The Multidimensional Aspects of Sleep Spindles and Their Relationship to Word-Pair Memory Consolidation

Abstract

Study Objectives:

Several studies proposed a link between sleep spindles and sleep dependent memory consolidation in declarative learning tasks. In addition to these state-like aspects of sleep spindles, they have also trait-like characteristics, i.e., were related to general cognitive performance, an important distinction that has often been neglected in correlative studies. Furthermore, from the multitude of different sleep spindle measures, often just one specific aspect was analyzed. Thus, we aimed at taking multidimensional aspects of sleep spindles into account when exploring their relationship to word-pair memory consolidation.

Design:

Each subject underwent 2 study nights with all-night high-density electroencephalographic (EEG) recordings. Sleep spindles were automatically detected in all EEG channels. Subjects were trained and tested on a word-pair learning task in the evening, and retested in the morning to assess sleep related memory consolidation (overnight retention). Trait-like aspects refer to the mean of both nights and state-like aspects were calculated as the difference between night 1 and night 2.

Setting:

Sleep laboratory.

Participants:

Twenty healthy male subjects (age: 23.3 ± 2.1 y)

Measurements and Results:

Overnight retention was negatively correlated with trait-like aspects of fast sleep spindle density and positively with slow spindle density on a global level. In contrast, state-like aspects were observed for integrated slow spindle activity, which was positively related to the differences in overnight retention in specific regions.

Conclusion:

Our results demonstrate the importance of a multidimensional approach when investigating the relationship between sleep spindles and memory consolidation and thereby provide a more complete picture explaining divergent findings in the literature.

Citation:

Lustenberger C, Wehrle F, Tüshaus L, Achermann P, Huber R. The multidimensional aspects of sleep spindles and their relationship to word-pair memory consolidation. SLEEP 2015;38(7):1093–1103.

INTRODUCTION

Numerous behavioral studies link sleep spindles, a unique electrophysiological characteristic of non-rapid eye movement (NREM) sleep, to declarative memory consolidation, thereby claiming that sleep spindles are involved in consolidation processes.^1,2 A task often used to demonstrate the relationship between sleep spindles and sleep dependent declarative memory consolidation is the verbal associate learning task (word-pair task). Several studies show that sleep related retention in a word-pair task was positively correlated with sleep spindle activity (e.g., electroencephalography [EEG] power 12–15 Hz) or density (number of spindles per min of sleep).^3–6 In addition, an increase of sleep spindle density was found after word-pair learning,⁷ in particular, when the encoding was difficult.⁶ This increase of spindle density was correlated with sleep related changes in recall performance.⁶ However, there are also numerous studies that failed to demonstrate positive associations^8–11 or even negative correlations¹² between sleep spindles and memory consolidation were reported using the same learning paradigm.

A possible explanation for such a discrepancy is that sleep spindles have multidimensional aspects, e.g., slow versus fast frequency spindles¹³ that may serve different functions in memory processes.^14–16 In addition, behavioral studies without manipulations can poorly distinguish between the possibility that the relationship between task learning and subsequent sleep spindles is a reflection of general learning abilities during wakefulness or the possibility that sleep spindles play a causal role in learning and memory. In this regard, sleep spindles were proposed to reflect trait- and state-like characteristics.¹³ Hereby a trait represents a biological fingerprint that is not restricted to specific situations and is fairly stable over time.^17,18 For example, absolute spindle measures per night are trait-like, because they are highly correlated from night to night.¹⁹ Trait-like spindle measures were associated with learning and intellectual abilities^1,12,20–24 and sleep related memory consolidation in different learning tasks.^3–5,10,25 It is difficult, however, to resolve whether performance changes after sleep purely reflect sleep related consolidation processes or rather mirror a general learning trait.^11,12,26 State-like aspects are situation-specific changes in sleep spindles and may be defined as the differences between nights. Thus, the increase of sleep spindles after learning is a typical state-like aspect of sleep spindles.^6,11 Hence, it was proposed that the interindividual baseline differences in sleep spindles (trait-like aspects) are correlated with learning potential, whereas learning-related increases in sleep spindles (state-like aspects) reflect processes specific to memory consolidation.^1,11 However, trait-like and state-like aspects of sleep spindles and their relation to memory consolidation were mostly independently investigated and many studies neglected a differentiation between trait and state-like aspects.

Another issue common to several studies is the selective choice of a specific sleep spindle measure. Thus, from the multitude of different sleep spindle measures, often just one specific aspect was selected (e.g., power or density), in a restricted frequency range (slow versus fast), for a certain sleep stage (NREM stage 2 versus slow wave sleep) or at a specific time during sleep (early versus late). Furthermore, analysis was mostly constrained to a few electrodes possibly explaining some divergent findings and confusions in literature. A striking example is the variety of frequency range definitions used for slow and fast sleep spindles (e.g. cutoff frequency 12 Hz^27–29 versus 14 Hz^4,30). This discrepancy calls for a more objective way to separate them.

In summary, no study combined all commonly used spindle measures and assessed topographical differences. Our study aim was to take the multidimensional aspects of sleep spindles into account when exploring trait- and state-like aspects of sleep spindles and their relationship to word-pair memory consolidation. The rational for this study was to increase our understanding of the divergent/confusing findings in the literature and thereby providing a more complete and conclusive picture about the relationship between sleep spindles and word-pair learning.

METHOD

Participants and Design

Twenty young male participants (age: 23.3 ± 2.1 y, mean ± standard error of the mean [SEM]) without sleep disorders, personal or family history of psychopathology, chronic diseases, and current use of psychoactive agents or other medications were recruited. The data used in this manuscript are part of a larger study that was approved by the cantonal ethic commission in Zurich (Switzerland) and for which all subjects gave written informed consent to participate. The study comprised different conditions, but for this manuscript only baseline nights were analyzed for all subjects. We only included male subjects because sleep spindles and sleep related memory consolidation are known to be influenced by the menstrual cycle.^31,32 Participants were right handed, nonsmokers, and free of medication and drugs. During a screening night, subjects were acclimatized to the laboratory and the high-density EEG (hdEEG) recordings. All included subjects were healthy sleepers with good sleep quality and no sleep disorders were detected. Thereafter, subjects underwent 2 study nights 2 w apart in the sleep laboratory of the Institute of Pharmacology and Toxicology, University of Zurich. In the evening subjects had to memorize and recall word-pairs in a paired associate learning task. Subsequently, hdEEG nets were applied and subjects went to bed either at 22:50 or 23:40. Eight hours later (06:50 or 07:40), subjects were awakened and retested on the paired associate learning task. We used the same procedure for both study nights. Three days before these 2 nights, subjects had to adhere to regular bedtimes (8 h time in bed, according to scheduled bedtime in the laboratory), and abstain from caffeine, naps, and alcohol. Compliance with the instructions was controlled by breath alcohol test, sleep logs, and wrist-worn actimeters.

Sleep EEG Recordings

A hdEEG (Electrical Geodesics Sensor Net for long-term monitoring, 128 electrodes, including electrooculogram and electromyogram, Electrical Geodesics, Eugene, OR) net was adjusted to the vertex (Cz) and filled with gel electrolyte. HdEEG provides good spatial resolution thereby allowing the analysis of topographical and local aspects of the sleep EEG.³³ During the 8 h of continuous sleep EEG recording, the analog signals were referenced to Cz, band-pass filtered (0.01–200 Hz), and digitized at 500 Hz. Preprocessing of the signal included filtering (0.5 Hz high-pass, 40 Hz low-pass filter) and downsampling to 128 Hz. Sleep stage scoring was performed on 20-s epochs according to standard criteria,³⁴ and artefacts were identified on a 20-s basis by visual inspection and semi-automatically. During the semiautomatic artefact detection the algorithm automatically excluded epochs with a power value exceeding 13 times the power of the sliding mean (average over 15 20-s epochs) in the 0.75–4.5 Hz frequency band and a power value exceeding six times the power of the sliding mean (average over 25 20-s epochs) in the 20–30 Hz frequency band. Please note that we used a user interface that enabled us to further exclude epochs that were clearly deviant from the background by manually adjusting the threshold.

In a next step, the EEG was re-referenced to average reference after exclusion of EEG channels of insufficient quality (on average, seven channels per subjects). For topographical analysis of sleep-EEG activities (“topoplots,” Figure S1, supplemental material), bad channels were interpolated using a spherical interpolation provided by the EEGLAB toolbox.³⁵ We only included 108 channels into the statistical analysis (excluding marginal electrodes which would lead to 128), but used interpolated values of bad channels that belonged to the 108-channel configuration.

Spindle Analysis

To address the multidimensional aspects of sleep spindles we included all electrodes of the hdEEG. As a consequence of including > 100 EEG channels, a visual detection of sleep spindles was not feasible and we automatically detected sleep spindles for each electrode using an established algorithm.^36,37 Detailed description about this procedure can be obtained from Ferrarelli et al.³⁶ Specifically, the signal was band-pass filtered between 12–15 Hz and rectified. A sleep spindle was detected from the signal if the amplitude exceeded an upper threshold that was defined relative to the mean signal amplitude (eight times mean signal). Beginning and end of sleep spindles were defined as the time points when the signal around the peak amplitude dropped below a lower threshold (two times mean signal). In order to optimize the analysis for our specific aim we adapted this original procedure as follows. Because it was important for our study to accurately distinguish between slow and fast spindle frequencies we needed to adapt the frequency range. As shown by preliminary analysis, the originally used 12–15 Hz band-pass filter provided best signal-to-noise ratios, but was too narrow in order to accurately detect slow spindles (around 12 Hz and lower). It is well established that sleep spindles exist within the frequency range of 10–16 Hz^38–40 (please also see Figure S2, supplemental material, for spectral plots). Previous studies also used band-pass filter settings between 10–16 Hz to investigate spindles.⁴¹ Thus, to detect sleep spindles in this broader frequency range, we used a Chebychev filter of order 20 with nominal passband corner frequencies at 10 and 16 Hz and stopband corner frequencies at 5 Hz (at least 80 dB dampening ≤ 5 Hz) and 32 Hz (at least 80 dB dampening ≥ 32 Hz) allowing an accurate and negligible dampening around 11–16 Hz (empirically determined dampening at 11 Hz was 1.9 % and 0.1 % at 16 Hz). Because this broadening of the filter-band decreased the signal-to-noise ratio, we had to adapt the upper and lower threshold accordingly. An upper threshold of five times and a lower threshold of 1.25 times the mean signal provided best spindle detection as verified by visual inspection. We used channel-wise threshold definition because signal amplitude varied significantly between channels. In addition this approach aims at taking factors into account that might vary across channels (e.g., skull thickness) and would influence EEG signal amplitude. Besides using channel-wise threshold approaches there are also multichannel studies that found good results using one common threshold for all channels, at least in analysis where spindle detections were performed separately for slow and fast spindles.⁴² Further studies are needed to investigate whether both approaches work equally well and provide comparable results of detected sleep spindles. Please note that a further broadening of the filter settings would lead to high noise, i.e., no spindle-specific activity (e.g., alpha activity in the low spindle frequency range), and as a result spindle detection becomes inaccurate. Using this adapted algorithm our spindle density values were in the range of other studies using visual and automatic spindle detection.^43,44 Moreover, they also showed the expected topographical distribution (see Figure S1 for the topographic distribution of slow and fast spindles). Our algorithm provides different spindle measures and we focused on sleep spindle density (number/ min), which has often been related to declarative memory consolidation,^2,5,6 and the averaged integrated spindle activity of individual sleep spindles (in μVs), which provides information about the shape or the “power” of the sleep spindles. The integrated spindle activity comprises spindle amplitude and duration and is calculated as follows: The band-pass filtered absolute EEG signal is first summed up over the course of each sleep spindle and then averaged for all included spindles. Both sleep spindle measures are quantified for different frequency bins from 11–16 Hz (spindles were detectable in this frequency range; see Figure S3, supplemental material, with a histogram of the spindle distribution). We used the histogram “histc” function provided in MATLAB (The Mathworks, Inc.) to divide the sleep spindles into specific frequency bins with a frequency resolution of 0.5 Hz (e.g. the 10-Hz bin covers all the spindles that are > 10 Hz and < 10.5 Hz).

Furthermore, sleep spindles with a duration lower than 300 ms were excluded.⁴⁰ This analysis was performed for the first (FH) and the last hour (LH) of artifact-free NREM sleep. By focusing on FH and LH of NREM sleep, we included the timing of sleep spindles during the sleep period and also accounted for different levels of sleep pressure (FH: high sleep pressure, in particular sleep stage N3; LH: low sleep pressure, in particular sleep stage N2).

Word-Pair Task

Based on several earlier studies, the paired associate learning or word-pair task was used to assess sleep dependent performance improvement in declarative memory.^{7,27,29,32,45} Forty semantically related word-pairs were presented on a computer screen for 4 s each, separated by an interstimulus interval of 100 ms. After the presentation, there was an immediate cued-recall, where the first words of the word pairs were presented in random order and the second one had to be recalled. The subjects were asked to guess in case they did not remember the word. No time limit for the answers was set and after the subjects entered their answer, a feedback for accuracy was provided and the correct word pair was presented again for 2 s. The subjects were instructed to memorize the word pair again during the feedback. In the morning (45 min after subjects woke up), there was a delayed recall, where the procedure was the same as in the immediate recall. Overnight retention was defined as the difference in correct answers between delayed (morning) and immediate recall (evening). It is important to mention that our retention measure does not allow us to differentiate between a sleep dependent memory consolidation and learning related improvement due to feedback that was given during the immediate recall as done in many previous studies.^{8,12,27,29,32,46–48} We used two parallel lists for the 2 nights in a randomized order.

Statistics

First, we compared the results of the word-pair task between night 1 and night 2 using paired t tests and bivariate Pearson correlations.

In order to investigate trait- and state-like aspects of sleep spindles and their association to overnight retention, we averaged our measures for both nights (trait) or took the difference between night 1 and night 2 (state). For both trait and state, Pearson correlation matrices between sleep dependent overnight changes of a word-pair task and spindle measures (density, integrated spindle activity) for each frequency bin (11–16 Hz, 0.5-Hz bins) and all electrodes (108, marginal electrodes were excluded) were performed. For correlations with trait-like aspects, we further controlled for evening performance in a partial correlation design (see Results for more details). For both trait- and state-like aspects we conducted hierarchical cluster analysis on the r-values of the two dimensional correlation matrices (channels/electrodes × frequency bins). We used this exploratory approach to elaborate in an objective way the frequency bins that show similar correlations with the word-pair variables for multiple EEG channels (clustering of a two-dimensional (2D) correlation matrix of frequency bins × channels) and are clearly separated from other frequencies. To do so, a 2D hierarchical cluster analysis of the Euclidean distances between all the 11 frequency bins and all 108 EEG channels was performed based on their correlation (r-values) with the word-pair task using the “clustergram” function of the bioinformatics toolbox provided by MATLAB. The 2D correlation matrix is grouped into a hierarchical cluster tree (dendrogram) according to their proximity or how close the correlation coefficients are. A dendrogram consists of upside-down U-shaped lines or branches that are also called clades. Each terminal end of the clade is called leave (with represent frequency bins on the x-axis and channels on the y-axis in our examples). The arrangement of the clades indicates which frequency bins and channels (leaves) are most similar to each other. The height of the branch/clade indicates how similar or different they are from each other: the smaller the height, the closer correlation coefficients are for the clustered frequency bins and channels. To calculate this height or distance between the clusters, the “single linkage method” (nearest neighbor defined as the smallest Euclidean distance: shortest connection line between two points) was applied. To define which similarity in correlation coefficients is useful to define frequency clusters in our data set, we performed visual inspection of the derived dendrograms and tested different Euclidean distances for the definition of clusters. Using a Euclidian distance of 3.5, which is an arbitrary cut-off derived from visual inspection, the derived frequency clusters were the most meaningful ones across all dendrograms.

An illustrative example of such a two-dimensional dendrogram is shown in Figure 1. Subsequently, neighboring frequency bins that clustered together (below Euclidean distance of 3.5) were averaged (integrated spindle activity) or summed up (density). Please note that the cluster analysis included all r-values and not just the significant ones. In order to control for multiple comparison we only included frequency bands that contained more than 5% (number of electrodes × included frequency bins above the chance level) of significant correlations (P < 0.05). In a next step, we tested for significant electrode clusters by calculating the correlation between the chosen frequency bands of sleep spindle measures and sleep related performance changes in the word-pair task. This topographical analysis provides further information about the spatial clustering (channel clustering instead of frequency bin clustering) of the correlations between sleep spindle measures and word-pair retention. To control for multiple comparisons and to define specific regions of interest^36,49–51 we applied statistical nonparametric mapping (SnPM⁵⁰) whenever appropriate. A suprathreshold cluster analysis was applied in which electrodes that were above/below a correlation coefficient of 0.39/−0.39 (corresponding to P = 0.1 for n = 19) and exceeded the 90th percentile cluster size given by the permutation analysis were considered for further analysis.

Figure 1.

Illustrative example of the two-dimensional hierarchical cluster trees (dendrogram) and the heat plot of the r-values of a partial correlation between trait-like sleep spindle density during the last hour of non-rapid eye movement sleep and overnight retention, controlled for immediate recall performance. Please note that colors of the heatplot encode differences from the mean r-value, with blue colors indicating lower and red higher values than the mean. Red/blue marked branches illustrate clusters with a Euclidean distance below 3.5. A clear differentiation between slow and fast spindle density is observable with a cut-off frequency at 13.5 Hz. Color-coded ellipses are used to code for the electrode location on the scalp: green-frontal, yellow-central, red-occipital, orange-temporal, blue-parietal.

One subject had to be excluded from the whole analysis due to bad sleep and data quality (movement artifacts, many bad channels).

RESULTS

Sleep Quality

All included subjects (n = 19) had good sleep quality in the first and second experimental night (Table 1). During night 1, subjects had a significantly longer sleep latency, more NREM stage 1 (N1) sleep and less rapid eye movement (REM) sleep (Table 1). When both nights are averaged, the first hour of NREM sleep (FHNREMS) includes significantly more N3 sleep (31.1 ± 2.0 versus 6.1 ± 1.3 min) and significantly less N2 sleep (28.9 ± 2.0 versus 53.9 ± 1.3 min) compared to the last hour of NREM sleep (LHNREMS).

Table 1.

Sleep variables derived from visual scoring.

Word-Pair Task

A summary of the performance in the associate verbal learning task is presented in Figure 2. Subjects recalled significantly more word pairs in the delayed recall (morning) compared to the immediate recall (evening) in both nights (P < 0.05). This significant overnight improvement is expected since feedback was given during the immediate recall in the evening allowing the subjects to re-encode the word pairs.¹² Immediate recall was positively correlated between the two nights (r₁₉ = 0.51, P = 0.02). Overnight retention (difference morning to evening) also tended to be correlated between the two nights (r₁₉ = 0.40, P = 0.09). These data indicate that immediate recall and overnight retention are partially reproducible. Thus, the average of the 2 nights likely represents a trait aspect. Number of recalled word pairs in the immediate and delayed recall was not different between the 2 nights (Figure 2A), whereas overnight retention was significantly higher in the first compared to the second night (Figure 2B). Thus, the difference between night 1 and night 2 can be used as a state measure for the overnight retention. Importantly, when both nights are averaged (trait), evening performance tended to negatively correlate with overnight retention (r₁₉ = −0.46, P = 0.05, we also found a tendency when both nights were analyzed separately; Night 1: r₁₉ = −0.42, P = 0.08; Night 2: r₁₉ = −0.40, P = 0.09). This was not the case for the state (night difference) variables (r₁₉ = −0.26, P > 0.25). Thus, we included the evening performance as a covariate (correlation of residuals that are controlled for evening performance) in the Pearson correlations that focus on trait-like aspects of sleep spindles (FH and LH) and word-pair learning using a partial correlation design.

Figure 2.

(A) Recall performance of associate verbal learning task of both nights. Stars indicate significant differences (P < 0.05, paired t test, n = 19) between evening and morning recall. (B) Overnight retention was also significantly different between the two nights as indicated by a star (P < 0.05, paired t test, n = 19).

Trait and State Aspects of Sleep Spindles and Their Association to Overnight Retention in a Word-Pair Task

Correlation matrices between overnight retention and the two sleep spindle measures (density, integrated spindle activity) for trait/state and FHNREMS/LHNREMS are illustrated in Figure 3. These matrices include all 108 electrodes of the hdEEG net and frequency bins from 11–16 Hz (0.5 Hz resolution). Based on these correlation matrices we performed a hierarchical cluster analysis (see Method for details). An illustrative example of such a two dimensional hierarchical tree (dendrogram) is provided in Figure 1. All neighboring frequency bins that grouped together as a cluster in the dendrogram (Euclidean distance < 3.5) and had significant electrodes above the chance level (number of significant correlations > 5% of electrodes × frequency bins) were used for topographical plots.

Figure 3.

Heat plots of the correlation coefficients and corresponding statistics for the relationship between spindle measures (density [number/min] and integrated spindle activity [μVs]) of the first (FH) and last hour (LH) of non-rapid eye movement sleep and overnight retention in word-pair learning for trait-(average of night 1 and 2) and state-like (difference between nights) aspects. In each plot the x-axis represents spindle frequency and the y-axis the 108 high density electroencephalography electrodes grouped into sets of electrodes that were close in distance to the 10-20 system configuration. C, central; F, frontal; O, occipital; P, parietal; T, temporal. Red rectangles highlight neighboring frequency bins that were grouped together in a hierarchical cluster analysis (see Figure 1) and include number of significant correlations above the chance level (see Method, white bars, P < 0.05).

When looking at trait aspects, we found that during the first hour of NREM sleep, fast sleep spindle density was inversely correlated with overnight retention (controlled for evening performance) on a rather global level (Figures 3 and 4). Conversely, during the last hour of NREM sleep, slow sleep spindle density was positively related to overnight retention (also controlled for evening performance).

Figure 4.

Top panel: Topographical representation of the correlation coefficients highlighted by red rectangles in Figure 3 (trait aspect; spindle density). White dots indicate electrodes with significant correlations (P < 0.05, partial Pearson correlation controlled for evening performance) and gray dots indicate trend level partial correlations (P < 0.1). Bottom panel: Reported correlation coefficients are based on a Pearson correlation (not controlled for evening performance difference, white dots P < 0.05, gray dots P < 0.1). Clusters (state aspect) that survived statistical nonparametric mapping (SnPM P < 0.1) are highlighted with blue circles.

State-like aspects (difference between the two nights) are illustrated in the second row of Figure 4. This analysis shows that slow sleep spindle measures were positively related to overnight retention, generally on a spatially more local level. To identify these local clusters of electrodes we performed statistical non-parametric mapping (q.v. Method). The significant clusters are highlighted with blue circles. Our analysis of state-like aspects revealed a cluster of frontal electrodes for the density of slow spindles and a cluster of left centroparietal and right temporo-occipital electrodes for integrated activity of slow spindles. A comparison of these clusters between the first and second night revealed that only the electrode clusters of slow integrated spindle activity in LHNREMS were significantly higher in night 1 compared to night 2 in subjects that showed superior positive overnight retention in night 1 compared to night 2 (Figure 5).

Figure 5.

Difference of spindle density (number/min) and integrated spindle activity (μVs) between night 1 and night 2 for clusters that survived statistical nonparametric mapping (SnPM) correction (Figure 4, 12–13.5Hz). Electrode clusters are color coded in the topographical outline of the high-density electroencephalography net. Individual differences were divided into two groups according to their word-pair retention performance; MI = higher overnight retention rate during night 1 than night 2 (n = 15); MD = lower retention rate during night 1 than night 2 (n = 4). Stars indicate significant differences between night 1 and night 2 (paired t test, * < 0.05, ** < 0.01). Please note that due to a low number of subjects no statistical analysis was performed for the MD group. However, in the MD group all four subjects showed reduced integrated spindle activity in the two clusters and three of four subjects had reduced spindle density. Gray-filled marginal electrodes were not included in the analysis.

DISCUSSION

Our results demonstrate the importance of a multidimensional approach when investigating the relationship between sleep spindles and memory consolidation. We found different associations between memory retention and spindle measures, depending on the type of the spindle measure, spatial location, and time of night. In addition, the differentiation between trait-and state-like aspects of sleep spindles and their relation to declarative verbal associate memory adds to the picture. Our approach shows the complexity of the relationship between sleep spindles and declarative memory and thereby provides a more conclusive picture that might explain the divergent findings reported in literature.

Trait-Like Aspects of Sleep Spindles and Overnight Retention

Most of the studies that performed correlations between sleep spindles and overnight/nap retention in word-pair learning focussed on trait-like aspects of spindles. These studies found controversial results with opposite signs of the correlations.^8–12 Our analyses showed a similar discrepancy. Most of the studies reported positive correlations of overnight/ nap retention in word-pair learning with spindle density,^25,52 sigma activity,^3,4 peak amplitude,²⁹ or a combination of these measures.^5,25 Interestingly, these positive associations were mainly restricted to stage 2 sleep^3,4,25,52 and/or spindle frequencies between 12–14 Hz.^3,4 In line with these observations, we found a positive correlation between trait-like slow spindle density (11–13.5 Hz) in the last hour of NREM sleep, a time window that is dominated by stage 2 sleep, and overnight retention. Notwithstanding, other studies found no correlation between trait-like spindle measures and overnight memory retention.^8–11 These findings can be explained by our results. Our analysis showed contradictory correlations when including different frequency ranges and different time windows. Moreover, we found no signifi-cant correlations with integrated spindle activity. Studies often averaged spindle measures for the whole night and frequency range, which may have resulted in a nullification of the correlations.

One of our main results was a negative correlation between overnight retention and fast spindle density in the first hour of NREM sleep. We reported a similar negative correlation already in an earlier study.¹² In this study with young adults, we observed that sigma activity (comprising both density and integrated spindle activity) was rather related to learning efficiency than sleep specific consolidation. In other words, subjects that encoded much of their capacity (maximal amount of correct word-pairs reached) in the first place (immediate recall performance relative to delayed performance) had more sigma activity and less overnight improvement.¹² This interpretation raises an important question: What do trait-like spindle measures and overnight retention reflect? Our and previous results indicate that these measures may reflect learning ability per se rather than sleep specific consolidation processes.^1,11 In line with this interpretation trait-like aspects of sleep spindles have been related to learning and cognitive abilities.^1,12,20–24 In addition, word-pair performance improvement/stabilization (retention) after a period of sleep partially depends on the feedback given during the immediate recall (second learning opportunity¹²) or the strength of encoding that might then be reflected in sleep spindle measures. Thus, divergent findings may be a result of the design of the word-pair task that substantially differs between the studies in terms of feedback, difficulty of word pairs, number of encoding opportunities, and the degree of semantic relation. Importantly, in our study, evening performance was negatively related to overnight retention and was included in the correlation as a confounding factor. Because there is good evidence that encoding performance is also related to sleep spindles,^7,11,53 future studies should carefully control for the contribution of this initial performance.

Significant correlations between trait-like aspects of word-pair overnight retention and sleep spindles were restricted to the spindle density measure. In addition, we found a clear, objective cutoff in the direction of the correlations at 13 Hz (FHNREMS) and 13.5 Hz (LHNREMS) separating our slow and fast spindles. Thus, we found significant correlations for the FH of NREM sleep between the frequencies 10–13 Hz and 13.5–16 Hz (r₁₉ = −0.57, P < 0.01), and also the LH tended to be negatively correlated between 11–13.5 Hz and 14–16 Hz (r₁₉ = −0.45, P = 0.05). Why these frequency bands revealed correlation coefficients with opposite directions is not evident. It was hypothesized that these two types of sleep spindles serve different functions¹³ and might result from two different generators, for example, from two different thalamic sources.⁵⁴ However, different spindle frequencies might result from the hyperpolarization-rebound sequence duration of thalamocortical neurons⁵⁵ or their level of hyperpolarization.³⁸ We found a negative correlation between the fast and slow sleep spindle density indicating that they might be related and derive from a single mechanism, as has previously been hypothesized, e.g., in intracranial measures of human subjects.^38,56 Hence, one might speculate that the duration of hyperpolarization-rebound sequences in the thalamocortical neurons or their level of hyperpolarization (either leading to slow or fast spindles) could be related to overnight retention or learning abilities.

State-Like Aspects of Sleep Spindles and Overnight Retention

Even though we had the same study design for both nights, overnight retention in the first night was significantly higher than in the second night while evening performance (immediate recall) was similar. The difference in overnight retention was also not related to the difference in immediate recall. This difference provided an important state-like aspect of declarative overnight retention that might be specifically related to consolidation processes during sleep rather than encoding per se. The causal role of sleep spindles can only be established using selective manipulation of sleep spindles, which is to date not possible. To overcome this problem, we need in a first step different approaches that investigate the role of sleep spindles in memory consolidation in more detail. A fruitful approach to investigate the importance of sleep spindles in memory consolidation is the comparison of learning and non-learning conditions where an increase of spindle density was found in a learning compared to a non-learning condition.⁷ To overcome the limitation of such an approach that alterations in sleep spindles might also be associated with task performance related general plasticity changes in the evening, such trait-like effects could be eliminated by using a within-subject design. We applied an alternative approach, taking into account a similar evening baseline performance and just comparing nights that had a difference in overnight changes (declarative memory retention) rather than evening learning performance. Thus, our experimental design represents an alternative approach not comparing learning versus non-learning but focusing on more versus less declarative memory retention that reflects memory consolidation. If indeed sleep spindles are causally involved in memory consolidation processes, we expect that day-to-day variances of declarative memory retention result in different amounts/shapes of sleep spindles, or vice versa, that fluctuations in sleep spindles should be related to differences in overnight retention measures. We therefore addressed the question, whether this difference in overnight retention is related to state-like aspects of sleep spindles. It had been proposed that learning related increases in sleep spindles reflect processes specific to memory consolidation.^1,11 We found significant positive correlations between overnight retention and slow sleep spindle density as well as integrated spindle activity in the first and last hour of NREM sleep, respectively. These correlations were restricted to specific regions, frontal for spindle density, and centro-parietal and parieto-temporal for integrated spindle activity. A few studies showed similarly that learning related difference in spindle activity (related to integrated spindle activity^6,11) during stage 2 were positively correlated with overnight retention. If indeed sleep spindles are causally involved in declarative memory consolidation processes one would expect higher spindle measures in night 1 compared to night 2. We found such increased spindle measures in night 1 compared to night 2 but only for slow integrated spindle activity and only in subjects that showed superior overnight retention in night 1. Interestingly, the four subjects that had worse overnight retention in night 1 compared to night 2 all had lower integrated spindle activity in night 1. Our results provide a first hint that especially state-like slow integrated spindle activity is localized to specific regions (e.g., left centro-parietal and right temporo-occipital), which might be involved in memory consolidation processes. However, because this study and most others are only correlative in nature we cannot conclude that integrated sleep spindle activity is causally involved in memory consolidation. To date, only a few studies enhanced/ reduced spindle measures by manipulation using electrical brain stimulation,^27,28 tones,²⁹ or pharmacological interventions^52,57,58 and showed superior/reduced declarative memory consolidation. However, these studies not only affected sleep spindles, but also modulated slow-wave activity,^27,29,58,59 slow wave sleep,⁵² or affected REM sleep.⁵² Thus, proof of a causal role of sleep spindles in memory consolidation is still missing.

Compared to other studies using single or few electrodes, our hdEEG recordings revealed regional differences, in particular for state-like aspects of sleep spindles. Why our correlations were restricted to these specific regions is unclear. Functional magnetic resonance imaging and positron emission tomography studies reported different frontal and posterior parts involved in word-pair retrieval and encoding.⁶⁰ However, interpretation of these regions is precarious, since we cannot directly deduce an underlying source of spindle activity from cortical EEG topography, especially because thalamo-cortical circuits are involved in their generation.

Finally, it is important to emphasize that we used a specific automated spindle detection algorithm since a visual detection of sleep spindles is not feasible with hdEEG. This might be a limitation of our study because it was recently reported that diverging results were obtained with different spindle detection algorithms and visual scoring.⁴⁰ For instance, our filter settings did not allow accurate (very low frequency) spindle detection below 11 Hz due to signal attenuation. However, lowering the low-frequency cutoff would also lead to a reduced signal-to-noise ratio because frequencies in the alpha and theta range would also substantially contribute to the filtered signal. Future studies will need to establish and compare valid algorithms to address the multidimensional aspects of sleep spindles. Thus, our multidimensional approach should further be extended including different algorithms as an additional level in the analysis.

In conclusion, our results favor a multidimensional approach in analysing sleep spindles and their relation to memory consolidation. We found clear differences in the association between spindles and word-pair memory consolidation depending on the type of spindle measure, frequency, timing, and localization we analyzed. Our results also show that trait-like sleep spindle density reflects learning traits rather than sleep dependent consolidation processes per se. However, state-like aspects of slow integrated sleep spindle activity might be involved in memory consolidation processes. Tools allowing to selectively manipulating sleep spindles are needed for establishing a causal relationship.

DISCLOSURE STATEMENT

This was not an industry supported study. This work was supported by the Swiss Research Foundation on Electricity and Mobile Communication (FSM-Project Nr. A2011-05 to R.H. and P.A.), the Swiss National Science Foundation (PP00P3_135438 to R.H.; 32003B_146643 to P.A.) and the Clinical Research Priority Program “Sleep and Health” of the University of Zurich. The authors have indicated no financial conflicts of interest. The work was performed at the University of Zurich, Zurich, Switzerland.

SUPPLEMENTAL MATERIAL

Figure S1

Topographical distribution (n = 19) of slow (11–13 Hz) and fast (13.5–16 Hz) spindle density averaged for both experimental nights. Topography is based on NREM sleep stages N2 and N3 of the entire night. Values are color coded between 0 and the maximum and plotted on the planar projection of the hemispheric scalp model. Electrode locations are indicated by black dots.

Figure S2

Individual power density spectra during NREM sleep (power density in μV²/Hz) for frequency bins between 0.5–20 Hz of 19 subjects, both experimental nights and different electrodes. Blue lines indicate frontal electrodes (F3, F4), green lines central electrodes (C3, C4), red lines parietal electrodes (P3, P4) and magenta lines occipital electrodes (O1, O2).

Figure S3

Distribution of spindle densities of 0.5-Hz frequency bins between 10 and 16 Hz. Spindle densities (number/min) were averaged over all 108 electrodes, subjects and experimental nights. Please note that below 11 Hz almost no spindles were detected and that at 16 Hz a clear drop is visible. This limited detection at the boarders is likely to be cau sed by the filter (10–16 Hz) that prevented accurate detection close to the cut-off frequencies.