Structure-function coupling of large-scale cortical networks across the lifespan is spectrally specific

Abstract

With the advent of large-scale neuroimaging datasets, expansive lifespan studies are making novel discoveries about foundational principles of human brain development and aging. In one of the largest lifespan magnetoencephalography (MEG) studies of its kind (N = 388, age 6–84 years), we document age-related coupling between local rhythmic neural activity and cortical thickness using cutting-edge analyses of whole-brain structure-function coupling. This work establishes normative linear and non-linear age-related changes that fluctuate in spectrally specific patterns and that follow a hierarchical cortical gradient from sensory to association networks. Moreover, these results demonstrate that canonical spectral bands tend to display strong lifespan coupling to structure in sensorimotor networks and diminished, more variable coupling in association networks with age. This reinforces accounts of less phylogenetic conservation of association networks due to their relative evolutionary novelty, with sensorimotor networks being more strictly conserved. The addition of spectral signatures to this account also reveals that dominant, spectrally specific oscillations (i.e., delta, theta, alpha, beta, gamma) emanating from local neural assemblies follow a hierarchical structure. Taken together, this landmark study provides a dynamic view of how age-related reorganization of large-scale networks unfolds across the human lifespan through structure-function coupling in a spectrally specific and hierarchically organized manner.

Structure-function coupling between local rhythmic neural activity and cortical thickness reveals how age-related reorganization of large-scale networks unfolds across the human lifespan in a spectrally-specific and hierarchically organized fashion.

Introduction

Decoding the organizational properties of neural structure and function is a central goal of the field of neuroscience. To date, extensive efforts have been made to define structure-function coupling (SFC) patterns within specific developmental windows, particularly using MRI-based modalities, with applications to disease states and highly plastic windows of development (e.g., adolescence)^1,2. However, with the accrual of larger datasets that capture the majority of the human lifespan, understanding the full concordance of functional and structural metrics is finally within reach. Such large-scale studies measuring brain development and aging across the lifespan are especially well-positioned to examine the non-linearity of age-related change that appears to be a prominent feature of many neuroimaging metrics^3–5. Nonetheless, work in this area remains limited and fragmented, particularly when considering millisecond-scale neural dynamics that are measurable by magnetoencephalography (MEG). Importantly, SFC has been operationalized in multiple ways, including correspondence between structural and functional connectivity profiles across regions¹, as well as spatial correspondence between structural and functional properties within the cortex⁶. These approaches capture distinct but complementary aspects of brain organization.

Following landmark conceptualizations of neural plasticity⁷, studies have repeatedly shown alignment between the development of functional specialization and underlying brain structure, such that cortical gradients are arranged within a functional hierarchy⁸. This hierarchy spans from sensorimotor (i.e., low-level, unimodal) to associative networks (i.e., higher-order, transmodal)¹, whereby cortical morphometry is more tightly coupled with functional topography in sensorimotor compared to associative networks⁹. Heritability estimates of these patterns support this notion, in that the networks with the lowest heritability tend to be higher-order networks¹⁰, suggesting that the development of SFC tends to be more phylogenetically conserved in sensorimotor networks. Importantly, the field’s ability to capture variability in the organizational properties of cortical gradients, especially the higher-order networks that are subject to greater variability, is essential to understanding human cognition¹¹ and neuropathology¹².

Extant SFC studies have focused on either young adult samples or youth, with only a few capturing broader windows of the life course¹³. Nonetheless, these approaches have yielded crucial insights into developmental cortical expansion, and subsequent refinement and declines with healthy aging^2,14. The present study extends this work by leveraging a dataset of 6 to 84-year-olds to garner a more complete representation of the lifespan. Further, to date, MRI methods have been the dominant modality for studying structure-function relationships, with limited use of tools like MEG that capture other facets of the neural signal, such as spontaneous rhythms. Recent work suggests that the hemodynamic response function changes with age in ways unrelated to neural activity (e.g., vascular changes) and this may confound some indices relying on the BOLD response¹⁵, although MEG-based measures should be unaffected. MEG offers exquisite temporal resolution (~1 ms) and high spatial precision (~3-5 mm), making it uniquely suited to quantify oscillations and regional spontaneous rhythms. Moreover, decades of animal and human work have highlighted the utility of quantifying spectral information emanating from behavior-dependent neural assemblies¹⁶, with demonstrations that large-scale networks exhibit distinct oscillatory signatures that are consequential for development^5,17, aging^13,18, and neuropathology^19,20.

The MEG signal primarily reflects coordinated neurophysiological activity in pyramidal neurons, inhibitory interneurons, and other neuronal cell types²¹, which are all local constituents of cortical thickness²². Although cortical thickness and spontaneous spectral power are distinct parameters that change on very different time scales (i.e., months/years vs. milliseconds), they are both strongly influenced by cortical maturation, including synaptic density, dendritic architecture, and circuit refinement^21,22. Thus, tracking how changes in cortical anatomy co-fluctuate with spontaneous cortical dynamics, and examining how this relationship varies across networks, spectral windows, and stages of life, offers a unique vantage into sensitive periods when cortical circuits may be especially plastic and open to reorganization. In this context, we quantify SFC as the spatial correspondence between cortical thickness and spontaneous spectral power within the same cortical tissue. While alternative metrics of brain structure (e.g., myelin-sensitive measures) and function (e.g., functional connectivity) can provide insight into inter-regional communication, spectral power offers a more direct measure of local neural population activity, which is consistent with prior work^6,23,24 and supports the present goal of capturing more local structure-function correspondence. This perspective builds on prior work showing that brain regions differ in their dominant spectral dynamics and in how structure and function are separately organized along hierarchically cortical gradients^8,25. Therefore, identifying systematic relationships between cortical thickness and spectral power may reveal how large-scale functional hierarchies emerge, interact with, and are even constrained by underlying structural architecture across development and aging. In this way, the integration of MEG and structural MRI is an ideal pairing to illuminate SFC, with the added benefit of spectral specificity, which has been linked to specific processes (e.g., gamma and prediction error)^26,27 and types of coding (e.g., cross-frequency coupling)²⁸ on the neural population scale, and is becoming increasingly recognized as critical to cortical computation on the level of columns²⁹.

To date, SFC has been rarely examined using MEG, with the few existing studies focusing on specific brain regions and/or examining restricted portions of the lifespan (e.g., young adults) in relatively small samples^6,30–33, rather than using more high-dimensional approaches across a wide swath of development. Thus, although foundational, the existing studies have notable limitations that hinder interpretability and generalizability. Therefore, in the present study, we use one of the largest MEG lifespan datasets (n = 388) reported to date to capture how spontaneous cortical activity relates to cortical thickness in the first-ever lifespan (i.e., 6–84-year-olds) study of this type. We leverage a robust MEG source reconstruction approach with age-specific cortical thickness maps to model linear and nonlinear structure-function associations from sensorimotor to association networks with spectral specificity. Recent MEG work focusing on functional changes alone has shown age-related decreases in slow wave activity (e.g., delta, theta) with increases in faster rhythms (e.g., alpha, beta, gamma), and that some of these associations may be best explained by nonlinear effects⁵. Finally, we also examined the extent to which SFC across the lifespan follows a cortical gradient similar to that reported in studies using other imaging modalities (i.e., conserved coupling in sensorimotor networks, less conserved coupling in association networks).

Results

Participants

A total of 388 participants (183 females; 47%) with an age range of 6–84 years (M_age = 25.32, SD = 20.86 years) were included in these analyses. Demographic characteristics are provided in greater detail in Table S1 and the age distribution is shown in Fig. S1. Individual SFC values were computed as the r-value of the correlation between cortical thickness and spontaneous power per spectral band for each parcel of the Schaefer 200 atlas (Fig. 1a). As proof of concept, we show an example of the SFC for alpha activity in one parcel of the visual cortex across all participants (Fig. 1b). Individual Fisher Z-transformed SFC values per parcel across age are presented in Fig. 1c as heatmaps, with more negative coupling in dark blue (e.g., increases in spontaneous power with decreasing cortical thickness) and more positive coupling in yellow.

Fig. 1. Structure-Function Coupling Analysis Schematic.

a The Schaefer 200 parcel, a seven-network atlas, was used to parcellate the data. Within each parcel, the cortical thickness (structure) and the spontaneous power (function) at each vertex were correlated. The resulting r-value was Fisher’s Z-transformed and used in subsequent statistical modeling as a measure of structure-function coupling for that parcel. This was performed across vertices within each parcel per participant resulting in one whole-brain map of structure-function coupling per participant per spectral band (e.g., alpha). b Scatterplot depicting an example of the structure-function coupling for alpha activity in one parcel of the visual cortex across all participants. Cortical thickness in mm is plotted on the x-axis, and spontaneous alpha power is in Pseudo-Neural Activity Index (PNAI) units on the y-axis. Participant age is reflected using the color scale. c Raw structure-function coupling values are displayed using heatmaps arranged by network (left to right) and age (youngest at the bottom; oldest at the top). Positive coupling is displayed with positive Z-values and negative coupling with negative Z-values according to the color scale at the bottom.

Linear and non-linear relationships between age and structure-function coupling

Hierarchical regressions showed robust linear and nonlinear relationships between SFC and age in all networks and all spectral bands at a stringent p_fdr < 0.001 threshold. Regression results, including the best fit polynomial (i.e., linear, quadratic, or cubic), predictor t values, adjusted r values, F values, intercept, slope, raw p-values, and p_{fdr –}values can be found in Tables S2–7. The proportion of parcels per network (Yeo 7)^34,35 exhibiting a significant relationship between SFC and age is shown in Fig. 2. Overall, all networks and frequency bands exhibited significant linear and quadratic relationships between SFC and age. Only delta and beta exhibited cubic relationships across all networks. All other bands showed more limited cubic associations between SFC and age. Notably, alpha exhibited overwhelmingly linear coupling patterns with age across all networks. Conversely, beta consistently showed a large proportion of network parcels with quadratic coupling with age across all networks (Fig. 2; all p_fdr <0.001).

Fig. 2. Age-related changes in structure-function coupling by network.

Proportion of significant (p_fdr < 0.001) relationships between age and structure-function coupling by network and spectral band. Best fit polynomial (i.e., linear, quadratic, cubic), as defined using a hierarchical regression approach, detected robust linear, quadratic, and cubic relationships in all networks and spectral bands. Bar plots display the proportion (%) of parcels per network, which showed a significant relationship between age and structure-function coupling, displayed by best-fit polynomial (linear—dark color, quadratic—medium color, cubic—light color) and the directionality of the relationship (staggered, with positive—right, top, and negative—left, bottom). Note that the ‘o’ symbol is shown for network and spectral band polynomials that did not show a significant relationship between structure-function coupling and age after correction for multiple comparisons (p_fdr < 0.001).

In terms of specific network patterns, the visual network showed strong linear relationships in delta, theta, and alpha bands, as well as strong quadratic relationships between SFC and age in the gamma spectral band (Fig. 3; all p_fdr <0.001). Specifically, in the delta and theta bands, patterns of linear coupling (i.e., decreasing power and cortical thickness) were observed across age in the visual network. Conversely, in alpha, decreased cortical thickness was related to increased alpha power in younger participants, with near-zero SFC in older participants, particularly in higher-order visual parcels. In the striate visual cortex, however, the opposite pattern of age-related increases in negative SFC emerged, whereby striate regions exhibited increased alpha power with older age as structure remained stagnant or became slightly thinner. In other words, more basic visual areas (striate) exhibited increased coupling with age, while the higher-order (extrastriate) visual areas showed looser coupling with age in alpha.

Fig. 3. Relationships Between Structure-Function Coupling and Age in the Visual Network.

Parcels showing a significant relationship between structure-function coupling and age (p_fdr < 0.001) within the visual network are displayed by spectral band (top to bottom: delta, theta, alpha, gamma). (Left): Brain plots display directionality and strength (log[p_fdr]) of the coupling-age relationship using the color scale bar in the top left corner; red colors display positive coupling with age (i.e., tighter structure-function link with increasing age, such that both indices are increasing or decreasing across age) and blue colors display negative coupling (i.e., one index is increasing as the other is decreasing with age). The colors of each parcel in the brain plots are also reflected in the corresponding scatterplots (e.g., a light blue parcel is also light blue in each scatter plot). Scatterplots display relationships between age in years (x-axes) and structure-function coupling (left, Z-value), spontaneous power (middle, Pseudo-Neural Activity Index [PNAI]), and cortical thickness (right, mm) on the y-axes. Note that the spontaneous power (middle) and cortical thickness (right) plots are for visualization purposes, as all analyses were conducted with the structure-function coupling metric (left). Shaded grey areas represent the standard error of the mean. Parcels with significant linear relationships between structure-function coupling and age are shown for delta, theta, and alpha bands, while parcels with significant quadratic relationships are shown for the gamma band.

In the somatomotor network, a high proportion of parcels exhibited significant linear relationships in the alpha band, with linear and quadratic relationships in the beta band (Fig. 4; all p_fdr <0.001). In particular, alpha exhibited stronger SFC across age with co-occurring decreases in spontaneous power and cortical thickness in most somatomotor parcels. In beta, a robust pattern of quadratic SFC revealed the weakest coupling at the youngest and oldest ages (i.e., near zero), with peak negative coupling (stronger power, reduced cortical thickness) during middle age.

Fig. 4. Structure-function coupling relationship with age within the somatomotor network.

Parcels showing a significant relationship between structure-function coupling and age (p_fdr < 0.001) within the somatomotor network are displayed by spectral band (top: alpha; bottom: beta). Layout and axis labels are the same as Fig. 3. Shaded grey areas represent the standard error of the mean. Parcels with significant linear relationships between structure-function coupling and age are shown for alpha, while parcels with significant quadratic relationships are shown for the beta spectral band.

Significant quadratic relationships between SFC and age were observed across many parcels and all spectral bands for the higher-level default-model network (Fig. 5; p_fdr <0.001). All parcels showed significant relationships between age and SFC, per polynomial type (e.g., linear, quadratic) and network, and these can be found in Figs. S2–22. A similar pattern emerged in lower spectral bands (delta, theta, and alpha) in juxtaposition to higher bands (beta and gamma). For example, in delta and theta, coupling was weakest (i.e., near zero) at younger and older ages, with the strongest coupling observed around middle age, with power decreasing as the cortex thins. Gamma showed a similar pattern in that the earlier and later years of life exhibit diminished coupling, with peak coupling around middle age, with age-related increases in spontaneous power being linked with decreasing cortical thickness.

Fig. 5. Relationship between structure-function coupling and age within the default mode network.

Parcels showing a significant relationship between structure-function coupling and age (p_fdr < 0.001) within the default mode network are shown by spectral band (top to bottom: delta, theta, alpha, beta, gamma). Layout and axis labels are the same as Figs. 3, 4. Shaded grey areas represent the standard error of the mean. Parcels exhibiting significant quadratic relationships between structure-function coupling and age are shown for all spectral bands.

In order to succinctly present the developmental and aging effects on SFC, the regression intercept representing the average SFC at the youngest ages and the regression slope representing the change in SFC per year of life have been plotted separately per spectral band for parcels showing significant linear, quadratic, and cubic relationships (Fig. 6; all p_fdr <0.001). Interestingly, parcels in specific networks tended to cluster together, and this was observed across multiple networks across the spectral bands. For example, the parcels of the visual network that exhibited a significant linear relationship between SFC and age in the alpha spectral band presented almost universally with a negative intercept and positive linear slope (i.e., extrastriate parcels), indicating strong negative coupling in early development with a shift toward zero coupling with age.

Fig. 6. Regression intercept and slope values across network, spectral band, and polynomial relationship.

Regression intercept (y-axis, coupling at 0 years) is plotted against regression slope (x-axis, change in coupling per year). Scatterplots are organized by spectral band (top to bottom: delta, theta, alpha, beta, and gamma) and polynomial relationship (left: linear; middle: quadratic; right: cubic). Dots represent each parcel, which showed a significant (p_fdr < 0.001) relationship between structure-function coupling and age within each network, defined by color.

Cortical hierarchy interacts with age in predicting structure-function coupling

Potential interactive effects of cortical hierarchy on the relationship between age and SFC were probed using linear-mixed effects (LME) modeling. Specifically, a cortical gradient variable was created and network assignment for each parcel was coded as one through seven, with one being sensory dominant (visual network) and seven being higher-order dominant (default mode network; Fig. 7). LME models were computed for each canonical spectral band, controlling for sex and MRI scanner (all MEG data was collected on the same scanner). Inspection of residual distributions and residuals versus fitted values indicated that assumptions of normality and homoscedasticity were met (Fig. S23). Significant interactions between network hierarchy and age on SFC were observed in delta (t[77210] = 12.60, p < 0.001), theta (t[77210] = 12.05, p < 0.001), alpha (t[77210] = −12.11, p < 0.001), and beta (t[77210] = −6.30, p < 0.001), but not gamma (t[77210] = −1.50, p = 0.133). These relationships are shown in Fig. 7 and plotted based on spectral band and network hierarchy. Simple slopes analysis revealed SFC in the delta band became weaker with increasing age in sensorimotor networks (t[77210] = −6.44, p < 0.001), while the opposite pattern was detected in more association networks (t[77210] = 2.80, p = 0.005). Similarly, in the theta spectral band, SFC became weaker with increasing age in sensorimotor networks (t[77210] = −9.04, p < 0.001), although there were no relationships with age in association networks (t[77210] = -0.67, p = 0.503). Conversely, in the alpha spectral range, sensorimotor networks exhibited increased SFC with advancing age (t[77210] = 2.11, p = 0.034), while association networks showed strong negative relationships (t[77210] = −7.32, p < 0.001). Regarding beta, all networks exhibited significantly weaker SFC with increasing age (p < 0.001), with the effects being stronger for sensorimotor networks (t[77210] = −10.40) than association networks (t[77210] = −6.25). Finally, within the gamma band, both sensorimotor networks (t[77210] = 3.41, p < 0.001) and association networks (t[77210] = 2.30, p = 0.021) showed increased SFC and increasing age, the strength of which did not significantly differ by position within network hierarchy. For post-hoc statistics of the effect of age on SFC within each network per band, see Table S7. Taken together, the effect of age on SFC patterns differed for networks depending upon their position in the cortical hierarchy in all frequency bands except gamma, with sensorimotor and association networks largely showing distinct, internally consistent age-related changes across the lifespan.

Fig. 7. Changes in Structure-Function Coupling Relationships with Age By Network Hierarchy.

(Top) Ranking of networks based on a cortical hierarchy, with the most sensory network on the left (visual) and the higher-order network toward the right (default mode network). LME models revealed significant interactions between network gradients and age on structure-function coupling in delta, theta, alpha, and beta, but not gamma. Each line represents the relationship between structure-function coupling (y-axis, Z-value) and age (x-axis), across all parcels assigned to each network, defined by color. Shaded areas around each line represent the standard error of the mean.

Discussion

Despite countless characterizations of how cortical structure changes across the lifespan, multimodal efforts to understand how these changes relate to developmental and aging fluctuations in cortical neurophysiology have been severely lacking. Herein, using a landmark lifespan MEG-MRI cohort with the most complete age range (6–84 years old) to date, we evaluated how spectrally specific SFC changes across the lifespan in large-scale cortical networks, with key findings that extend from sensorimotor to association networks and broadly expand upon and complement prior MRI-based studies. Overall, we showed robust linear and nonlinear age-related changes in SFC from childhood through older adulthood, and that these fluctuations are spectrally and network-specific, with intriguing patterns consistent with prior work examining age-related changes in network hierarchy with other imaging modalities. We discuss these findings and their implications for future directions below.

There is a rich literature demonstrating that microstructural and functional cortical gradients follow a hierarchical sensorimotor to association (S-A) axis organization, with evidence across time and spatial scales, modalities, and species³⁶. Although largely separate literatures, there is growing convergence between s/fMRI and EEG/MEG in demonstrating the S-A axis as an organizing principle of intrinsic networks, with variability across networks following a S-A hierarchy scaling in complexity and evolutionary novelty. The dominant spectral rhythms (i.e., delta, theta, alpha, beta, gamma) are also thought to be organized in a hierarchical fashion with neural assemblies oscillating at preferential rhythms that flexibly facilitate brain states and processes²⁵, which the data presented here largely support. That is, we extend this literature by providing one of the first structure-function studies to integrate cortical thickness and MEG measures across the lifespan, with consideration of the S-A hierarchy. With respect to a gradient structure, we found distinct spectrally-specific relationships between SFC and age across the cortical hierarchy. For example, theta range SFC was strongly linked with age at the lowest end of the hierarchy (i.e., visual network), but much less so in association networks. Similarly, beta SFC varied with age much more strongly in sensorimotor networks compared to association networks. There is longstanding evidence that slower rhythms (e.g., delta, theta, alpha) support synchronized, large-scale, distributed networks for integrating information and exerting top-down control¹⁶. Conversely, faster rhythms (e.g., gamma) are likely to be instigated more locally and are crucial for early sensory and motor transformations¹⁸. Thus, the theta finding is consistent with the notion that while SFC is present in all networks, the extent to which this is conserved across the human lifespan tends to follow the S-A axis. In other words, lesser coupling of structure and theta rhythmic activity with age in association cortices may be indicative of theta’s standing as a conductor of higher-order processing within the spectral hierarchy. This could provide further support for the role theta is thought to play in facilitating long-range connections in the service of increasingly complex cognitive functions that develop across the lifespan. In contrast, beta is the dominant rhythm of sensorimotor cortices, particularly motor cortices, and is known to emerge early in development and be consistently present across the lifespan^5,17. Thus, it is unsurprising that our findings revealed tight beta SFC with age in sensorimotor networks, particularly because these networks are expected to be highly conserved³⁷ and thus less variable across the lifespan once initially programmed early in development. Overall, the addition of spectral specificity to the known age-related topographic reorganization of the S-A axis provided herein affords a more complete picture of the reconfiguration of networks that occur across the lifespan.

These spectral- and hierarchical-specific age-related changes may also reflect underlying circuit-level constraints that shape intrinsic neural dynamics during development and evolve through aging. Recent cross-species work has identified conserved spectrolaminar motifs in which gamma oscillations tend to dominate processing in superficial cortical layers, while alpha and beta are stronger in deeper layers²⁹. These patterns are observed across cortical areas and are present even at rest, suggesting that they reflect computational parameters likely due to intrinsic circuit architecture rather than task-evoked dynamics. Viewed through this compelling framework, the macroscale SFC relationships observed here may indicate that variations in cortical thickness due to laminar composition and dendritic architecture²² may influence the types and hierarchy of rhythms that a region is predisposed to generate. In this view, the SFC patterns we report across the lifespan could reflect the cumulative output of regionally distinct circuit motifs, whose influence may be more stable in sensory cortices and more variable or developmentally dynamic in association cortices. Although the modalities used in this study cannot speak to circuit-level processes, it may be the case that the macro-scale patterns we observed are supported by more localized architectural constraints, though future work using invasive methods should more directly test these mechanistic hypotheses.

We found that in alpha, SFC strengthened with age across most cortical networks, with the key exception of the visual network, where alpha power is strongest. Specifically, within extrastriate regions of the visual network, increased cortical thickness was related to decreased alpha power in younger participants, with a shift toward decoupling in older participants (near zero SFC) across most visual parcels. Interestingly, striate visual cortex showed the opposite pattern of greater SFC in older age, mainly driven by striate regions exhibiting increased alpha power in older age, while structure remained stable or displayed typical cortical thinning patterns. These results indicate that visual alpha activity persists despite expected age-related cortical thinning, and that alpha activity shows regionally-specific associations with cortical thinning among higher- and lower-order parcels of the visual cortex. Thus, alpha may be tied to different mechanisms across different parcels, such that primary sensory visual cortices show robust lifelong SFC. Indeed, alpha oscillations may arise through multiple cortical mechanisms³⁸ that differ between regions even in primary and secondary visual regions³⁹, and this may reflect variability in cellular architecture and/or local neurotransmitters⁴⁰. Moreover, visual alpha differences could be due to local variation in thalamic afferent connections⁴¹, particularly in striate visual cortex, where anatomical thalamocortical connections are numerous⁴². In this context, the heterogeneity of alpha SFC across the visual network may be due to distinct cortical and thalamocortical mechanisms with differing capacities to generate strong alpha rhythms, which, of course, may also be differentially affected by aging. Furthermore, the regionally-specific coupling patterns are consistent with evolutionary accounts of brain development such that higher-order extrastriate regions tend to show less persistent coupling, while primary visual cortices exhibit continual coupling from early age into the last decades of life, suggesting that circuits serving primary sensory functions tend to be more strictly conserved.

In the beta range, a high proportion of parcels in the somatomotor network showed quadratic SFC with age. That is, there was the least SFC in the earliest and latest years, while peak coupling appeared during middle age (i.e., ~45 y), with stronger spontaneous beta activity being related to reduced cortical thickness. Increases in spontaneous beta with thinner cortices in older age have been well established across multiple overlapping topographies^6,23,43–45. As prior work has shown²³, it may be the case that a stronger level of coupling is necessary for efficient cortical functioning and motor performance. Regarding the quadratic nature of this relationship, high proportions of quadratic relationships were also present in beta across many of the other networks, highlighting that the nonlinear nature of this coupling may be a broader motif of transient optimization followed by declines later in the lifespan. Interestingly, a previous function-only analysis of this dataset revealed that spontaneous beta activity in the left and right precentral gyri also showed significant quadratic components in their relationship with age⁵, suggesting that the quadratic component is not necessarily specific to SFC.

Parallels were observed in the gamma range, where several networks, including visual and default mode, exhibited U-shaped coupling, with peak coupling in midlife and decoupling at early and older ages. The visual network, which showed linear coupling in slower spectral bands, was uniquely characterized by a quadratic gamma pattern, indicating that gamma-related SFC may be particularly dynamic in sensory cortices. In contrast, the DMN exhibited quadratic coupling across all frequency bands, but with an opposite pattern in gamma: a midlife coupling peak whereby higher gamma power tracked with increased cortical thickness. The specificity of such gamma changes aligns with known age-related changes in γaminobutyric acid (GABA) signaling. GABA activity, which is critical to synchronous cortical gamma activity⁴⁶, sharply increases during development, supporting the acquisition and refinement of increasingly complex skills, and then declines throughout adulthood⁴⁷. These shifts closely mirror the quadratic relationships observed here, particularly in visual and DMN networks where gamma rhythms are robust and developmentally dynamic. Further, gamma oscillations have been linked with GABA concentrations and GABA receptor density⁴⁸, and local GABA activity also modulates spontaneous beta, particularly in sensorimotor cortices^49,50. Thus, age-related changes in GABA may contribute to nonlinear SFC patterns across both beta and gamma bands, though more systematic work is needed to decipher these associations. Together, these findings may reflect a broader developmental arc of spectral and structural optimization, punctuated by peak coupling during middle adulthood, followed by the emergence of compensatory neural mechanisms later in life, potentially driven by shifts in the excitatory/inhibitory balance⁵¹.

One of the most consistent findings in studies of large-scale network development is that the DMN is among the most gradual to develop⁵², with nonlinearities in the first two decades of life⁵³. Although most of these conclusions have relied on MRI-based studies, the present study offers convergence and adds important nuance to this literature. Specifically, our results show that, by and large, the DMN showed a predominantly quadratic SFC pattern across nearly all spectral bands examined. Our results are broadly consistent with the notion that correspondence between function and underlying structure in association networks may become “decoupled” over the lifespan⁵⁴, due to less phylogenetic conservation⁵⁵ and that looser coupling may be an inherent property of more flexible and integrative networks; we further delineate these previously established patterns by demonstrating that this decoupling follows a dramatic quadratic function. That is, in the beginning of life, DMN structure-function is largely uncoupled, by middle age (~45–50 years old), there is strong DMN coupling across spectral bands, with distinct decoupling (i.e., near zero) in late life in most regions and spectral bands. Though it should be noted that the nature and directionality of these associations are spectral-band dependent in terms of which DMN regions showed the strongest relationships. For example, delta and theta both showed quadratic associations with age, with spontaneous power decreases corresponding to greater cortical thinning in the same tissue across the life course. Although delta and theta exhibited overlapping spatial specificity in some areas (e.g., superior temporal sulcus), each also showed distinct focal patterns, with delta being most pronounced in the posterior cingulate and theta being most pronounced in the ventromedial prefrontal. These results support longstanding evidence that neural assemblies tend to preferentially oscillate in specific spectral bands, and that delta and theta are often implicated in higher-order cognition supported by association regions⁵⁶. The present findings also significantly contribute to our understanding of how age-related changes in spontaneous cortical activity co-occur with the anatomical alterations to grey matter thickness that are known to occur with age (e.g., synaptic pruning, dendritic shrinkage)^3,4. Also in line with previous work is that the variability in the SFC in all spectral bands was notably greater with older age, which would support prior work suggesting greater genetic variance⁵⁷ and extensive heterogeneity in predominantly association networks¹⁰. Taken together, these neurophysiological results reinforce and extend the mounting evidence from other functional imaging modalities that association regions like those in the DMN exhibit increasing individualization over the lifespan.

Although the present study adds foundational knowledge to the nascent but growing SFC literature using MEG, there are limitations to be acknowledged. First, the approach to quantifying brain structure is limited to cortical thickness, which is only one of many facets of neuroanatomy known to change across the lifespan. Future studies should integrate additional structural measures, such as those sensitive to myelin changes, which are known to undergo radical reorganization across the human lifespan⁵⁸. Moreover, connectome-based approaches, such as those integrating MEG functional connectivity with diffusion-weighted imaging-derived structural connectivity, would offer a complementary framework for capturing how inter-regional SFC changes across the lifespan. Second, by modeling SFC as the correlation between cortical thickness and spectral power, we inherently assume linearity in this relationship, which may not hold uniformly across regions, spectral bands, or developmental stages. Thus, future work should more systematically characterize the potential non-linearity in these associations. In the same vein, studies building on this work would benefit from more nuanced modeling schemes, such as generalized additive models that are sensitive to dynamic changes and curvature in data. Third, it is plausible that the choice of atlas could have implications for the observed developmental patterns. While the Schaefer 200 atlas provides a widely used, functionally defined parcellation for examining large-scale cortical organization, the parcellation choice and resolution could potentially influence estimates of age-dependent structure-function correspondence. In the current study, SFC was computed using vertex-wise information within parcels, which helps mitigate some loss of spatial detail. In addition, most available parcellations are derived from adult data, which likely do not fully capture developmental variation in cortical organization. Although we attempted to partially address this through more developmentally-sensitive MRI preprocessing (i.e., age- and sex-specific tissue probability maps), future work should evaluate the robustness of these findings across parcellation resolutions, including developmentally informed atlases (e.g., Dev-Atlas)⁵⁹. Fourth, despite capturing the better part of eight decades of life, the youngest participants in our study were 6 years old, which fails to capture the first few years of life. There is an abundant and rapidly growing literature demonstrating extensive functional and structural changes that happen in the first years of life, with changes in infants occurring on the order of days⁶⁰. Lifespan studies reaching down into these ages are rare, and there are limitations in acquiring MEG from infants with standard systems (e.g., distance from scalp). However, alternative solutions with measurements comparable to cryo-MEG are emerging, with wearable technology capable of capturing neural activity in awake and behaving infants (e.g., optically-pumped magnetometers). Finally, the present sample was cross-sectional, which provides only a snapshot of an individual’s development and thus constrains interpretations regarding change across time. Longitudinal work is needed to make stronger conclusions about structure-function relationships and to chart intraindividual changes across the lifespan.

For the first time, we document dynamic age-related changes occurring in the coupling between dominant brain rhythms and the thickness of the underlying cortical tissue in the largest MEG cohort to date, spanning eight decades of life. Given MEG’s sensitivity to activity emanating from large populations of pyramidal cells⁶¹, and the known correlates of age-related cortical growth and thinning in areas rich with pyramidal cells²², the pairing of MEG-based cortical activity with cortical thickness garners insight into the refinement of neural assemblies supporting large-scale network development across the lifespan. Our results indicate spectrally- and regionally-specific patterns of SFC across the lifespan, with remarkable nonlinearities. These findings also extend prior work on cortical gradients by highlighting that hierarchies in the coupling of canonical spectral bands with structure also exhibit patterns consistent with strong lifespan coupling in sensorimotor networks and lesser, more variable coupling in associative networks with advancing age.

Methods

Participants

A total of 388 participants (183 females) between the ages of 6 and 84 years old (M_age = 25.32, SD = 20.86) across five different studies with MEG data collected on the same instrument were included in this analysis. Information on the demographics and protocols of each study are summarized in the SI Appendix (Table S1). Adult participants were cognitively healthy controls in studies examining neurological and psychiatric conditions, while pediatric participants were included from studies of normative neurodevelopment. Exclusion criteria included an inability to complete the full resting-state scan, any medical illness affecting the CNS, any neurological or psychiatric disorders, history of head trauma, current substance abuse, and standard MEG/MRI exclusion criteria (e.g., dental braces, battery-operated implants, and/or any type of ferromagnetic implanted material). All demographic data were reported by the participant or a guardian as part of the intake process. For child participants, parents signed informed consent forms, and child participants signed assent forms before participating in the study. For adult participants, written informed consent was obtained from each participant. The local Institutional Review Board (IRB) approved the study. All ethical regulations relevant to human research participants were followed.

MEG data acquisition

MEG recordings took place in a one-layer magnetically shielded room (MSR) with active shielding to compensate for environmental noise. Participants were seated in a nonmagnetic chair within the MSR, with their head positioned within the sensor array. A 306-sensor MEGIN MEG system (Helsinki, Finland), equipped with 204 planar gradiometers and 102 magnetometers, was used to sample neuromagnetic responses continuously at 1 kHz with an acquisition bandwidth of 0.1–330 Hz. The same instrument was used across all recordings. Participants were instructed to rest with their eyes closed for 5 to 8 minutes and were monitored by a real-time audio-video feed from inside the shielded room throughout MEG data acquisition.

Structural MRI acquisition, processing, and MEG-MRI coregistration

Prior to MEG acquisition, four coils were attached to the participants’ heads and localized, together with the three fiducial points and scalp surface, using a 3-D digitizer (Fastrak, Polhemus Navigator Sciences, Colchester, VT, USA). Once the participant was positioned for MEG recording, an electrical current with a unique frequency label (e.g., 322 Hz) was fed to each of the coils. This induced a measurable magnetic field and allowed each coil to be localized in reference to the sensors throughout the recording session. Since coil locations were also known in head coordinates, all MEG measurements could be transformed into a common coordinate system. With this coordinate system (including the scalp surface points), each participant’s MEG data were co-registered with their structural T1-weighted MRI data prior to source space analyses using Brainstorm. MRI acquisition parameters can be found in the SI Appendix. The 3 T scanner on which each individual’s MRI data were acquired was included in all models as a covariate of no interest.

MEG data Pre-Processing

Similar to our previous work⁵, each MEG dataset was individually corrected for head motion and subjected to noise reduction using the signal space separation method with a temporal extension (MaxFilter v2.2; correlation limit: 0.950; correlation window duration: 6 seconds). Noise-reduced MEG data underwent standard data preprocessing procedures using the Brainstorm software. A high-pass filter of 1 Hz, a low-pass filter of 200 Hz, and notch filters at 60 Hz and its harmonics were applied. Cardiac artifacts were identified in the raw MEG data and removed using an adaptive signal-space projection approach, which was subsequently accounted for during source reconstruction. Data were then divided into four-second epochs for detection and rejection of bad segments of data. The maximum peak-to-peak amplitude and numerical gradient for each epoch were computed. To account for variance across participants in the distance between the brain and the MEG sensor array, as well as other sources of variance, we used an individually-determined threshold based on the signal distribution for both amplitude and gradient to reject artifacts. Following conventions in the literature, thirty epochs were removed from the end of each scan of our older adults to harmonize the length of their resting scan with the remaining dataset, which were shorter in length (Table S1).

MEG source imaging and spectral power maps

Source reconstruction methods and processing followed the analysis pipeline previously detailed in our prior work⁵. Briefly, source analysis of neuromagnetic fields used an overlapping-spheres forward model, unconstrained to the cortical surface. This approach models a single sphere per sensor, which in this study included 204 gradiometers (minus any channels previously marked as “bad”) for each participant. A linearly constrained minimum variance (LCMV) beamformer implemented in Brainstorm was used to spatially filter the epoched data based on the data covariance computed from the resting-state recording. Using these source estimates, we then estimated the power of cortical activity in each of five canonical spectral bands: delta (2–4 Hz), theta (5–7 Hz), alpha (8–12 Hz), beta (15–29 Hz), and gamma (30–59 Hz). We focused on these five canonical spectral bands (Fig. 1) to maximize comparability to previous studies. We used Welch’s method for estimating power spectrum densities (PSD) per four-second epoch across each MEG recording, with one-second sliding Hamming windows overlapping at 50%. These raw PSD maps were averaged across epochs to obtain an absolute PSD map per participant. The units in these absolute maps correspond to the Pseudo Neural Activity Index (PNAI), which is a modified version of Van Veen’s Neural Activity Index⁶². The norms of the three unconstrained orientations per location were then projected onto a common FSAverage template surface and used for statistical analysis.

Cortical thickness processing

Prior to calculating cortical thickness, we first created customized tissue probability maps (TPM) using the CerebroMatic Toolbox to allow for more precise spatial registration that is developmentally sensitive, rather than using common templates that are biased toward adult samples. CerebroMatic uses tissue priors from >1900 brains to statistically model demographic variables (i.e., participant age, sex) in each voxel to segment six tissue types (i.e., gray matter, white matter, cerebrospinal fluid, bone, soft tissue, and background), rather than simply averaging across images. In our dataset, we created custom TPMs separately for ages 6–15.99 years, 16–44.99 years, and 45–84 years to minimize distortion of images during spatial normalization, particularly in children and older adults. To examine voxel-wise cortical thickness, T1-weighted MRI data were processed using additional surface-based morphometry calculations in the computational anatomy toolbox (CAT12 v12.6; at a resolution of 1 mm³). This method uses a projection-based thickness approach to estimate cortical thickness and reconstruct the central surface in one step. Briefly, the white matter (WM) distance is estimated, and the local maxima are projected onto other gray matter voxels using a neighboring relationship described by the WM distance. This method accounts for partial volume correction, sulcal blurring, and sulcal asymmetries. Topological defects are corrected based on spherical harmonics, and the cortical surface mesh was reparametrized into a common coordinate system via an algorithm that reduces area distortion. Finally, the resulting maps were reviewed for quality prior to resampling and smoothing using a 15 mm FWHM Gaussian kernel, consistent with CAT12 recommendations. These maps of cortical thickness were ultimately utilized in subsequent analyses.

Structure-function coupling (SFC)

Whole-brain PSD maps per canonical band and whole-brain cortical thickness maps were used in subsequent SFC analyses. The Schaefer 200 parcel atlas³⁴ was applied to both functional and structural maps. Within each individual, the structure and function values for the vertices assigned to each atlas parcel were extracted and correlated to derive an SFC metric (i.e., the r-value of the correlation). This was performed for each parcel, spectral band, and individual. Therefore, five atlas-based maps of SFC were created per individual, with one for each of the five canonical spectral bands. These parcellated maps were used for statistical analysis.

Statistics and Reproducibility

All statistical analyses were performed in R. The SFC r-value for each parcel was Fisher Z-transformed and submitted to regression analyses. Hierarchical regression analyses were performed to determine the relationship between SFC and age in each parcel, as well as potential nonlinear relationships with age. Linear age was included in the base model, quadratic age in the second step, and cubic age in the third and final step. Sex, as well as age x sex interactions, were initially included as effects of interest, but did not survive multiple comparisons correction; thus, sex was included as a covariate of no interest in all models. Within each spectral band, each parcel was assigned a polynomial based on the hierarchical regression, with the best model selection based on ΔR². Only the results of this assigned polynomial were examined. To account for multiple comparisons, false discovery rate (FDR)-corrected p-values were used with a threshold of p_FDR < .001. To aid interpretation, the resulting maps were divided and presented based on the Yeo 7 Networks^34,35. Finally, to assess the interactive effect of network-level cortical gradients (i.e., sensory to association axis) on the relationship between age and SFC, linear mixed-effects (LME) models were used. A cortical gradient variable was created by coding each network as 1 through 7 (termed NetGradient below), with 1 being the lowest-level (Visual network) and 7 being the highest-order (Default Mode), and then treated as a continuous variable³⁵. LME modelling was performed using the nlme package in R with the following form: Structure-Function Coupling ~ Age*NetGradient + Sex + MRI Scanner, random = (~1 | SubID).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: Y.W., V.C., J.M.S., E.H.G., M.K. and T.W.; data collection: M.S., AW, CE, HJ, HO, and MW; methodology and software: G.P., M.R., N.P., J.J.S., C.C., A.W., C.E., S.P. and L.O.; formal analysis: M.R., G.P., C.C., L.O. and S.P.; writing- original draft: G.P., M.R., N.P. and J.S.; reviewing and editing: all authors; supervision: TW. All authors approved the final version for publication and can certify that no other individuals not listed as authors have made substantial contributions to the paper.

Peer review

Peer review information

Communications Biology thanks Tyler Santander and the other anonymous reviewer(s) for their contribution to the peer review of this work. Primary Handling Editors: Sahar Ahmad and Jasmine Pan. A peer review file is available.

Data availability

All data are publicly available via the Collaborative Informatics and Neuroimaging Suite (COINS; https://coins.trendscenter.org/) upon request.

Code availability

The MEG and MRI analyses reported here used Brainstorm (https://neuroimage.usc.edu/brainstorm/), Matlab 2018b, and CAT12 v12.6. R programming in R Studio was used for statistical analyses and visualization. The code and batch functions used in these analyses are available upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-026-10445-z.