Divergent Cortical Generators of MEG and EEG during Human Sleep Spindles
Sleep spindles are among the most recognizable oscillatory events in human neurophysiology. They appear as brief, waxing-and-waning bursts of approximately (10)–(15\,\mathrm{Hz}) activity, typically lasting on the order of one second, and are most prominently observed during stage 2 non-rapid eye movement sleep. Since the early days of electroencephalography, spindles have served as a central example of thalamocortical rhythmicity: a macroscopic oscillation that reflects the interaction of cortical circuits, thalamic relay cells, and the thalamic reticular nucleus.
But the apparent simplicity of the spindle waveform hides a deeper question. Is the human sleep spindle a globally synchronous cortical event, generated by a broadly phase-locked thalamocortical system? Or is the scalp-visible spindle a macroscopic projection of multiple local, partially independent generators whose activity happens to fall in a similar frequency range?
This paper addressed that question by comparing simultaneous electroencephalography (EEG) and magnetoencephalography (MEG) recordings during naturally occurring human sleep spindles, and by projecting both modalities into cortical source space using distributed source modeling. The result was not simply that EEG and MEG differ at the sensor level. The more important finding was that the divergence persists after source reconstruction: EEG-derived sources appeared spatially widespread and highly synchronous, whereas MEG-derived sources were more focal, variable, and dynamically shifting across cortical regions and hemispheres.
The paper therefore sits at the intersection of sleep neurophysiology, statistical source modeling, cortical biophysics, and the physics of inverse problems. It asks a deceptively basic question: when we observe a neural oscillation at the scalp, what kind of cortical process are we actually seeing?
The classical view: spindles as distributed thalamocortical synchrony
Sleep spindles have long been interpreted through the framework of thalamocortical synchronization. In animal preparations, spindle-like rhythms can emerge from interactions between inhibitory neurons in the thalamic reticular nucleus and thalamocortical relay cells. These rhythms can then recruit cortical circuits through thalamocortical projections, producing oscillatory activity visible in cortical field potentials.
A natural extrapolation to human EEG is that scalp spindles reflect a broadly synchronized generator. This interpretation is encouraged by the high coherence of spindle discharges across spatially separated EEG electrodes. If frontal, central, and parietal scalp electrodes all show spindle activity at similar times and frequencies, it is tempting to infer a globally phase-locked cortical source.
However, several observations complicate that interpretation.
First, human spindles are not always homogeneous across the scalp. Frontal and parietal spindles can differ in frequency, with slower components often more prominent frontally. Second, MEG studies have suggested that spindle activity may arise from multiple sources that vary across time, spindles, and subjects. Third, simultaneous EEG and MEG recordings do not always show a one-to-one correspondence: some spindles are apparent in EEG but not MEG, others in MEG but not EEG, and some in both.
This raises a fundamental issue. EEG and MEG are generated by related physiological currents, but they are not identical measurements. They are filtered differently by tissue geometry, source orientation, skull conductivity, cortical folding, and the sensor physics of electric versus magnetic field detection. Thus, similarity or dissimilarity at the sensor level cannot by itself determine whether the underlying cortical generators are the same.
The paper asked whether the apparent EEG–MEG divergence would remain after solving the inverse problem: after estimating the cortical sources that could have generated the observed sensor-level signals.
The inverse problem is not a technical detail
In neuroimaging, source modeling is often treated as a computational step between data acquisition and interpretation. But for EEG and MEG, the inverse problem is not a neutral transformation. It is mathematically ill-posed.
The forward problem can be written schematically as
[ \mathbf{y}(t) = \mathbf{G}\mathbf{x}(t) + \boldsymbol{\epsilon}(t), ]
where (\mathbf{y}(t)) is the measured sensor activity, (\mathbf{x}(t)) is the unknown source activity, (\mathbf{G}) is the lead field or gain matrix determined by head geometry and tissue conductivities, and (\boldsymbol{\epsilon}(t)) is noise. The inverse problem asks for (\mathbf{x}(t)) given (\mathbf{y}(t)) and (\mathbf{G}).
But there are many possible source configurations that can produce the same or nearly the same sensor pattern. A unique inverse solution is impossible without assumptions. These assumptions may concern source depth, spatial smoothness, dipole orientation, noise covariance, anatomical constraints, or the norm of the solution.
In this study, the inverse solution was based on dynamic statistical parametric mapping, or dSPM. The source space was constrained to the reconstructed cortical surface of each subject, derived from high-resolution structural MRI. The forward model used a realistic three-shell boundary element model representing the inner skull, outer skull, and scalp. This was important because EEG is strongly affected by skull and scalp conductivity, while MEG is less sensitive to some of these compartments. To compare modalities as fairly as possible, both EEG and MEG were modeled using the same anatomical framework.
The inverse solution was a noise-normalized minimum-norm estimate. In simplified terms, a minimum-norm solution selects the source configuration with relatively small overall source power that can explain the measured sensor data. dSPM then normalizes this estimate by noise sensitivity, so that the resulting map is closer to a statistical map of source significance than to a raw estimate of dipole moment.
The cortical surface was represented by approximately (6500) dipole locations, spaced at roughly (7\,\mathrm{mm}). For each spindle, source estimates were computed separately from EEG, separately from MEG, and also from combined MEG+EEG data.
This architecture of analysis matters because the paper was not merely comparing sensor waveforms. It was comparing the inferred cortical dynamics of the same sleep spindles under the same anatomical and statistical modeling framework.
Simultaneous recordings during natural sleep
The study recorded natural sleep from seven healthy adults. The recordings included (306) MEG channels and (60) EEG channels. MEG was acquired with a whole-head system containing magnetometers and planar gradiometers; EEG was recorded simultaneously. Structural MRI was collected for each participant and used for cortical reconstruction and realistic forward modeling.
Stage 2 sleep and spontaneous spindles were identified using standard criteria. The final analysis used (85) spindles across the seven subjects. The mean spindle duration was approximately (721\,\mathrm{ms}), with variability across events.
This simultaneous design was essential. If EEG and MEG are acquired separately, differences between the modalities can always be attributed to different spindles, different sleep states, different subjects, or different recording contexts. Here, the comparison was made during the same physiological events. The EEG and MEG were observing the same brain at the same time.
Time-averaged similarity hides dynamic divergence
If one averages across the full duration of spindles, and then across spindles, the EEG- and MEG-derived source maps do not look completely unrelated. Both modalities tend to show broad cortical involvement, with strong activity in medial cortical regions, including cingulate, parahippocampal, medial parietal, central, and frontal regions. This agrees broadly with previous EEG and MEG source-localization studies of sleep spindles, many of which implicated medial frontal, parietal, and central regions.
But this time-averaged view is misleading.
A spindle is a dynamical event. It is not simply a static spatial map with an oscillatory label attached to it. The critical question is not only where spindle-related activity appears across an event, but how the estimated sources evolve over tens of milliseconds within the event.
When the source estimates were examined dynamically, the EEG and MEG solutions showed strikingly different temporal organization.
EEG-derived source activity appeared broadly synchronous across the cortical surface. The estimated equivalent current dipoles rose and fell together across large regions of cortex. The spatial pattern changed in amplitude, but remained relatively stable in its large-scale organization.
MEG-derived source activity behaved differently. The estimated sources shifted rapidly in phase, amplitude, hemisphere, and cortical location. Within a single spindle, maximal MEG-derived activation could move from one cortical region to another, sometimes from one hemisphere to the other, across successive cycles of the oscillation.
Thus, a spindle that appears coherent in EEG can correspond, in MEG source space, to a sequence of more local and partially independent events.
Quantifying synchrony in source space
The distinction between EEG and MEG source dynamics was quantified by measuring correlations among source time courses.
For each spindle and each modality, the analysis considered the time courses of all pairs of cortical dipoles. With approximately (6500) source locations, this produces millions of pairwise comparisons. Averaging across these pairwise correlations provided a measure of within-modality synchrony.
The result was clear:
[ r_{\mathrm{within,EEG}} \approx 0.64, ]
whereas
[ r_{\mathrm{within,MEG}} \approx 0.13. ]
The EEG-derived sources were therefore highly correlated across cortical locations, consistent with large-scale synchrony. The MEG-derived sources were much less correlated, consistent with spatially heterogeneous and asynchronous generators.
The between-modality comparison was even more revealing. For each cortical location, the time course estimated from EEG was compared with the time course estimated from MEG at the same location. If EEG and MEG were simply two sensor-level views of the same reconstructed cortical process, these time courses should have been strongly correlated.
They were not. The average EEG–MEG source-space correlation was approximately
[ r_{\mathrm{EEG,MEG}} \approx 0.09. ]
This is remarkably low given that the signals were recorded simultaneously during the same spindles and reconstructed using the same anatomical model.
However, the modalities were not completely unrelated. Spectral coherence in the spindle band was moderate. Using MVDR spectral estimation, the average coherence between EEG- and MEG-derived source time courses in the (10)–(15\,\mathrm{Hz}) band was approximately
[ C_{\mathrm{EEG,MEG}} \approx 0.44. ]
Using Welch’s method, the corresponding value was higher, approximately (0.54).
This combination is important: low time-domain correlation but moderate frequency-domain coherence. It suggests that EEG and MEG source estimates share rhythmic content in the spindle frequency range, but differ in phase, timing, and spatial expression. In other words, the two modalities are not measuring unrelated phenomena, but neither are they simply redundant measures of the same cortical generator.
Why EEG and MEG need not see the same spindle
At first glance, the divergence between EEG and MEG may seem paradoxical. Both signals ultimately arise from neural transmembrane currents. Active transmembrane currents and passive return currents are part of the same physiological circuit. Why, then, should EEG and MEG suggest different cortical generators?
The answer lies in the physics of field generation and measurement.
EEG is sensitive to extracellular currents and is strongly influenced by volume conduction through brain tissue, cerebrospinal fluid, skull, and scalp. MEG is sensitive to magnetic fields generated primarily by intracellular currents and is less distorted by the skull, but it has its own selectivity. In particular, MEG is relatively insensitive to radially oriented dipoles, whereas EEG can detect both radial and tangential components.
Cortical folding further complicates the relationship between source activity and measured fields. The cortex is not a flat sheet. It is a highly folded surface, and neighboring patches of cortex can have dipoles with opposing orientations. When distributed sources are coactivated, their fields can cancel before reaching the sensors. Simulations of realistic cortical architectures have shown that even coactivation of a small fraction of cortical dipoles can lead to cancellation of most of the measurable extracranial signal.
This means that inverse estimates should not be interpreted as maps of all active cortical tissue. They are estimates of the cortical activity that survives the biophysical filtering imposed by geometry, tissue, orientation, and sensors.
A distributed oscillatory process may therefore project differently into EEG and MEG. EEG may preferentially reflect a large-scale synchronous component that produces a coherent scalp potential. MEG may preferentially reveal more focal tangential components that are less visible in EEG or that are partially canceled in the electric field. Conversely, some EEG-visible diffuse components may generate little measurable MEG because their magnetic fields cancel or because their geometry is unfavorable.
The paper therefore does not treat EEG–MEG divergence as a nuisance. It treats it as evidence about the physical structure of spindle generation and measurement.
The core–matrix interpretation
The neurophysiological interpretation proposed in the paper draws on the distinction between two thalamocortical projection systems: the core and matrix systems.
The matrix system consists of broadly projecting thalamocortical neurons, especially associated with nonspecific and intralaminar thalamic nuclei. These neurons project diffusely across cortical areas and terminate prominently in superficial cortical layers, including layer I. Such a system is well suited to generating or coordinating widespread cortical synchrony.
The core system, by contrast, consists of more specific thalamocortical projections, often associated with sensory relay nuclei. These projections are more focal and terminate in middle cortical layers, especially layer IV. Such a system is better suited to producing localized, modular thalamocortical activity.
Classical animal studies of spindle generation described both distributed and focal spindles. Some spindle events were broadly synchronous, while others were restricted to local thalamocortical modules with independent onset, duration, phase, and frequency. The core–matrix distinction provides a plausible anatomical and physiological framework for understanding this coexistence.
The EEG–MEG divergence observed in the paper can be interpreted through this framework. EEG-derived source estimates, being widespread and synchronous, may preferentially reflect activity of a diffuse matrix-like thalamocortical system. MEG-derived source estimates, being more focal, variable, and asynchronous, may preferentially reflect local core-like thalamocortical modules.
This interpretation should not be read as a claim that EEG equals matrix and MEG equals core in a literal one-to-one sense. The mapping between physiology and extracranial measurement is too indirect for that. Rather, the claim is that the two modalities may have differential sensitivity to different components of the spindle-generating system. The scalp EEG spindle may emphasize the globally synchronized component, while MEG may reveal a more fragmented and locally organized component of the same broader thalamocortical event.
Spindles as structured heterogeneity, not a single global oscillator
One of the broader lessons of the paper is that oscillatory labels can hide mechanistic heterogeneity. Calling an event a “spindle” identifies a frequency band, a sleep stage, and a waveform morphology. It does not uniquely identify a generator.
A spindle may be composed of several interacting components:
- a diffuse component that is spatially broad and highly synchronous;
- local cortical or thalamocortical modules with distinct phases and amplitudes;
- frequency gradients across cortical regions;
- modality-dependent visibility determined by source orientation and field cancellation;
- temporal evolution across successive cycles within the same spindle.
From this perspective, a spindle is not necessarily a unitary cortical event. It is a macroscopic signature of thalamocortical dynamics, and different measurement modalities may reveal different projections of that dynamics.
This point is especially important for neurophysiology because synchrony is often inferred from field coherence. But coherence at the sensor level does not necessarily imply a single coherent source. Conversely, heterogeneity at the source level does not imply the absence of a global rhythm. The paper shows that both can be true: EEG can show large-scale synchrony while MEG reveals local heterogeneity during the same physiological event.
Statistical interpretation: correlation, coherence, and phase
The distinction between correlation and coherence is central to the interpretation of the results.
Correlation measures time-domain linear similarity. If two source time courses rise and fall together with the same phase and amplitude relationship, their correlation is high. If they oscillate at the same frequency but with shifting phase, their correlation can be low.
Coherence measures frequency-domain association. Two signals may have moderate coherence in a given frequency band even if their time-domain correlation is low, especially when their phase relationship is nonzero or variable.
This is exactly the pattern observed in the paper. EEG- and MEG-derived source time courses were poorly correlated but moderately coherent in the spindle band. This indicates that both modalities are connected to the spindle rhythm, but they do not express that rhythm with the same spatial and temporal structure.
For source modeling, this matters because a purely spatial comparison of average activation maps would miss the key result. The important divergence is dynamical. It appears in the phase, timing, and movement of source activity across the cortical surface.
Why the combined MEG+EEG solution is not the final answer
The paper also computed source estimates using combined MEG and EEG data. One might think that the combined solution should resolve the disagreement between modalities. But the situation is more subtle.
A combined inverse solution attempts to find sources that can explain both MEG and EEG measurements. However, accurate combination of the two modalities requires correct relative scaling of MEG and EEG amplitudes. This is not trivial. Without an external calibration source, such as a known single tangential dipole response, the combined solution may fit spatial patterns without fully resolving the absolute amplitude relationship between the modalities.
This is particularly important for spindles because the expected EEG amplitude from a focal MEG-visible source may be far smaller than the observed EEG spindle amplitude. Conversely, a diffuse EEG-visible source may contribute little to MEG. Therefore, a physiologically accurate combined model might need to include both a diffuse synchronous generator and multiple focal asynchronous generators.
In that sense, the combined solution should not be interpreted as proving that there is one compromise generator. It may instead reflect the difficulty of forcing distinct modality-weighted components into a single inverse estimate.
Limitations and caution
The paper is careful about the limitations of the inference. The electromagnetic inverse problem remains ill-posed, even with realistic head models and cortical constraints. dSPM imposes assumptions, and different inverse methods may emphasize different features of the data. The study used a three-shell BEM, but future models could improve the treatment of cerebrospinal fluid, skull conductivity, and white-matter anisotropy. These factors are especially important for EEG.
The study also did not include simultaneous intracranial recordings, which would provide a more direct test of the inferred cortical sources. Intracranial EEG studies have shown that some spindles are local and that intracranial spindles do not always have a simple relationship to scalp spindles. But simultaneous scalp EEG, MEG, and intracranial recordings would be the stronger test.
Thus, the central conclusion is not that the exact cortical sources are definitively known. The stronger conclusion is that EEG and MEG, even during the same spindle events and under the same distributed source modeling framework, exhibit markedly different source-space dynamics. Any theory of human spindle generation must account for that divergence.
Why this matters beyond sleep spindles
Although the paper focuses on sleep spindles, the broader issue applies to many neural oscillations. Alpha rhythms, beta bursts, gamma activity, slow waves, epileptiform discharges, and evoked responses are all often interpreted through scalp-level EEG or MEG patterns. But scalp-level oscillations are projections of cortical and subcortical dynamics through a complex physical measurement apparatus.
The lesson is not that EEG or MEG is superior. The lesson is that they are complementary physical measurements. Each modality emphasizes different aspects of the same underlying neural system. EEG is highly sensitive to large-scale synchronous potentials and volume-conducted fields. MEG is more selective for certain source orientations and may reveal focal tangential generators with less skull distortion. Neither modality provides a complete view by itself.
For statistical neuroimaging, this means that source modeling should not be treated as a mere localization tool. It should be treated as a generative modeling problem constrained by physics, anatomy, noise, and physiology. The goal is not only to ask where an oscillation is generated, but what kind of spatially distributed dynamical system could have produced the measured fields.
For sleep neurophysiology, the paper supports a view of spindles as structured, multiscale events: globally rhythmic, but locally heterogeneous; coherent in frequency, but not necessarily phase-locked across all generators; visible through EEG as a broad synchronized field, but through MEG as a shifting set of focal sources.
Concluding perspective
The central message of the paper is that the human sleep spindle should not be reduced to a single global oscillator. Simultaneous EEG and MEG reveal different aspects of spindle dynamics, and distributed source modeling suggests that these differences persist in cortical source space.
EEG-derived sources are widespread, stable, and highly synchronous. MEG-derived sources are more focal, less mutually correlated, and dynamically shifting across the cortex. The two modalities share spindle-band rhythmicity, but they do not express that rhythm with the same phase, spatial organization, or temporal evolution.
This divergence is not merely a methodological complication. It is a clue. It suggests that human sleep spindles may involve both diffuse thalamocortical synchronization and local modular generators, possibly related to the matrix and core thalamocortical systems. It also illustrates a broader principle: macroscopic neural oscillations are shaped not only by neural circuits, but by the physics of how those circuits project to sensors.
A serious theory of brain rhythms must therefore be simultaneously physiological, statistical, and physical. Sleep spindles are not just waves in a frequency band. They are dynamical events generated by structured neural circuits and observed through modality-specific filters. Understanding them requires treating synchrony not as an assumption, but as an object of measurement.