Sleep-related consolidation and generalizability of motor skills learned by physical practice, motor imagery and action observation

Sleep benefits the consolidation of motor skills learned by physical practice, mainly through periodic thalamocortical sleep spindle activity. However, motor skills can be learned without overt movement through motor imagery or action observation. Here, we investigated whether sleep spindle activity also supports the consolidation of non-physically learned movements. Forty-five electroencephalographic sleep recordings were collected during a daytime nap after motor sequence learning by physical practice, motor imagery or action observation. Our findings reveal that a temporal cluster-based organization of sleep spindles underlies motor memory consolidation in all groups, albeit with distinct behavioral outcomes. A daytime nap offers an early sleep window promoting the retention of motor skills learned by physical practice and motor imagery, and its generalizability towards the inter-manual transfer of skill after action observation. Findings may further have practical impacts with the development of non-physical rehabilitation interventions for patients having to remaster skills following peripherical or brain injury.


Introduction
Repeated practice is critical for the learning and mastering of motor skills. Training procedures encouraging the ability to exploit the features of a learned skill for its transfer from one situation to another is fundamental across diverse contexts, such as in sports sciences and rehabilitation 1 . Although the common way to learn a movement is by performing the task physically (see 2,3 for reviews), other forms of practice can contribute to motor-skill learning.
Compelling evidence shows that motor skills can be learned without overt movement through motor imagery (MI) or action observation (AO) [4][5][6][7] . While MI refers to the process of mentally rehearsing a movement without physically performing it, AO consists in observing another actor performing the movement. Numerous neuroimaging studies reported that neural structures are commonly activated during MI, AO, and physical practice (PP), thus providing evidence of a relative "functional equivalence" between practice modalities [8][9][10] .
Both covert modalities of practice (MI, AO) engage a cognitive demand upon sensorimotor networks, boosting activity-dependent neuroplasticity and enhancing motor performance and learning 4,5,11 . Traditionally, to evaluate the beneficial effects of MI and AO practice on motor skill learning, participants observe a model or imagine themselves performing the motor task before being evaluated on a post-training test requiring the overt practice of the movement 5,12 . It is commonly reported that MI and AO practice lead to enhanced motor performance and learning, albeit to a lesser degree than PP [13][14][15][16] . However, motor learning is not limited to task-specific learning but also concerns the ability to transfer or generalize the newly acquired skill to another one or another effector (i.e., inter-limb transfer), which may depend on the modality of practice 1 . For instance, while the amount of PP influences the development of an effector-specific or unspecific representation of the sequence 17 , PP has been shown to rely more on an effector-specific representation of the motor skill after extensive practice than the two other modalities 5,18 , thus leading to impaired inter-limb skill transfer 17 . In contrast, MI and AO practice have mostly been revealed to develop an effector-unspecific representation of the learned motor task, thus allowing effective skill transfer from one limb to another 5,18 . Therefore, the representation of a motor skill acquired through PP, MI, and AO practice partly relies on distinct coding systems for movement production, leading to specific skill learning and transfer capacities [19][20][21] .
Although the repeated practice of a motor skill is crucial for its initial acquisition, the development of an effective movement representation is not only a result of practice 2 . The newly formed memory continues to be processed "offline" over several temporal scales, from a rapid form of consolidation at a microscale of seconds 22 to longer forms of consolidation occurring primarily during the waking and sleeping hours following practice 2,23 . This offline period offers a privileged time window for memory consolidation, which relates to the process whereby newly acquired and relatively labile memories are transformed into more stable and enhanced memory traces 24 . The memory trace is thought to be dynamically maintained during wakefulness and actively reprocessed during a subsequent sleep period. A night of sleep and a daytime nap have been shown to play a crucial role in the strengthening and transformation of motor memory representations developed through PP during consolidation (see 25 for a review), which behaviorally results in either performance stabilization or improvement 26 . In the same way, a few studies reported sleep-dependent consolidation of motor skills learned through MI practice 11,12,27,28 or AO 29 , and it has been recently emphasized that the stability of the newly formed motor memory through consolidation processes during wake and sleep episodes may modify its generalizability 1 . However, whether similar consolidation mechanisms are engaged during sleep following PP, MI and AO practice, and how sleep affects the generalizability of motor skills remains to be determined.
At the brain level, and in the context of PP, motor memory consolidation is thought to be mediated by transient thalamocortical sleep spindle activity -an electrophysiological hallmark of non-rapid eye movement stage 2 (NREM2) sleep involving brief 0.3-2 s bursts of waxing and waning 11-16 Hz oscillations. Sleep spindles have been suggested to support the offline reactivation of newly acquired motor memories, resulting in post-night and post-nap motor memory improvements [30][31][32][33][34] . Boutin and Doyon (2020) 25 recently proposed a theoretical framework for motor memory consolidation following PP that outlines the essential contribution of the clustering and hierarchical rhythmicity of spindle activity during this sleepdependent process. They posited that the rhythmic expression of sleep spindles over taskrelevant cortical and subcortical brain regions is critical for the efficient reprocessing and consolidation of motor memory traces following PP. Specifically, it is suggested that the occurrence of sleep spindles follows two periodic rhythms: an infra-slow rhythm that corresponds to a 0.02 Hz periodicity of spindle-enriched periods, called spindle trains, and in which spindles iterate at an intermediate rhythm of about 0.2-0.3 Hz. Current theoretical models posit that this 0.2-0.3 Hz rhythmic occurrence of spindles during trains defines the sequential alternance of spindles and associated refractory periods, thus regulating the cyclic reinstatement and interference-free reprocessing of memory traces for their efficient consolidation 25,35-38 . However, whether a similar temporal cluster-based organization of sleep spindles underlies the consolidation of motor skills acquired by MI and AO practice remains to be addressed.
Hence, by combining behavioral and electroencephalographic (EEG) sleep measures, the aim of the present study was twofold: (i) to determine whether similar consolidation mechanisms are engaged during sleep following PP, MI and AO practice, and (ii) to investigate the specific contribution of sleep spindle activity and its rhythmic organization in motor skill consolidation and inter-manual transfer, depending on the modality of practice.

Results
Methods: overall experimental approach. Forty-five participants were required to learn a motor sequence task either by physical practice, motor imagery or action observation. Response times (i.e., interval between two keypresses; RT) were measured to evaluate skill performance during test blocks ( Fig. 1): the pre-test and post-test blocks were respectively performed before and after the training phase, while the retention and inter-manual transfer tests were delayed and performed after a 90-min daytime nap following training. Motor skill acquisition is reflected by changes in RT performance (in percentages) from the pre-test to the post-test. Motor skill consolidation was assessed by analyzing the percentage of RT changes from the post-test to the retention test, while motor skill transfer was measured as the difference in RT performance (in percentages) between the retention test and the transfer test.
Behavioral data. Fig. 2   Sleep and spindle characteristics. For all groups, sleep architecture and spindle metrics are respectively summarized in Table 1 and Table 2        seconds (~0.2-0.3 Hz), irrespective of the practice mode. We also conducted independent sample student's t-tests to evaluate between-groups differences of the spectral power variations.
Similarly, TF maps were corrected for multiple comparisons using the Benjamini-Hochberg procedure (Ntest = 60020). No significant difference was found between the groups, underlying the modality-unspecific pattern of the spindle-power rhythmic variations over time. Noteworthy, TF maps also reveal a rhythmic pattern of power increases in the theta frequency band (4-8 Hz), which accords with current trends suggesting that cross-frequency interactions between sleep spindles and theta waves may be relevant for sleep-related memory consolidation (see also 41 ).

Discussion
In the current study, we examined (i) whether similar consolidation mechanisms are engaged during sleep following PP, MI and AO practice, and (ii) the contribution of sleep spindles in motor skill consolidation and transfer, depending on the modality of practice. Our findings confirmed that participants acquired the motor sequence through PP, MI and AO practice, with an advantage for PP. Our results further revealed that sleep, and more specifically the time spent in NREM2 sleep, is related to motor skill consolidation following PP and MI practice and motor skill transfer following AO practice, hence pointing towards potential modality-specific effects of sleep upon skill consolidation and its generalizability. In addition, we found that spindles occurring in trains (grouped spindles) during NREM2 sleep are primarily involved in the sleep consolidation process, in comparison to isolated ones, leading to enhanced skill retention following PP and MI practice and improved skill transfer following AO practice. During training, and as expected, the rehearsal of the motor sequence increased performance for all practice groups 4-6,13,16,42-44 . As expected, though, participants in the PP group expressed higher performance improvements than their MI group counterparts 13,16,42,45 .
Traditionally, it is also assumed that PP leads to greater practice-related gains than AO 5,14,15 .
Albeit not significant, a clear tendency emerged in our results in accordance with this latter assumption. Previous studies have shown that AO learners may reach similar performance levels than PP learners when only a few physical practice trials are provided 5,15 11 , who demonstrated that an early sleep window following AO is crucial for the emergence of offline performance gains.
However, in their study, the benefits of sleep were only observed in behavioral performance at the group level, and the absence of polysomnographic monitoring prevented any analysis of sleep architecture or spindle activity in relation to motor skill consolidation. Thus, our results rather accord with studies suggesting that AO practice may trigger different consolidation processes than those triggered by PP, leading to different behavioral outcomes 14,52 .
In this way, and very interestingly, our results confirmed the potential engagement of distinct consolidation mechanisms during sleep following PP, MI and AO. Indeed, our findings revealed a positive relationship between the time spent in NREM2 sleep and inter-manual skill transfer for participants in the AO group only. We further found a positive correlation between the number of NREM2 sleep spindles and the magnitude of skill transfer following AO practice, revealing the role of sleep spindle activity in the memory consolidation process following AO.
Hence, we conjecture that sleep differently affects the representation of a motor skill depending on prior training experience. This assumption is also supported by transfer performance at the group level since we showed greater post-nap skill transfer following both MI and AO compared to PP, which accords with previous studies 5,17,18,20,45 . It is now well accepted in the literature that PP of long complex sequences would mainly engage an encoding of the motor task in visuo-spatial coordinates (i.e., effector-unspecific learning) 2,5,53 . In contrast, shorter sequences would mainly be encoded in motor coordinates (i.e., effector-specific learning) 2,54,55 .
Therefore, physical practice of a short 5-element motor sequence in our study may have primarily led to the development of an effector-specific learning of the motor skill, as reflected by impaired transfer performance compared to other practice modalities. In contrast, the higher inter-manual skill transfer capacity found for the MI and AO groups indicates that both non-physical practice conditions may have led to the development of a motor skill representation mostly at an effector-unspecific level, in comparison to the PP group (see also 5,45 ).
Our behavioral results also accord with previous neuroimaging findings showing more consistent brain activations during MI and AO compared to PP, in an effector-unspecific manner within a predominant premotor-parietal network 8,9,56,57 . In a recent neuroimaging metaanalysis, Hardwick and colleagues 8 compared the pattern of brain activations following PP, MI and AO, and identified brain areas involved in both the simulation of actions (MI and AO) and actual motor execution (PP) through a cortical-dominant network. This network comprises essentially premotor, parietal, and sensorimotor regions, albeit to a lesser extent for MI and AO.
Hence, while a set of common neural structures is activated during PP, MI and AO, leading to enhanced motor performance and learning, the additional and specific recruitment of brain regions with respect to the modality of practice may be responsible for the distinct encoding as for patients with motor deficits having to remaster skills following physical or brain injury.

Limitations of the study
In this study, we used a nap paradigm and recruited exclusively young healthy participants. It would be interesting then to study the potential involvement and temporal organization of sleep spindles in the consolidation of motor skills learned by PP, MI and AO practice during a whole night of sleep or with clinical populations facing spindle abnormalities or altered rhythmicity of spindles, for instance.
We conducted our main analyses on the parietal Pz derivation of the EEG cap, described as sensitive to sleep spindle activity in relation to motor memory consolidation 31,32,62 .
Relationships with other derivations are described in the supplemental material (see Fig. S2 and S3). However, further analyses may be relevant to assess and precise the spatial dynamics of the cluster-based organization of sleep spindles over different brain regions.
Regarding the evaluation of motor performance, we decided to administer only a single block of practice during the retention and transfer tests, leaving little room for additional physical practice. It was necessary in our study to limit the physical execution of the sequence to prevent additional training during the test phases, and especially for the non-physical MI and AO practice groups. However, such a design implies reduced measures of RT performance and therefore greater performance variability compared to more common designs using two or more blocks during testing phases 31,49 .

Declaration of interests
The authors declare no competing interest.

Author Contributions
Author contributions: A.C., A.B. and U.D. conceived the experiment; A.C. collected the data; A.C. and A.B. analyzed the data and discussed the results; A.C., A.B. and U.D. wrote the manuscript; All authors revised the manuscript.

Lead contact
Further information should be directed to and will be fulfilled by the lead contact, Arnaud Boutin (arnaud.boutin@universite-paris-saclay.fr)

Data and code availability
The data that support the results of this study are available from the corresponding author upon reasonable request and under a formal data-sharing agreement.
Sleep EEG data were processed using the MATLAB R2019b software from The Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS
Forty-five healthy volunteers (18 females, mean age: 23.7 ± 4 years) were recruited by local advertisements and were randomly and equally assigned to either a PP group (8 females, informed consent before inclusion. Participants were asked to maintain a regular sleep-wake cycle and refrain from all caffeine-and alcohol-containing beverages 24h before the experimentation. Participants in the MI group were also required to complete the French version of the Movement Imagery Questionnaire-Third version (MIQ-3f) 64 before starting the experimental protocol. The MIQ-3f is a twelve-item self-report questionnaire, in which participants are asked to perform a given movement followed by its mental execution either by external visual imagery, internal visual imagery or kinesthetic imagery. Participants rate the difficulty with two 7-point scales respectively for visual or kinesthetic imagery, ranging from 1 (very hard) to 7 (very easy). A higher average score represents a greater imagery capacity.

Experimental design and motor sequence task
Participants sat on a chair at a distance of 50 cm in front of a computer screen, equipped with a 64-channel EEG cap. The motor task consisted of performing as quickly and accurately as possible a 5-element finger movement sequence by pressing the appropriate response keys on a standard French AZERTY keyboard using their left, non-dominant hand. The sequence to be performed (C-N-V-B-C, where C corresponds to the little finger and N to the index finger) was explicitly taught to the participant prior to training.
Physical practice blocks consisted of 16 repetitions of the 5-element sequence (i.e., a total of 80 keypresses). Each block began with the presentation of a green cross in the center of the screen accompanied by a brief 50-ms tone. In case of occasional errors, participants were asked "not to correct errors and to continue the task from the beginning of the sequence" (see 65 for a similar procedure). At the end of each block, upon completion of the 80 keypresses, the color of the green-colored imperative stimulus turned red, and participants were then required to look at the fixation cross during the 30-s rest period. This protocol controlled the number of movements executed per block to ensure that the same amount of practice with the task was afforded to participants during a particular session. Stimuli presentation and response registration were controlled using the MATLAB R2016b software from The MathWorks (Natick, MA) and the Psychophysics Toolbox extensions 66 .
The study started at 1.00 pm to minimize the putative impact of both circadian and homeostatic factors on individual performance levels and sleep characteristics 60,61,67 . The experimental procedure was composed of seven main phases: familiarization, pre-test, training, post-test, 90-minute nap, retention and transfer tests (Fig. 1). Before training, participants underwent a brief familiarization phase during which they were instructed to repeatedly and slowly perform the 5-element sequence until they accurately reproduced the sequence three consecutive times. This familiarization was intended to ensure that participants understood the instructions and explicitly memorized the sequence of movements.
During the pre-test, all participants physically performed one block of the 5-element motor sequence. The ensuing training phase consisted of 14 blocks performed with physical, observational or mental practice. Participants in the PP group were asked to physically execute the sequence task with their left-hand fingers, as previously described. Participants in the AO group were instructed to keep their fingers on the corresponding response keys. Following the imperative green-cross stimulus and audio cue, a video of a model performing the motor task was displayed on the screen. The model was depicted so that the observers could see both the finger movements of the model and the green cross appearance on the screen. This viewing angle was adopted in order to closely match the perspective view of the AO participants. An additional window inset zooming on the left-hand fingers of the model was implemented so that participants could precisely watch fine finger movements. Participants in the AO group were free to observe both perspectives in an active and conscious manner while avoiding any concurrent muscular execution of the movement (controlled online by electromyography (EMG) recording electrodes placed on the left flexor digitorum superficialis; see section EEG-EMG data acquisition and pre-processing for details). Based on previous motor sequence learning findings 31 , performance improvements of the model across training blocks followed the power-law of practice (mean response time between consecutive keypresses ranging from RTBlock1 = 616 ms to RTBlock14 = 223 ms). Participants in the MI group were instructed to keep their left-hand fingers on the corresponding response keys and their thumb on the keyboard's space bar. When they heard the imperative audio cue, they had to imagine themselves performing the sequence using a combination of internal visual and kinesthetic imagery, while avoiding any associated overt movements (controlled online by similar EMG procedure as for the AO group). After completing each mentally rehearsed sequence, they were asked to press the space bar with their thumb to objectively control for the amount of MI practice. EEG data were bandpass filtered between 0.5 and 50 Hz to remove low-frequency drift and high-frequency noise, down-sampled to 250 Hz, and re-referenced to the linked mastoids (i.e., TP9 and TP10). EOG and EMG data were bandpass filtered between 0.3-35 Hz and 10-100 Hz, respectively.

QUANTIFICATION AND STATISTICAL ANALYSIS
All statistical analyses used every participant in each experimental group (PP: n = 15; MI: n = 15; AO: n = 15). These sample sizes were determined based on previous studies 32, 36,49 .
All error measurements indicate the standard error of the means (SEM).

Behavioral analysis
RT performance was measured as the interval between two consecutive keypresses during each test block. Also, since participants were asked to start over from the beginning of the sequence if they made any error during task production, RTs from error trials (i.e., erroneous key presses) were excluded from the analyses. Across all test blocks, the PP group performed 14.6 (± 1.6) accurate sequences, 14. were normalized by multiplying the power at each frequency bin with the frequency value (1/f compensation). To assess the statistical significance of the power variation at each frequency bin over time, we used a two-tailed student's t-test against a predefined baseline window for each practice group with a significance threshold set at 0.05. The baseline was set from -2 s to -0.5 s before spindle onset. We also conducted two-tailed independent sample student's t-tests to evaluate between-group differences of the power variation over time. All statistical maps were then corrected for multiple comparisons using the Benjamini-Hochberg procedure to control the false discovery rate (Ntest = 60020 for each analysis) 39     Relationship between the number of grouped spindles (left column) and isolated spindles (right column) detected over main scalp derivations with the magnitude of skill consolidation and transfer following physical practice (orange, first row), motor imagery (blue, second row) and action observation (green, third row). Pearson correlation coefficients (r) are reported for each correlation. Colored circles correspond to significant p-value (p < 0.05).