 24 blocks of training (repeated measures) showed a significant effect of block (F_{23, 1081} = 51.397, p < 0.00001), with no significant block × group interaction (F_{23, 1081} = .448, p = 0.99), or main effect of group (F_{1, 47} = .195, p = 0.66). These results suggest that, while all participants showed improvement in performance across the training session, both groups expressed similar overall performance and learning rates.

Additionally, a second mixed repeated measures ANOVA performed on the last four blocks of the training session did not reveal any significant main effect of block (F_{1, 47} = 2.162, p = 0.15), block × group interaction (F_{1, 47} = .234, p = 0.63), nor any main effect of group (F_{1, 47} = .974, p = 0.33). The latter results demonstrate that acquisition performance leveled off by the end of the training session, with all participants reaching an asymptotic level of performance.

Consolidation was measured by comparing the mean GPI score of the first four blocks of retest to the last four blocks of training using a repeated-measures ANOVA. This analysis resulted in a main effect of session (F_{1, 141} = 44.234, p < 0.00001) confirming that both groups showed an offline consolidation effect (pairwise comparisons: Cond, p < 0.001; NoCond, p = 0.01). Importantly, post hoc pairwise comparisons reported a significant session × group interaction (F_{1, 141} = 8.672, p = 0.005) demonstrating that cued participants (Cond; mean = .036 ± .01) reached greater offline improvement in performance than those who were not conditioned to the olfactory stimulus (NoCond; mean = .014 ± .01). Complementary analyses performed on speed only (mean time per block) yielded the same pattern of results.

Sleep architecture and stimulation

Independent samples t-tests conducted on sleep architecture did not show any difference between the two groups with regard to the total sleep time (TST), total recording time (TRT), wake duration, SE, nor any of the sleep stages (see Table 1), hence confirming that the experimental manipulation had no differential effect (p > 0.05) on sleep architecture between these two groups. Another independent samples t test was performed to compare the stim period duration between both groups, but again the latter analysis did not reveal any differences (t₍₄₇₎ = 1.325, p = 0.19; Cond: 38.7 ± 13.8 min; NoCond: 33.7 ± 12.8 min), hence showing that both groups received similar periods of stimulation.

Table 1.

Sleep architecture

	Cond		NoCond		t(47)	P
	Mean	SE	Mean	SE
Wake	56.3	5.3	53.0	6.5	0.387	0.70
NREM1	20.1	10.4	20.3	9.4	−0.075	0.94
NREM2	197.3	34.0	194.3	41.5	0.275	0.78
SWS	123.1	36.8	115.7	35.8	0.714	0.48
REM	93.2	19.3	94.9	6.0	−0.246	0.81
Movement	3.3	0.8	3.5	0.8	−0.262	0.80
TRT	493.2	26.5	481.8	32.4	1.353	0.18
TST	433.6	30.1	425.2	46.1	0.763	0.45
Sleep efficiency (%)	88.0%	4.9	88.2%	6.4	−0.129	0.90

All sleep measurements are presented in minutes, except for sleep efficiency, which corresponds to a percentage calculated from the ratio of TST on TRT. Standard errors (SE) are reported. Independent t-tests were conducted for each sleep characteristic between groups to identify whether there were any significant differences in sleep architecture. As expected, there were no significant group differences in any of the sleep periods and characteristics.

Sleep spindle selection and characteristics

Spindle filtering preceding analyses included discarding of false positives, artifacts, and multiple events occurring within a 5-s window. After detection and filtering of all sleep spindles, a total of 12235 sleep spindles were selected for analyses. These events had a mean duration of 0.68 s (±0.35 s) and their distribution showed that 81.7% were shorter than 1 s. In each condition (pre-stim, stim), no difference was found (p > 0.05) between groups in spindle duration, amplitude, frequency, or density. Assessment of rejected events numbers through a univariate general linear model identified no significant differences. We further investigated the difference in results between the current and initial study [20]. Independent samples t-test comparing groups were conducted on changes between conditions of spindle characteristics from previously rejected events—because occurring concomitantly to other spindles in a 5-s window. These analyses identified a significant difference in sleep spindle frequency change between condition (t₍₄₇₎ = 2.359, p = 0.02; Cond: .011 Hz ± .02; NoCond: −.001 Hz ± .02) but not in amplitude (t₍₄₇₎ = 1.491, p = 0.14) nor duration (t₍₄₇₎ = −0.731, p = 0.47).

Sleep spindle time–frequency decomposition

Average spindle description

An ERSP average of all sleep spindle events included in the analyses is shown in Figure 3A. This time–frequency map confirmed that most events ranged from 11 to 16 Hz, although some appeared to exceed those boundaries (~10–19 Hz), and that spindle frequency tended to decrease over the course of the spindle event [35]. The average time–frequency decomposition showed that spindle onsets, as defined by the detection algorithm, did not represent the real event onset. Indeed, the actual spindle onsets appeared to occur earlier compared with detected onsets. This shift is to be expected, however, as the detection algorithm established onsets based on reaching a predefined threshold. When this threshold is reached, the actual spindle oscillation has already started for a few milliseconds. Further, it is likely that part of the detected changes in amplitude exceeding the expected boundaries in time, and particularly in frequencies, could be due to frequency smearing. Interestingly, the average spindle ERSP identified another component outside the usual sigma (11–16 Hz) range, spanning from ~ 24 to 30 Hz in the frequency domain and from 0 to 1000 ms in the time domain. This simultaneous high-beta activity, which can be described as a secondary spindle component, appeared consistently on individual spindle time–frequency decomposition.

Figure 3.

Changes in coherence between regions during sleep spindles. Significant changes (p < 0.00001) in imaginary coherence metrics during Pz-detected sleep spindles between the pre-stim and stim periods are represented by colored link between pairs of sites—increases in red and decreases in blue—for each group. Analyses were processed over four different bands of frequencies: (A) SWA (0.25–4 Hz), (B) theta (4–8 Hz), (C) sigma (11–16 Hz), and (D) high-beta (20–30 Hz).

Localized changes in spectral power between conditions and groups

Group differences maps of ERSP changes between the pre-stim and stim periods were generated using a linear regression corrected for multiple independent analyses (false discovery rate [FDR]) (see Supplemental Material SFigures 1–4 for contrast maps between conditions and SFigure 5 for group comparison maps). Only significant beta values (p < 0.001) were reported and detailed in the following section.

The group contrast map on Pz (Figure 2B) showed that no major significant change occurred in the sigma band during spindles (0–1000 ms). However, it identified significant ERSP differences (Cond > NoCond) in the theta band prior to the event onset (~−500 to −300 ms) and, in both delta and theta bands following the end of the spindle (~1200−1600 ms; based on the average spindle in Figure 2A). Comparatively, still in the theta band, the Pz map showed higher ERSP changes in the NoCond compared with the Cond group during spindles (~50–600 ms). Altogether, these results suggest that cued memory reactivation influenced the timing, and thus power (ERSP), of lower frequencies (delta–theta) in relation to cued spindle events. Indeed, comparing both groups, while there seems to be a decrease in lower frequencies during the cued events, significant increases were identified before the onset and at the offset of cued sleep spindles. Furthermore, the regression analyses also detected segregated clusters of significant higher differences for the Cond compared with the NoCond group in the high-beta band (20–30 Hz) from the onset to the end of the epoch’s window (0–2000 ms).

Figure 3C provides a topographical summary of the time–frequency linear regression significant results for all the analyzed electrodes (F3, Fz, F4, C3, Cz, C4, P3, Pz, P4, and Oz) during Pz sleep spindles.

High-beta. 

Multiple significant changes in every time segments were identified in high-beta (20–30 Hz). First, before the event onset, the topographical map showed significantly lower ERSP differences over Pz, as well as higher difference for the Cond compared with the NoCond group in the right parietal region. Then, during the first 1000 ms following the onset, significant higher differences for the cued group compared with the NoCond group were observed in parietal (Pz) and right central (C4) areas. Finally, during the last segment (i.e. 1000–2000 ms), results showed that the Cond group had significantly higher differences over several posterior (P3, Pz, and Oz) and central (Cz and C4) areas than the NoCond group.

Sigma. 

As previously seen in Pz (Figure 2B), no major significant differences were identified in the sigma frequency band in any time period apart from a lower difference before the even onset for the Cond compared with the NoCond group.

Delta–theta. 

The topographical ERSP map demonstrated strong differences in the delta and theta bands for the cued group in contrast to the non-cued group before the event onset (−1000 to 0 ms) and after the average spindle offset (1000–2000 ms) in central and parietal regions. Notably, the differences found before the spindle onset were predominantly lateralized in the right hemisphere and covered the primary motor (C4) and somatosensory cortex (Pz and P4). The opposite effect appeared during the event (0–1000 ms) as shown by lower changes for the Cond compared with the NoCond group across parietal and right central areas. Analyses on the time–frequency maps from the Pz derivation comparing (pre-stim[Cond] vs pre-stim[NoCond]) did not identify any differences between groups prior to the stimulation onset. The same analysis was also conducted on the stim condition and did not yield any significant difference. Associations between local activity changes related to spindles and gains in performance were assessed by including the overnight gains as a regressor. The analysis did not identify any significant relationship between changes in performance and frequential activity.

Overall, these results suggest that cuing with a conditioned stimulus is associated with increases in signal power at higher (hi-β) and lower (δ–θ) frequencies time-locked to spindle events. Additionally, these increases appear to be stronger in posterior and contra-lateral regions (in reference to the hand that performed the MSL task), and thus, may be regionally specific and use-dependent. Moreover, the contrast maps identified a stimulation-dependent decrease in delta and theta frequencies within the 0- to 1000-ms time window, which corresponds roughly to the mean spindle duration. It also highlights that the spindles main sigma component seems to be embedded in time (δ–θ) and frequency (hi-β) by oscillations that are sensitive to the reactivation of a memory trace with a conditioned stimulus.

Changes in connectivity during sleep spindles

Coherence analyses were carried out over all derivations on four frequency bands (delta, theta, sigma, and high-beta) using the same dataset of filtered events used for the ERSP analyses in order to test for changes in connectivity between cortical regions during Pz spindle events (Figure 3; p < 0.00001).

Delta

The results showed that the Cond group coherence in delta increased between several pairs of regions along the anterior–posterior axis. In contrast, the NoCond group showed interhemispheric, left frontal-right posterior (C4 and P4) decreases in coherence.

Theta

Analyses revealed a clear difference in pattern between the two groups. While coherence in the theta band for the Cond group seemed to increase across widespread pairs of regions covering all deviations, the NoCond group showed a sparser mix of a few increases (three pairs) and decreases in coherence (two pairs) between regions. In the Cond group, the occipital region seemed to be a hub for connectivity between sites, as it was part of no less than seven significant increases in coherence between parietal (Pz), central (C3, Cz, C4), and frontal (F3, Fz, F4) regions.

Sigma

Coherence results demonstrated that the TMR had a significant effect on multiple regions in the sigma band. Indeed, only the Cond group showed significant changes in coherence where increases in coherence were predominantly evident on centro-parietal regions. Comparatively, the NoCond group did not show any evidence of change in connectivity in sigma.

High-beta

Changes detected in the high-beta band were only found in the NoCond group. However, three of these four significant changes were identified as decrease in coherence (C4–P3, C4–Oz, P3–Pz).

In sum, coherence analyses demonstrated that connectivity with central–parietal regions significantly increased in the sigma band during cued NREM2 sleep spindles. This is in contrast to the absence of local changes in this band as reported above through ERSP analyses and suggests that sleep spindles might be synchronizing activity between regions during memory reactivation. Furthermore, results from this analysis reinforce the implication of theta oscillations by showing large-scale increases in a global connectivity network occurring during cued sleep spindles. No significant correlation was identified between specific changes in connectivity and overnight gains in performance.

Discussion

This study builds upon previous behavioral and electrophysiological findings [20], and suggests that sleep spindle activity interacts with multiple types of oscillations in a time-locked manner during reactivation of a newly acquired memory trace following MSL. We first showed that cued reactivation through olfaction produced a greater increase in performance for the group that was conditioned during the evening training period versus the unconditioned control group. Then, time–frequency decompositions of parietal sleep spindles revealed an increase in local low-frequency activity before and after spindle oscillations. The same analyses also identified high-beta activity increases during sleep spindles when subjects were reexposed to the odor during NREM2 sleep. Finally, our results demonstrate the presence of large increases in connectivity generated during cued spindles in the theta and sigma bands only. Our present findings thus highlight the emergence of a global mechanism, temporally locked to sleep spindles, and possibly supporting MSL consolidation. This mechanism involves the dynamic interplay of spindle-related activity with multiple EEG frequencies in the delta/theta and high-beta bands.

As already shown in our previous study [20], analyses of MSL changes revealed that both groups elicited overnight gains in performance and that cuing during post-training NREM2 sleep potentiated the consolidation process, resulting in significantly better performance at retest for the Cond compared with the NoCond group. Moreover, the absence of difference in sleep architecture between both groups suggests that the offline gains in performance were produced by the experimental manipulation rather than factors related to sleep quality.

Analyses of sleep spindle characteristics (i.e. amplitude, frequency, duration, density) on selected events did not yield any difference between groups, contrary to our previous findings [20]. Further analyses suggested that this discrepancy stemmed from two main factors. First, different sleep spindle detection algorithms were used in each study, using a different methodological approach to detect and extract characteristics [51]. This factor seems to account for the difference in results concerning spindle amplitude changes following cuing. Second, in this study, in order to allow for clean and unperturbed signal around each detected event, a pre-processing step eliminated segments containing more than a single spindle during a 5-s period. This resulted in sparse clusters of events that were not analyzed. Analyses of clustered spindles yielded similar results to those in our previous study; hence, suggesting that these clusters of spindles might be more reactive to the reactivation of the memory trace. However, whether these clusters of spindles following learning are associated to stronger consolidation and motor sequence performance was beyond the scope of the aims of this study and remains to be confirmed.

High-frequency component during sleep spindles

The averaged spindle time–frequency decomposition (Figure 2A) allowed to identify a secondary component, nested in the high-beta band, in addition to the common sigma oscillation from spindles. This component occurred a few milliseconds after the event onset and faded out before the end of the sigma oscillation. In contrast to other studies that used time–frequency decompositions, for example [35], we did not filter out the EEG signal outside the sigma band, hence allowing us to detect this high-beta activity. The presence of a high-beta secondary component during sleep spindles could explain one report of beta and sigma activity occurring within the same timeframe, both time-locked to slow positive half-waves [26]. Although this is, to our knowledge, the first report of this spindle high-beta component, it is not the first characterization of such a synchronization phenomenon between sigma-band spindles and higher frequencies. Previous studies using intracranial [57] and cortical [32] recordings in epileptic patients and animal models [33] have also identified evidence of hippocampal ripple (80–120 Hz) bursts of activity, time-locked to spindle troughs. Their findings suggest that the synchronization of oscillations at different frequencies resulting from communications between cortical and subcortical brain structures does underlie memory consolidation. It is thus possible that the high-beta component described here represents the activity of a population of neurons synchronized to spindle oscillations. In that case, task-related changes in this component would occur when participants are cued with a conditioned stimulus.

However, there is also the possibility that the high-beta activity concomitant to sleep spindles is the results of a non-perfect sinusoidal spindle waveform [58]. In this particular case, the high-beta activity could be the result of neuronal activity within the spindle frequency range in a neighboring but somewhat independent region under the same electrode detection field. During a spindle event, this other region would oscillate at a similar frequency range than the main recorded region but with enough differences in amplitude and frequency as to create slight distortions in the spindle waveform. Then, wavelet transformations would detect activity in both sigma and high-beta bands. Thus, more precise recordings (e.g. intracranial EEG) are needed to identify the origin of this high-beta activity. The interaction of high-beta frequencies with sleep spindles and other frequency bands during cuing will be discussed further below.

Localized changes in power during sleep spindles

Parietal sleep spindle time–frequency comparison maps revealed that, in contrast to the non-cued group, spindles from the cued group were associated with higher changes in the delta band following event offset, the latter difference being possibly caused by a higher presence of SO in proximity to sleep spindles. In support of this conjecture, the time-locked association between SO and sleep spindles is well documented. Previous studies describing the relation between SO and spindles [24, 25, 37, 38] have proposed that a change in SO state (e.g. from down to up-state) provides a neuronal dynamic propitious to the formation of sleep spindles and other fast frequency events (e.g. ripples). Furthermore, before spindle onset, the theta difference between our groups was mostly lateralized in the right hemisphere over an area comprising the sensorimotor regions (contralateral to the hand that performed the task), hence suggesting that cuing increased activity in task-related cortical regions. This is in line with previous studies showing that localization of spindle activity correlates with cortical regions activated during a newly learned task [17, 59]. Thus, our results suggest that cued memory reactivation increases the amount of slow frequency activity (theta activity and SWA) time-locked to sleep spindles.

Moreover, time–frequency decomposition maps demonstrated that the cued group had a significant increase in the high-beta band during sleep spindles. This implies that the high-beta activity cluster straddling the main sigma component and described earlier might be involved in the cued reactivation of the motor memory trace. A closer look at the regions where these changes took place revealed that sensorimotor and motor areas relevant to the newly learned task showed increases in high-beta, especially after spindle onset. No localized difference was detected in the sigma band between groups, however.

Cue-related functional connectivity changes during sleep spindles

Change in connectivity was assessed through pairwise comparisons of EEG-coherence between pre-stim and stim periods. As with the time–frequency analyses, results were divided into frequency bands. The most striking differences between the cued and non-cued groups were found in the theta and sigma bands. Surprisingly, results showed that participants who were reexposed to the conditioned stimulus demonstrated significant and widespread increases in connectivity within the theta frequency band, with occipital regions acting as a hub for connectivity. As predicted, connectivity was also significantly higher in the sigma band following reexposure to the conditioned stimulus. Together, these results suggest that cuing produced an increase in connectivity in lower frequencies and sigma, but not in high-beta. Past studies investigating normal sleep in a non-learning context have shown that synchrony between regions increases in sigma frequencies across NREM cycles [60] and during sleep spindles [61, 62]. The present report extends these findings by demonstrating that the reactivation of a motor sequence memory trace increases cortico-cortical functional connectivity during sleep spindles within specific frequency bands (e.g. delta, theta, and sigma) over and above normal changes.

In addition, the overall delta and theta patterns suggest that cuing increased antero-posterior connectivity with less emphasis on inter-hemisphere (left-right) changes, while in sigma, most increases were found between parietal and central sites with a mix of inter-hemisphere and antero-posterior changes. This is in line with previous studies showing that delta and theta oscillations correlate with long-range communications between cortical regions, while sigma and beta bands were associated with short-range transmission and gamma bands were related to very localized exchanges [63]. Our data suggest that long-range antero-posterior connectivity is enhanced during sleep spindles when subjects are exposed to a contextual cue conditioned to a motor task, and that this effect is led by low frequencies (delta and theta). Further, short-range, central and parietal coherence seems to be increased by the olfactory cue during sleep spindles over and above the normally high coherence previously reported between these regions [64]. This also complements previous findings on regional task-specific changes in spindles density after motor skill learning [13]. Thus, our results support the notion that sleep spindles may be actively participating in motor sequence consolidation by reactivating complex memory traces involving different cortical regions and that this multisite interaction seems to rely on activity within the delta, theta, and sigma bands.

Theta activity and memory consolidation

Our assessment of cortical electrical activity time-locked to the onset of sleep spindles indicated that theta activity significantly changed both locally (change in power) and between regions (increased connectivity) during cuing. The theta band has been widely studied and consistently associated to memory processing [19, 65–67]. For example, it has been shown that both theta synchrony and power increase during recollection of past experiences [68]. It is thought that in the context of a motor learning protocol using a TMR stimulation design, as in this study, the cued stimulus represents an episodic component triggering the motor memory trace. Hence, in such case, the reactivation of a motor memory trace associated to a contextual cue could explain the similarities in findings in this research with studies looking at declarative memory. Indeed, recent TMR studies using declarative tasks also reported changes in the theta band following cue onset [69–72]. Further, following theta increase, one of these studies using language memory tasks reported increases in sigma [72], while another, using a visuo-spatial test, reported an increase in beta (15–30 Hz) [70].

Although recognized as a key feature of REM, theta is also present during NREM sleep. It has been thought to correspond to a hippocampal hallmark oscillation playing a pacemaker role in memory encoding and retrieval processes [66], as well as to a long-range communication mechanism connecting segregated brain regions [67]. A recent report suggested that theta rhythms serve as a functional binding between the hippocampus, prefrontal cortex, and the striatum during recollection of visual memories, a subtype of declarative memory [73]. Importantly, this frequency band is also thought to play a central role in the integration process of multiple mnemonic representations into a coherent whole [67, 74, 75]. In relation to our results, this particular function of the theta band could thus be instrumental in promoting the MSL memory representation integration within a given context (presentation of an odor) effectively leading to memory consolidation at a system level. The current findings are thus in line with the working model proposed by Schreiner and Rasch [76], which portrayed the reactivation of a memory during sleep as a dynamic interaction, first involving hippocampally driven theta and gamma, followed by sleep spindles providing the necessary support for the stabilization, strengthening, and integration of the memory trace. However, in this study, we did not find a direct correlation between changes in motor performance and theta (or any other frequency bands). Still, significant changes in several frequency bands, including theta, were observed in the cued compared with the non-cued group. Hence, the details of the possible underlying association between regional and interregional changes induced by reactivation and memory consolidation are still unknown. Thus, studies using TMR paradigms with intracranial recordings should be carried out in order to shed light on the possible interaction of the hippocampus, striatum, and cortex during cued sleep spindle.

From local to interregional changes

In sum, this study yielded two major findings. First, altogether time–frequency decomposition and coherence analyses demonstrate that the reactivation of a conditioned memory trace engage, not only the sigma oscillations related to sleep spindles, but also the boundary frequencies at specific stages of the spindle event. In the Cond group, the presence of significant increases in theta before spindle onset, and delta/theta at the end of spindles, is in line with aspects of the active system consolidation model [21, 77]; more specifically, the facilitation by slower oscillations (e.g. SWA) of higher frequency activity such as sleep spindles. One of the key concepts proposed by this model is that newly learned memory traces are repeatedly reactivated during sleep, thus promoting synaptic connections within the associated neuronal network and reinforcing the associated mnemonic representation. It is thought that these reactivations are driven by SO and that the temporal grouping of neuronal depolarization during these SO up-states constitutes a favorable timeframe for memory reactivation through hippocampal sharp-wave ripples and thalamocortical spindles [22]. Although our time–frequency analyses did not look at SO (<1 Hz), they did highlight the presence of local power increases in lower frequencies before and after the events. This supports the notion that the mechanisms underlying cued reactivation of a newly acquired motor memory trace involve spindle activity time-locked to lower frequencies comprising theta and delta. Furthermore, despite the fact that the increase in high-beta straddling the main sigma component is composed of lower frequencies than sharp-wave ripples (which are not measurable from scalp derivations in humans), it is possible that it follows a similar generation pattern, nested inside spindle troughs [29, 32]. Thus, the high-beta component could be the result of reverberating activity within local neuronal networks triggered and synchronized by spindle oscillations. While this is an interesting possibility, the high-beta increases triggered by olfactory cuing could also reflect the underlying changes in activity of nearby but independent local cortical regions oscillating also in the sigma range.

Second, our results show a possible distinction between local and functional changes in terms of frequency bands. Indeed, while local sigma power was not significantly modulated by the reexposure to a conditioned stimulus, coherence analyses identified increases in sigma connectivity between central and parietal cortical regions in this same condition. On the other hand, local increases in high-beta were observed during spindles, while connectivity changes between regions were not evident. Interestingly, both significant changes (interregional sigma; local high-beta) were identified over the sensorimotor regions. Yet these results do not allow us to clearly determine the exact underlying neuronal mechanisms triggering these changes. However, it could be hypothesized that during a cued spindle event, the synchrony between distinct cortical regions involved in the same motor representation is regulated by the spindle’s sigma oscillations. Still, no current model has yet discussed high-beta’s involvement in memory consolidation. Hence, as theorized for the relationship between SO and spindles [77], as well as for theta and gamma activity [78], we propose that sleep spindles may be the basis for local activation of task-related networks through higher frequency activity such as SWR and possibly high-gamma. Still, investigations using metrics such as phase amplitude coupling are needed to confirm this hypothesis and to test the involvement of high-beta activity in motor sequence consolidation. Further, studies using declarative tasks could also help defining if the sigma and high-beta topographic distributions are indeed task-dependent.

Conclusions

Here we provide evidence that sleep spindles are associated to the reactivation of newly learned motor sequence memory traces during sleep. Following the previous demonstration of the critical role of NREM2 sleep in MSL consolidation [20], we demonstrate that cued and locally expressed reactivations increases the activity in the high-beta band during spindle events. Our results also show that lower frequencies are temporally locked to the onset of cued sleep spindles through spectral power increases before the onset and at the end of spindle events. Finally, connectivity between several cortical regions increases as a result of cuing. While low-frequency bands increased connectivity throughout the cortex in an antero-posterior pattern, sigma activity seemed to influence coherence between task-related cortical areas. Still, the relation between these changes related to the cuing of a memory trace during NREM2 sleep and the improvement in motor performance remains unclear since no correlation was found between these measures. Our findings, in relation with current theories on memory consolidation, suggest that sleep spindles are part of a complex and broad reactivation, and communication system involving multiple frequencies and brain regions. Yet, the interactions of sigma, high-beta, and possibly gamma oscillations during cued sleep spindles still need to be investigated in order to elucidate the neuronal mechanisms binding these frequency bands together and their potential implication in motor sequence memory consolidation.

Supplementary material

Supplementary material is available at SLEEP online.

Funding

Grant received by JD from the Canadian Institutes of Health Research (CIHR, Grant number: MOP-97830; URL: http://www.cihr-irsc.gc.ca/e/193.html). The CIHR had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Conflict of interest statement: None declared.