Location specific sleep spindle activity in the early visual areas and perceptual learning

Abstract

Visual perceptual learning (VPL) is consolidated during sleep. However, the underlying neuronal mechanisms of consolidation are not yet fully understood. It has been suggested that the spontaneous brain oscillations that characterize sleep stages are indicative of the consolidation of learning and memory. We investigated whether sleep spindles and/or slow-waves are associated with consolidation of VPL during non-rapid eye movement (NREM) sleep during the first sleep cycle, using magnetoencephalography (MEG), magnetic resonance imaging (MRI), and polysomnography (PSG). We hypothesized that after training, early visual areas will show an increase in slow sigma, fast sigma and/or delta activity, corresponding to slow/fast sleep spindles and slow-waves, respectively. We found that during sleep stage 2, but not during slow-wave sleep, the slow sigma power within the trained region of early visual areas was larger after training compared to baseline, and that the increase was larger in the trained region than in the untrained region. However, neither fast sigma nor delta band power increased significantly after training in either sleep stage. Importantly, performance gains for the trained task were correlated with the difference of power increases in slow sigma activity between the trained and untrained regions. This finding suggests that slow sigma activity plays a critical role in the consolidation of VPL, at least in sleep stage 2 during the first sleep cycle.

1. Introduction

Our visual system remains plastic even after early postnatal development. This plasticity manifests itself, e.g., in visual perceptual learning (VPL), which is defined as a long-term performance improvement on a visual task after repeated experience (Sasaki, Nanez & Watanabe, 2010). VPL is highly specific to the visual location of the trained stimulus (Crist et al., 1997; Fahle & Edelman, 1993; Fiorentini & Berardi, 1980; Karni & Sagi, 1991; McKee & Westheimer, 1978; Poggio, Fahle & Edelman, 1992; Sagi & Tanne, 1994; Shiu & Pashler, 1992) and features of the stimulus (Ball & Sekuler, 1987; Fiorentini & Berardi, 1980; Koyama, Harner & Watanabe, 2004; Poggio, Fahle & Edelman, 1992; Schoups et al., 2001; Vaina et al., 1998; Watanabe et al., 2002). These perceptual specificities suggest that neuronal changes associated with VPL occur within visual areas, which are highly organized with respect to location and feature. Previous studies have investigated the possible cortical sites of VPL and reported neuronal changes in the lowest areas of visual cortex, such as V1 (Hua et al., 2010; Li, Piech & Gilbert, 2004; Schoups et al., 2001; Schwartz, Maquet & Frith, 2002; Shibata et al., 2011) or higher-level visual cortical areas such as V4 (Adab & Vogels, 2011; Raiguel et al., 2006; Yang & Maunsell, 2004) and MT (Gu et al., 2010; Zohary et al., 1994). It should be noted, however, that the location specificity in VPL has become controversial in some cases (Xiao et al., 2008; Zhang et al., 2010a; Zhang et al., 2010b) and sensory adaptation plays a critical role in location specificity (Harris, Gliksberg & Sagi, 2012).

It has been shown that consolidation of VPL occurs during sleep (Gais et al., 2000; Karni & Sagi, 1993; Stickgold, James & Hobson, 2000; Stickgold et al., 2000; Yotsumoto et al., 2009). Karni & Sagi (1993) have demonstrated that learning of a texture discrimination task (TDT) does not improve after as much as eight hours of wakeful rest, but improves significantly after a full night sleep. Notably, research has shown that sleep must occur within 30 hours of practice in order for subsequent performance improvements to develop (Stickgold, James & Hobson, 2000). Sleep is a multifaceted process with distinct neuronal activity patterns nested within each sleep stage. Thus, it has been assumed that each sleep stage plays a different role in the consolidation of VPL (Stickgold et al., 2000).

Sleep can be broadly categorized into non-rapid eye movement (NREM) sleep and rapid eye movement (REM) sleep: NREM sleep can be further divided into stage 2 and slow-wave sleep (SWS). According to Gais et al. (2000), the neuronal processes related to the early sleep period, which contains an abundance of NREM sleep and very little REM sleep, facilitate consolidation of VPL, whereas the later sleep period, which has abundant REM sleep, does not facilitate consolidation, if the later sleep period alone occurs. In addition, Yotsumoto et al. (2009) has shown that brain activity in the trained region of V1 during NREM sleep is significantly greater after training compared to before training. Interestingly, the brain activity within the trained region of V1 observed during NREM sleep was highly correlated with later performance improvements on the trained task. These studies suggest that NREM sleep plays a crucial role in the development of performance improvements. Notably, the latter study indicates that the consolidation processes of NREM sleep occur specifically within trained region of early visual areas.

However, the results so far obtained do not reveal which specific neuronal activities are specifically involved in the consolidation of VPL. There are two leading models which attempt to explain the neuronal mechanism of consolidation during sleep: an active system consolidation model (Born, Rasch & Gais, 2006; Born & Wilhelm, 2012; Hasselmo, 1999; Rasch et al., 2007) and a synaptic homeostasis model (Tononi & Cirelli, 2003, 2006). The active system consolidation model assumes that memory traces that are involved in learning are reactivated and redistributed during subsequent NREM sleep, so that synaptic connections in the neocortex are strengthened (Born, Rasch & Gais, 2006; Born & Wilhelm, 2012; Hasselmo, 1999; Rasch et al., 2007). According to this model, the reactivation and redistribution processes are mediated by sleep spindle and slow-wave activity, which are the most prominent features of sleep stage 2 and SWS, respectively (Born, Rasch & Gais, 2006; Born & Wilhelm, 2012). Sleep spindles are generated from the thalamic nucleus reticularis and spread to the entire neocortex (Steriade, McCormick & Sejnowski, 1993), whereas slow-waves are primarily generated in the neocortex and extend to other areas such as the hippocampus and thalamus (Buzsaki & Draguhn, 2004; Steriade, 1999). Experiments have demonstrated that spike trains similar to spindle activity induce long-term potentiation (LTP) in cortical neurons (Rosanova & Ulrich, 2005). Research in humans has shown that spindle activity increases after intense learning of a declarative word pair task (Gais et al., 2002; Molle et al., 2009; Schabus et al., 2004; Schmidt et al., 2006) and a procedural motor task (Fogel & Smith, 2006; Morin et al., 2008). In addition, it has been shown that performance improvements on a trained task correlate with spindle activity in cortical areas primarily associated with the task; for example, the parietal cortex and a visuospatial task (Clemens, Fabo & Halasz, 2006), the prefrontal cortex and a word pair task (Clemens, Fabo & Halasz, 2005), and the motor cortex after a finger tapping task (Nishida & Walker, 2007). Slow-wave activity is also thought of as a signature of consolidation during sleep. Slow-waves temporally bind neuronal activity, including sleep spindles, and ripple into depolarizing up-states during which neurons’ firing rates increase to a level similar to a waking state (Steriade, 2006). Previous studies have shown that ripples and spindles nested in depolarizing up-states of slow-waves support LTP (Buzsaki, Haas & Anderson, 1987; King et al., 1999; Rosanova & Ulrich, 2005). However, it should be noted that the role of slow-waves is still controversial; it has been shown that slow-waves with T-type Ca²⁺ influx lead to long-term depression (LTD) (Czarnecki, Birtoli & Ulrich, 2007).

On the other hand, the synaptic homeostasis model proposes that global synaptic downscaling, which occurs during sleep, generates performance improvements after sleep as a by-product, and does not necessarily support the existence of specific consolidation processes during sleep (Tononi & Cirelli, 2003, 2006). According to this theory, synapses become potentiated as information is encoded in the brain during the waking state. Sleep then globally downscales synaptic strength to an energetically sustainable level. As a result, weak synaptic connections are removed, and the remaining relatively strong synaptic connections are preserved. This process improves the signal-to-noise ratio of the encoded information and leads to memory enhancement. Slow-waves are assumed to be related to synaptic downscaling because slow-waves associated with Ca²⁺ channel activation induce LTD (Czarnecki, Birtoli & Ulrich, 2007). There is some evidence to support the synaptic homeostasis model. A local increase in slow wave activity was found in cortical motor areas during sleep after a motor learning task, and this activity was correlated with performance gain (Huber et al., 2004). Subsequent research showed that slow wave activity was reduced when information encoding was prevented (Huber et al., 2006). This indicates that changes in synaptic strength are directly related to slow wave activity: greater synaptic strength is associated with greater downscaling.

It remains to be seen whether sleep spindles and/or slow-waves are involved in the consolidation of VPL, which is thought to involve early visual areas (Hua et al., 2010; Li, Piech & Gilbert, 2004; Schoups et al., 2001; Schwartz, Maquet & Frith, 2002; Shibata et al., 2011). Based on the aforementioned findings, this study aims to investigate which oscillatory activity plays a role in the consolidation of VPL, specifically within the first cycle of NREM sleep. To address this question, we trained subjects to perform a texture discrimination task (TDT) in one specific quadrant of the visual field, a paradigm that is well-known to induce location-specific learning (Karni & Sagi, 1993; Yotsumoto et al., 2009; Yotsumoto, Watanabe & Sasaki, 2008). The location specificity of the task allows us to examine activity in the corresponding retinotopic area on the cortex, as well as a control area. We recorded brain activity during sleep using magnetoencephalography (MEG) and polysomnography (PSG). In order to source-localize the spontaneous brain oscillations measured by MEG to individuals’ cortical space, we also collected the subjects’ brain structures using magnetic resonance imaging (MRI). The combination of MEG and MRI provides fine spatio-temporal resolution (Ahveninen et al., 2007; Lin et al., 2004). Using a retinotopic mapping technique (Choi et al., 2012; Engel et al., 1994; Fize et al., 2003; Sereno et al., 1995; Shibata et al., 2012; Shibata et al., 2011; Yotsumoto et al., 2009; Yotsumoto, Watanabe & Sasaki, 2008), we localized two different regions of the early visual areas which retinotopically correspond to trained and untrained visual field quadrants. A systemic frequency analysis was used to calculate power for oscillations from these identified cortical regions. Since we were interested in slow-waves and sleep spindles only, we selected the delta (0.5-1.5 Hz), slow sigma (11.5-12.5 Hz), and fast sigma (13.5-14.5 Hz) bands. The delta activity is considered to correspond to the frequency of slow-waves, and each sigma to slow and fast sleep spindles. The separation of slow and fast sigma activity (faster or slower than 13 Hz) is based on results from previous studies (Anderer et al., 2001; Schabus et al., 2007; Schabus et al., 2008; Tamaki et al., 2009; Werth et al., 1997; Zeitlhofer et al., 1997; Zygierewicz et al., 1999).

If neuronal activity in the sigma or delta bands is involved in the consolidation of TDT during NREM sleep, changes in those activities should fit the following three criteria. First, due to the location specificity of TDT learning, the power increase in the involved frequency band should be higher in the trained region compared to the untrained region. Second, due to the consolidation process, the power of the involved frequency band within the trained region of early visual areas should be higher during post-training sleep compared to pre-training sleep. Third, the difference in the oscillation power increase between trained and untrained regions of early visual areas should be positively correlated with performance improvements.

2. Methods

2.1. Subjects

Fifteen healthy subjects (9 females and 6 males, mean age 22.53 ± 2.95) with no sleep disorders, abnormal sleep habits, or use of medication participated in the experiment. All subjects had normal or corrected-to-normal vision and refrained from caffeinated drinks and alcohol for a period of 24 hours before the experiment. Subjects were not allowed to take daytime naps on the day of the experiment. All subjects were informed of the purpose and the procedures of the experiment and gave written informed consent, which was approved by the Institutional Review Board of Massachusetts General Hospital, where the experiment was conducted.

2.2. Experimental design

The experiment consisted of three non-consecutive nightly sessions for MEG recording and one additional daytime session for fMRI recording (Fig. 1). The first night served as an adaptation session to mitigate the “first night effect” (FNE) (Tamaki et al., 2005). FNE refers to the alteration of sleep structure that occurs when sleeping in a new environment. The second and third nights of sleep were used as the pre-training and the post-training sleeps, respectively.

Fig. 1. Experimental procedure.

There was an interval of 3-8 days between the pre- and post-training sleep sessions, during which subjects were required to maintain their regular wake-sleep pattern. The subjects’ daily sleep quality and wake-sleep rhythms were monitored in self-reported sleep logs from 3 days before the experiment until the last experimental session. No subjects were excluded based on their sleep logs, since no irregularity in wake-sleep cycle was found.

In all sleep sessions, subjects slept inside a magnetically shielded room each night lying on a non-magnetic bed with their heads positioned inside the helmet-shaped tip of the dewar containing the MEG sensor array. They were allowed to sleep at their regular sleep onset time. MEG and PSG were recorded and monitored simultaneously during sleep. Subjects were awakened when they began showing traces of the first REM sleep stage, which typically occurs 60-90 minutes after sleep onset. Waking the subjects after the first NREM sleep cycle precludes REM sleep and the following sleep cycles from exerting any possible effects on VPL. After the sleep session was over, the subjects could catch up on their sleep at home so that they would not be deprived of sleep (but see below for the timing of re-test on the third night).

On the third night, subjects conducted TDT (see below for details) before and after sleep. The initial TDT test and training of TDT were conducted five hours before sleep with no time interval between them. It is important to note that we had separated the initial TDT test, which served as a pre-training test with fewer trials, from intensive training, which involved a larger number of trials, so that the initial test would not be obscured by any possible fatigue effect (Censor, Karni & Sagi, 2006; Censor & Sagi, 2008; Mednick, Arman & Boynton, 2005; Mednick et al., 2002). The re-test was administered 30 minutes after sleep termination on the third night, before the subjects would catch up with their sleep. Thus, the re-test was conducted before any REM sleep and following sleep cycles appeared. For each subject, one specific visual field quadrant was assigned as the trained quadrant, either upper left or upper right, randomly. In addition, we measured subjects’ sleepiness during the initial and re-tests by using the Stanford Sleepiness Scale (SSS) (Hoddes et al., 1973).

On the fourth day, MRI scans were conducted to obtain anatomical brain images for visualization and MEG source modeling, as well as fMRI data for retinotopic mapping to determine the boundaries between the early visual areas, including the V1, V2, and V3. This method was consistent with previous researches (Engel et al., 1994; Fize et al., 2003; Sereno et al., 1995; Shibata et al., 2011; Yotsumoto et al., 2009).

2.3. Texture discrimination task

TDT was used for the behavioral training session that took place on the third night (Fig. 2). The task was conducted in a dimly lit room. Subjects’ heads were restrained in a chin rest and visual stimuli were displayed on the computer screen at a viewing distance of 57 cm. Each trial began with the presentation of a fixation point at the center of the screen (1000 ms). Then the TDT stimulus was briefly presented (13 ms), followed by a blank screen of varying duration, and finally, a mask stimulus (100 ms). The blank interval between the TDT and mask stimuli is referred to as the stimulus-to-mask onset asynchrony (SOA). When the mask stimulus is presented, new information processing in the retina overrides what remains of the TDT stimulus. Therefore, the difficulty of the task increases with decreasing SOAs. The size of the TDT display was 19° and contained a 19 × 19 array of horizontal background bars. Each bar was jittered by 0.2°. Each TDT stimulus had two components: at the center of the display, either a letter ‘L’ or ‘T’ was presented, and within the trained quadrant at a 5-9° eccentricity, three diagonal bars were presented. The diagonal bars formed either a horizontal or vertical array on the background of horizontal bars. After each trial, subjects used a button box to report whether the central letter was ‘L’ or ‘T’, and whether the three diagonal bars were aligned ‘horizontally’ or ‘vertically’. The letter task was designed to control subjects’ fixation. Mask stimuli were composed of randomly rotated v-shaped patterns. After the subject’s response, a feedback sound was delivered to indicate whether the central letter task was correct or incorrect. No feedback was given for responses on the diagonal bar task.

Fig. 2. Texture discrimination task.

In the initial test and re-test, there were seven SOAs (400, 300, 200, 160, 130, 100, 80 ms), each presented in a block of 20 trials. The difficulty of the task thus increased from the beginning to the end of the task. After the initial test, subjects participated in an intense training session. The training session was composed of 1,620 trials with 13 SOAs (400, 300, 230, 200, 180, 170, 160, 150, 130, 120, 110, 90, 80 ms). The first 2 easiest SOAs were presented in the first 2 blocks, while the remaining 11 SOAs were distributed over 7 blocks. Each block consisted of 20 trials. SOA decreased sequentially. The training session took approximately 90 minutes.

2.4. MEG and PSG acquisition

MEG signals created by spontaneous brain oscillations were collected with a 306-channel Neuromag Vectorview system (Elekta Neuromag, Helsinki, Finland), which has a magnetometer and two planar gradiometers at 102 sites. Along with MEG, PSG was recorded to score the sleep stage. PSG included electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), and electrocardiogram (ECG). EEG was collected from four sites (C3, C4, O1, O2) referenced to the nasion. For EOG recording, electrodes were attached both above and below the left eye and at the outer canthi of the two eyes. EMG and ECG were recorded from the mentum and the chest, respectively. To co-register the anatomical MRI and the MEG data, four head position indicators (HPI) were attached on the subject’s scalp. The initial head position was measured at the beginning of the MEG recording and tracked during the entirety of the sleep session (Uutela, Taulu & Hamalainen, 2001).

Subjects were asked to lie down in the bed with their head positioned inside the MEG dewar. Five minutes of data from the room void of a subject were acquired either before or after the MEG recording to estimate the noise covariance matrix for MEG source analysis. The sampling rate for MEG and EEG was 601Hz and the electrode impedance was maintained at smaller than 15kΩ.

2.5. MR image acquisition and analysis

Subjects were scanned in a 3-Tesla MRI scanner (Trio, Siemens). For the anatomical reconstruction, high-resolution T1-weighted MR images (MPRAGE; TR = 2.53 sec, TE = 3.28 ms, flip angle = 7°, TI = 1100 ms, 256 slices, voxel size: 1.3 × 1.3 × 1.0 mm³) were acquired. For retinotopic mapping, functional MR images were acquired using gradient echo EPI sequences (TR = 2 sec, TE = 30 ms, flip angle = 90°, voxel size: 3 × 3 × 3.5 mm³). Thirty-three contiguous slices oriented parallel to the AC-PC plane were acquired to cover the whole brain.

T2 functional images were corrected for motion (Cox & Jesmanowicz, 1999), smoothed with a Gaussian kernel of 5.0 mm (FWHM), and normalized individually across scans. These functional data were co-registered to the individual T1-weighted image (Dale, Fischl & Sereno, 1999; Fischl, Sereno & Dale, 1999) and the average signal intensity maps were calculated in each condition for each subject. To conduct voxel-by-voxel statistical tests, we computed contrasts based on a univariate general linear model. We projected significance levels onto the flattened cortex map individually (Dale, Fischl & Sereno, 1999; Fischl, Sereno & Dale, 1999) and localized ROIs (see below) based on these maps.

2.6. Retinotopic mapping and regions of interest (ROI)

To localize the regions of interest (ROI), we used an fMRI retinotopic mapping technique for each subject. It is important to note that our ROIs were determined functionally, not anatomically. While subjects were scanned in the MRI scanner, a flickering checkerboard pattern was presented at vertical and horizontal meridians in a block design (Choi et al., 2012; Engel et al., 1994; Fize et al., 2003; Sereno et al., 1995; Shibata et al., 2012; Shibata et al., 2011; Yotsumoto et al., 2009; Yotsumoto, Watanabe & Sasaki, 2008). Based on the blood-oxygen-level-dependent (BOLD) signal contrast between these two meridians, the four visual quadrants of V1, V2, and V3 were localized for every subject (Engel et al., 1994; Fize et al., 2003; Sereno et al., 1995; Shibata et al., 2011; Yotsumoto et al., 2009). Additionally, annulus stimuli were used to localize the visual field of 5°- 9° eccentricity within each quadrant (Yotsumoto et al., 2009; Yotsumoto, Watanabe & Sasaki, 2008). This 5°- 9° eccentricity corresponds to the size of the trained visual field quadrant. Through this analysis, four sub-areas of V1, V2, and V3, which correspond to the visual field representations of upper left, upper right, lower left and lower right within 5°- 9° eccentricity, were acquired. Among the four sub-areas, one ROI corresponding to the subject’s trained visual quadrant was labeled as the trained region, whereas the other three ROIs corresponding to the three untrained visual quadrants were combined into one ROI and labeled as the untrained region. We also combined the trained and untrained regions as a reference to normalize the activation of the separate trained and untrained regions (see below). Therefore, there were 9 ROIs in total: the trained, the untrained, and the reference regions in V1, V2, and V3.

2.7. PSG and MEG data analysis

The sleep stages from PSG data were scored as wake, stage 1, stage 2, SWS or REM sleep for every 30-second epoch in accordance with the American Academy of Sleep Medicine (AASM) scoring manual (Iber, 2007). In the subsequent analysis, only the epochs scored as NREM sleep (stage 2 and SWS) were used.

We analyzed MEG data recorded from 306 channels. To estimate MEG power sources in the individual cortical surface using anatomical MRI, we used recently developed technique, which combines a cortically constrained minimum norm estimate (MNE) and spectral analysis utilizing Morlet wavelets (Ahveninen et al., 2007; Lin et al., 2004). To compensate for head motion and to reduce noise, the MEG data were processed with the maxfilter software (Elekta-Neuromag, Helsinki, Finland). Then, the data were band-pass filtered between 0.1 Hz and 50 Hz. The continuous Morlet wavelet transform was applied to the filtered data and the spectral power was calculated for every 30 second-epoch per each of the chosen frequency bands (Ahveninen et al., 2007; Lin et al., 2004). The frequency bands that we analyzed were delta (0.5-1.5 Hz), slow sigma (11.5-12.5 Hz) and fast sigma (13.5-14.5 Hz), as mentioned above.

In order to calculate a location-specific power for each ROI, we averaged spectral power across every voxel within each ROI per epoch. Then we averaged each ROI’s spectral power across every epoch within stage 2 sleep and SWS, respectively. To normalize the mean power of trained and untrained ROIs for each sleep stage between the pre- and post-training sleep sessions, we divided it by the mean power of reference ROI of the corresponding night session’s sleep stages. Then we computed the power change index for trained and untrained ROIs by subtracting the normalized mean power of each ROI of pre-training sleep from the normalized mean power of the same ROI of post-training sleep:

$$\text{Power change index}=\mathrm{p}(\mathrm{i},\text{post})∕\mathrm{p}(3,\text{post})-\mathrm{p}(\mathrm{i},\mathrm{pre})∕\mathrm{p}(3,\mathrm{pre})$$

where p(i, post/pre) represents mean power within region i (1 = trained, 2 = untrained, 3 = reference) computed from either post/pre- training. The power change index was calculated for trained and untrained V1, V2, and V3 during stage 2 sleep and SWS, separately. Since we were interested in delta, slow sigma, and fast sigma bands, the power change index was obtained for these three frequency bands.

3. Results

3.1. Sleep structure for pre- and post-training sleep sessions

To confirm that the first night effect did not confound the sleep structure between the pre- and post-training sleep sessions, a paired t-test was conducted for each variable. As shown in Table 1, there were no significant differences in any of these sleep variables between the pre- and post-training sleep sessions. Thus, we confirmed that the first night effect did not confound the sleep structure in the pre- and post-training sleep sessions.

Table 1. Sleep variables. Sleep efficiency was measured as the ratio of total sleep time to the amount of time spent in bed. No significant difference was found between pre-training and post-training sleep for any of the sleep variables (paired t-test).

	Pre-training sleep		Post-training sleep		P value
	mean	s.e.m.	mean	s.e.m.
Total sleep time (min)	80.1	4.16	77	2.53	0.47
Wake (min)	10.8	2.96	6.88	1.69	0.12
Stage 1 (min)	19.3	3.53	16	3.78	0.44
Stage 2 (min)	22.5	2.69	22.5	2.83	0.99
SWS (min)	37.1	4.68	36.8	5.17	0.94
REM sleep (min)	1.17	0.81	1.66	0.92	0.71
Waking after sleep onset (min)	3.92	1.62	3.04	1.48	0.63
Wake (%)	11.4	3.2	7.79	1.78	0.19
Stage 1 (%)	21	3.92	18.5	4.21	0.59
Stage 2 (%)	24.8	2.93	26.7	3.01	0.61
SWS (%)	41.4	5.3	45	6.35	0.55
REM sleep (%)	1.19	0.74	1.95	1.05	0.58
Waking after sleep onset (%)	4.18	1.81	3.41	1.59	0.70
Sleep efficiency (%)	88.7	3.22	92.2	1.78	0.22

3.2. Behavioral results and sleepiness

We calculated the percentage of correct responses for the diagonal bar task for each SOA. A logistic function was fitted to each individually measured psychometric data to determine the threshold SOA corresponding to 80% correct performance. This was done for both the initial test and re-test using the psignifit toolbox (ver. 2.5.6) for Matlab (http://bootstrap-software.org/psignifit/). This software package implements the maximum-likelihood method described by Wichmann and Hill (2001). To test whether learning occurred in the trained quadrant, the 80% threshold SOAs in the initial test and re-test were compared (Fig. 3). There was a significant difference between the initial and re-tests, indicating that learning occurred after sleep (2-tailed paired t-test, t(14) = 3.571, p = 0.003). The result of the training session is shown in Fig. 4.

Fig. 3. The performance improvement after sleep.

The 80% threshold SOA for the initial (red) and re-test (blue). Error bars are SEM (n=15). *** p = 0.003

Fig. 4. Line plot representing the percentage of correctness (%) as a function of SOA (ms) in the training session. Error bars are SEM (n=15).

The results of the SSS at the initial test and re-test did not indicate a significant difference. The average (± SE) score of the SSS was 2.6 ± 0.36 for the initial test and 2.9 ± 0.27 for the re-test (2-tailed paired t-test, t(14) = −0.576, p = 0.57). The higher the score of the SSS is, the sleepier the subject is (Hoddes et al., 1973). The score of the SSS suggests that the subjects were sleepier at the re-test than at the initial test, although it was not statistically significant. In contrast, the performance of TDT at the re-test was significantly better than at the initial test. Thus, the SSS results suggest that the performance improvement was not caused by higher alertness at the re-test.

3.3. MEG results

The power change index of each frequency band (delta, slow sigma, and fast sigma) was calculated for each of the trained and untrained regions of early visual areas (V1, V2, and V3) and sleep stages (stage 2, and SWS). We used a four-factor linear mixed-model regression to examine whether there were main effects of sleep stage (stage 2, and SWS), frequency (delta, slow sigma, and fast sigma), training (trained, and untrained) and location (V1, V2, and V3), as well as interactions between them during the first NREM sleep cycle, on a power change index, as the dependent measure. In this model, sleep stage, frequency, training, and location were fixed factors, and subject was a random factor. We found significant interactions in the sleep stage × frequency (F_2,486 = 3.179, p = 0.042) and sleep stage × frequency × training (F_2,486 = 7.491, p = 0.001), while none of significant main effects of sleep stage, frequency, training and location was found.

Since these two- and three-factor interactions were significant, we applied a two-factor linear mixed-model regression (factors: frequency and training) to the data for each sleep stage (stage2 and SWS), separately. During sleep stage 2 (Fig. 5A), we found a significant main effect of frequency (F_2,264 = 3.360, p = 0.036), and training (F_1,264 = 4.387, p = 0.037). In addition, there was a significant interaction between training and frequency on the power change index (F_2,264 = 7.531, p = 0.001). Given the interaction between the training and frequency, Tukey’s post-hoc test, which corrects for an experiment-wise error rate, revealed that the power change index of slow sigma activity was significantly greater in the trained region compared to the untrained region (p = 0.007), whereas no significant difference was found in the power change index of delta and fast sigma activity between the trained and untrained regions.

Fig. 5. Mean (± 1 SEM) change index for the trained (red) and untrained (blue) regions in early visual areas during stage 2 (A) and during SWS (B). ** p < 0.01, *** p < 0.005.

In contrast, no significant main effect was observed during SWS (Fig. 5B), although there was only a marginal interaction between training and frequency (F_2,246 = 2.629, p = 0.074). This suggests that none of three frequencies’ activity changed meaningfully between the trained and untrained regions during SWS.

Among three frequency bands during sleep stage 2, only the slow sigma activity satisfied the first criterion: the power increase should be greater in the trained region of early visual areas than in the untrained region. Therefore, we further tested whether slow sigma activity satisfies the second criterion, which is that the power of the trained region should be higher in the post- than in the pre-training sleep. The following test showed that the normalized mean power of slow sigma activity within the trained region during sleep stage 2 was significantly greater in the post- than in the pre-training sleep (2-tailed paired t-test, t(44) = 3.125, p = 0.003). This indicates that the slow sigma activity increased after training specifically within the trained region of early visual areas.

Finally, we tested whether the changes in the slow sigma band were correlated with the performance improvement after sleep. Interestingly, we found a significant correlation between the performance improvement and the difference of slow sigma power change indices between the trained and untrained regions during sleep stage 2 (2-tailed Pearson’s correlation, r = 0.6934, p = 0.004, n = 15, Fig. 6). The Grubbs’ test for outliers in the performance improvement and in the difference of slow sigma power change indices (p = 0.608 and p = 0.609, respectively) indicated that there is no outlier in the plot. This implies that the larger the discrepancy of power increases between the trained and untrained regions, the more the subject learned. In contrast, none of the difference of delta, and fast sigma’s power change indices during sleep stage 2, and the difference of delta, slow, and fast sigma’s power change indices during SWS, showed any correlation with performance improvement (r = −0.183, p = 0.514; r = 0.375, p = 0.169; r = −0.274, p = 0.342; r = 0.051, p = 0.863; r = 0.185, p = 0.527, respectively).

Fig. 6. Scatter plot representing the correlation between threshold improvement and the difference in sigma band power indices between trained and untrained regions within early visual areas (averaged across V1, V2, V3) during sleep stage 2.

4. Discussion

In the present study, we took advantage of the location specificity that occurs in VPL of the TDT and investigated whether slow/fast sigma or delta band activity was involved in the consolidation of VPL during NREM sleep, using MEG, MRI and PSG in combination. As mentioned earlier, we tested whether the slow/fast sigma or delta activity satisfied three criteria: (1) greater power increase in the trained compared to the untrained region of early visual areas, (2) greater power in the post- compared to the pre-training sleep sessions, and (3) positive correlation between performance improvement and the difference of power increase between the trained and untrained regions of early visual areas. The behavioral test indicated significant off-line learning after sleep, as has been shown previously (Gais et al., 2000; Karni & Sagi, 1993; Stickgold, James & Hobson, 2000; Stickgold et al., 2000; Yotsumoto et al., 2009). Among the slow/fast sigma and delta activities, we found that only slow sigma activity, specifically within sleep stage 2, satisfied the criteria. On the other hand, none of slow/fast sigma and delta activity satisfied the criteria during SWS.

Our results indicate that slow sigma activity during sleep stage 2 is involved in the consolidation of TDT, at least during the first sleep cycle. It has been suggested that sleep spindles play a critical role in sleep-dependent consolidation of various kinds of learning and memory (Clemens, Fabo & Halasz, 2005, 2006; Fogel & Smith, 2006; Gais et al., 2002; Molle et al., 2009; Morin et al., 2008; Nishida & Walker, 2007; Rosanova & Ulrich, 2005; Schabus et al., 2004; Schmidt et al., 2006; Wamsley et al., 2012). Since sigma activity corresponds to sleep spindles, our results are consistent with the other studies that have suggested that sleep spindle activity is involved in consolidation (Clemens, Fabo & Halasz, 2005, 2006; Fogel & Smith, 2006; Gais et al., 2002; Molle et al., 2009; Morin et al., 2008; Nishida & Walker, 2007; Rosanova & Ulrich, 2005; Schabus et al., 2004; Schmidt et al., 2006; Wamsley et al., 2012). Increased spindle activity was found after the learning of a declarative memory task (Gais et al., 2002; Molle et al., 2009; Schabus et al., 2004; Schmidt et al., 2006), a motor procedural task (Fogel & Smith, 2006; Morin et al., 2008) and was correlated with overnight gains in verbal and motor tasks (Clemens, Fabo & Halasz, 2005, 2006; Nishida & Walker, 2007; Schabus et al., 2004). Our finding adds to the growing body of evidence suggesting that spindle activity is involved in the consolidation of a visual task, and is associated with the overnight improvement in TDT performance.

Because spindle-associated spike trains can induce LTP in cortical synapses (Rosanova & Ulrich, 2005), it has been suggested that newly acquired memories are “replayed” during sleep via sleep spindles (Bergmann et al., 2012; Born, Rasch & Gais, 2006; Born & Wilhelm, 2012; Hasselmo, 1999; Rasch et al., 2007). The replay of memory traces during sleep has been reported in both animal and human studies. In rats and zebra finches, the bursting patterns from large numbers of neurons observed during maze tasks or song learning were reactivated during subsequent NREM sleep (Dave & Margoliash, 2000; Ji & Wilson, 2007; Margoliash, 2010; Wilson & McNaughton, 1994). In human studies, reactivation-like enhanced activity was observed in the hippocampus after learning a spatial task (Peigneux et al., 2004), in the neocortex and hippocampus after learning face-scene associations (Bergmann et al., 2012), and in the trained region of V1 during NREM sleep after learning TDT (Yotsumoto et al., 2009). Importantly, it has been found that enhanced BOLD signals are time-locked with sleep spindles (Bergmann et al., 2012). These findings suggest that the trained visual representations used in our study were reactivated through sleep spindles during sleep stage 2, although the question of which representation the neurons were “replaying” presents an exciting avenue of future research.

There are two major models of consolidation processes during sleep: the active system consolidation model and the synaptic homeostasis model. The results of the present study favor the active system consolidation model, since the enhanced slow sigma power we observed corresponds to slow sleep spindle activity. It implies that an active reactivation process occurred in the trained region of early visual areas during sleep. However, several discrepancies between this model and our results must be noted. First, significant power increases in the slow sigma band were restricted to sleep stage 2, whereas the model assumes that reactivation occurs during both sleep stage 2 and SWS (NREM sleep). Therefore, it is not clear whether this present stage 2 specific slow sigma power increase is consistent with the active system consolidation model. Second, delta band activity, corresponding to slow-waves, did not increase during either sleep stage 2 or SWS, whereas the model suggests that slow-waves also play a role in the reactivation and redistribution process. Slow-waves temporally group neuronal activity to depolarizing up-states not only in the neocortex, but also in various brain structures via efferent pathways (Steriade, 2006). Thus, slow-waves may provide a temporal frame globally, during which the up-states drive reactivation in neocortex in parallel with other brain regions (Born & Wilhelm, 2012). The active system consolidation model highlights this feature of slow-waves because this process is thought to transfer reactivated memory information, particularly in the declarative memory system between the temporary and long-term storages. However, this redistribution concept is yet to be confirmed in VPL. Therefore, future studies are needed to examine how the consolidation process differs between VPL and declarative memory systems.

Our results do not directly support the synaptic homeostasis model, which assumes that slow-waves are a key component of consolidation (Tononi & Cirelli, 2003, 2006), because we found no significant delta power increase. However, it is still possible that the process proposed by the synaptic homeostasis model acts in concert with reactivation at the neuronal level in subsequent NREM sleep cycles, which were not measured in the present study.

This study separated sleep spindle activity into slow and fast spindles based on previous findings (Anderer et al., 2001; Schabus et al., 2007; Schabus et al., 2008; Tamaki et al., 2009; Werth et al., 1997; Zeitlhofer et al., 1997; Zygierewicz et al., 1999). Slow and fast sleep spindles differ not only by frequency but also by the scalp distribution. Slow sleep spindles appear dominantly over the frontal part of the brain, whereas fast sleep spindles prevail over the centro-parietal region (Anderer et al., 2001; Schabus et al., 2007; Werth et al., 1997; Zeitlhofer et al., 1997). The present study found the involvement of the slow sigma band from the occipital region in the consolidation of VPL. Thus, the locus of slow sigma oscillations in the present study may seem inconsistent with the previous studies that reported predominance of slow spindles over the frontal region. However, it is questionable whether each type of sleep spindle serves to consolidate learning only within their known dominant loci of the brain. In the case of fast spindles, their predominant area was consistent with the region where they consolidate learning: fast sleep spindles were found to be involved in the consolidation of motor learning through the activation in motor related areas, which are located at the central region (Barakat et al., 2011; Schabus et al., 2007; Tamaki et al., 2013; Tamaki et al., 2008, 2009). On the other hand, slow spindles’ dominant loci seem to be different from the region where they consolidate VPL. This implies that roles of slow spindles in consolidation could go beyond the dominant loci depending on the task and the brain areas where the task related consolidation occurs.

One may wonder whether the present results are inconsistent with a previous study that showed a negative relationship between sleep spindles and perceptual learning (Mednick et al., 2013). We argue that these studies are not necessarily inconsistent with each other, given that there are several significant methodological differences. First, a pharmacological intervention was used to enhance sleep spindle activity during the daytime nap in the previous study (Mednick et al., 2013), whereas the present study investigated the endogenous sigma oscillations during nocturnal sleep. It is an open question whether the spindle activity caused by a pharmacological intervention during a daytime nap has the same neural origin or plays the same role in learning as endogenous spindles during the night.

Second, brain regions where the spindle activities originated from may differ between the present and previous studies (Mednick et al., 2013). Mednick et al. (2013) visually counted the number of sleep spindles observed in EEG channels C3 and C4. In contrast, we investigated the sigma power, which was source-localized in early visual areas. Indeed, Mednick et al (2013) has suggested that spindle activity observed in the central region in their study is crucially involved in hippocampal activity. In contrast, the sigma activity investigated in the current study represents activity in the early visual cortex. It has been found that activity in the early visual areas change as VPL of TDT occurs (Schwartz, Maquet & Frith, 2002; Walker et al., 2005; Yotsumoto, Watanabe & Sasaki, 2008), especially during the following sleep after the initial training (Yotsumoto et al., 2009). Given these differences, it would be premature to state that the previous (Mednick et al., 2013) and present studies are inconsistent in terms of the relationship between VPL and spindle activity.

One may also wonder whether the NREM effect shown in the present study reflects mere amelioration of deterioration, rather than improvements as a result of the consolidation process. The deterioration here refers to the worsening in performance compared to the previous session, due to fatigue caused by the intensive training session (Censor, Karni & Sagi, 2006; Censor & Sagi, 2008; Mednick, Arman & Boynton, 2005; Mednick et al., 2002). However, it is unlikely that the current results reflect only amelioration of deterioration for the following reasons. First, in our study, the initial TDT test was separated from training, so that the initial performance threshold would not be masked by any fatigue effects, which may occur with a large number of trials during training (Censor, Karni & Sagi, 2006; Censor & Sagi, 2008; Mednick, Arman & Boynton, 2005; Mednick et al., 2002). In addition, there were only 140 trials in the initial test of the present study. A previous study suggested that a total of 450 trials per session may not cause deterioration (Censor & Sagi, 2008). Second, in our earlier work (Yotsumoto et al., 2009) that included careful control experiments, the performance improvement, but not deterioration, was found in TDT with a training period of 1760 trials, which was larger than the current number of trials (1620 trials) in training. Third, more importantly, we have found a significant and high correlation between performance improvement and the difference in slow sigma power change indices, indicating a tight connection between the sleeping brain and learning (Fig. 5). Thus, the NREM effect in the present study is highly likely to represent a neural process involved in the consolidation of TDT.

It should be noted that an accumulating body of evidence suggests a multifaceted consolidation process of sleep in facilitating VPL (Stickgold et al., 2000). Although we restricted our study to the effect of the first sleep cycle’s NREM sleep only, this does not necessarily indicate that we do not acknowledge possible roles of REM sleep (Karni et al., 1994; Mednick, Nakayama & Stickgold, 2003; Stickgold et al., 2000). The possible roles of REM sleep and later sleep cycles in facilitating VPL need to be addressed in the near future.

The present study investigated which component of NREM sleep is involved in the consolidation of TDT, as a representative task of VPL. However, the neural mechanism of VPL in general has been controversial. Perhaps, this is because the neural process involved in VPL is affected by various factors (Lu et al., 2011; Sagi, 2011; Sasaki, Nanez & Watanabe, 2010). The neural site of VPL may shift and change over the course of training. Yotsumoto et al. (2008) showed that early visual areas were significantly involved in the early phase of training. However, in the later phase of training, the activation level of early visual areas returned to the baseline. Other researchers argue that VPL reflects deployment of attention (Xiao et al., 2008; Zhang et al., 2010a; Zhang et al., 2010b) or changes in the process to read out sensory information by decision areas (Huang, Lu & Dosher, 2012; Law & Gold, 2008; Petrov, Dosher & Lu, 2006). It should be noted that in the present study, we focused on the brain activation of early visual areas during sleep, but not higher brain regions. Therefore, integrative studies are needed to fully understand the mechanisms of VPL in the near future.

5. Conclusions

The present study suggests that slow sigma band activity corresponding to slow spindles is involved in the consolidation of a texture discrimination task, at least in sleep stage 2 during the first sleep cycle. This finding is not entirely consistent with either the active system consolidation model or the synaptic homeostasis model, suggesting that further theoretical and experimental work is necessary to investigate this neuronal process.

Highlights.

• Which sleep component serves to consolidate visual perceptual learning (VPL)? (79/85 chrct)

• Slow sigma activity in the early visual areas was significantly changed after VPL (82/85)

• The change in slow sigma activity was correlated with improvement of VPL (74/85)

• Slow sigma activity during sleep stage 2 was involved in the consolidation of VPL (84/85 chrct)

• Which sleep component serves to consolidate visual perceptual learning (VPL)?

• Slow sigma activity in the early visual areas was significantly changed after VPL

• The change in slow sigma activity was correlated with improvement of VPL

• Slow sigma activity during sleep stage 2 was involved in the consolidation of VPL