Abstract
How sleep leads to offline performance gains in learning remains controversial. A use-dependent model assumes that sleep processing leading to performance gains occurs based on general cortical usage during wakefulness, whereas a learning-dependent model assumes that this processing is specific to learning. Here, we found evidence that supports a learning-dependent model in visual perceptual learning (VPL) in humans (both sexes). First, we measured the strength of spontaneous oscillations during sleep after two training conditions that required the same amount of training or visual cortical usage; one generated VPL (learning condition), while the other did not (interference condition). During a post-training nap, slow-wave activity (SWA) and sigma activity during non-rapid eye movement (NREM) sleep and theta activity during REM sleep were source localized to the early visual areas using retinotopic mapping. Inconsistent with a use-dependent model, only in the learning condition, sigma and theta activity, not SWA, increased in a trained region-specific manner and correlated with performance gains. Second, we investigated the roles of occipital sigma and theta activity during sleep. Occipital sigma activity during NREM sleep was significantly correlated with performance gains in presleep learning; however, occipital theta activity during REM sleep was correlated with presleep learning stabilization, shown as resilience to interference from postsleep learning in a trained region-specific manner. Occipital SWA was not associated with offline performance gains or stabilization. These results demonstrate that sleep processing leading to performance gains is learning dependent in VPL and involves occipital sigma and theta activity during sleep.
SIGNIFICANCE STATEMENT The present study shows strong evidence that could help resolve the long-standing controversy surrounding sleep processing that strengthens learning (performance gains). There are two conflicting models. A use-dependent model assumes that sleep processing leading to performance gains occurs because of general cortical usage during wakefulness, whereas a learning-dependent model assumes that processing occurs specifically for learning. Using visual perceptual learning and interference paradigms, we found that processing did not take place after general cortical usage. Moreover, sigma activity during non-rapid eye movement (REM) sleep and theta activity during REM sleep in occipital areas were found to be involved in processing, which is consistent with the learning-dependent model and not the use-dependent model. These results support the learning-dependent model.
Introduction
Numerous studies have demonstrated that sleep is beneficial for various types of learning and memory (Walker et al., 2002; Mednick et al., 2003; Born and Wilhelm, 2012). Individuals show better performance in visual perceptual learning (VPL), which is defined as long-term enhanced performance on a visual task (Sasaki et al., 2010; Lu et al., 2011; Sagi, 2011), after sleep than before sleep. Such improvements in learned skills achieved during a sleep period are called offline performance gains (Korman et al., 2007; Albouy et al., 2013; Lugassy et al., 2018; Tamaki et al., 2020b).
There are two major classes of models that have been proposed to account for offline performance gains through sleep: learning-dependent models and use-dependent models. Learning-dependent models postulate that offline performance gains are generated by a neural process that is specifically involved in learning and memory during post-training sleep (Stickgold, 2005; Aton et al., 2009; Rasch and Born, 2013). In contrast, use-dependent models assume that the process resulting in offline performance gains is triggered by a general function of sleep, such as homeostasis. In the synaptic homeostasis hypothesis model (Tononi and Cirelli, 2003), a use-dependent model, the process resulting in offline performance gains simply depends on brain usage during presleep wakefulness and is not specific to learning. Thus, in this model, the process resulting in performance gains occurs in response to brain usage, not to presleep learning per se.
Moreover, the range of the involved spontaneous oscillatory activity underlying offline performance gains is controversial. In learning-dependent models, various ranges of spontaneous oscillatory activity are suggested to play a role, as follows: slow-wave activity (SWA; 1–4 Hz; Mascetti et al., 2013; Tamaki et al., 2013); and sigma activity (13–16 Hz; Wamsley et al., 2012; Bang et al., 2014; Manoach et al., 2016; Tamaki et al., 2019) during non-rapid eye movement (NREM) sleep and theta activity (5–9 Hz) during REM sleep (Nishida et al., 2009; Boyce et al., 2016). In a use-dependent model (Tononi and Cirelli, 2003), SWA is suggested to play a role in offline performance gains. The model assumes that synaptic potentiation during wakefulness is homeostatically downregulated by SWA during sleep because SWA has been known to increase in proportion to the duration of prior wakefulness and decrease during sleep (Borbély, 2001). However, sigma activity during NREM sleep and theta activity during REM sleep are not assumed to play specific roles in the use-dependent model.
Here, we examined whether a learning-dependent or use-dependent model is likely to account for the generation of offline performance gains through sleep in VPL and investigated which spontaneous oscillations are involved in the process. In experiment 1, we created the learning and interference conditions, to dissociate a learning component and a brain usage component by taking advantage of a VPL interference paradigm (Seitz et al., 2005; Yotsumoto et al., 2009b; Shibata et al., 2017). Since the same numbers of trials were conducted before sleep in both conditions, visual usage could be considered equivalent between these conditions. However, learning occurred only in the learning condition and not in the interference condition (Table 1). A learning-dependent model predicts that the process during sleep that leads to offline performance gains occurs only in the learning condition, while a use-dependent model predicts that the process occurs during sleep in both conditions. The process would be represented by increased spontaneous activity in early visual areas.
The brain usage and learning effect in the learning condition versus the interference condition
The results of experiment 1 supported learning-dependent models that involve increased sigma activity during NREM sleep and theta activity during REM sleep in early visual areas. Next, we clarified the roles of occipital sigma and theta activity in experiment 2 by combining the findings with those of our previous studies (Tamaki et al., 2020b). The results of experiment 2 suggest that occipital sigma activity during NREM sleep is involved in offline performance gains in VPL, whereas occipital theta activity during REM sleep is involved in stabilization in VPL.
Materials and Methods
Participants
The present study initially included data from a total of 61 young healthy human subjects: n = 24 for experiment 1 and n = 37 for experiment 2 from two datasets from the previous study (Tamaki et al., 2020b). However, the from a total of 8 subjects were omitted, which made the total number of subjects analyzed 53 (see sections Eligibility and Data omission). All subjects provided written informed consent for their participation in the experiments. This study was approved by the Institutional Review Board at Brown University.
In experiment 1, the subjects were randomly assigned to one of the two conditions, the learning condition (n = 12, 8 females; mean age, 23.8 ± 0.80 years) and the interference condition (n = 10, 8 females; mean age, 24.1 ± 1.03 years; not significant between conditions, independent-samples t test: t(20) = 0.21, p = 0.837). To estimate the number of subjects required to reliably indicate the effect of a nap on offline performance gains in VPL, we applied the G*Power program (Faul et al., 2007, 2009)—set to a power of 0.8, a required significance level (α) of 0.05, and a two-tailed distribution—to our published behavioral data (Tamaki et al., 2019) for a t test. Offline performance gains in VPL were set as the dependent variable, and the t test was set as the test family. The result indicated nine subjects as the minimum sample size. However, we recruited 12 subjects for each condition to ensure that there would be ≥9 data points per condition in experiment 1 even when some data had to be dropped because of ineligibility or being an outlier.
For experiment 2, the total number of subjects analyzed was 31. The data from our previously published experiments (Tamaki et al., 2020b), termed here as prev-Exp. 1 (previous experiment 1) and prev-Exp. 2 (previous experiment 2), were reanalyzed and included in the current study (n = 19 from prev-Exp. 1; n = 12 from prev-Exp. 2; 22 females; mean age, 22.9 ± 0.54 years). We chose these data because they allowed us to measure the correlation between the degree of stabilization of presleep learning and theta activity during REM sleep, or correlations between the degree of performance gains in presleep learning and SWA and sigma activity during NREM sleep (i.e., the data from experiments 1 and 2 in the study by Tamaki et al., 2020b, but not experiments 3 and 4).
Eligibility
We conducted a careful screening process for participation eligibility because various factors are known to influence visual sensitivity and sleep stages (Tamaki et al., 2020b). The eligibility criteria were as follows. First, subjects were required to have normal or corrected-to-normal vision. Second, all subjects had to have no prior experience in VPL tasks, as prior experience in such tasks may cause a long-term visual sensitivity change (Karni and Sagi, 1991; Schwartz et al., 2002; Seitz and Watanabe, 2005; Yotsumoto et al., 2008; Sasaki et al., 2010; Lu et al., 2011; Sagi, 2011). Third, people who frequently play action video games were excluded because extensive video game playing has been shown to affect visual and attention processing (Green and Bavelier, 2003; Li et al., 2009; Berard et al., 2015). Frequent gamers were defined as those who played action video games at least 5 h a week for a continuous period of ≥6 months, as defined by previous research (Green and Bavelier, 2003; Berard et al., 2015). Fourth, subjects who were included had a regular sleep schedule [i.e., differences in average bedtimes and wake-up times between weekdays and weekends were <2 h, and the average sleep duration regularly ranged from 6 to 9 h using a sleep–wake habits questionnaire (Tamaki et al., 2019, 2020b) and the Munich Chronotype Questionnaire (Roenneberg et al., 2003)]. Fifth, based on a self-report questionnaire (Tamaki et al., 2019, 2020b), anyone who had a physical or psychiatric disease, was currently under medication, or was suspected to have a sleep disorder was excluded. No subjects were suspected of having insomnia, sleep apnea, restless leg syndrome, sleep walking, narcolepsy, or REM sleep behavior disorder based on a self-report questionnaire that asked whether the participant had any symptoms known for these sleep disorders (Tamaki and Sasaki, 2019; Tamaki et al., 2020a). Sixth, they had to be between 18 and 30 years of age. This is because age could influence sleep (Mander et al., 2017). We excluded individuals >30 years old because of their possible altered sleep structures from the younger populations (Ehlers and Kupfer, 1997).
Data omission
In experiment 1, two sets of data were removed from the interference condition for the following reasons. One dataset was an outlier (n = 1, performance change on task A (see section Texture discrimination task protocol) approximately −125% at the post-training test, exceeding the criteria by Grubbs' test). The other dataset (n = 1) was because of an ineligibility to participate because of irregular sleep–wake habits. This reduced the total number of subjects in the interference condition to 10.
Experiment 2 originally included 37 subjects but excluded 6 subjects. One subject's data were excluded from the analysis because the electroencephalogram (EEG) data were too noisy to be used. The data from the other five subjects were excluded because these subjects did not show REM sleep; hence, no correlation analysis between theta activity and resilience to interference was possible. In addition, these data were collected under a design that lacked a postsleep test session, thus providing no usable data for offline performance gains.
Procedures
First, we will describe the procedures common to Experiments 1 and 2, including the overall designs and TDT (Fig. 1A; see section Common procedures for experiments 1 and 2) and the procedures specific to each experiment (see sections Specific procedures for experiment 1 and Specific procedures for experiment 2). These sections will also describe how spectral analysis was performed in each experiment. Finally, we will describe how we combined the results of spectral analysis data across experiments (see section Integration of spectral analysis data).
Schematic of the TDT and experimental design in experiment 1. A, Left, An example of the target display. The TDT stimulus consisted of a target array of three-line segments aligned vertically or horizontally (e.g., top left, highlighted in a yellow oval that is not shown in experiments, just for illustration purposes) with the background line segments. There were two tasks: the orientation task and the letter task. The orientation task was the main task, whereas the letter task (highlighted in light blue) was designed to control participants' fixation. Right, A mask stimulus shown for 100 ms was composed of randomly rotated v-shaped patterns. The time interval between the target onset and mask is referred to as the SOA. The SOA was modulated across trials to control the task difficulty. B, Experimental design for the learning (LRN) and interference (INT) conditions. In both conditions, an adaptation sleep session (Ad, yellow) was conducted before the main session (see Materials and Methods). An MRI session was conducted in a separate session to obtain structural brain information and retinotopy mapping (orange). In the LRN condition, the background line orientation of the TDT stimuli was consistent across the two consecutive training blocks (training on task A, pink). In the INT condition, the background line orientation was switched from the first training block (task A, pink) to the second training block (task B, cyan). In both conditions, there were three test sessions: before training (pretraining test), after training, and before sleep (post-training test) and after sleep (postsleep test). In both conditions, subjects took a 90 min nap with PSG for the sleep session.
Common procedures for Experiments 1 and 2
Overall experimental designs.
After each participant consented to participate in the present study, he or she completed the Pittsburgh Sleep Quality Index (PSQI; Buysse et al., 1989) and the Morningness-Eveningness Questionnaire (MEQ; Horne and Ostberg, 1976). The PSQI results indicated that none of the subjects fell into the category of poor sleepers (Extended Data Fig. 3-1). The MEQ results indicated that none of the subjects showed extreme degrees of morningness or eveningness (Extended Data Fig. 3-1).
Before the experiments started, subjects were instructed to maintain their regular sleep–wake habits (i.e., their daily wake/sleep time and sleep duration) until the study was over. To achieve this, first, the sleep–wake habits of subjects were monitored by requiring them to keep a sleep log for 3–7 d before the experiment. If a subject's sleep/wake time differed by >2 h between weekdays and weekends, a scheduled experiment was postponed. Second, we ensured that subjects had not taken a trip to a different time zone within a week before the experiment. On the day before the experiments, all subjects were instructed to refrain from alcohol consumption, unusually excessive physical exercise, and naps. Caffeine consumption was not allowed on the day of the experiments.
An adaptation sleep session was conducted before the main sleep session to mitigate the first-night effect (FNE; Agnew et al., 1966; Carskadon and Dement, 1981; Tamaki et al., 2005a,b, 2014, 2016). The FNE refers to degraded sleep quality that subjects experience when they sleep in a sleep laboratory for the first time (Agnew et al., 1966; Carskadon and Dement, 1981; Tamaki et al., 200a5,b ,2014, 2016). Subjects slept in the same fashion as they did in the main sleep session. The adaptation session was conducted ∼1 week before the main experimental sleep session so that any effects because of sleeping during the adaptation nap would not carry over to the experimental sleep session.
The main sleep session started in the early afternoon for both of the experiments. A sleep session lasted for 90 min followed by a 30 min break. The 30 min break ameliorated sleep inertia (Lubin et al., 1976), which is known to deteriorate performance. During the 30 min break, a questionnaire was administered to obtain subjects' introspection about their sleep (Tamaki et al., 2016). Then, subjects took off the electrodes for polysomnography (PSG; see section PSG measurement) and washed their heads.
Subjective and behavioral sleepiness were measured before each test session. For subjective sleepiness measures, the Stanford Sleepiness Scale (SSS) rating (Hoddes et al., 1973) was used. For behavioral sleepiness measures, the psychomotor vigilance task (PVT; Dinges and Powell, 1985) was used. The PVT lasted for ∼2 min.
Sleep stages were scored every 30 s epoch, following the standard criteria (Rechtschaffen and Kales, 1968; Iber et al., 2007), into stage wakefulness (stage W), NREM stage 1 sleep (stage N1), NREM stage 2 sleep (stage N2), NREM stage 3 sleep (stage N3), and REM sleep (stage R).
Standard sleep parameters were obtained (Extended Data Fig. 3-1) to indicate the general sleep structure for each experiment. Sleep parameters included the duration of each sleep stage, sleep-onset latency (SOL; the latency to the first appearance of stage N2, defined as the sleep onset from lights off), and sleep efficiency (the total percentage of the time spent asleep; Iber et al., 2007).
Texture discrimination task protocol.
The texture discrimination task (TDT; Fig. 1A), a standard and widely used VPL task (Karni and Sagi, 1991), was used. The TDT was conducted in a dimly lit room. The subject's head and chin were restrained by a chin rest. Visual stimuli were displayed on the computer screen at a viewing distance of 57 cm. Stimuli were generated using MATLAB software with Psychtoolbox (Brainard, 1997; Pelli, 1997).
Each trial began with the presentation of a fixation point at the center of the screen (1000 ms). Then, a target display was briefly presented (17 ms), followed by the presentation of a blank screen for a varied duration and then by the presentation of a mask stimulus (100 ms), which was composed of randomly rotated v-shaped patterns. Subjects were instructed to fixate their eyes on the center of the display throughout the stimulus presentation. The size of the target display was 19°, which contained a 19 × 19 array of background lines. Each background line was jittered by 0.2°. The orientation of the background lines was either horizontal or vertical. Each target display had two components: a letter (either “L” or “T”) presented at the central location of the display to help subjects maintain eye fixation; and three diagonal lines (the target array, as the main task) presented in a peripheral location within one of the visual field quadrants (the trained visual field quadrant) at a 5–9° eccentricity. The target array was aligned either horizontally or vertically on the background in the trained visual field quadrant. The trained visual field quadrant was either the upper left or upper right visual field quadrant; the trained quadrant was randomly assigned for the subjects. The trained visual field quadrant was consistent throughout the trials and sessions. After the mask display, subjects used a keyboard to report whether the central letter was 'L' or 'T' and whether the target array was aligned “horizontally” or “vertically.” After the subject's responses for the letter and orientation tasks were recorded, a feedback sound was delivered to indicate whether the letter task was correct (1000 Hz pure beep) or incorrect (400 Hz pure beep). No feedback was given for participants' responses in the orientation task.
The time interval between the target onset and mask, or the stimulus-to-mask onset asynchrony (SOA), was modulated across trials to control task difficulty. There were five or six SOA presentations in both test and training sessions ranging from 316 to 33 ms. In a test session, there were 15 or 20 trials for each SOA. Thus, there were a total of 75–120 trials in a test session. The presentation order of the SOA presentations in test sessions was pseudorandomized to reduce the amount of participant learning and fatigue (Machizawa et al., 2014). In a training session, there were 60 trials for each SOA, and the training session consisted of two blocks. Thus, there were a total of 600–720 trials in a training session. The SOA values decreased in the training sessions.
The background orientations in the two training blocks in the learning condition in experiment 1 were either horizontal or vertical. When orthogonal background orientations were required for the two blocks of TDT (the interference condition in experiment 1 and all data in experiment 2), horizontal and vertical orientations were used. In this case, the background orientation was either horizontal or vertical for the first block (task A), and the background orientation for the second block (task B) was orthogonal to that for task A.
The performance of the TDT was defined as the threshold SOA at which a subject provided 80% correct responses in the orientation task. The threshold SOA values were obtained for each test session for each subject in the following way. The percentage of correct responses for the orientation task was calculated for each SOA in a test session. A cumulative Gaussian function was fitted to obtain a psychometric curve to determine the threshold SOA that corresponded to a correct response rate of 80% using the psignifit toolbox (version 2.5.6) in MATLAB (https://www.nip.uni-tuebingen.de/research/software/psignifit.html; Wichmann and Hill, 2001). All trials in which the letter task was incorrect were removed from the calculation of the threshold SOA.
The performance changes in the TDT were computed based on the relative changes in the threshold SOA values (milliseconds) between test sessions. For instance, the performance change because of training was calculated as [100 (%) × (SOApretraining – SOApost-training)/(SOApretraining)].
Before the first test session, to explain to participants how to perform a TDT, an introductory session was presented. The introductory session was conducted until the subjects reached a certain level of performance. During the introductory session, three long SOA values (800, 600, and 400 ms) were used. The introductory session started with the longest SOA (800 ms) and was followed by 600 and 400 ms SOA values in a blocked fashion, where a set of 20 trials was conducted for each SOA (a total of 60 trials). The introductory session was repeated until the subject could perform the orientation task with at least a 95% accuracy (only one miss in 20 trials) for the 400 ms SOA trials to ensure that all the subjects knew how to perform the task. The average number (±SEM) of introductory sessions was 1.4 ± 0.26 for the learning condition and 1.9 ± 0.53 for the interference condition in experiment 1 (learning vs interference condition: Mann–Whitney U test, U = 54.5, p = 0.674) and 1.5 ± 0.14 in experiment 2 (experiment 1 vs 2: Mann–Whitney U test, U = 327.5, p = 0.776).
In addition, when there was a test session after the sleep session, subjects completed a reminder session (Tamaki et al., 2020b) so that they could remember how to perform a TDT. The reminder session had the same three sets of SOA values as the introductory session.
Specific procedures for experiment 1
Experimental design.
There were three test sessions (Fig. 1B). The first (pretraining) and second (post-training) test sessions were conducted before and after TDT training before the sleep session, and the third (postsleep) test session was conducted after the sleep session.
After the pretraining test session, the subjects underwent intensive training for the TDT. The training session consisted of two blocks (tasks A and B), and the stimuli used for these two blocks differed between conditions, although the total number of trials was the same between conditions (see section Texture discrimination task protocol). In the learning conditions, the same background orientation (horizontal or vertical) was used for the two blocks so that learning occurred in the TDT. In the interference condition, orthogonal background orientations (horizontal and vertical) were used so that no learning occurred in the TDT. After the training session, the post-training test session was conducted. Shortly after participants completed the post-training test session, a sleep session was conducted. During the sleep session, which lasted 90 min, PSG data were recorded (see below). After the 30 min break (see section Common procedures for experiments 1 and 2), the postsleep test session was conducted. There was a short 2 min break between the test and training sessions.
Spectral analysis.
In experiment 1, EEG data were source localized to the cortical mantle using structural brain information obtained from magnetic resonance imaging (MRI; Lin et al., 2004; Ahveninen et al., 2007; Tamaki and Sasaki, 2019). Detailed procedures for PSG measurements (see section PSG measurements), MRI acquisition (see section MRI acquisition), and source localization of EEG signals (see section Source localization of EEG signals), used for computing spontaneous brain activity, are described below.
PSG measurements
The electrodes for PSG measurements were attached, taking ∼45 min, before the introductory session. The PSG measurements consisted of EEG, electro-oculogram (EOG), and electromyogram (EMG) measurements. EEG signals were recorded at 64 scalp sites according to the 10–10 system for electrode positioning (Sharbrough et al., 1991). The data were later source localized using structural brain information (see section Source localization of EEG signals). PSG data were measured using active electrodes (actiCap, Brain Products) with a standard amplifier (BrainAmp Standard, Brain Products). The reference, which was Fz online, was measured with active electrodes and rereferenced to the average of the left (TP9) and right (TP10) mastoids after the recording for analysis. The sampling frequency was 500 Hz. The impedance was kept to <20 kΩ. Passive electrodes (BrainAmp ExG, Brain Products) were used for EOG and EMG recordings. Horizontal EOG signals were recorded using two electrodes placed at the outer canthi of both eyes. Vertical EOG signals were measured using two electrodes 3 cm above and below the left eye. EMG signals were recorded from the mentum. The impedance was maintained at ∼10 kΩ for the passive electrodes. Brain Vision Recorder software (Brain Products) was used to record the data. The data were filtered between 0.1 and 100 Hz. PSG signals were recorded in a soundproof and shielded room.
MRI acquisition
Anatomical MRI data were acquired to determine the conductor geometry for the boundary element model (BEM) of the participant's head (Hämäläinen and Sarvas, 1989) and to register the EEG sensor locations with the individual subject's anatomy (Dale et al., 1999; Fischl et al., 1999). Subjects were scanned in a 3 T Siemens Trio MR Scanner with a 32-channel head coil. Three T1-weighted MR images (MPRAGE) were acquired. Based on these T1-weighted MR images, the cortical surface was reconstructed for each subject for brain parcellation to localize individual gyri and sulci (Fischl et al., 2004).
Functional MRI data for retinotopic mapping (see below) were also acquired to localize the early visual areas using a gradient echoplanar imaging sequence (TR = 2 s, TE = 30 ms, flip angle = 90°). Twenty-five contiguous slices (3 × 3 × 3.5 mm3) oriented orthogonal to the calcarine sulcus were acquired covering the occipital to parietotemporal cortices. Data were analyzed with FSFAST and FreeSurfer (http://surfer.nmr.mgh.harvard.edu) software. All functional images were motion corrected (Cox and Jesmanowicz, 1999), spatially smoothed with a Gaussian kernel of 5.0 mm full-width at half-maximum (FWHM), and normalized individually across scans. Functional data were registered to the individual reconstructed brains (Dale et al., 1999; Fischl et al., 1999).
A standard retinotopic mapping technique (Engel et al., 1994; Yotsumoto et al., 2008) was used to determine the trained and untrained regions in early visual areas of each subject. While the subjects were scanned in the MRI scanner, a flickering checkerboard pattern was presented at the vertical and horizontal meridians with a block design. Based on the blood oxygenation level-dependent signal contrast between these two meridians, the early visual areas (V1 and V2) were localized for each subject. Additionally, annulus stimuli were used to localize the upper left and upper right visual fields of 5–9° eccentricity of the early visual areas, which correspond to the size of the trained visual field (Yotsumoto et al., 2008). Because the present task is known to show trained-location specificity in learning (Karni and Sagi, 1991; Schwartz et al., 2002; Yotsumoto et al., 2009a), the trained region was defined as the cortical region that retinotopically corresponded to the target visual field quadrant (5–9° eccentricity) in the early visual areas and the ventral part of the early visual areas in the contralateral hemisphere to the trained visual field quadrant. The untrained region was defined as the ventral part of the early visual areas (5–9° eccentricity) in the hemisphere ipsilateral to the trained visual field quadrant.
Source localization of EEG signals
To compute the strength of the different brain activity components (SWA, sigma activity, and theta activity) during sleep in the trained and untrained regions in early visual areas (for the definition of the trained and untrained regions see section MRI acquisition), the EEG data were subjected to Morlet wavelet analysis and source localization using the minimum-norm estimate (MNE) of individual MRI information (see section MRI acquisition), as in our previous study (Tamaki and Sasaki, 2019). Morlet wavelet analysis was applied to the raw EEG data (Lin et al., 2004; Ahveninen et al., 2007; Tamaki and Sasaki, 2019) every 30 s to obtain the spectral strength from 1 to 4 Hz (SWA) and from 13 to 16 Hz (sigma activity) during stages N2 and N3, and from 5 to 9 Hz (theta activity) during REM sleep. To localize the current sources underlying the EEG signals, the cortically constrained MNE was used on the EEG data using individual anatomic MRI scans, and the current locations were constrained to the cortical mantle (Lin et al., 2004; Ahveninen et al., 2007; Tamaki and Sasaki, 2019). Information from the EEG sensor locations and the structural MRI segmentation were used to compute the forward solutions for all source locations using a three-layer model of the BEM (Hämäläinen and Sarvas, 1989). The individual forward solutions constituted the rows of the gain (lead-field) matrix. The noise covariance matrix was computed from the raw EEG data for 30 s during wakefulness. These two matrices were used to calculate the inverse operator to obtain the estimated source activity during sleep, as a function of time, on a cortical surface (Lin et al., 2004; Ahveninen et al., 2007). In the first sleep cycle, each of the source-localized strengths of the SWA and sigma bands was averaged across stages N2 and N3, and the strength of the theta band was averaged across REM sleep.
Finally, the power for each frequency band in the untrained region was subtracted from the trained region to obtain the “trained-untrained activity.” We used the trained-untrained activity for various analyses according to previous studies (Tamaki et al., 2020b). Previous studies suggested that spontaneous activity in the trained region consists of both learning-related processes and baseline activity, whereas spontaneous activity in the untrained region consists of only baseline activity during sleep. Thus, if there is a learning-dependent process during sleep, the trained-untrained activity is expected to reveal the learning-dependent process better than using spontaneous activity in the trained region alone.
Exploratory analysis for the frontal and parietal regions.
Additionally, we conducted two kinds of exploratory analyses for the frontal and parietal regions regarding SWA and sleep spindle waves.
First, we tested whether SWA in the frontal and parietal regions differed between the learning and interference conditions. This analysis was performed because SWA has been shown to originate in frontal and parietal regions (Murphy et al., 2009) and may travel to the occipital region. We used a multitaper spectral analysis (Mitra and Pesaran, 1999; Prerau et al., 2017) on frontal and parietal EEG data (for more detail, see section Spectral analysis of EEG). The F3, F5, F7, F4, F6, and F8 channels were treated as channels that represented the frontal region, and the P3, P5, P7, P4, P6, and P8 channels were treated as channels that represented the parietal region (Tamaki et al., 2016).
Second, we tested whether sleep spindle waves in the frontal and parietal regions differed between the learning and interference conditions. We identified sleep spindle waves whose duration was ≥0.5 s, with at least five peaks, and ≥15 µV. It has been shown that slower sleep spindles are dominant over the frontal region, whereas faster sleep spindles are dominant over the parietal regions (Werth et al., 1996; Mölle et al., 2011). Thus, we identified sleep spindle waves in the 10–13 Hz band in the frontal region and sleep spindle waves in the 13–16 Hz band in the parietal region. We counted the mean number, amplitude, duration, and frequency of sleep spindle waves in the frontal and parietal regions in each condition.
The detection of sleep spindle waves was in accordance with previous studies (Tamaki et al., 2008, 2009). First, a Butterworth-type zero-phase digital filter (10–16 Hz) was applied to the EEG recordings to locate troughs and peaks of EEG waves for each electrode during stages N2 and N3. Then, in the filtered EEG data, the points at which the gradient value changed from negative to positive were defined as the troughs of EEG waves; likewise, the points at which the value changed from positive to negative were defined as the peaks of EEG waves. The start and end of consecutive EEG waves were determined by the interval between the troughs of the EEG waves and used as the duration of the waves. The difference between the trough and peak voltages was used as the amplitude of the wave. The waves that fulfilled the criterion (≥15 µV, ≥0.5 s, and ≥5 waves, 10–13 Hz for the frontal or 13–16 Hz for the parietal regions) were defined as the frontal or parietal sleep spindles.
Specific procedures for experiment 2
Experimental design.
Because the purpose of experiment 2 in the present study was to distinguish the roles of occipital sigma and theta activity during NREM sleep and REM sleep, respectively, in VPL, we combined available data from prev-Exp. 1 and prev-Exp. 2 in the previous study (Tamaki et al., 2020b). The experimental designs were based on the interference paradigm of the TDT (Yotsumoto et al., 2009b), which is similar to the interference condition in experiment 1 in the present study. However, the configurations of the two interfering blocks were different: the two blocks were performed before the sleep session in the interference condition in experiment 1, whereas the two blocks were separated by the sleep session in experiment 2. The latter design allowed us to measure the degree of stabilization of presleep learning by sleep. See the previously published study (Tamaki et al., 2020b) for detailed procedures.
The following data were included in the present analyses from prev-Exp. 1 and prev-Exp. 2 (Tamaki et al., 2020b). Among all 31 subjects analyzed from prev-Exp. 1 and prev-Exp. 2, data from 24 subjects were used for the correlation between SWA and sigma during NREM sleep and performance gains (19 subjects from prev-Exp. 1; 5 subjects from prev-Exp. 2), and data from 22 subjects were used for the correlation between theta activity during REM sleep and stabilization (10 subjects from prev-Exp. 1; 12 subjects from prev-Exp. 2. The manner in which we integrated the results of spectral analysis is shown in the section Integration of spectral analysis.
Spectral analysis.
Experiment 2 reanalyzed data from the published study (Tamaki et al., 2020b, prev-Exp. 1 and prev-Exp. 2). Because the primary purpose of the previous study (Tamaki et al., 2020b) was to investigate the magnetic resonance spectroscopy data during the sleep stages in relation to VPL, necessary information for spectral EEG source localization (including individual retinotopic mapping and structural data for surface construction) was not collected. As such, the source localization of EEG signals was not performed in experiment 2. A summary of the procedures for PSG (see section PSG measurements) and spectral analysis (see section Spectral analysis of EEG) is described below.
PSG measurements
The electrodes for PSG measurement were attached, which took ∼30 min, before the introductory session of TDT. There were notable differences in the procedures used to conduct the spectral analysis between prev-Exp. 1 and prev-Exp. 2; these differences were as follows. Because the primary purpose of prev-Exp. 1 was to collect magnetic resonance spectroscopy data while subjects were asleep, EEG data were collected in an MRI scanner. Thus, specific postprocessing for EEG data because of scanner noise and ballistocardiogram artifacts was necessary (for more details see, Tamaki et al., 2020b), as was an additional normalization process in spectral analysis (see below). However, such postprocessing was not necessary for the data from prev-Exp. 2, as these EEG data were collected in a sound-attenuated electrically shielded chamber outside the MRI scanner.
Spectral analysis of EEG
The following procedures were applied to all data in experiment 2 (n = 31). First, a multitaper spectral analysis (Mitra and Pesaran, 1999; Prerau et al., 2017) was performed on the EEG data during the first sleep cycle in 30 s windows (frequency resolution = 1 Hz, the time-half-bandwidth product = 15, the number of tapers = 29) to compute the spectral power from 1 to 4 Hz (SWA) and 13–16 Hz (sigma activity) during stages N2 and N3 and from 5 to 7 Hz (theta activity) during REM sleep. We calculated the power of each frequency band during sleep in occipital regions, including O1, PO3, PO7, O2, PO4, and PO8 EEG channels. Because the present task is known to show trained-location specificity in learning (Karni and Sagi, 1991; Schwartz et al., 2002; Yotsumoto et al., 2009a), when targets were presented in the upper left visual field quadrant in the TDT, the O2, PO4, and PO8 channels in the right occipital area were treated as channels that represented the trained region, and O1, PO3, and PO7 channels in the left occipital area were treated as the untrained region. Analogously, when the TDT targets were presented in the upper right visual field quadrant, O1, PO3, and PO7 channels were treated as the trained region, and O2, PO4, and PO8 channels were treated as the untrained region. Although the MT area was used as the untrained control region in some of the analyses in our previous article (Tamaki et al., 2020b), we did not use the MT area as the untrained region; therefore, the definition of the untrained regions was unified in the present study. Each of the powers of SWA and sigma activity was averaged per sleep stage (N2 and N3) and per region (trained and untrained). Then, each power of SWA and sigma activity in each region was averaged across stages N2 and N3 to obtain the mean power for NREM sleep. Theta power was averaged only within REM sleep.
In addition, the following procedures were performed on the data (n = 19) that were acquired simultaneously with MRI [corresponding to the data from the prev-Exp. 1 in the previous study (Tamaki et al., 2020b)]. We first obtained the power of each frequency band during stage W and then divided the power in each frequency band per region (trained and untrained) per sleep stage (NREM sleep and REM sleep) by the power of each frequency band during stage W to further control the noise level based on a procedure in our previous article (Tamaki et al., 2020b). This normalization was performed because the amounts of MRI-derived noise on EEG data were not strictly consistent across subjects and electrodes because of the known influences of strong and rapidly changing gradient fields and radiofrequency pulses (Ives et al., 1993; Goldman et al., 2000) that cause noise specific to individual electrodes even if cables and equipment were carefully placed in the MRI room.
Integration of spectral analysis data.
Because of the differences in the methods used to obtain spectral data, trained-untrained activity values were transformed to z scores within each experiment and then combined. Trained-untrained sigma activity and SWA were transformed to z scores within each of the prev-Exp. 1 and prev-Exp. 2 datasets and in the learning condition in experiment 1 in the present study. Analogously, trained-untrained theta activity was transformed to z scores within each of the prev-Exp. 1 and prev-Exp. 2 datasets. Then, correlation coefficients between the behavioral measures and z-scored spectral data (see Fig. 4A,C) were obtained.
In addition, forest plots were made for trained-untrained sigma and theta activity in correlation with behavioral measures (see Fig. 4B,D), although the numbers of independent datasets and each sample size were small: three independent datasets (prev-Exp. 1, prev-Exp. 2, and the learning condition in experiment 1 in the present study) for the sigma analysis, and two independent datasets (prev-Exp. 1 and prev-Exp. 2). for the theta analysis.
Statistical analyses
An α-level (type I error rate) of 0.05 was set for all statistical analyses. The Shapiro–Wilk test was conducted before each statistical test to investigate whether the data were normally distributed. If the data were not normally distributed, nonparametric tests, including the Mann–Whitney U test, were used. In experiment 1, the Box's M test was conducted to investigate the equality of covariance before conducting multivariate ANOVA (MANOVA), and Levene's test was conducted to investigate the homogeneity of variance before conducting ANOVA and independent-samples t tests. In experiment 2, before the meta-analysis, we performed the Cochran's Q test to investigate the heterogeneity of the data. Then, we performed the Hunter–Schmidt method, which gives a weighted mean of the raw correlation coefficient (Hunter and Schmidt, 1990; Viechtbaue, 2010).
The assumption of homoscedasticity was met for all the comparisons. When a correction for multiple comparisons was needed, the Bonferroni correction was applied, and the adjusted α-level was shown.
All statistical tests conducted in this study were two tailed. When a test indicated statistical significance, the effect size was shown using Cohen's d (Cohen, 1988) for t tests and partial η2 for ANOVA.
Statistical tests were conducted with SPSS (version 25; IBM), MATLAB (versions R2016b and R2019a; MathWorks), and RStudio (version 1.1.456; RStudio) with the metafor Package (Viechtbaue, 2010).
Results
Experiment 1
We used the TDT (Fig. 1A; see section Texture discrimination task protocol in Materials and Methods), which is one of the standard tasks in VPL (Karni and Sagi, 1991, 1993; Karni et al., 1994; Schwartz et al., 2002; Yotsumoto et al., 2008). Sequential blocks of TDT training with orthogonal background orientations are known to cause anterograde and retrograde interference with each other in learning, resulting in no learning (Yotsumoto et al., 2009b), even when the same trained visual field is consistently used across the sequential training sessions.
Two blocks of TDT training were conducted sequentially in both the learning and interference conditions before the sleep session (Fig. 1B). In the learning condition, the same background orientations were repeated in the two blocks (tasks A and A). In the interference condition, two sequentially presented blocks displayed orthogonal background orientations (tasks A and B). The numbers of trials and the trained visual field quadrant in the training block were the same in these two conditions. The only difference in stimuli between the learning and interference conditions was the background orientation in the second training session; the background orientation was the same as the first block in the learning condition, whereas it was orthogonal to the first block in the interference condition.
Both the learning and interference conditions included a pretraining test, two blocks of TDT training, and a post-training test before participants took a 90 min nap, which was followed by the postsleep test (Fig. 1B). Each of the test sessions measured performance for both tasks A and B. In the test sessions, six SOA values were used (see Materials and Methods). Our metric of performance was the threshold SOA at which subjects gave 80% correct responses in the orientation task. The performance changes from the pretraining to the post-training test ([100 × (SOApretraining – SOApost-training)/SOApretraining]) should show the effect of the two blocks of the training session, while those from the post-training to postsleep test ([100 × (SOApost-training – SOApostsleep)/SOApost-training]) should show the offline performance gains generated by sleep, if there were any.
Performance changes
First, we tested the accuracy of the premise that the learning condition, but not the interference condition, yielded learning by training and that offline performance gains occurred during sleep in the learning condition but not in the interference condition. Two-way mixed-model ANOVA with a between-subjects factor of condition (learning vs interference) and a within-subjects factor of test session (post-training vs postsleep) was conducted on the performance changes for tasks A and B.
For the performance on task A, only the main effect of condition was significant (F(1,20) = 12.21, p = 0.002, partial η2 = 0.379). Neither the main effect of test session (F(1,20) = 0.07, p = 0.791) nor the condition × test session interaction (F(1,20) = 0.02, p = 0.901) was significant. Because of training (Fig. 2A), a significant performance change was obtained in the learning condition but not in the interference condition [one-sample t test; learning condition: t(11) = 3.13, p = 0.010, d = 0.9; interference condition: t(9) = 0.17, p = 0.866, with a Bonferroni-adjusted α-level of 0.025 (0.05/2)]. As a result of sleep (Fig. 2B), the learning condition showed a significant performance change or offline performance gains (one-sample t test: t(11) = 3.16, p = 0.009, d = 0.9), but the interference condition did not [one-sample t test: t(9) = 0.05, p = 0.962, with a Bonferroni-adjusted α-level of 0.025 (0.05/2)].
Behavioral results in experiment 1. A, Mean performance changes on task A (±SEM) from the pretraining to post-training test sessions for the learning (LRN; black, n = 12) and interference (INT; white, n = 10) conditions. B, Mean performance changes on task A (±SEM) from the post-training to postsleep test sessions for the LRN (black, n = 12) and INT (white, n = 10) conditions. C, Mean performance changes on task B (±SEM) from the pretraining to post-training test sessions for the LRN (black, n = 12) and INT (white, n = 10) conditions. D, Mean performance changes on task B (±SEM) from the post-training to postsleep test sessions for the LRN (black, n = 12) and INT (white, n = 10) conditions. E, Boxplots for individual data at the pretraining test session in the LRN (n = 12) and INT (n = 10) conditions for Task A. Each dot represents individual data. In each boxplot, the bottom and the top of the box correspond to the 25th and 75th percentiles (the lower and upper quartiles), respectively. The inner thick gray horizontal line represents the median, and the plus sign represents the mean. The whiskers show the maximum and minimum of the data. F, Boxplots for individual data at the pretraining test session in the LRN (n = 12) and INT (n = 10) conditions for Task B. See the main text for details of statistical results. **p < 0.01.
We also confirmed that the performance on task B did not change in either condition (Fig. 2C,D). A two-way mixed-model ANOVA with the between-subjects factor of condition (learning vs interference) and the within-subjects factor of test session (post-training vs postsleep) was conducted on the performance changes on task B. Neither the main effects (condition: F(1,20) = 1.04, p = 0.321; test session: F(1,20) = 0.11, p = 0.747) nor their interaction (F(1,20) = 0.32, p = 0.576) was significant. Training brought no significant performance change in the learning condition (one-sample t test: t(11) = 0.86, p = 0.411) or in the interference condition (one-sample t test: t(9) = 0.77, p = 0.462). Additionally, sleep brought no significant performance change in either the learning condition (one-sample t test: t(11) = 1.51, p = 0.161) or the interference condition (one-sample t test: t(9) = 0.28, p = 0.787).
These results indicate that our experimental design worked as expected: while the same number of trials were performed during training in both the learning and interference conditions, the effects of training and offline performance gains on presleep VPL occurred only in the learning condition and not in the interference condition.
Spontaneous brain activity
We measured SWA and sigma activity during NREM sleep and theta activity during REM sleep from participants' EEG data in the trained and untrained regions in the early visual areas (V1/V2), which were retinotopically identified (Yotsumoto et al., 2008); importantly, the early visual areas have been indicated to be involved in learning TDT (Karni and Sagi, 1991; Schwartz et al., 2002; Yotsumoto et al., 2008). Moreover, since the TDT shows a retinotopically specific learning effect (Karni and Sagi, 1991; Schwartz et al., 2002; Yotsumoto et al., 2008), the trained and untrained regions in the early visual areas corresponded contralaterally to the trained and untrained visual fields, respectively. The untrained region served as a control region. Then, we calculated the trained-untrained activity by subtracting the activity in the untrained region from the activity in the trained region separately for the SWA and the sigma and theta bands (for more details about how the individuals' cortical structures were incorporated from their MRI data, see section Source localization of EEG signals in Materials and Methods).
First, we tested whether the three bands of trained-untrained activity were significantly different between the learning and interference conditions by one-way MANOVA with a between-subjects factor of condition on trained-untrained activity in three bands (Fig. 3A). We used MANOVA to protect against inflating the type I error rate in the post hoc comparisons. After we confirmed that the Box's M value of 13.054 was not significant (p = 0.100), suggesting equality of covariance, a statistically significant MANOVA effect was obtained (Pillai's trace = 0.617, F(3,16) = 8.59, p = 0.001). 20 subjects' data (n = 10 for the learning and n = 10 for the interference conditions), of a total of 22 subjects, were included in this analysis because two subjects did not show REM sleep (thus, no theta activity data were available for these participants) in the learning condition and their data were excluded. Next, the results of follow-up post hoc one-way ANOVA with the between-subjects factor of condition showed that SWA was not significantly different between the conditions (F(1,18) = 0.99, p = 0.332), whereas sigma activity (F(1,18) = 27.16, p < 0.001, partial η2 = 0.601) and theta activity (F(1,18) = 14.48, p = 0.001, partial η2 = 0.446) were significantly different between the conditions.
Results for trained-untrained activity in experiment 1. A, Trained-untrained activity during NREM and REM sleep (mean ± SEM) in the early visual areas in the learning (black bars, n = 12 for sigma; SWA, n = 10 for theta) and interference (white bars, n = 10) conditions. B, Scatter plots for the offline performance gains on Task A against trained-untrained sigma activity during NREM sleep in the learning condition (n = 12). C, Scatter plots for the offline performance gains on task A against trained-untrained theta activity during REM sleep in the learning condition (n = 10). D, Scatter plots for the offline performance gains on Task A against trained-untrained SWA during NREM sleep in the learning condition (n = 12). Note that the number of subjects in the learning condition was 12 for B and D, and 10 for C, since 2 of 12 subjects did not show REM sleep. See the main text for the detailed results of the statistical tests. ****p < 0.001; ***p < 0.005, **p < 0.01. See also Extended Data Figures 3-1 and 3-2.
Figure 3-1
Sleep parameters for experiments 1 and 2 (mean ± SEM). SE, Sleep efficiency (percentage), defined by the stage W time (minutes) divided by the time interval between lights-off and lights-on; PSQI ranging from 0 to 21, where a score >5 indicates poor sleep quality (Buysse et al., 1989); MEQ, ranging from 16 to 86, where scores from 16 to 30 indicate a definite evening type, scores from 31 to 41 a moderate evening type, scores from 42 to 58 neither type, scores from 59 to 69 a moderately morning type and scores from 70 to 86 a definite morning type (Horne and Ostberg, 1976). Experiment 1: learning condition, n = 12 (n = 10 only for REM sleep); interference condition, n = 10; experiment 2, n = 31 (n = 22 for REM sleep). Here, we used nonparametric tests (Mann–Whitney U test) because normality was not assumed for most of the parameters (stage W, N1, N3, R, SE%, and MEQ). Download Figure 3-1, DOCX file.
Figure 3-2
Frontal and parietal sleep spindles in experiment 1 (mean ± SEM). We detected frontal and parietal sleep spindles of participants in stages N2 and N3 in accordance with previous studies (Tamaki et al., 2008, 2009). Because we could not confirm a normal distribution in several spindle measures, we applied a nonparametric Mann–Whitney U test to investigate whether any spindle measures significantly differed between the learning and interference conditions. There was no significant difference between the conditions in any of the spindle measures. Download Figure 3-2, DOCX file.
In addition, as a supplementary analysis, we performed a one-way ANOVA with a between-subjects factor of condition (learning vs interference) for SWA and sigma activity, including the data of the two subjects who were excluded in the above analysis, as they did not show REM sleep but did show NREM sleep, which meant that SWA and sigma data during NREM sleep were available for these participants. The overall statistical results were the same in essence. SWA was not significantly different between the conditions (F(1,20) = 1.26, p = 0.274), whereas sigma activity was significantly different between the conditions (F(1,20) = 22.07, p < 0.001, partial η2 = 0.525).
Moreover, one-sample t tests against 0 showed that sigma activity [t(11) = 6.37, p < 0.001, d = 1.8, with a Bonferroni-adjusted α-level of 0.017 (0.05/3)] and theta activity [t(9) = 3.86, p = 0.004, d = 1.2, with a Bonferroni-adjusted α-level of 0.017 (0.05/3)], but not SWA (t(11) = 0.48, p = 0.639), were significantly different from zero in the learning condition. The results indicated that spontaneous oscillatory activity in the sigma and theta bands, but not SWA, was increased in the trained regions of the early visual areas in the learning condition.
Performance and spontaneous brain activity
We tested whether the strengths of three bands of trained-untrained activity were correlated with offline performance gains in the learning condition. Both trained-untrained sigma and theta activity were significantly correlated with offline performance gains, but the latter did not survive Bonferroni correction for multiple comparisons [Fig. 3B, sigma activity: r = 0.72, p = 0.009, n = 12, with a Bonferroni-adjusted α-level of 0.017 (0.05/3); no outlier detected by Grubbs' test; Fig. 3C, theta activity: r = 0.70, p = 0.024, n = 10, not significant with a Bonferroni-adjusted α-level of 0.017 (0.05/3); no outlier detected by Grubbs' test]. SWA was not significantly correlated with offline performance gains (Fig. 3D; r=−0.07, p = 0.826, n = 12).
Furthermore, we tested whether increasing the sample sizes would make the correlation between SWA and offline performance stronger. Thus, we merged all the data from the learning and interference conditions and computed the correlation coefficient between trained-untrained SWA and offline performance gains; regardless, SWA was still only weakly correlated with offline performance gains (r = 0.172, p = 0.444, n = 22).
These results clearly demonstrate that the strengths of spontaneous oscillations in the early visual areas are consistent with a learning-dependent model. In particular, sigma activity during NREM sleep, which was significantly increased retinotopically in the early visual areas in the learning condition, was robustly correlated with offline performance gains. While theta activity during REM sleep was not significantly correlated with offline performance gains after Bonferroni correction, the effect size of the correlation was rather large (r = 0.70).
Control tests
We conducted several control analyses to rule out the possibility that these significant differences in performance and brain activity between the learning and interference conditions were because of factors other than experimental manipulations.
First, we tested whether the initial performance levels were different between the two conditions. We found no significant difference in pretraining performance between the conditions (independent-samples t tests; task A: t(20) = 0.56, p = 0.583; task B: t(20) = 0.293, p = 0.772; Fig. 2E,F, Table 2). Thus, it is unlikely that no significant offline performance gains in the interference condition were because of a ceiling effect.
Threshold SOA at test sessions
Second, we examined whether sleepiness was different between the two conditions. We used two measures to assess sleepiness: the SSS (Hoddes et al., 1973) and the PVT (Dinges and Powell, 1985). We found no significant differences in sleepiness between the conditions in terms of the SSS scores (Mann–Whitney independent U test, as the sleepiness measures were not normally distributed; pretraining: U = 55, p = 0.741; post-training: U = 47.5, p = 0.402; postsleep: U = 54, p = 0.674; no corrections for multiple comparisons) or the reaction times in the PVT (independent-samples t tests; pretraining: t(20) = 0.18, p = 0.856; post-training: t(20) = 0.12, p = 0.906; postsleep: t(20) = 0.77, p = 0.452; no corrections for multiple comparisons).
Third, we examined whether any of the sleep parameters were significantly different between the conditions. We did not find any significant differences between the conditions in any of these parameters (Extended Data Fig. 3-1).
Fourth, we tested whether daily sleep habits were different between the conditions. We compared daily sleep quality by the PSQI (Buysse et al., 1989) and individual differences in morningness or eveningness by the MEQ (Horne and Ostberg, 1976). There was no significant difference in participants' PSQI or MEQ scores between the conditions (Extended Data Fig. 3-1).
The results of the series of control analyses showed that the significant differences in behavioral and spontaneous oscillatory measures between the two conditions were unlikely to be attributed to differences in the initial performance levels, sleepiness, sleep quality, or morningness-eveningness between the conditions.
Exploratory analysis of the frontal and parietal regions
One may wonder whether SWA in the frontal and parietal regions might be different between the learning and interference conditions, as SWA originating in the frontoparietal regions might travel toward occipital regions (Murphy et al., 2009) to play a role in VPL. However, this occurrence was not the case. In an exploratory analysis, we obtained SWA power in the frontal and parietal regions and compared the outcomes of the conditions. We found no significant difference in either of the regions (unpaired t tests; frontal: t(20) = 1.84, p = 0.08; parietal: t(20) = 1.03, p = 0.32). Additionally, we found no significant correlation between performance and SWA power (frontal: r = 0.24, p = 0.46; parietal: r = 0.19, p = 0.56). The present results suggested that SWA may not play a role during sleep after VPL in any regions of the brain.
We also performed an exploratory investigation of whether several measures of sleep spindle waves in the frontal and parietal regions differed between the learning and interference conditions. We measured the mean number, amplitude, duration, and frequency of sleep spindle waves (Extended Data Fig. 3-2). Since these measures were not normally distributed, we performed a Mann–Whitney independent U test for each variable. However, none of these measures was significantly different between the conditions (Extended Data Fig. 3-2).
Experiment 2
The results of experiment 1 are consistent with those of learning-dependent models, as only the learning condition showed retinotopically increased spontaneous activity in the early visual areas. The increased spontaneous activity included occipital sigma activity during NREM sleep and theta activity during REM sleep. Interestingly, our previous study (Tamaki et al., 2020b) demonstrated that REM sleep is involved in the stabilization of presleep learning, shown as resilience to retrograde interference from new postsleep learning. An animal study also showed that REM sleep is involved in maintaining new synapses that emerge during development or learning (Li et al., 2017). Together, these results suggested that the learning-dependent process during sleep has a subprocess that is involved in the stabilization of presleep learning during REM sleep in addition to a process involved in offline performance gains. The previous study (Tamaki et al., 2020b) further suggested that occipital theta activity during REM sleep was involved in the process of presleep learning stabilization but not in the process of offline performance gains.
Experiment 2 addressed the question of whether occipital sigma activity during NREM sleep and theta activity during REM sleep play different roles in learning-dependent processes during sleep. First, we tested whether occipital theta activity during REM sleep is involved in the stabilization of presleep learning. Second, we tested whether occipital sigma activity during NREM sleep is replicated and thus expected to be involved in offline performance gains in presleep learning. We sought answers to these questions by including a larger number of data samples than the previous study (Tamaki et al., 2020b). Because the correlation analysis in the previous study focused on spontaneous oscillations and magnetic resonance spectroscopy data, only spontaneous oscillation data obtained simultaneously with magnetic resonance spectroscopy were included in the correlation analysis. Experiment 2 combined all other data generated in the previous study, including data obtained by different procedures (see Materials and Methods for demographic information of participants and the experimental designs). We selected these data because the same visual task (TDT) was used to investigate the stabilization of presleep learning.
The previous experiment (Tamaki et al., 2020b) used an interference paradigm that was slightly different from that in experiment 1, as follows: task A was performed before a nap, whereas task B was performed after a nap (presleep and postsleep training). This design allowed us to calculate the degree of stabilization of presleep learning, shown as resilience to postsleep learning. Then, we tested whether occipital theta activity during REM sleep and the degree of stabilization were correlated using a total of 22 data samples from the previous two experiments. For replication purposes in occipital sigma activity and SWA during NREM sleep and performance gains, we combined 36 data samples in total; that is, 24 data samples from two experiments in the previous study (prev-Exp. 1 and prev-Exp. 2) and 12 data samples from the learning condition in current experiment 1. We obtained trained-untrained sigma activity, SWA, and theta activity, similar to the analysis performed in experiment 1. All data were standardized in each experimental dataset available since slightly different procedures were used to collect spontaneous EEG signals. Then, we obtained correlation coefficients between the degree of stabilization of presleep learning and trained-untrained activity in each of the three activity bands. We also measured correlation coefficients between the degree of offline performance gains and trained-untrained activity in each of the three activity bands [for details about the data from the previous study, see sections Experimental design and Spectral analysis in Materials and Methods (Tamaki et al., 2020b) that were included and how trained-untrained sigma, SWA and theta activity were calculated; for the manner in which we combined data from different experiments, see also section Integration of spectral analysis].
First, trained-untrained theta activity was robustly and significantly correlated with the degree of stabilization of presleep learning (r = 0.49, n = 22, p = 0.021; Fig. 4A). However, none of the trained-untrained sigma activity or trained-untrained SWA was correlated with the degree of stabilization (sigma activity: r = 0.11, n = 31, p = 0.562; SWA: r = 0.15, n = 31, p = 0.427). Second, the combined data corroborated that trained-untrained sigma activity was robustly and significantly correlated with the degree of offline performance gains (r = 0.54, n = 36, p = 0.001; Fig. 4C), whereas trained-untrained SWA activity was not correlated with offline performance gains (r = 0.04, n = 36, p = 0.808) of presleep learning. However, interestingly, trained-untrained theta activity was correlated with offline performance gains (r = 0.45, n = 25, p = 0.023). Note that for the sake of simplicity, corrections for multiple comparisons were not performed with the α-level being 0.05 here and for the results below.
The relationship of occipital theta and sigma activity with behavioral performance in experiment 2. A, Scatter plots between the degree of stabilization and the trained-untrained theta activity during REM sleep (r = 0.49, p = 0.021, n = 22). B, Forest plots for the correlation coefficients between trained-untrained theta activity and the degree of stabilization for each experiment and the mean estimates by the Hunter–Schmidt method. C, Scatter plots between the degree of offline performance gains and the trained-untrained sigma activity during NREM sleep (r = 0.54, p = 0.001, n = 36). D, Forest plots for the correlation coefficients between trained-untrained sigma activity and the degree of offline gains for each experiment and the mean estimates by the Hunter–Schmidt method. Note that the confidence interval of r is not symmetrical when the sample size is small in forest plots. prev-Exp. 1, Data from experiment 1 in the previous study (Tamaki et al., 2020b); prev-Exp. 2, data from experiment 2 in the previous study (Tamaki et al., 2020b); current experiment 1, data from the learning condition in the current experiment 1.
These results raise the possibility that not only occipital sigma activity during NREM sleep, but also occipital theta activity during REM sleep is involved in offline performance gains. Thus, we tested whether theta activity is correlated with offline performance gains when sigma activity is controlled by a partial correlation. We used 25 subjects' data, as these subjects had all three measures for offline performance gains, trained-untrained sigma activity during NREM sleep, and trained-untrained theta activity during REM sleep. Trained-untrained theta activity was not significantly correlated with offline performance gains while controlling trained-untrained sigma activity (r = 0.23, n = 25, p = 0.287). In contrast, trained-untrained sigma activity was significantly correlated with offline performance gains while controlling trained-untrained theta activity (r = 0.52, n = 25, p = 0.010). Because trained-untrained sigma and theta activity were correlated with each other in this dataset (r = 0.48, n = 25, p = 0.015), the results of this analysis suggested that the seemingly significant correlation between trained-untrained theta activity and offline performance gain may have originated from trained-untrained sigma activity.
Analogously, we estimated how much trained-untrained sigma activity contributed to stabilization by performing a partial correlation. For this analysis, we used 22 subjects' data, as they had all measures for stabilization, trained-untrained sigma activity, and trained-untrained theta activity. We found that a partial correlation between trained-untrained theta activity and stabilization while controlling trained-untrained sigma activity tended to be significant (r = 0.43, n = 22, p = 0.054). In contrast, when trained-untrained theta activity was controlled, a partial correlation between trained-untrained sigma activity and stabilization was not significant (r = 0.13, n = 22, p = 0.589). Thus, it is suggested that while the contribution of trained-untrained sigma activity to stabilization is not negligible, trained-untrained sigma activity alone is not enough to account for stabilization. Perhaps the combination of trained-untrained sigma and theta activity plays a role in stabilization, with theta activity having more weight.
Thus far, we used trained-untrained activity, which was calculated by subtracting spontaneous oscillatory activity in the untrained region from that in the trained region during sleep. One may wonder whether spontaneous oscillatory activity in each of the trained and untrained regions during sleep is correlated with the relevant performance measures. We thus computed the following four correlation coefficients: the correlation between offline performance gains and sigma activity in each of the trained and untrained regions, and the correlation between stabilization and theta activity in each of the trained and untrained regions. We found that only sigma activity in the trained region was correlated with offline performance gains (r = 0.43, n = 36, p = 0.008). The remaining correlation coefficients were not significant (correlation between offline performance gains and sigma activity in the untrained region: r = 0.26, n = 36, p = 0.132; correlation between stabilization and theta activity in the trained region: r = 0.23, n = 22, p = 0.313; and correlation between stabilization and theta activity in the untrained region: r = −0.10, n = 22, p = 0.653). The results of this analysis suggested that our method of computation of trained-untrained activity, which subtracted spontaneous activity in the untrained region from that in the trained region, captured the nature of the learning-dependent process and replicated the findings of previous studies (Tamaki et al., 2020b).
As a supplementary analysis, we also tested whether the mean r value for each of the above theta and sigma analyses was significant in the meta-analysis, although the total number of independent experiments included was only two and three for the theta and sigma analyses, respectively. First, we performed the Cochran Q test to investigate the heterogeneity of the dataset for the theta and sigma analyses. Heterogeneity was not assumed for either the theta or the sigma analyses (theta: Q1 = 0.001, p = 0.9805; sigma: Q2 = 1.090, p = 0.589). Then, we performed the Hunter–Schmidt method to obtain a weighted mean of the raw correlation coefficient (Viechtbaue, 2010). The Hunter–Schmidt method indicated that the correlation between trained-untrained theta activity during REM sleep and stabilization (Fig. 4B; r = 0.52, z = 3.20, p = 0.0014, 95% CI = 0.202, 0.841) and the correlation between trained-untrained sigma activity during NREM sleep and performance gain (Fig. 4D; r = 0.54, z = 4.37, p < 0.0001, 95% CI = 0.298, 0.782) significantly differed from 0 in these datasets. Thus, the results of the supplementary analysis were consistent with the results of the above theta and sigma analyses.
In summary, the results of experiment 2 suggested that occipital trained-untrained theta activity during REM sleep following presleep training was involved in the stabilization of presleep learning, whereas occipital trained-untrained sigma activity during NREM sleep was involved in offline performance gains in presleep learning.
Discussion
The results of experiment 1 indicated that the offline performance gains in VPL are in accordance with learning-dependent models. The strength of spontaneous oscillatory activity during post-training sleep in the early visual areas was significantly different between the learning and interference conditions, while visual cortex usage was matched between the conditions. These results are consistent with learning-dependent models, which assume a learning-specific process during post-training sleep in response to presleep learning.
In contrast, use-dependent models were inconsistent with the present results. A use-dependent model would predict that SWA increases to the same degree in the trained region in the early visual areas in both conditions. However, we did not find a significant increase in SWA in the trained region compared with that in the untrained region in any of the conditions. We did not find a significant correlation between SWA and offline performance gains in the learning condition, either. Because the correlation coefficient was very close to zero, the statistical insignificance in the correlational analysis was unlikely because of a small effect size or insufficient statistical power. The additional analysis with increased power in experiment 2 also did not find a significant relevance of SWA in offline performance gains. Moreover, behavioral data may argue against use-dependent models in experiment 1 because a significant performance improvement was not observed in the interference condition. A use-dependent model may predict significant offline performance gains during sleep following brain usage because of multiple training blocks during wakefulness before sleep, even when these training blocks are designed to interfere with each other, thus hindering learning. However, the present study demonstrated that offline performance gains did not occur following brain usage alone without learning before sleep. Thus, use-dependent models are unlikely to be consistent with offline performance gains, at least in VPL.
We found that sigma activity during NREM sleep and theta activity during REM sleep in early visual areas are involved in the learning-dependent process during sleep for VPL. However, the roles of these types of activity may be different. The results of the present study suggested that theta activity during REM sleep is mainly involved in the stabilization of presleep learning, while sigma activity during NREM sleep is mainly involved in offline performance gains. Overall, it is suggested that sleep has a learning-dependent process that enhances and stabilizes presleep learning in association with local sigma activity during NREM sleep and with local theta activity during REM sleep. We speculated that local sigma activity is associated with increased plasticity by long-term potentiation (Rosanova and Ulrich, 2005) or a replay of circuit or neuron activity involved in presleep learning (Born and Wilhelm, 2012), for which local theta activity stabilizes, as was shown in other types of memory (Boyce et al., 2016; Li et al., 2017; Tamaki et al., 2020b). However, our partial correlation analysis suggested that sigma activity may also play a minimum role in stabilization, in addition to theta activity, which plays a major role. Future studies are necessary to investigate the exact roles of sigma activity and theta activity in the learning-dependent process during sleep.
It is important to note that we applied different types of spectral analysis to EEG data in experiments 1 and 2. It is possible that different spectral analysis methods led to different outcomes. However, overall, the results of the present experiments 1 and 2 were consistent, showing the roles of sigma and theta activity in the learning-dependent process during sleep, but not SWA. Thus, the influence of the difference in spectral analysis methods may not be large in our study.
One may wonder whether the degree of cortical usage during training was successfully equated between the learning and interference conditions in experiment 1. Such a concern may arise at the orientation column level because the number of trials in tasks A and B differed between the conditions, although the total number of trials in the training periods including both tasks was equal between the conditions. Regarding the background orientation, one orientation was used as the background of the TDT in the learning condition, whereas two orientations were used in the interference condition. However, each trial was followed by a mask display that contained multiple orientations in the TDT. Thus, each condition should have excited multiple orientations; hence, we argue that the overall visual usage was equated between the conditions.
Some studies have suggested that SWA is involved in a learning-dependent process for offline performance gains (Tononi and Cirelli, 2003; Huber et al., 2004; Diekelmann and Born, 2010; Tamaki et al., 2013; Wilhelm et al., 2013). However, we did not find that SWA was involved in the learning facilitation process in this study. Our exploratory analysis did not find a significant involvement of SWA in the frontal or parietal regions in the learning facilitation process during sleep. While the present results are at odds with some of the previous studies, our finding is consistent with our previous study (Bang et al., 2014) that did not find that the strength of SWA is involved in the offline performance gains in VPL during sleep, but the strength of sigma activity was involved. We speculate that the reason we did not find SWA involvement in offline performance gains in VPL is related to the modality of learning and its involved neural circuits. It has been shown that the distribution of SWA is not consistent over brain regions during NREM sleep (Morikawa et al., 1997; Finelli et al., 2001; Iber et al., 2007): SWA appears predominantly over the frontocentral region, with a smaller magnitude over the occipital lobe, which is involved in the current visual task. Thus, one possibility is that SWA is more sensitive to the facilitation of learning or memory, which involves the frontocentral brain regions to a greater degree than it involves the occipital regions. Another possibility, although less likely, is that because the current experiment was designed to include only naps as sleep sessions, the strength of SWA was weaker in general than it is in nocturnal sleep, making the contribution of SWA to the facilitation process unclear. However, a previous study (Bang et al., 2014) investigated this matter in nocturnal sleep but did not observe contributions of SWA in the facilitation process of VPL. Thus, even in nocturnal sleep, the strength of SWA may not explain the facilitation process of VPL during sleep. Future studies are needed to clarify whether the lack of involvement of SWA in the facilitation process of VPL is because of the modality of learning.
In a similar vein, the present study did not show that typical sleep spindle waves found in the frontal and parietal regions were not significantly involved in the learning facilitation process during sleep. This may be because that not only the frontoparietal region, but also the hippocampus, is hardly involved in the learning-dependent process of VPL during sleep. Typical sleep spindle waves are coupled with hippocampal activity (Schabus et al., 2007; Skelin et al., 2021). However, it has been shown that the hippocampus is not significantly involved in visual learning, since amnesic patients with hippocampal lesions have been shown to be capable of visual learning (Levy et al., 2005; Rungratsameetaweemana et al., 2019). Thus, while our results corroborated that occipital sigma activity is involved in offline performance gains in VPL (Bang et al., 2014; Tamaki et al., 2020b), occipital sigma activity may not be directly correlated with typical sleep spindle waves found in the frontal and parietal regions.
In conclusion, the present study demonstrates that sleep has a learning-dependent process that involves local sigma activity during NREM sleep and theta activity during REM sleep. At the very least, offline performance gains in VPL are not consistent with a use-dependent model.
Footnotes
This work was supported by National Institutes of Health Grants R21-EY-028329 (Y.S.) and R01-EY-031705 (Y.S.); Japan Society for the Promotion of Science KAKENHI Grant JP20K22297 (M.T.); and the Brain Science Foundation, Japan (M.T.).
The authors declare no competing financial interests.
- Correspondence should be addressed to Yuka Sasaki at yuka_sasaki{at}brown.edu
