Sleep Oscillations Across Cortical, Subcortical and Cerebellar Structures in Magnetoencephalography

ABSTRACT

Sleep involves widespread changes in neural activity, with distinctive oscillatory patterns emerging across frequency bands and brain regions. Characterising these dynamics is essential for understanding their functional roles in health and their disruption in sleep‐related disorders. However, most work on healthy humans has used techniques with limited temporal or spatial resolution, focusing mainly on the cerebral cortex. Growing evidence suggests that subcortical and cerebellar structures contribute to sleep dynamics, yet these regions remain largely unexplored in human neuroimaging due to methodological limitations. Magnetoencephalography (MEG) offers millisecond temporal resolution with spatial precision to localise activity across cortical, subcortical and cerebellar regions. Recent evidence demonstrates that MEG can detect signals from deep brain structures, challenging assumptions about its spatial limitations, but systematic validation and whole‐brain mapping of oscillatory activity during sleep remain lacking. In this study, we provide comprehensive maps of oscillatory power across the whole brain during non‐rapid eye movement (NREM) sleep using source‐localised MEG. We first validated signal differentiability across cortical, subcortical and cerebellar regions using spectral fingerprinting analysis. We then characterised frequency‐specific and stage‐specific changes in oscillatory power across six frequency bands and three NREM sleep stages. Finally, we examined sigma‐band dynamics during spindle‐rich stage 2 sleep to investigate spindle‐related activity across brain regions. Our results reveal structured, region‐specific patterns of sleep modulation that extend beyond traditional cortical‐thalamic circuits, including novel evidence for cerebellar engagement in fast spindle frequencies. These findings expand models of sleep‐related brain activity and demonstrate the utility of whole‐brain MEG for understanding distributed sleep networks.

This study uses source‐localised MEG to characterise oscillatory whole brain activity across NREM sleep stages, including the cerebellum. Findings reveal distinct, stage‐ and frequency‐specific patterns, with cerebellar signals reliably differentiable and engaged during sigma‐band activity in N2 sleep, highlighting the cerebellum's contribution to whole‐brain sleep dynamics beyond traditional cortical focus.

Abbreviations

AAL3

Automated Anatomical Labelling Atlas, version 3

AICHA

Atlas of Intrinsic Connectivity of Homotopic Areas

ANOVA

analysis of variance

ECG

electrocardiography

EEG

electroencephalography

EMG

electromyography

EOG

electrooculography

FDR

false discovery rate

fMRI

functional magnetic resonance imaging

IRASA

irregular‐resampling auto‐spectral analysis

LME

linear mixed‐effects

MCTQ

Munich Chronotype Questionnaire

MEG

magnetoencephalography

MMHQ

Montreal Musical History Questionnaire

MNE

minimum norm estimation

MNI

Montreal Neurological Institute

MRI

magnetic resonance imaging

MSL

motor sequence learning

N1, N2

sleep stages 1, 2, 3

NREM

non‐rapid eye movement

PET

positron emission tomography

PSQI

Pittsburgh Sleep Quality Index

REM

rapid eye movement

ROI

region of interest

RS1

RS2

RS3

RS4 Resting‐state recordings 1 to 4

1. Introduction

The brain remains highly dynamic during sleep, engaging in organised patterns of neural oscillations that support a range of physiological and cognitive functions. Rather than representing a passive or quiescent state, sleep reflects a continuation of the brain's intrinsic dynamics. These dynamics are thought to underlie key processes including memory consolidation (Stickgold 2005), synaptic homeostasis (Tononi and Cirelli 2014), metabolic clearance (Zielinski et al. 2016) and emotional regulation (Goldstein and Walker 2014; Zielinski et al. 2016). Similar to neural activity during wake, sleep dynamics unfold through interactions across distributed cortical and subcortical networks.

It is increasingly clear that patterns of frequency‐specific activity change systematically across sleep stages, with distinct regional contributions emerging as sleep deepens. Much of the foundational work on sleep oscillations has relied on electroencephalography (EEG), which provides excellent temporal resolution and is well‐suited for sleep staging and detecting sleep events such as spindles and K‐complexes. Despite EEG's strengths, its spatial limitations have prompted the growing use of magnetoencephalography (MEG) to study the spatio‐temporal dynamics of sleep. Unlike EEG electric fields, MEG magnetic fields are not distorted by tissue conductivity of the scalp, skull and cerebrospinal fluid (Gross et al. 2013), enabling source‐localised reconstructions of whole‐brain activity. Combined with subject‐specific anatomical modelling, this approach allows cortical activity to be localised with high accuracy (e.g., tonotopic maps in the auditory cortex; Falet et al. 2021). MEG has been used to investigate transient sleep events such as spindles (Ioannides et al. 2017; Ayoub et al. 2012; Klinzing et al. 2016; Dehghani et al. 2010), slow waves (Klinzing et al. 2016) and rapid eye movement (REM)–related eye movements (Ioannides et al. 2004), as well as more continuous spectral dynamics. However, few studies have examined how the spectral profile of spontaneous brain oscillations, as measured with MEG, systematically varies across sleep stages and across the whole brain.

Recent work by Fan et al. (2023) and Brancaccio et al. (2020) demonstrates that MEG can capture fine‐grained spectral transitions across sleep stages, revealing spatially specific patterns of oscillatory power that extend established EEG findings on frequency and amplitude changes during sleep (Ogilvie 2001; De Gennaro and Ferrara 2003; De Andrés et al. 2011). Despite these advances, most MEG studies of sleep have focused exclusively on cortical sources, leaving subcortical and cerebellar contributions remaining largely unexplored. Expanding the anatomical scope of sleep studies is important as sleep involves coordination across distributed brain networks, including subcortical and cerebellar regions (De Andrés et al. 2011; Dang‐Vu et al. 2008; Canto et al. 2017).

While historically regarded primarily as a motor control structure, the cerebellum is now understood to play a broader role in coordinating both motor and higher order functions, including language, working memory, emotional regulation and social cognition (Schmahmann et al. 2019; Mariën et al. 2014; Adamaszek et al. 2017) during wake. These functions are thought to arise from the cerebellum's dense and reciprocal connectivity with cortical regions, including the prefrontal cortex, parietal cortex and temporal lobes (Ramnani 2006; Ishikawa et al. 2016). In addition to its integrative network role during wake, emerging evidence from anatomical, clinical and neuroimaging studies suggests that the cerebellum contributes meaningfully to sleep regulation. Anatomical studies have demonstrated robust connections between the cerebellum, brainstem sleep‐regulatory centres and thalamo‐cortical circuits (De Andrés et al. 2011), suggesting that cerebellar structures may influence both the initiation and maintenance of sleep states. Clinical evidence further supports this view: Individuals with primary sleep disorders often show reduced cerebellar volume, while patients with cerebellar disorders frequently report disrupted sleep and excessive daytime sleepiness (Canto et al. 2017), and animal lesion studies show that cerebellar damage alters both non‐REM (NREM) sleep and REM sleep architecture. Additional animal studies reveal that cerebellar activity during sleep is sleep‐stage‐dependent (Canto et al. 2017), with cerebellar participation in slow‐wave, spindle and delta‐band activity (Jackson and Xu 2023; Torres‐Herraez et al. 2022). Coordinated cerebello‐hippocampal activity has been observed during both NREM and REM sleep (Torres‐Herraez et al. 2022), while broader interactions between the cerebellum and neocortical, thalamo‐cortical and hippocampal circuits during sleep have also been reported (Benarroch 2023). In humans, multimodal imaging studies combining EEG with functional magnetic resonance imaging (fMRI) or positron emission tomography (PET) have revealed that cerebellar activity during sleep broadly mirrors cortical states: Cerebellar signals are attenuated during sleep stage 1 (N1), co‐occur with K‐complexes and spindles during sleep stage 2 (N2) and align with cortical slow waves during sleep stage 3 (N3) (Benarroch 2023). Together, these findings suggest that the cerebellum is not passive during sleep but may participate in the coordination of sleep‐related neural dynamics.

One sleep stage in which cerebellar involvement may be especially relevant is N2 sleep, which is characterised by a high density of sleep spindles (Fernandez and Lüthi 2020). Spindles are brief bursts of oscillatory activity (typically 0.5–2 s) classically defined within the 12–15‐Hz sigma band (De Gennaro and Ferrara 2003; Brancaccio et al. 2020), though this range is often extended to 10–15 Hz and subdivided into slow (< 13 Hz) and fast (> 13 Hz) subtypes (Fernandez and Lüthi 2020). Spindle subtypes differ spatially: Slow spindles are most prominent over frontal regions, while fast spindles peak over parietal and central areas, suggesting separable neural generators or network configurations. While spindles have traditionally been viewed as thalamocortical events, intracranial recordings in nonhuman primates by Xu et al. (2021) provide evidence for a broader origin. Using simultaneous cortical, thalamic and cerebellar recordings, they showed that during spindles, the cerebellum exerted directional influence on the motor cortex via the thalamus, reversing the cortical‐to‐cerebellar direction observed during slow‐wave activity. These findings led Xu et al. (2021) to propose a dual‐oscillator framework in which cerebellar and thalamic generators jointly contribute to spindle dynamics, positioning the cerebellum as an active node in spindle generation and propagation.

Direct electrophysiological evidence for human cerebellar spindle activity comes from Niedermeyer and Uematsu (1974), who performed depth electrode recordings from the fastigial and dentate nuclei in three patients with epilepsy undergoing pre‐surgical evaluation. Sleep spindles were detectable in both scalp and cerebellar recordings. Importantly, spindle activity at the two sites showed partial asynchrony, leading the authors to propose that the cerebellum may exhibit a relatively independent type of sleep spindle. Despite this early and compelling evidence, the cerebellum has received limited attention in human sleep research in the decades since, and the spatial and spectral characteristics of cerebellar oscillatory activity during NREM sleep remain poorly characterised. This gap likely stems in part from the reliance on imaging techniques such as fMRI and PET for studying the cerebellum, which, while useful for mapping haemodynamic changes, lack the temporal resolution needed to resolve the oscillatory dynamics central to sleep research. MEG, with its millisecond‐scale temporal precision, is better suited for this purpose but has traditionally been used to model activity generated by cortical sources. Nonetheless, prior work has demonstrated successful MEG detection of activity in deep brain structures, including the hippocampus (Attal and Schwartz 2013), amygdala (Dumas et al. 2013), thalamus (Rattray et al. 2025) and brainstem (Coffey et al. 2016, 2021).

Two recent papers have specifically addressed the detectability of cerebellar signals using MEG. Andersen et al. (2020) reviewed the electrophysiological literature and argue that, although the cerebellum's folded structure and deep location have contributed to scepticism about its detectability with non‐invasive methods, this view is outdated. They presented several examples of successful cerebellar signal detection using MEG (and EEG) and argue that with appropriate source modelling and analysis, cerebellar activity can be reliably detected. They also highlighted that sensor placement may be a contributing factor, suggesting that even simple adjustments, such as positioning the head optimally in the MEG system, may improve signal detection. Further support comes from simulations by Samuelsson et al. (2020), who showed that cerebellar signals in both MEG and EEG are only 30%–60% weaker than cortical signals, despite greater cancellation effects. Importantly, the authors caution that studies omitting the cerebellum from their source models may inadvertently misattribute cerebellar activity to nearby occipital regions. This possibility runs counter to common assumptions that any cerebellar signal merely reflects cortical ‘leakage’. Furthermore, using the same dataset as the present study, Rattray et al. (2025) demonstrated that MEG can resolve deep brain networks during NREM sleep, showing that coherence‐based connectivity increased during spindles across a distributed thalamocortical–hippocampal network. Using a differentiability metric adapted from neural fingerprinting approaches (Amico and Goñi 2018; da Silva Castanheira et al. 2021), they further demonstrated that thalamic nuclei are reliably differentiable in source space, supporting the validity of MEG for investigating deep subcortical structures during sleep.

The present study addresses a gap in our understanding of the spatial and spectral characteristics of whole‐brain activity during NREM sleep, with particular attention to the cerebellum. We recorded simultaneous MEG and EEG from 19 healthy young adults during wake and a 2.5‐h daytime nap, with participants positioned to optimise sensor coverage of the cerebellum. Recordings included periods of wake as well as NREM sleep stages N1, N2 and N3, allowing us to characterise oscillatory dynamics across the sleep–wake continuum. We first assessed and compared the differentiability of MEG signals source‐localised to cerebellar regions relative to that of cortical and other subcortical structures. Upon confirming that cerebellar signals were reliably differentiable, we pursued three primary aims. First, we characterised whole‐brain oscillatory power—including the cerebellum—across wake and NREM sleep stages, focusing on classical frequency bands commonly used to describe sleep: slow (0.5–2 Hz), delta (2–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), sigma (12–15 Hz) and beta (15–30 Hz). Second, we examined stage‐specific spectral changes by statistically comparing power in each frequency band between wake and the corresponding NREM stages. Finally, we focused on spindle‐rich N2 sleep to investigate cerebellar involvement in sigma‐band dynamics. This analysis included contrasting spindle‐band activity between cerebellar and occipital regions, mapping cortical and cerebellar sigma power in 1 Hz increments and examining peak frequency topographies within the spindle range. This work expands current models of sleep‐related brain activity and demonstrates the utility of whole‐brain MEG for understanding distributed sleep networks.

2. Methods

2.1. Experimental Design

The data analysed in this study were collected as part of a larger project examining neuroplasticity across states of consciousness. Although the dataset includes task and resting‐state recordings, only the sleep data are used in the present work. Methods concerning the present work are described in full below; for completeness, we include a brief outline of the full data collection sequence, including elements not used in the present analyses. All procedures were approved by both the Montreal Neurological Institute (MNI) and Concordia University's Research Ethics boards. Participants were compensated for their time.

To increase sleep pressure, participants were asked to stay up 1 h past their habitual bedtime the night before and avoid caffeine on the morning of the experiment. The session began at approximately 12:00 pm, aligning the nap opportunity with the mid‐afternoon circadian dip in alertness (Milner and Cote 2009). During the first 30 min, participants gave informed consent, completed questionnaires and changed into comfortable sleepwear before entering the magnetically shielded MEG chamber. One of 20 recruited participants was excluded due to a metal dental appliance that caused magnetic artefacts, resulting in a final sample size of 19.

Participants were positioned in the MEG to maximise sensor coverage of the whole head, including the cerebellum (Andersen et al. 2020), with the head fully inserted and the chin angled slightly downward; head and body position was stabilised using foam padding. Following an initial resting‐state recording (RS1), a motor sequence learning (MSL) task and a second resting‐state recording (RS2), participants underwent a 2.5‐h nap opportunity in darkness, during which they were encouraged to rest peacefully even if unable to sleep. After the nap, participants were gently awakened and allowed to regain alertness before completing a further resting‐state recording (RS3), MSL task and final resting‐state recording (RS4). An overview of the data collection sequence is shown in Figure 1a. In a separate session, participants returned for a T1‐weighted MRI scan for MEG source reconstruction.

FIGURE 1.

Experimental design and whole‐brain parcellation. (a) Overview of the data collection sequence. After head positioning in the MEG, each participant completed a series of recordings: an initial resting‐state session (RS1), a motor sequence learning (MSL) task, a second resting‐state recording (RS2), a 2.5‐h nap opportunity, a third resting‐state session (RS3), a post‐nap MSL task and a final resting‐state recording (RS4). The present analyses focus exclusively on data recorded during the nap opportunity (highlighted in blue). (b) Whole‐brain source space parcellation. Cortical and subcortical regions were defined using the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA), while cerebellar regions were derived from the Automated Anatomical Labelling Atlas, version 3 (AAL3). The combined atlas included 416 regions in total (44 subcortical, 32 cerebellar).

2.1.1. Questionnaires

Participants completed the Pittsburgh Sleep Quality Index (PSQI) (Buysse et al. 1989) and the Munich Chronotype Questionnaire (MCTQ) (Roenneberg et al. 2003) to assess sleep habits and screen for sleep disorders, along with a Subjective Fatigue Scale and a custom health questionnaire to screen for medical conditions that could affect sleep quality. Participants also completed the Montreal Musical History Questionnaire (MMHQ) (Coffey et al. 2011), a self‐report measure of musical training history.

2.1.2. Subject Preparation

EEG, electrooculography (EOG) and electromyography (EMG) were used to assess sleep depth and quality, following standard sleep research practices (Iber et al. 2007). Five EEG electrodes were placed at Cz, C3, C4 and the left and right mastoids (10–20 International System), with a forehead ground and left mastoid reference. Bipolar EMG electrodes on the chin and neck aided sleep scoring, bipolar EOG electrodes beside the eyes helped identify REM sleep and detect MEG artefacts and electrocardiography (ECG) electrodes on the chest were used to monitor and remove cardiac artefacts. Head shape and head position indicator coil locations (forehead, left and right) were digitised using a Polhemus Isotrak system (Polhemus Inc., VT, United States) for co‐registration of MEG data with T1‐weighted anatomical images and correction of head movement during recording.

2.2. Participant Characteristics

Data were collected from 19 neurologically healthy young adults, a sample size comparable with or larger than previous MEG studies on sleep neurophysiology (Brancaccio et al. 2020; Tamaki et al. 2014; Mylonas et al. 2022). All participants had normal or corrected‐to‐normal vision and no history of neurological or sleep disorders. The sample comprised 13 females and 18 right‐handed individuals, with a mean age of 21.7 years (SD = 2.8). Musical experience was documented for completeness. Twelve of the 19 participants had no experience playing instruments; cumulative practice hours among those with experience ranged from 12.5 to 8236.8 h (mean = 2478.5, SD = 3157.4). This convenience sample lacks a sufficient range of musical experience to support further analyses related to musicianship (see Figure S1 for the distribution of instrument hours).

2.3. Data Acquisition

Two hundred and 70 MEG channels (axial gradiometers), three EEG channels (Cz, C3, C4) and EOG and ECG data (pairs of bipolar electrodes) were simultaneously recorded using a CTF MEG System with an integrated EEG setup (Omega 275, CTF Systems Inc.). All data were sampled at 2400 Hz. Separate runs were created for each resting‐state recording (5 min), each MSL task recording (12 min) and the nap opportunity, which was divided into five 30‐min segments. Head position within the scanner was recorded before each run to ensure accurate source reconstruction. High‐resolution anatomical T1‐weighted images (176 slices, FOV = 256 × 256, TR/TE/TI: 2300/2.98/900 ms, voxel size = 1 mm isotropic) were acquired using a MPRAGE pulse sequence on a 3T Siemens Magnetic Resonance Imaging (MRI) scanner for subject‐specific reconstruction of neural sources.

2.4. EEG Data Processing

Although MEG and EEG can measure related brain dynamics, EEG data were used for sleep scoring to ensure comparability with established norms in sleep research. Data were prepared for sleep scoring using Brainstorm (Tadel et al. 2011) and custom Matlab scripts (The Mathworks Inc., MA, United States). Band‐pass filters were applied (EEG: 0.1–20 Hz; EOG: 0.1–5 Hz; EMG: 10–58 Hz), and data were resampled to 150 Hz. Sleep scoring, in which 30 s windows of data are visually inspected and categorised into wake, non‐REM sleep stages 1–3, and REM sleep, was completed according to standard AASM practices (Iber et al. 2007) based on the C3 and C4 EEG channels, EOG and EMG.

2.5. MEG Data Processing

We performed MEG data processing in Brainstorm (Tadel et al. 2011) and using custom Matlab scripts (The Mathworks Inc., MA, United States). Eye blink and heart‐beat artefacts were removed from MEG data using Brainstorm's built‐in source signal projection algorithm (Hämäläinen 2009; Tesche et al. 1995), using the recommended procedure: Projectors were removed when they captured at least 12% of the signal and the topography of the components matched those of ocular or cardiac origin upon visual inspection. A band‐pass filter (0.1–37.5 Hz) was applied to the data, which were then resampled to 150 Hz.

2.5.1. Source Modelling

FreeSurfer (Fischl et al. 2002; Fischl 2012) was used to prepare cortical surfaces and automatically segment subcortical structures from each subject's T1‐weighted anatomical MRI scan. The results were imported into Brainstorm (Tadel et al. 2011), and the FreeSurfer automatic subcortical segmentation (aseg) atlas (which includes the brainstem, thalamus and cerebellar structures) was combined with the cortex surface as a basis for modelling the neural sources of brain activity. We created a volume‐based model with a three‐dimensional 5 mm isotropic grid within all structures. An overlapping‐sphere head model was computed for each run, and a noise covariance matrix from two‐minute empty‐room recordings (taken before each participant session) was used to capture and remove sensor noise. We computed the minimum norm estimation (MNE) source distribution with unconstrained source orientations for each run using Brainstorm default parameters (Gramfort et al. 2014; Tadel et al. 2011), with the exception of depth weighting (increased from 0.5 to 0.7), following previous work in which subcortical sources were of interest (Gorina‐Careta et al. 2021; Lin et al. 2006).

2.5.2. Defining Brain Regions and Signal Extraction

We created a whole‐brain volume atlas consisting of 416 regions for each subject (see Figure 1b). Cortical and subcortical regions were defined using the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA), a functional brain atlas derived from resting‐state fMRI data (Joliot et al. 2015). This atlas provides structurally and functionally meaningful cortical parcellations of relatively small, homogeneous size, reducing the risk of signal degradation from averaging over larger regions. Cerebellar regions were added using the Automated Anatomical Labelling Atlas, version 3 (AAL3) (Rolls et al. 2020), which parcellates the cerebellum (including vermis) into 32 regions. All regions were delineated by importing the AICHA or AAL3 in MNI space into Brainstorm and mapping ROIs (region of interest) onto individual MRIs, using an affine co‐registration with the MNI ICBM152 template (Ashburner and Friston 2005). For each ROI and each 30‐s epoch, we extracted three timeseries of mean amplitude corresponding to the x, y and z source orientations.

2.5.3. Spectral Decomposition

Neural signals captured by MEG consist of both rhythmic and arrhythmic components (Wen and Liu 2016). The rhythmic component arises from oscillatory network activity with characteristic time scales. It is the primary focus of this study, as it reflects organised neural activity (Buzsaki and Draguhn 2004). In contrast, the arrhythmic component, often referred to as ‘1/f activity’, is scale‐free and reflects fractal dynamics (He et al. 2010). To separate these components, we employed irregular‐resampling auto‐spectral analysis (IRASA) (Wen and Liu 2016) and subtracted the fractal 1/f component from the original power spectrum to compute the 1/f‐corrected oscillatory component. Following the approach of Feld et al. (2021), we then computed a normalised measure of oscillatory power by expressing it relative to the fractal component:

$$\text{Relative oscillatory power}=\frac{\text{oscillatory component}}{\text{fractal component}},$$

where the oscillatory component is defined as the original signal minus the fractal component. This step normalises oscillatory power across subjects and frequencies, revealing the less pronounced oscillations typically masked by 1/f noise and reducing the effects of intersubject variability in EEG/MEG signal strength (Feld et al. 2021). Spectral values for each epoch and region were averaged across the three spatial directions (x, y, z).

2.6. Analytical Approaches

To characterise sleep‐related brain activity across cortical, subcortical and cerebellar regions, we employed a series of complementary analyses: MEG signal differentiability, median power topographies, linear mixed‐effects (LME) models of sleep–wake contrasts and frequency‐specific investigations of N2 spindle‐band dynamics. N2 was selected as the primary focus for spindle‐band analyses given its high density of sleep spindles, making it the most relevant stage for investigating cerebellar involvement in sigma‐band dynamics. For these analyses, we first confirmed that cerebellar signals were distinguishable from neighbouring cortical sources before characterising spindle‐band activity and peak frequency distributions across brain regions.

2.6.1. MEG Signal Differentiability

We conducted a neural fingerprinting analysis to assess whether MEG power spectra could reliably distinguish signals between brain regions in whole‐brain source models, including the cerebellum. Neural fingerprinting is typically used to differentiate individual subjects based on brain features such as spectral information or connectivity patterns (da Silva Castanheira et al. 2021). The approach involves splitting each subject's data into two datasets, computing correlations between each subject's features in the first and second datasets and those of all other subjects. A continuous differentiability metric is defined as the z‐score of a subject's self‐correlation relative to their correlations with all other subjects (Amico and Goñi 2018; da Silva Castanheira et al. 2021). If a subject is easily differentiable, its differentiability metric is high, whereas lower differentiability scores indicate a subject that is more difficult to differentiate from the other subjects.

We applied the same logic to differentiate signals between brain regions rather than participants, with brain regions serving as our ‘subjects’. We randomly split data from sleep stages N1–N3 and wake into two separate datasets. For each dataset, power spectral features were computed for each frequency point by averaging normalised oscillatory power across epochs within each sleep stage or wake condition separately, preserving condition‐specific spectral profiles rather than a combined average. We then computed correlations between brain regions across both datasets, producing a per‐subject correlation matrix representing the similarity of power spectral patterns between regions, with rows corresponding to dataset 1, columns to dataset 2 and the diagonal representing self‐correlations for each brain region across datasets. The continuous differentiability metric (D _region ) was then computed for each region. To assess the overall differentiability of a region across subjects, a null distribution was created by computing a D _region score for every other region‐to‐region correlation. An empirical p‐value was generated by identifying the upper quantile of D _region within this subject‐specific null distribution. We used a threshold (uncorrected p < 0.05) to classify a region as differentiable for a given subject. The overall differentiability of a region was therefore summarised as the proportion of subjects for whom the region met this criterion.

2.6.2. Power Spectra Topographies and Models

2.6.2.1. Topographies

To generate brain topographies of oscillatory power, we first computed mean normalised oscillatory power within each frequency band (slow‐wave: 0.5–2 Hz; delta: 2–4 Hz; theta: 4–8 Hz; alpha: 8–12 Hz; sigma: 12–15 Hz; beta: 15–30 Hz) for each subject and sleep stage. Values were scaled within each subject and frequency band across brain regions and sleep stages. Group‐level medians were computed for each frequency band, sleep stage and brain region. Colour scales were set within each frequency band and rescaled so that the minimum value was set to zero.

2.6.2.2. Sleep Stage Contrasts With Wake

LME models were used to determine statistically significant changes in oscillatory power between each of the sleep stages and wake, across brain regions and frequency bands. These models account for the repeated‐measures structure of the data, with each subject contributing varying numbers of 30‐s epochs per sleep stage. One model was fit per brain region and frequency band, with normalised oscillatory power as the dependent variable, and fixed effects for sex and sleep stage (categorical: wake, N1, N2, N3), with wake as the reference level. Two random‐effects structures were compared using analysis of variance (ANOVA): a random intercept per subject and a random intercept with random slope for sleep stage; the better‐fitting model was retained for each region and frequency band.

All statistical analyses were conducted in R (R Core Team 2023). LME models were fit using the lmerTest package (Kuznetsova et al. 2017), which extends lme4 (Bates et al. 2015) to provide t‐statistics and p‐values for fixed effects using Satterthwaite's degrees of freedom method (Satterthwaite 1946). Where residuals were substantially non‐normal, a Box‐Cox transformation (Osborne 2010) was applied using the bcmixed package (Maruo et al. 2021) and the method described by Maruo et al. (2015). Models were refit to transformed data. We corrected for multiple comparisons across brain regions using the Benjamini–Hochberg false discovery rate (FDR) method (Benjamini and Hochberg 1995).

2.6.3. N2 Sleep Dynamics

2.6.3.1. Cerebellar Versus Occipital Power Spectra

To confirm that cerebellar signals were distinguishable from neighbouring cortical sources, we compared oscillatory power spectra of the cerebellum and occipital lobe during N2 sleep. Composite signals were generated for each region by averaging normalised oscillatory power across three closely neighbouring subregions per hemisphere per subject (see Figure 6a). Differences between regions were tested at each frequency point using paired t‐tests or Wilcoxon signed‐rank tests, depending on normality as assessed by the Shapiro–Wilk test (Lam and Longnecker 1983). Multiple comparisons across the frequency range were controlled using FDR correction (Benjamini and Hochberg 1995).

FIGURE 6.

Cerebellar versus occipital oscillatory power. This figure compares oscillatory activity in the cerebellum and occipital lobe during N2 sleep. (a) Composite signals are generated by averaging normalised oscillatory power across three closely neighbouring subregions per area, shown in dark green (cerebellum) and dark purple (occipital) against lighter backgrounds representing the full structures. (b) Mean normalised oscillatory power (with standard error) shows greater occipital activity in the theta range (4–7 Hz) and greater cerebellar activity at low (1.5–3 Hz) and sigma (12–15 Hz) frequencies in both hemispheres. The bottom bar displays log‐transformed FDR‐corrected p‐values from paired statistical tests; nonsignificant differences are shown in grey.

2.6.3.2. Spindle‐Band Oscillatory Power and Peak Frequency Topographies

To examine regional variations in spindle‐band activity during N2 sleep, we computed mean oscillatory power within 1 Hz intervals across the 11–18‐Hz range, following the same approach as the median topographies (Section 2.6.2) but at finer frequency resolution. Topographical maps were generated for each frequency, with cerebellar maps independently scaled relative to within‐cerebellum values to highlight regional differences. To summarise the dominant spindle frequency for each brain region, we identified the frequency between 11 and 18 Hz at which each region exhibited maximal oscillatory power for each subject and computed the median peak frequency across subjects. Results were visualised as a topographical brain map, allowing examination of spatial gradients in peak spindle frequency across the brain.

3. Results

3.1. Sleep Architecture

Total sleep time (N1, N2, N3) ranged from 29 to 146 min (mean = 74.3; SD = 40.0). Time spent in sleep stages 1 through 3 were as follows: Stage 1 sleep ranged from 3.0 to 25.5 min (mean = 12.8; SD = 7.0), Stage 2 from 8.0 to 96.0 min (mean = 39.3; SD = 26.5) and Stage 3 from 0.0 to 73.0 min (mean = 22.3; SD = 18.7). No participants entered REM sleep during their nap. Results are displayed in Figure 2.

FIGURE 2.

Minutes of sleep and wake during the 2.5 h nap opportunity. The time spent in each sleep stage and the wake and total sleep time are displayed; points represent individual participants. Across participants, total sleep time ranged from 29 to 146 min, demonstrating that all participants successfully napped in the scanner.

3.2. MEG Signal Differentiability Across Brain Regions

A region was considered differentiable if its empirical p‐value was less than 0.05, reflecting a strong differentiability index relative to its null distribution. The percentage of regions that were differentiable varied across subjects, ranging from 41.1% to 98.6% (mean = 79.5%, SD = 16.3%; see Table S1), indicating notable inter‐individual variability. The primary goal of this analysis was to assess whether certain regions, particularly the cerebellum, exhibited lower differentiability. We calculated the percentage of subjects in which each region showed good differentiability and visualised the results as topographical brain maps (Figure 3). Differentiability of subcortical sources, including the cerebellum, was similar to the majority of cortical regions. Slightly lower values of differentiability were observed in the anterior temporal poles relative to other cortical regions. Mean differentiability percentages were relatively consistent across cortical (79.6%, SD = 11.1%), subcortical (77.4%, SD = 11.0%) and cerebellar (80.4%, SD = 7.6%) areas (Table 1), with no significant differences detected across brain areas (Kruskal–Wallis test, p = 0.526). Overall, these results suggest that the cerebellum does not exhibit disproportionately lower differentiability compared with other brain regions, supporting its inclusion in MEG‐based analyses of sleep dynamics. However, individual variability remains, as indicated by the range of differentiability scores across participants (see Table S1). This variability informs our choice of summary measures to improve the consistency and reliability of interpretations; specifically, it motivated the use of medians for group‐level topographical summaries to ensure robustness against the influence of potential subject outliers.

FIGURE 3.

Topographical brain maps of differentiability. Each region is coloured according to the percentage of subjects in which it showed good differentiability (p < 0.05), providing a region‐wise index of differentiability. (Left) Whole–brain differentiability distribution. (Right) Subcortical and cerebellar differentiability. The differentiability of subcortical sources, including the cerebellum, is similar to the majority of cortical regions.

TABLE 1.

Mean differentiability percentage, standard deviation (SD), and number of regions in each brain area. Differentiability percentages were relatively consistent across cortical, subcortical and cerebellar areas, with no significant differences detected (p = 0.526, Kruskal–Wallis test). This suggests that MEG signal differentiation is comparable across brain areas, including subcortical regions that are infrequently studied using MEG.

Area	Mean	SD	Number of regions
Cortical	79.6	11.1	340
Subcortical	77.4	11.0	44
Cerebellum	80.4	7.6	32

3.3. Oscillatory Power Across Sleep Stages

Figure 4 shows median normalised oscillatory power for each frequency band across wake and NREM sleep stages (N1–N3). Slow‐wave power increased progressively with sleep depth, peaking in N3 with a characteristic frontal‐dominant topography. The cerebellum followed a similar trajectory of increasing power with sleep depth, but at lower magnitudes. Cortical delta power was low during wake across most regions but increased and became more spatially diffuse across N1 and N2, peaking in N2 across frontal, parietal and temporal regions before declining slightly in N3. In contrast, cerebellar delta was relatively elevated during wake and diminished progressively across sleep stages, diverging notably from the cortical trajectory. Theta power showed a progressive regional shift with deepening sleep: frontal activity during wake gave way to increasing occipital and cerebellar activity across N1–N3, with parietal power diminishing as sleep deepened. Alpha (8–12 Hz) and beta (15–30 Hz) power were both suppressed across sleep stages relative to wake, though their trajectories differed. Alpha power was highest during wake in occipital and cerebellar regions, with widespread decreases across all regions from N1 onward. Beta declined more gradually, with frontal and parietal regions remaining relatively elevated into N1 and N2; cerebellar beta followed a similar progressive reduction but remained lower in magnitude than cortical regions throughout. Sigma power (12–15 Hz) increased during N2 and N3, most prominently in frontal and parietal cortices and subcortical structures, consistent with the presence of sleep spindles. Cerebellar sigma power also increased during N2, though to a lesser extent than cortical regions.

FIGURE 4.

Oscillatory power across frequency bands and NREM sleep stages. Median normalised oscillatory power across six frequency bands (slow‐wave: 0.5–2 Hz; delta: 2–4 Hz; theta: 4–8 Hz; alpha: 8–12 Hz; sigma: 12–15 Hz; beta: 15–30 Hz) during wake and NREM sleep stages (N1–N3). Colour ranges are scaled within each frequency band across sleep stages to highlight state‐dependent changes in topography. Slow‐wave and sigma power increase during sleep, while alpha and beta power are suppressed relative to wake; delta and theta show more complex stage‐ and region‐specific profiles. Cerebellar activity largely follows cortical trends but generally remains lower in magnitude. Mean oscillatory power trajectories across sleep stages for selected regions are shown in Figure S2.

Mean oscillatory power trajectories across sleep stages for selected cortical, subcortical and cerebellar regions are shown in Figure S2, further illustrating the frequency‐ and region‐specific patterns observed in the median topographies.

3.4. Sleep–Wake Oscillatory Power Modulations: Regional and Frequency‐Specific Changes

To formally assess sleep‐related changes in oscillatory power, we applied LME models to quantify differences between each sleep stage and wake for each brain region and frequency band. Residuals from initial LME models revealed deviations from normality and homoscedasticity. A Box‐Cox transformation (Osborne 2010) was applied to the dependent variable. Comparisons of t‐statistics and coefficients before and after transformation indicated strong agreement, suggesting that non‐normality had minimal practical impact on model outputs; all reported results are based on models fit to the transformed data. Figure 5 displays model coefficients for contrasts reaching statistical significance (FDR‐corrected p < 0.05), with warmer colours indicating higher oscillatory power relative to wake and cooler colours indicating lower oscillatory power.

FIGURE 5.

Oscillatory power in sleep stages N1, N2, N3 versus wake. Topographical distributions of coefficients from linear mixed‐effects (LME) models, illustrating differences in oscillatory power between each sleep stage and wake across six frequency bands (slow‐wave: 0.5–2 Hz; delta: 2–4 Hz; theta: 4–8 Hz; alpha: 8–12 Hz; sigma: 12–15 Hz; beta: 15–30 Hz). Red indicates increased oscillatory power in sleep relative to wake, while blue represents decreases; all changes are thresholded at an FDR‐corrected significance level of p < 0.05. Regions that do not meet this threshold are shown in grey. Colours for each frequency band are scaled relative to its maximum absolute value.

Slow‐wave power showed minimal changes in N1, with no significant increases or decreases across cortical or cerebellar regions. By N2, widespread increases emerged across frontal and parietal areas and subcortical regions, though the cerebellum showed limited change. In N3, increases became larger and more spatially widespread, extending throughout the cerebellum and indicating broad engagement of brain regions in slow‐wave activity during deep sleep. Delta power increased in frontal and temporal regions in N1, becoming more widespread across frontal, parietal and temporal cortices by N2. However, portions of the inferior cerebellum and occipital lobe showed significant decreases relative to wake, a pattern that persisted and grew in magnitude through N3 as cortical increases became more spatially restricted. This divergence between slow‐wave and delta dynamics is notable: Whereas slow‐wave activity continued to increase through N3, delta‐band increases became more spatially restricted, with cerebellar and occipital decreases intensifying. Theta power showed a distinctive pattern of simultaneous increases and decreases across sleep stages. From N1 onward, occipital and cerebellar regions showed strong increases, while frontal regions showed significant decreases that persisted and grew in magnitude through N2 and N3. Cortical increases became more spatially restricted from N1 to N2, with parietal and temporal contributions diminishing, while occipital and cerebellar increases remained robust through N3. This pattern of region‐specific increases and decreases suggests different functional roles of theta activity during wake and sleep. Alpha power was progressively suppressed across all sleep stages, with widespread decreases emerging from N1 and deepening through N3 across cortical, subcortical and cerebellar regions. The strongest reductions were in occipital, parietal and cerebellar regions. Notably, some frontal areas did not differ significantly from wake in N2 and N3, overlapping spatially with regions showing theta‐band decreases across sleep stages. Sigma power increased primarily in frontal regions and subcortical structures in N1, with the cerebellum showing no significant changes at this stage. In N2, increases became stronger and more spatially extensive across much of the cortex and subcortical structures, and emerged prominently across large portions of the cerebellum. In N3, increases remained broad across frontal and parietal regions and subcortical structures, though cerebellar sigma changes were less prominent than in N2. Beta power showed a progressive reduction across sleep stages, with decreases emerging in occipital regions and portions of the cerebellum in N1. By N2, reductions extended across most of the cortex and throughout the cerebellum. In N3, decreases became the most spatially extensive of all bands, covering the entire cortex, subcortical structures and cerebellum, suggesting that beta suppression is a consistent and widespread feature of NREM sleep.

Together, these patterns reveal frequency‐specific trajectories of sleep‐related modulation, characterised by sleep stage 1 decreases in higher‐frequency bands (alpha, beta), stage‐dependent increases in lower‐frequency bands (slow, delta, theta) and distinct regional signatures of oscillatory engagement. Notably, the cerebellum exhibits both increases (in slow, theta and sigma bands) and decreases (in alpha, beta and delta bands), depending on frequency and sleep stage.

3.5. N2 Sleep Dynamics

3.5.1. Cerebellar Versus Occipital Oscillatory Profiles During N2 Sleep

We compared cerebellar and occipital oscillatory power spectra during N2 sleep to confirm that cerebellar signals reflect genuine local dynamics rather than leakage from adjacent cortical sources. Composite cerebellar and occipital signals were computed by averaging across three spatially adjacent subregions per area (Figure 6a). The two regions exhibited divergent power profiles: The occipital lobe showed enhanced power in the theta range (approximately 4–7 Hz), while the cerebellum showed higher power at low frequencies (approximately 1.5–3 Hz) and in the sigma range (12–15 Hz), which may reflect its specific role in spindle‐related processes (Figure 6b). These differences were statistically significant (FDR‐corrected p < 0.05), supporting the interpretation that cerebellar signals reflect different local dynamics rather than spread from occipital sources.

To further illustrate the cerebellar‐occipital dissociation in spindle‐band activity, Figure S3a,b in the Supporting Information show examples of time‐series from a cerebellar ROI and a nearby occipital cortical ROI, highlighting cases where spindle events are detected in one region but not in the other despite their close anatomical proximity. Figure S3c shows the anatomical location of the selected ROIs.

3.5.2. Spindle Band Oscillatory Dynamics During N2 Sleep: Cortical and Cerebellar Profiles

To characterise the spatial distribution of spindle‐band rhythms, median oscillatory power was computed for each frequency from 11 to 18 Hz using 1 Hz frequency bins, resulting in eight frequency‐specific spatial maps. This frequency‐resolved analysis extended beyond the classical sigma band (12–15 Hz) to capture the full progression of spindle‐related activity. This approach allowed us to observe transitions from lower amplitude activity to peak spindle engagement, particularly within the cerebellum, where spindle dynamics remain largely unknown. Cerebellar maps were independently scaled to highlight within‐cerebellum differences. Figure 7 displays results from 11 to 16 Hz, as this range captures the full progression from the onset to the decline of spindle‐related activity, with power substantially reduced beyond 16 Hz. During N2 sleep, spindle‐band activity exhibited frequency‐specific spatial patterns across cortical and cerebellar regions. Cortical power was low at 11 Hz but increased from 12 Hz onward, particularly in frontal areas, intensifying in anterior and parietal regions by 13 Hz. By 14 Hz, cortical activity became more spatially focused in parietal regions, before declining at 15–16 Hz. Cerebellar activity emerged prominently at 12 Hz and peaked broadly across the vermis and hemispheres at 13–14 Hz, coinciding with peak cortical engagement. By 16 Hz, activity was substantially reduced across both structures. Overall, cortical involvement was strongest and peaked in the 13–14‐Hz range, while the cerebellum showed a similar peak but with activity emerging at lower frequencies, suggesting that cerebellar contributions to spindle activity vary systematically across the spindle frequency range.

FIGURE 7.

Spindle band oscillatory dynamics in N2 sleep: whole‐brain and cerebellar patterns. Each panel shows median oscillatory power for the whole brain (top row) and cerebellum (bottom row) at a given frequency. Colour scales were chosen to highlight the evolution of spindleband activity across frequencies and to emphasise differences between regions. Enlarged cerebellar maps were independently scaled relative to values observed within the cerebellum. Cortical involvement is strongest overall and peaks in the 13‐ to 14‐Hz range. Cerebellar activity emerges prominently at 12 Hz, but peaks with cortical activity through 13 to 14 Hz.

3.5.3. Regional Distribution of Peak Spindle Band Frequencies

Topographical maps of median peak frequency across subjects (Figure 8) revealed that frontal regions exhibited lower peak frequencies and parietal regions higher peak frequencies, consistent with established anterior–posterior gradients of slow and fast spindles (Kumral et al. 2023). Most cerebellar regions showed peak frequencies within the high 13 Hz range (mean = 13.85, SD = 0.23), with the lowest peak frequency observed in right Crus II at 13.22 Hz (cyan, Figure 8) and the highest in left Crus II at 14.17 Hz (orange, Figure 8). Overall, cerebellar hemispheric differences were minimal (left: mean = 13.89, SD = 0.21; right: mean = 13.80, SD = 0.25) and regions fell spectrally between the lower frontal and higher parietal peaks. To investigate a potential lateralisation in cerebellar activity, we conducted additional exploratory paired t‐tests, comparing left and right peak frequency values across cerebellar ROIs. No results were significant after FDR correction, see Table S2.

FIGURE 8.

Regional variations in peak spindle band frequencies. Topographical brain maps show the median peak frequency (11–18 Hz) across regions during N2 sleep. Frontal areas exhibit lower peak frequencies, parietal regions higher peaks and cerebellar regions fall in the high 13 Hz range (mean = 13.85, SD = 0.23). The lowest cerebellar peak is in right Crus II at 13.22 Hz (cyan), and the highest in left Crus II at 14.17 Hz (orange). Overall, cerebellar hemispheric differences are minimal.

4. Discussion

Sleep involves widespread changes in neural activity across distributed brain networks, yet comprehensive characterisations of oscillatory dynamics have traditionally been limited to cortical regions. The present study addressed this gap by providing a systematic whole‐brain characterisation of oscillatory activity during NREM sleep using source‐localised MEG, explicitly including cerebellar regions. Our results reveal structured, frequency‐ and region‐specific patterns of sleep modulation that extend beyond traditional cortical‐thalamic circuits, including novel evidence for cerebellar engagement in sleep dynamics.

4.1. Methodological Validation: Reliable Differentiation of Deep Brain Signals

Our differentiability analysis confirmed that cerebellar regions exhibit differentiability scores comparable to those of cortical and subcortical areas, validating their inclusion in whole‐brain sleep analyses. This challenges the common assumption that deep brain structures are inaccessible to MEG, a view increasingly contested by empirical and simulation studies demonstrating that cerebellar activity can be detected with appropriate source modelling (Andersen et al. 2020; Samuelsson et al. 2020). The differentiability metric, adapted from neural fingerprinting approaches (Amico and Goñi 2018; da Silva Castanheira et al. 2021), quantifies how consistently a region's spectral profile can be matched across datasets, applied here to spatial rather than subject‐level comparisons. As noted by Dalal et al. (2013), the cerebellum's posterior regions lie at approximately the same depth as the occipital cortex, and its layered cellular organisation is capable of generating detectable magnetic fields. Omitting the cerebellum from source models may therefore lead to misattribution of cerebellar activity to nearby occipital regions (Samuelsson et al. 2020). These results support MEG as a tool for studying cerebellar activity in sleep and underscore the importance of whole‐brain modelling in MEG sleep research, even when cortical activity is the primary focus.

4.2. Characterisation of Neural Oscillations Reveals Frequency‐ and Region‐Specific Sleep Modulation

Slow‐wave power increased progressively with sleep depth, peaking in N3, consistent with the literature (Brancaccio et al. 2020). Our results show a more anterior cortical focus than the pronounced temporal increases reported by Brancaccio et al. (2020), revealing a spatial gradient in slow‐wave modulation, with occipital regions showing smaller differences and frontal regions exhibiting more pronounced increases. In the cerebellum, all regions showed significantly increased slow‐wave power in N3, aligning with EEG‐fMRI evidence linking slow waves to transient cerebellar activity (Dang‐Vu et al. 2008). Although the cerebellum does not show the largest increases, this likely reflects its characteristically lower signal amplitude (Dalal et al. 2013) rather than absence of engagement.

Delta power increased significantly from wake to N1 and N2, particularly in frontal and temporal areas, consistent with previous findings linking delta activity to frontal cortical engagement (Dang‐Vu et al. 2008; Brancaccio et al. 2020). By N3, however, this effect became more spatially restricted. The spatial restriction may reflect a functional shift toward the more globally synchronised slow‐wave dynamics characteristic of deep sleep, with activity redistributing toward lower frequencies as N3 deepens. Cerebellar delta showed the opposite pattern: Activity was highest during wake and declined progressively across NREM stages. This reduction was most pronounced in inferior posterior cerebellar regions. Notably, this pattern diverged from both cortical delta and cerebellar slow‐wave dynamics, suggesting a frequency‐specific modulation of cerebellar activity in the lower frequency range.

Theta power showed a mixed spatial profile, increasing in parietal and occipital areas while simultaneously decreasing in frontal regions from N1 onward, a pattern that persisted into N2 and N3. The posterior increases are consistent with Brancaccio et al. (2020), although the frontal decreases were not reported in that study. While theta is classically associated with REM sleep (Adamantidis et al. 2019), recent work has documented a frontal‐to‐occipital shift in theta power during NREM sleep (Cui et al. 2024), consistent with our observations. In the cerebellum, theta power increased across all sleep stages, closely mirroring posterior cortical increases and suggesting cerebellar engagement as part of a broader posterior network shift during NREM sleep. The co‐occurrence of occipital and cerebellar theta increases diverges from classical EEG characterisations of theta as a hallmark of both light and REM sleep and confined to midline frontal regions, underscoring the spatial specificity afforded by source‐modelled MEG. Whether this posterior theta activity during NREM supports processes analogous to those associated with theta during REM sleep warrants further investigation.

Alpha and beta power were both progressively suppressed across sleep stages, though their trajectories differed. Alpha showed the largest reductions in occipital regions where it is strongest during wake (Sadaghiani and Kleinschmidt 2016), consistent with classical descriptions of alpha attenuation at sleep onset (Brancaccio et al. 2020). Our results reveal a broad spatial gradient of suppression, extending beyond occipital into central regions. Notably, some frontal regions did not show significant reductions in N2 and N3, suggesting alpha anteriorisation (Cui et al. 2024) and the possibility of frontal alpha preservation or re‐emergence in deeper sleep stages. This effect has also been observed during propofol‐induced loss of consciousness (Purdon et al. 2013), potentially reflecting shared spatial dynamics across states of reduced consciousness. Beta declined more gradually than alpha. During wake, beta was strongest in frontal and parietal regions around the somatosensory and motor cortices, consistent with its role as a marker of sensorimotor processing during quiet wakefulness (Barone and Rossiter 2021). Despite this frontal–parietal wake baseline, reductions in these regions were not significant until N2 and N3, whereas decreases in occipital regions and the cerebellum emerged from N1 onward.

Sigma power increased most prominently during N2 across frontal, parietal and subcortical regions, consistent with the well‐established distribution of spindles as thalamo‐cortical events. Elevated sigma was also observed in N3, though more spatially restricted to parietal regions. Importantly, sigma power also increased significantly in the cerebellum during N2, without comparable increases in neighbouring occipital regions, aligning with emerging evidence that spindles reflect broader circuit‐level synchronisation involving cortico‐thalamic‐cerebellar networks (Xu et al. 2021).

Collectively, these findings reveal frequency‐specific trajectories of oscillatory modulation across the whole brain, with cerebellar regions participating in patterns that sometimes diverged from neighbouring cortical areas. These results challenge models that focus exclusively on cortical‐thalamic circuits and set the stage for a more detailed examination of cerebellar involvement in N2 spindle dynamics.

4.3. Spindle‐Band Activity During N2 Sleep

Our results revealed significant increases in sigma power in the cerebellum during N2, with no comparable increase in neighbouring occipital regions. This spatial dissociation provides important evidence that cerebellar spindle‐related activity reflects genuine local dynamics rather than signal leakage from nearby cortical sources, a common concern in MEG studies of deep brain structures. The pattern of results suggests that cerebellar sigma activity may reflect independent local generation or targeted coordination via subcortical structures, rather than being a by‐product of posterior cortical spread. Our interpretation is supported by prior findings demonstrating separate functional contributions of spindles depending on their cortical location. For example, Lutz et al. (2021) showed that the amplitude of spindles over the occipital cortex predicted improvements in visual sequence learning, indicating that spindles may support experience‐specific consolidation in perceptual circuits. In contrast, our observed cerebellar sigma activity may reflect coordination of motor‐related learning or internal model updating, in line with the cerebellum's known role in adaptive control and prediction.

Fine‐grained analysis of spindle‐band activity in 1 Hz intervals from 11 to 18 Hz revealed a systematic anterior‐to‐posterior shift in cortical power, from frontal regions at 12 Hz to parietal regions at 13–14 Hz. These findings align with longstanding observations of two spindle subtypes: slow spindles (10–13 Hz), typically occurring over frontal regions, and fast spindles (13–16 Hz), most commonly observed in central–parietal areas (Kumral et al. 2023). However, the majority of prior work has focused exclusively on cortical topographies. In our cerebellar‐specific analysis, power was scaled relative to cerebellar activity. Spindle power began to emerge at 12 Hz with a clear peak at 13–14 Hz. Frontal regions showed lower frequency median peaks (12–13 Hz), parietal areas showed higher peaks (14–15 Hz) and cerebellar regions fell in the intermediate range (13–14 Hz), suggesting a spectrally distinct cerebellar contribution. This frequency dissociation echoes the early intracranial observations of Niedermeyer and Uematsu (1974), who noted that sleep spindles recorded from cerebellar depth leads showed lack of synchrony with scalp spindles and proposed that cerebellar spindles may be based upon different neuronal loops. Consistent with this view, Figure S3 shows examples of time series from a cerebellar ROI and a nearby occipital ROI in which spindle detections were region‐specific, with spindle events present in one region but absent in the other. Our findings extend the interpretation proposed by Niedermeyer and Uematsu: The peak frequency of cerebellar spindles, falling between the slow frontal and fast parietal subtypes, suggests not merely a timing difference but a separate circuit, pointing to partially independent cerebellar oscillatory dynamics within the spindle band rather than only entrainment to cortical rhythms.

4.4. Limitations and Future Directions

Several limitations of the present study should be noted. The data were collected during a daytime nap, which lacks the cyclical architecture of nocturnal sleep and did not include REM sleep, limiting generalisability to full‐night sleep architecture. Future work should extend these findings using full‐night MEG recordings to capture the complete NREM–REM cycle and examine whether REM‐specific rhythms exhibit similar regional specificity.

One potential caveat of the present study concerns the MSL task performed prior to the nap opportunity, as part of the larger study. Because waking activity can affect the strength of neural oscillations in subsequent sleep (e.g., Sjøgård et al. (2025)), it is possible that the present results reflect a bias toward motor‐related memory processing. However, the sleeping brain produces neurophysiological activity including slow oscillations and sleep spindles irrespective of specific task involvement, and sleep is naturally preceded by varied cognitive and sensorimotor experiences. The patterns reported here are broadly consistent with prior work studying sleep oscillations in the absence of a specific task, and we therefore expect the averaged patterns to be representative of sleep physiology more generally.

Our aim was to characterise the spatial and spectral structure of sleep oscillations rather than to link them to specific behavioural outcomes. The question of how prior learning modulates cerebellar and subcortical contributions to sleep remains compelling and is the focus of ongoing work. The current study characterises physiological patterns as a necessary precursor to establishing functional significance.

In addition to investigating the relationship between neural oscillations, behavioural performance and sleep‐dependent memory consolidation, future directions include investigating individual variability in peak spindle frequencies, examining directed connectivity between cerebellar, thalamic and cortical sources using methods such as Granger causality or dynamic causal modelling, and exploring cross‐frequency coupling to better characterise coordination mechanisms across the sleep network. Comparing MEG with fMRI signals from the same cohort could further validate the spatial precision of MEG and support integrative models of cortico‐thalamic‐cerebellar interactions during sleep.

5. Conclusion

Our systematic characterisation reveals structured patterns of sleep‐related oscillatory activity across the whole brain, spanning six frequency bands and three NREM sleep stages. Traditionally, MEG and EEG studies of sleep have focused on cortical sources, often excluding cerebellar regions due to concerns about signal detectability. Our results challenge these assumptions by demonstrating that cerebellar activity is not only measurable with source‐modelled MEG but also functionally structured, exhibiting frequency‐specific and stage‐dependent patterns.

Successfully mapping cerebellar signals during sleep expands the spatial scope of noninvasive electrophysiological sleep research, addresses longstanding methodological concerns and establishes a foundation for whole‐brain sleep studies. Capturing cerebellar dynamics opens new opportunities to investigate how subcortical regions contribute to sleep‐dependent processes such as memory consolidation, predictive coding and motor learning. Characterising these dynamics in relation to specific frequency bands, alongside well‐established cortical patterns, can help guide future research in refining models of sleep physiology.

The implications are particularly relevant for spindle activity. The cerebellum shows significant sigma‐band power increases during N2 sleep, independent of neighbouring occipital cortex and cerebellar oscillatory power peaks in the mid‐fast spindle range, positioned spectrally between lower‐frequency frontal spindles and higher frequency parietal spindles. Together, these findings will facilitate deeper investigation of the existence, structure and function of cerebellar spindles in the human brain.

Author Contributions

Keelin Greenlaw: conceptualisation, data curation, formal analysis, methodology, software, visualisation, writing – original draft, writing – review and editing. Anne Calvel: visualisation, writing – review and editing. Camille Bouhour: data curation, project administration. Christopher J. Steele: conceptualisation, data curation, formal analysis, funding acquisition, methodology, project administration, resources, software, supervision, visualisation, writing – review and editing. Emily B. J. Coffey: conceptualisation, data curation, formal analysis, funding acquisition, methodology, project administration, resources, software, supervision, visualisation, writing – review and editing.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

FIGURE S1: Distribution of total instrument hours across participants. Musical experience in our sample reflects a typical young adult convenience sample: 12 of 19 participants reported no formal musical training, with only seven reporting any instrument experience.

TABLE S1: Percentage of regions with good differentiability (p < 0.05) for each subject.

FIGURE S2: Mean oscillatory power from selected regions across wake and NREM sleep stages. (a) Five representative regions selected to illustrate oscillatory dynamics across cortical, subcortical and cerebellar structures: Cerebellum 4/5, Thalamus 5, Frontal mid 1 (frontal cortex), Rolando 4 (sensorimotor cortex) and Parietal Sup. 3 (parietal cortex), shown in lateral (left) and dorsal views. (b) Mean normalised oscillatory power (with standard error) across wake and NREM sleep stages (N1–N3) for each frequency band. Values represent averages across bilateral homologues. Each band captures a distinct pattern of sleep‐related modulation, illustrating the frequency‐ and region‐specific dynamics described in the results.

FIGURE S3: Dissociated spindle detection in nearby cerebellar and occipital regions. (a) Spindle detected in Cerebellum 4/5 but not Occipital Lat. 3; (b) spindle detected in Occipital Lat. 3 but not Cerebellum 4/5. The red patch marks the detected spindle, while the grey patch shows the corresponding time window in the other region, where no spindle is detected. Spindle detection is performed by first inspecting the broadband signal (0.5 to 40 Hz), then combining the three directions, filtering the resulting signal in the sigma band (9 to 18 Hz) and computing its RMS envelope with a 0.2‐s window. Candidate spindles are identified using an ROI‐specific threshold defined as mean RMS + 1.5 SD and retained when their duration fell between 0.5 and 3 s. The corresponding time‐frequency representation is shown for visualisation. (c) Anatomical location of Cerebellum 4/5 and Occipital Lat. 3. Cerebellum 4/5 is shown in dark green and Occipital Lat. 3 in dark purple. Cerebellar subregions are displayed in light green, and occipital regions in light purple. Shaded backgrounds represent the full anatomical structures.

TABLE S2: Left–right hemisphere comparisons of peak spindle band frequency across cerebellar ROIs. Paired t‐tests comparing left and right hemisphere peak frequencies across participants (n = 19). Two regions showed uncorrected differences (p < 0.05) in opposite directions; neither survived FDR correction, providing no robust evidence for consistent cerebellar lateralisation in peak spindle band frequency.

Greenlaw, K. , Calvel A., Bouhour C., Steele C. J., and Coffey E. B. J.. 2026. “Sleep Oscillations Across Cortical, Subcortical and Cerebellar Structures in Magnetoencephalography.” European Journal of Neuroscience 64, no. 1: e70615. 10.1111/ejn.70615.

Data Availability Statement

Anonymised, preprocessed MEG data (in the form of normalised oscillatory power) used in the main analyses are freely available on the Open Science Framework: https://osf.io/ewmgc/?view_only=0c70d62da2924de697ecf30e7be3ee8a. Raw data may be available upon reasonable request, pending ethical approval.