Abstract
Functional connectivity (FC) has attracted significant interest in the identification of specific circuits underlying brain (dys)function. Classical analyses to estimate FC (i.e., filtering electrophysiological signals in canonical frequency bands and using connectivity metrics) assume that these reflect oscillatory networks. However, this approach conflates nonoscillatory, aperiodic neural activity with oscillations, raising the possibility that these functional networks may reflect aperiodic rather than oscillatory activity. Here, we provide the first study quantifying, in two different human electroencephalography (EEG) databases (n = 59, 30 females and 29 males; n = 103, 62 females and 41 males), the contribution of aperiodic activity on reconstructed oscillatory functional networks in resting state. We also followed the same approach on cognitive task recordings (n = 59, 30 females and 29 males) as a complementary analysis. We found that ∼99% of delta, theta, and gamma functional networks, over 90% of beta functional networks, and between 23 and 61% of alpha functional networks were actually driven by aperiodic activity. While there is no universal consensus on how to identify and quantify neural oscillations, our results demonstrate that oscillatory functional networks may be drastically sparser than commonly assumed. These findings suggest that most FC studies focusing on resting state data actually reflect aperiodic networks instead of oscillations-based networks. We highly recommend that oscillatory network analyses first check the presence of aperiodicity-unbiased neural oscillations before estimating their statistical coupling to strengthen the robustness, interpretability, and reproducibility of FC studies.
Significance Statement
Assessing how brain regions communicate is critical for understanding behavior and cognition. In electroencephalography and magnetoencephalography, neural networks are commonly identified through functional connectivity estimated under the assumption that interregional coupling between brain regions reflects oscillatory networks. Our findings demonstrate that a substantial portion of presumed oscillatory networks are instead driven by aperiodic activity, thereby challenging a central methodological assumption in the field. By explicitly disentangling oscillatory and aperiodic components, this work calls for a reassessment of existing approaches, showing that oscillatory networks are far less widespread than commonly assumed and provides a refined framework to improve the robustness and reproducibility of functional connectivity research, with implications for both cognitive and clinical neuroscience.
Introduction
Functional connectivity (FC) is extensively used in cognitive neurosciences to identify brain networks, with over 30,000 articles referenced on PubMed. The knowledge of specific networks provides insights into (patho-)physiological processes (Medaglia et al., 2015), establishing FC as a tool ranging from predicting depressive states (Fingelkurts et al., 2007) to proposals regarding its role in consciousness mechanisms (Wang et al., 2020). Large-scale brain networks, originally studied using functional magnetic resonance imaging (fMRI; Biswal et al., 1995), have been extended to magnetoencephalography (MEG; Brookes et al., 2011a) and electroencephalography (EEG; Smit et al., 2008), which provide superior temporal resolution to capture rapid neural dynamics and frequency-specific contributions to functional interactions (Britz et al., 2010; Sockeel et al., 2016). In M/EEG, FC consists of identifying the statistical couplings of neural oscillations from different brain regions of interest (ROIs; Friston et al., 1993), in frequency bands such as delta (2–4 Hz), theta (5–7 Hz), alpha (8–12 Hz), beta (13–29 Hz), and gamma (30–40 Hz). Oscillation-based FC is often computed using amplitude- and phase-based measures. Despite extensive studies, there remains no consensus on which FC metrics reliably estimate connectivity across conditions (Fraschini et al., 2020; Allouch et al., 2023).
Classical M/EEG FC analyses usually filter data within a frequency band for each ROI and then compute a FC metric for each pair of signals, resulting in a N × N symmetrical matrix that provides the “strength” of statistical coupling between ROIs (Fraga González et al., 2018; Sion et al., 2020; Xie et al., 2022). Oscillations often behave in a bursty rather than tonic manner (Jones, 2016; Sherman et al., 2016). Although this could influence coupling between ROIs, a fundamental assumption is that oscillations drive FC results and filtered data reflects oscillatory networks. Network architecture is then analyzed, often using graph theory metrics to quantify organization and efficiency. This approach has been widely applied in resting-state EEG, revealing robust FC across oscillatory frequencies. For example, alpha oscillations, prominent even in raw EEG, dominate intrinsic activity at rest and have been linked to the default mode network (Tang et al., 2017; Samogin et al., 2019). Beta and gamma networks have been reported to support interactions between closer ROI pairs (Samogin et al., 2019). Delta-theta band FC has been considered a robust signature of consciousness (Bourdillon et al., 2020), while gamma FC has been associated with creativity (Chhade et al., 2024). Connectivity in these bands has also been investigated in clinical contexts, where alterations are studied to identify potential biomarkers such as of vegetative states or neurodegenerative diseases (Iyer et al., 2015; Nardone et al., 2018; Duszyk-Bogorodzka et al., 2022).
The emerging consensus, however, is that the usual estimation of oscillatory power is biased by aperiodic activity dynamics (Donoghue et al., 2020). Long considered meaningless noise (Freeman, 2000), aperiodic activity captures nonrhythmic, irregular features in neural data and exhibits a 1/f-like distribution, with power decreasing across frequencies. In contrast, oscillatory activity corresponds to narrowband, rhythmic activity appearing as peaks above the aperiodic background (Fig. 1). Interpreting power in canonical frequency bands without separating periodic and aperiodic components leads to a misrepresentation and misinterpretation of the underlying physiological phenomena (Donoghue et al., 2020). In this study, oscillations not conflated with aperiodic activity are called “aperiodicity-unbiased oscillations.” We argue that, considering aperiodic activity is essential for accurate FC interpretation, as networks driven by oscillations have different physiological meaning than those driven by aperiodic activity (Wolff et al., 2022). To date, the contribution of aperiodic activity in estimating FC has not yet been studied.
Oscillatory and aperiodic components in neural data. A, In the time domain. A simulated neural time series with an aperiodic component in blue and alpha oscillations in green. B, In the frequency domain. Power spectrum showing oscillatory peaks above the aperiodic component estimated by the offset and the exponent.
We found that aperiodic activity significantly impacts FC networks estimation. We show this in two independent human EEG resting state databases (n = 59, n = 103), first analyzed using a common FC pipeline assuming that all power in each frequency band reflects oscillations. We compared these results to our new approach that first verifies the presence of “aperiodicity-unbiased oscillations” before FC computation.
Materials and Methods
In this study, we applied the same analysis pipeline to two independent EEG datasets (Fig. 2).
Analysis pipeline of the study. After EEG recordings, signals were preprocessed to remove artifacts. The EEG inverse problem was solved using the weighted minimum norm estimate method and cortical sources were projected on the Desikan–Killiany Atlas. Then, two approaches were conducted: the classical pipeline and our approach that separately considers the periodic and aperiodic components of the signal. This approach permits identifying regions with aperiodicity-unbiased oscillations before the estimation of FC. After computing FC with five different metrics, a proportional threshold was applied to retain the top 5% of the strongest connections. Finally, the functional networks were described using graph theory metrics.
Participants
Dataset A
Thirty HC (15 females, 15 males) aged between 45 and 70 years old (mean, 61.7; SD = 7.3) and 29 patients (15 females, 14 males) diagnosed with idiopathic PD (Hughes et al., 1992), aged between 45 and 73 years (mean, 60.4; SD = 7.3), participated in this study. HC and patients did not significantly differ in age, sex, or education. All patients were recruited from the Neurology Department of Rennes University Hospital (France). All healthy participants were recruited from the general population by advertising. All participants were free from any major cognitive impairment (Montreal Cognitive Assessment Scale, MoCA scores <22; Nasreddine et al., 2005) or severe neurocognitive disorder according to the Diagnostic and Statistical Manual of Mental Disorders-V (DSM-V). Participants with moderate or severe psychiatric symptoms and present or past neurological pathology (other than PD for patients) were not included in this study. For further details, see Monchy et al. (2024). This study was conducted in accordance with the Declaration of Helsinki and was approved by a national ethics committee (CPP ID-RCB: 2019-A00608-49; approval number: 19.03.08.63626). After a complete description of the study, all participants gave their informed written consent.
Dataset B
One hundred and three HC (62 females, 41 males) aged between 17 and 71 years old (mean, 37.9; SD = 13.9) who participated in an experimental paradigm used in the Stimulus-Selective Response Modulation (SRM) project at the Dept. of Psychology (University of Oslo, Norway) were studied (Hatlestad-Hall et al., 2022). Subjects were recruited through social media platforms and local advertisement. None of the subjects reported severe psychiatric or neurological symptoms. For further details, see Hatlestad-Hall et al. (2022). The study was conducted in accordance with the Declaration of Helsinki and was approved by the Regional Ethics Committee of South-Eastern Norway (reference number: 2016/2003). All participants provided written informed consent.
HD-EEG recording and processing
Recording
Dataset A
EEG data was recorded using a High-Definition EEG (HD-EEG) system (EGI, Electrical Geodesics, 256 channels) with a sampling frequency of 1,000 Hz. Electrode impedance was maintained below 25 kΩ. The Cz electrode was used as reference and the ground close to Pz. Electrodes on the net were placed using the standard 10-10 geodesic montage. Most jaw and neck electrodes were removed due to excessive muscular artifacts, resulting in a total of 199 exploitable electrodes, as shown in the channel file available in the GitHub repository (see below, Data and code availability). This dataset includes both resting-state and task-related EEG data. Resting-state recordings consist of a 5 min eyes-closed resting period. Task-related activity corresponds to the poststimulus interval (0–1,000 ms) following stimulus presentation during a Simon task (for further details, see Monchy et al., 2024). Congruent and incongruent trials were processed separately.
Dataset B
EEG signals were recorded from a 64-channel (Ag-AgCl electrodes) BioSemi ActiveTwo system. Electrodes were positioned according to the extended 10-20 system (10-10). Data were sampled at 1,024 Hz without online filters, except for the default hardware anti-aliasing filter. EEG data was recorded during 4 min in resting-state with eyes closed.
Preprocessing
Dataset A
All EEG preprocessing and subsequent analyses were performed manually using the Brainstorm toolbox (Tadel et al., 2011) in Matlab (MathWorks). First, DC offset removal was applied. Second, a notch filter (50 Hz) and a bandpass filter (1–90 Hz) were applied (Finite Impulse Response; FIR filter using Kaiser window). Third, signals were visually inspected, and bad channels were removed before being interpolated using Brainstorm’s default parameters. Fourth, independent component analysis (ICA, jade method) was used to remove artifacts following visual inspection of the independent components. Fifth, signals were segmented into 4 s epochs (Góngora et al., 2022). Finally, a visual inspection was performed to manually reject remaining bad epochs. As a result, an average of 59 (SD = 14) resting-state epochs were studied per subject and 326 (SD = 13) task epochs.
Dataset B
Preprocessing of EEG data has already been achieved upon download. EEG data was preprocessed using a basic and fully automated pipeline on EEGLAB (Delorme and Makeig, 2004) and NoiseTools (De Cheveigné and Arzounian, 2018) in Matlab (MathWorks). First, the pipeline identified and removed bad channels. Second, a high-pass filter of 1 Hz was applied. Third, power line noise (50 Hz) was removed with Zapline (De Cheveigné, 2020). Fourth, ICA was calculated using the SOBI algorithm (Belouchrani et al., 1997). All components of ocular or muscular origin with a certainty above 85% were subtracted from data. Fifth, removed channels were interpolated. Sixth, a low-pass filter of 45 Hz was applied. Then, bad channels characterized by repeating small signal glitches were removed. Finally, data were referenced to the average reference and segmented into nonoverlapping 4 s epochs. Epochs containing discontinuities were removed. This included fully automated cleaning procedure, considered strict in terms of the data rejection, produced a derived dataset in which 64.1% of epoched data files kept >90% of their channels, and 23.5% retained between 75 and 90% of their channels.
Source reconstruction
FC was computed at the ROI level to enhance anatomical interpretability and reduce scalp volume conduction effects, in line with the literature favoring source-level analyses for investigating large-scale brain networks (Michel and Murray, 2012; Hassan and Wendling, 2018). We used a realistic head model and electrode position. Here, we used the boundary element method (BEM) head model fitted to the International Consortium for Brain Mapping Magnetic Resonance Imaging (ICBM MRI) template (Mazziotta et al., 2001) and the OpenMEEG toolbox (Gramfort et al., 2010). The EEG inverse problem was solved using the weighted minimum norm estimate (wMNE) method (Lin et al., 2006). Cortical sources were projected on the Desikan–Killiany Atlas parcellation (68 regions of interest; ROIs; Destrieux et al., 2010), with their orientation constrained normally to the cortical surface (Dale and Sereno, 1993). A noise covariance matrix was computed on all the recording. Signal-to-noise ratio (SNR) and depth weighting were set to default Brainstorm values.
Spectral parameterization
In our approach, we specifically parameterized the EEG signal into both aperiodic activity and oscillations in order to identify oscillations that are not conflated with aperiodic activity. The frequency bands considered in our approach were as follows: delta (2–4 Hz), theta (5–7 Hz), alpha (8–12 Hz), beta (13–29 Hz), and gamma (30–40 Hz). For each subject, frequency band, and ROI, we checked whether “true” oscillations, not conflated with aperiodic activity, were found by specparam. In this study, we used the term “aperiodicity-unbiased oscillations” to refer to oscillations not conflated with aperiodic activity according to the specparam algorithm.
Power spectrum density (PSD) was computed for each epoch using Welch’s method, which is a necessary step for subsequent estimation of aperiodic parameters. A sensitivity analysis was performed to identify the optimal time window, testing various lengths: 0.5, 1, and 1.5 s. The goodness-of-fit metrics, including R2 and mean squared error (MSE), indicated that the specparam algorithm performed best with a time window length of 0.5 s and a 50% window overlap. Spectra were then averaged by subject. For each resulting power spectrum, we applied the spectral parameterization algorithm in the 1–40 Hz frequency band (specparam toolbox version 1.0; Donoghue et al., 2020), which considers the PSD as a linear combination of two different types of components: aperiodic and periodic (oscillatory) components (Donoghue et al., 2020). The aperiodic component is estimated through aperiodic exponent and offset while the periodic component is assessed with the frequency and amplitude of oscillatory peaks. Both periodic and aperiodic parameters were estimated at the ROI level. Optimized parameters included the following: peak width limits, [0.5–6]; max number of peaks, 4; minimum peak height, 1.0; peak threshold, 2.0; and aperiodic mode, “fixed.” To tailor the specparam algorithm to our data, we compared the goodness-of-fit metrics of models with varying maximum peak values (3, 4, and 5). The model with a maximum of 4 peaks was ultimately selected, as it provided the best fitting metrics: (dataset A: ROI-averaged R2 = 0.86; ROI-averaged mean squared error, MSE = 0.02; dataset B: ROI-averaged R2 = 0.90; ROI-averaged mean squared error, MSE = 0.04). Optimized parameters included the following: peak width limits, [0.5–6]; max number of peaks, 4; minimum peak height, 1.0; peak threshold, 2.0 and aperiodic mode, “fixed.” Aperiodic parameters were estimated at the ROI level.
Functional connectivity
Numerous approaches have been proposed for calculating FC between reconstructed regional time series. Here, we chose to compute different widely used EEG FC estimation methods based on phase or amplitude. For each subject, each frequency band, and each epoch, FC was calculated using these different methods to obtain a 68 × 68 connectivity matrix, where each entry
Phase-based metrics
FC can be defined by the relative instantaneous phase between two time series. Phase-based metrics are robust against variations in signal amplitude, but low signal-to-noise values remain challenging for these measures (Lachaux et al., 1999; Mormann et al., 2000).
Phase locking value
The most commonly used phase measure is the phase locking value (PLV), defined as the absolute value of the mean phase difference between two signals, expressed as a complex unit-length vector (Lachaux et al., 1999; Mormann et al., 2000). When the distribution of the phase difference between two signals is uniform, PLV is 0. Conversely, if the phase difference between the two signals has a preferred angle, PLV approaches 1.
Considering a pair of narrowband analytic signals,
i and
j, PLV is defined as:
Corrected imaginary PLV (ciPLV)
PLV has a major limitation in assessing functional brain connectivity, namely, its sensitivity to volume conduction (Stam et al., 2007) and source-leakage effects. In source-level EEG data, neighboring sources may still share some activity due to data low spatial resolution (source leakage). To face this limitation, Bruña et al. (2018) proposed phase-based metrics that discard zero-lag connectivity, thereby limiting volume conduction.
Considering a pair of narrowband analytic signals,
i and
j, ciPLV is defined as:
Weighted phase lag index
Stam and colleagues also proposed the phase lag index (PLI) as another phase-based metric (Stam et al., 2007). PLI measures the asymmetry of the distribution of phase differences between two signals, is an efficient method to detect “true” changes in phase-synchronization, and has a low sensitivity to volume-conducted noise. However, its sensitivity to noise and volume conduction could be impacted by its discontinuity, particularly affecting synchronization effects of minimal magnitude.
Considering a pair of narrowband analytic signals,
i and
j, the complex cross-spectrum
C for two signals is computed by Fourier-transforming them into
Amplitude-based metrics
Another coupling mode between neuronal oscillations from different brain regions operates purely on the amplitude envelope correlation, reflecting the temporal comodulation of amplitude (Bruns et al., 2000; Siegel et al., 2012). Amplitude correlation is considered as an efficient index of the large-scale cortical interactions mediating cognition (Mazaheri and Jensen, 2010).
Amplitude envelope correlation
Amplitude envelope correlation (AEC) is computed by correlating the amplitude envelopes of two oscillatory signals. This normalized measure of amplitude coupling is based on linear correlations between signals’ envelopes derived from the Hilbert transform (Brookes et al., 2011b; Hipp et al., 2012). High AEC values (close to 1) correspond to synchronous amplitude envelope fluctuations between oscillations.
Orthogonalized amplitude envelope correlation (orthoAEC)
This metric is a variant of AEC that removes zero-lag signal overlaps due to spatial leakage through a multivariate symmetric orthogonalization approach (Colclough et al., 2015). Signals are orthogonalized before derivation of their envelopes, and then the linear correlation between these envelopes are computed.
After source reconstruction, FC was estimated using both the traditional pipeline and using our aperiodicity-unbiased oscillations approach. In the classical approach, FC estimation is performed directly after source reconstruction. In our new approach, FC is estimated only in ROIs exhibiting aperiodicity-unbiased oscillations. Each FC matrix coefficient
Thresholding method
After estimating FC, for each different FC method and for each frequency band, we computed the epoch-averaged connectivity matrix for each subject. A usual step in network estimation is FC matrix thresholding, thereby removing spurious connections and resulting in sparsely connected matrices which are necessary for the computation of most graph theoretical metrics. This proportional threshold retains the strongest correlations of the connectivity matrix. This selection is often mentioned as an analysis in which the “density” (Van Den Heuvel et al., 2008; Jalili, 2016) or “network cost” (Achard and Bullmore, 2007; Bassett et al., 2008; Ginestet et al., 2011) is fixed across groups. Thresholding value choice is arbitrary. Here, according to standard literature values, we set the proportional threshold of connectivity matrices at 5%.
After thresholding connectivity matrices, subject-averaged FC matrices were calculated for each FC estimation method and for each frequency band.
Comparison metrics
Descriptive aspects
After estimating cortical FC networks, the number of nodes and the percentage of connections were computed for each frequency band in both approaches to describe the architecture of networks.
Graph metrics
Graph theory metrics are a standard and well-established approach in neuroscience for quantifying the topology of brain networks (Rubinov and Sporns, 2010; De Vico Fallani et al., 2014). They are routinely applied to functional connectivity derived from EEG and MEG, both at rest and during cognitive tasks, as they provide meaningful insights into fundamental properties of brain organization such as functional segregation (e.g., clustering coefficient) and functional integration (e.g., characteristic path length, global efficiency; Bullmore and Sporns, 2009; Stam and Reijneveld, 2007). By using these well-established metrics, our goal was to leverage these common descriptors to make our findings comparable to the extensive literature on brain network connectivity. More importantly, in our context, we specifically applied these widely used metrics to assess whether such standard descriptors of network topology can be heavily biased when oscillatory activity is not disentangled from the aperiodic component. To ensure the reliability and interpretability of these measures, graph metrics were computed only for alpha-band networks with adequate density, given that sparse graphs can compromise the reliability of these measures. These metrics were computed using the Brain Connectivity Toolbox (BCT; Rubinov et al., 2009).
Clustering coefficient
The clustering coefficient is the fraction of triangles around a node which corresponds to the fraction of the node’s neighbors that are also neighbors of each other (Watts and Strogatz, 1998). This measure is indicative of the prevalence of clustered connectivity around individual nodes and more globally of network segregation.
Characteristic path length
Paths are sequences of distinct linked nodes that never visit a single node more than once and represent potential routes of information flow between brain regions. The length of these paths indicates the potential for functional integration between brain regions, shorter paths underlying stronger potential for integration. The average shortest path length between all pairs of nodes in the network is called characteristic path length (Watts and Strogatz, 1998), which is the most common measure of functional integration.
Global efficiency
Global efficiency corresponds to the average inverse shortest path length (Latora and Marchiori, 2001). This measure may be computed on disconnected networks, as paths between disconnected nodes are defined to have infinite length (zero efficiency). Global efficiency is often considered as a superior measure of integration (Achard and Bullmore, 2007).
Betweenness centrality
Betweenness centrality is the fraction of all shortest paths in the network that contain a given node. Nodes with high betweenness centrality values indicate participation in a large number of shortest paths. This centrality measure is based on the principle that central nodes participate in many short paths within a network, thereby acting as important centers of information flow (Freeman, 1977).
Statistical analysis
All statistical analyses were conducted in R v.4.1.3. (R Core Team, 2023) implemented with the tidyverse and lme4 packages (Bates et al., 2015; Wickham et al., 2019). The significance threshold was set at 0.05. In order to assess the quantitative contribution of aperiodic activity on the estimation of FC networks, we compared graph metrics values obtained from connectivity matrices derived from the classical pipeline versus those obtained from our approach identifying the aperiodicity-unbiased oscillations. The effect of the selection of aperiodicity-unbiased oscillations was assessed with paired t tests. Several aperiodicity-unbiased oscillation data showed a very large majority of values equal to 0. In such cases, statistical tests were not carried out. Bonferroni’s corrected p values were computed due to the multiple comparisons performed.
Data and code availability
Dataset A
The national ethics committee did not authorize data sharing, but anonymized FC matrices may be available upon request.
Dataset B
SRM resting-state EEG data are publicly available (https://OpenNeuro.org).
All Matlab and R codes used for processing and data analysis are publicly available at (https://github.com/noemiemonchy/1-f-corrected_functional_networks).
Results
Because the results were consistent across the different FC methods, only the findings derived from the PLV and wPLI are presented herein, with results from the other methods provided in Figures S1–S8. Results are first presented for resting state recordings, followed by those obtained in the task context.
Sparsity of functional networks computed on aperiodicity-unbiased oscillations
Quantitative contribution of aperiodic activity on FC estimation was evaluated by comparing, for each dataset, FC matrices derived from the standard pipeline with those obtained from our approach considering only aperiodicity-unbiased oscillations (Figs. 3–6; Figs. S1, S2). To characterize the sparsity of functional networks, we studied the number of nodes (cortical regions from Desikan–Killiany atlas) and the percentage of connections that were maintained with the selection of aperiodicity-unbiased oscillations in each dataset (Fig. 4, Table S1).
Subject-averaged functional connectivity matrices in resting state for each dataset and frequency band. Results are presented for each dataset: the first row shows FC matrices from the classical pipeline and the second row displays FC matrices from the approach verifying the presence of aperiodicity-unbiased oscillations. The top triangular portion of the matrices presents FC results from the PLV method and from the wPLI method for the bottom portion. In each connectivity matrix, rows and columns correspond to cortical ROIs from the Desikan–Killiany atlas. The color scale indicates connectivity strength, as estimated by the metric used.
Number of nodes (A) and percentage of kept connections (B) in resting-state functional networks computed on aperiodicity-unbiased oscillations, for each frequency band and dataset.
Subject-averaged functional connectivity matrices in task context for each dataset and frequency band. Results are presented for each dataset: the first row shows FC matrices from the classical pipeline and the second row displays FC matrices from our approach verifying first the presence of aperiodicity-unbiased oscillations. The top triangular portion of matrices presents FC results from the PLV method and from the wPLI method for the bottom portion. In each connectivity matrix, rows and columns correspond to cortical ROIs from the Desikan–Killiany atlas. The color scale indicates connectivity strength, as estimated by the metric used.
Number of nodes (A) and percentage of kept connections (B) in task context functional networks computed on aperiodicity-unbiased oscillations, for each frequency band and dataset.
Resting state
First, subject-averaged FC matrices derived from the classical pipeline exhibited a similar pattern across all frequency bands for each FC method (Fig. 3, Fig. S1). These results were consistent across all studied datasets. Additionally, FC was overall lower when looking at methods correcting for zero-lag. FC matrices computed on aperiodicity-unbiased oscillations greatly differed depending on the frequency band. Almost no connectivity was found in delta and gamma bands. A few FC values persisted in the theta and beta bands; however, overall connectivity in these frequency bands was greatly reduced. Although also reduced, only the alpha band maintained FC values similar to those derived from the classical pipeline. The selection of aperiodicity-unbiased oscillations resulted in sparser connectivity matrices across frequency bands. These results were consistent across all studied datasets.
The number of nodes (out of 68 possible nodes) was highest in the alpha band with 41.9 (SD = 11.9) nodes and 36.5 (SD = 22.1) nodes for the A-HC and PD data, respectively, and 58.0 (SD = 14.5) nodes for the dataset B (Fig. 4A; Table S1). In the beta band, fewer nodes were found across all datasets: 9.00 (SD = 5.9) nodes were kept in dataset A-HC, 7.2 (SD = 5.8) nodes in dataset A-PD, and 13.3 (SD = 13.7) nodes in dataset B. Nodes in delta, theta, and gamma bands were almost nonexistent. On average, in the theta band, no nodes (SD = 0) were kept in A-HC, ∼1 (SD = 3) node was found in dataset A-PD, and only 1 (SD = 5) node was maintained in dataset B. Finally, in the gamma band, ∼1 node (A-HC: SD = 3; A-PD: SD = 2; B-HC: SD = 3) was found in all datasets, while no node (SD = 0) was found in both datasets A-PD and B-HC in the delta band (SD = 0) and ∼1 node (SD = 1) was found in A-HC.
Similarly, the percentage of kept connections of functional networks computed on aperiodicity-unbiased oscillations was higher in the alpha band across all datasets (Fig. 4B, Table S1). Connections in the alpha band were reduced by about half after the selection of aperiodicity-unbiased oscillations: in dataset A-HC, 40.6% (SD = 20.7) of connections were kept, 38.7% (SD = 33.1) in dataset A-PD, and 77.0% (SD = 28.1) for the dataset B. In the beta band, the average percentage of connections did not exceed 10%: 2.3% (SD = 2.8) and 1.7% (SD = 2.5) of connections were kept in dataset A-HC and A-PD, respectively. In dataset B, ∼7.7% (SD = 14.6) of connections remained. In delta, theta, and gamma bands, the percentage of preserved connections was <1% across all datasets.
Cognitive task
Results obtained during the cognitive task were globally consistent with those observed during resting state. Similarly, subject-averaged FC matrices derived from the standard approach exhibited a comparable pattern across frequency bands for each FC method, with notably reduced values in the beta and gamma bands in methods correcting for zero-lag (Fig. 5, Fig. S2). As with resting-state FC matrices, cognitive task-related FC matrices exhibited significant changes across frequency bands. Notably, no connectivity was observed in the gamma band. Some FC values were still present in the delta, theta, and beta bands, depending on the dataset, though overall connectivity in these bands was significantly diminished. The alpha band, however, retained FC values similar to those derived from the classical pipeline. Identifying aperiodicity-unbiased oscillations led to sparser connectivity matrices across all frequency bands, with consistent results across all datasets.
Regarding the number of nodes and the percentage of connections of task-related networks estimated with aperiodicity-unbiased oscillations, values in delta, theta, and gamma bands were also all close to zero, regardless of the condition (Fig. 6, Table S1). Similar to the resting state, both the number of nodes and the percentage of connections were high in the alpha band, regardless of condition. In dataset A-HC, 15.00 (SD = 17.8) nodes were found in the congruent condition and 14.00 (SD = 12.3) nodes in the incongruent condition. The percentage of connections retained in the alpha band was 11.3% (SD = 19.6) in the congruent condition and 10.2% (SD = 18.2) in the incongruent condition. In dataset A-PD, 17.00 (SD = 12.3) nodes were observed in the alpha band in the congruent condition, and 16.7 (SD = 13.5) nodes in the incongruent condition. The percentage of connections maintained was 9.2% (SD = 10.6) in the congruent condition and 9.6% (SD = 13.0) in the incongruent condition. In the beta band, some nodes were also found in both datasets, regardless of condition, although the average percentage of connections did not exceed 5%. Specifically, in dataset A-HC, 2.5% (SD = 3.2) of connections were observed in the beta band, in the congruent condition, and 3.1% (SD = 3.6) in the incongruent condition, while in dataset A-PD, 1.3% (SD = 2.0) and 1.6% (SD = 2.3) were retained in the congruent and incongruent conditions, respectively.
Taken together, these results constitute the main message of our study: standard estimates of FC are heavily biased by the contribution of aperiodic activity, leading to networks that appear biasedly dense when oscillatory and aperiodic components are not disentangled. In the subsequent section, we further examine the consistency of these networks across subjects to demonstrate that FC is much less commonly shared than previously assumed and that even when oscillations are present, considerable intersubject variability remains.
Reduced cross-subject consistency of functional networks computed on aperiodicity-unbiased oscillations
To evaluate whether FC was consistent across subjects, we computed matrices encoding the percentage of participants exhibiting each connection for each frequency band (Figs. 7, 9; Figs. S3, S5). Next, we then focused on the alpha and beta bands, as these were the frequency ranges where FC was most preserved, to examine which networks remained after selecting only aperiodicity-unbiased oscillations and how these were consistent across participants (Figs. 8, 10; Figs. S4, S6).
Matrices encoding the percentage of participants exhibiting functional connectivity in resting state for each dataset and for each frequency band. Results are presented for each dataset: the first row presents matrices from the standard pipeline, and the second row displays matrices from our approach first verifying the presence of aperiodicity-unbiased oscillations. The top triangular portion of the matrix shows FC results from the PLV method and the wPLI method for the lower portion. In each connectivity matrix, rows and columns correspond to cortical ROIs from the Desikan–Killiany atlas. The color scale indicates the percentage of participants exhibiting functional connectivity, as estimated by the connectivity metric used.
Matrices encoding the percentage of participants exhibiting functional connectivity during the cognitive task for each dataset and for each frequency band. Results are presented for each dataset: the first row presents matrices from the standard pipeline, and the second row displays matrices from our approach verifying first the presence of aperiodicity-unbiased oscillations. The top triangular portion of the matrix shows FC results from the PLV method and the wPLI method for the bottom portion. In each connectivity matrix, rows and columns correspond to cortical ROIs from the Desikan–Killiany atlas. The color scale indicates the percentage of participants exhibiting functional connectivity, as estimated by the connectivity metric used.
Cortical functional networks in alpha and beta bands during resting state across subjects of each dataset for PLV and wPLI. Results are presented for each dataset: the first row shows cortical functional networks from FC matrices from the standard pipeline, and the second row displays cortical functional networks from FC matrices from our approach verifying first the presence of aperiodicity-unbiased oscillations. Edges coded the percentage of participants exhibiting the connection. For ease of reading, connections presented were shared by at least 20% of participants.
Cortical functional networks in alpha and beta bands during the cognitive task across subjects of each dataset for PLV and wPLI. Results are presented for each dataset: the first row shows cortical functional networks from FC matrices from the classical pipeline, and the second row displays cortical functional networks from FC matrices from our approach verifying first the presence of aperiodicity-unbiased oscillations. Edges coded the percentage of participants exhibiting the connection. For ease of reading, connections presented were shared by at least 20% of participants.
Decreased intersubject consistency of functional networks
Resting state
In the classical pipeline, the percentage of participants featuring each connection was similar across frequency bands, with some connections present in 100% of the population (Fig. 7, Fig. S3). When computing FC with aperiodicity-unbiased oscillations, the percentage of participants exhibiting connections dramatically decreased across frequency bands but the alpha band. No connectivity was observed in the delta band, beta networks were strongly reduced with only a few surviving connections, and theta/gamma results were sparse and inconsistent across datasets. In contrast, alpha connectivity remained relatively preserved, though shared by fewer participants overall.
Cognitive task
During the cognitive task, the overall pattern of FC consistency across participants was comparable with that observed at rest (Fig. 8, Fig. S5). Connectivity in delta, beta, theta, and gamma bands was again strongly reduced when computed on aperiodicity-unbiased oscillations, remaining sparse and inconsistently expressed across participants. Alpha networks were relatively preserved across approaches, though, as in resting state, they were shared by fewer individuals overall.
Reduced intersubject consistency and altered topography of functional networks
Resting state
In the alpha band, PLV functional networks derived from the classical pipeline were diffuse and widely distributed, with some connections present in all participants (Fig. 9). Using our approach, alpha functional networks exhibited a similar architecture, particularly in HC, but were shared by fewer individuals. Beta networks, in contrast, collapsed to only a few medial-parietal connections present in <40% of participants, a result that was consistent across all datasets. Regarding wPLI functional networks derived from the standard pipeline, a parieto-occipital alpha network was observed and shared by <50% of participants in each dataset, though PD patients exhibited a more diffuse distribution compared with HC (Fig. 8). Alpha networks computed on the aperiodicity-unbiased oscillations were more restricted in PD patients, but maintained a parieto-occipital architecture across all datasets. These networks were also shared by <50% of participants. Beta networks followed the same trend as for PLV, reducing to only a few medial-parietal connections expressed in a minority of participants.
Cognitive task
As in resting state, PLV alpha networks from the classical pipeline were diffuse and widely distributed across the cortex, with some connections present in nearly all participants (Fig. 10). When restricting to aperiodicity-unbiased oscillations, alpha networks became more restricted and were shared by fewer individuals. In HC, connections remained broadly distributed, whereas in PD patients they were localized to the occipital region. Beta networks were again sparse, with <20% of participants showing surviving connections. wPLI alpha functional networks from the standard pipeline were also distributed across the cortex and shared by <50% of participants, regardless of condition. Applying our approach resulted in alpha functional networks drastically reduced with few connections mainly in the medial region in HC and a network localized in the occipital region in PD. Beta networks followed the same pattern as PLV, being diffuse in the classical approach but reduced to very few connections shared by <20% of participants across datasets.
Graph theory metrics
Then, we computed graph theory metrics in the alpha band, the only frequency band that provided a sufficient amount of nonzero data to characterize functional networks in terms of segregation, integration, and centrality.
Resting state
Graph metrics values of alpha networks estimated with the PLV method reduced significantly in networks based on aperiodicity-unbiased oscillations as compared with the standard approach (Fig. 11, Table S2). Clustering coefficient values were significantly lower after selection of aperiodicity-unbiased oscillations (dataset A-HC: t = 5.33, p < 0.0001; dataset A-PD: t = 5.66, p < 0.0001; dataset B: t = 5.18, p < 0.0001). The clustering coefficient being a measure of segregation, this suggests that these networks were less segregated. Similarly, a significant reduction of characteristic path length values was found after the selection of aperiodicity-unbiased oscillations in the alpha band and in each dataset (dataset A-HC: t = 5.99, p < 0.0001; dataset A-PD: t = 4.42, p = 0.001; dataset B: t = 3.49, p = 0.001). This reduction implies integration between local clusters after selection of aperiodicity-unbiased oscillations. Betweenness centrality values were diminished in our approach in all datasets (dataset A-HC: t = 7.34, p < 0.0001; dataset A-PD: t = 6.99, p < 0.0001; dataset B: t = 2.76, p = 0.03). This significant reduction in betweenness centrality values implies that networks computed on aperiodicity-unbiased oscillations had fewer centrality. Last, global efficiency values were reduced in our approach in dataset A-HC (t = 7.34, p < 0.0001), dataset A-PD (t = 6.26, p < 0.0001), and dataset B (t = 2.88, p = 0.02). These results suggest that networks exhibited less communication efficiency when aperiodicity-unbiased oscillations were selected for dataset A.
Graph theory metrics of FC alpha networks in resting state estimated with the PLV method according to the approach. The darkest color (left part of the violin plots) indicates values from FC matrices derived from the classical pipeline while the lightest color (right part of the violin plots) corresponds to values from matrices computed on aperiodicity-unbiased oscillations.
Graph metrics of functional networks from other methods generally showed similar results (Fig. S7, Table S2); however, nonsignificant differences in clustering coefficient or characteristic path length values in dataset A-PD and dataset B can be observed in the alpha band according to the FC method. Overall, alpha-band networks based on aperiodicity-unbiased oscillations, although being the most preserved ones, were less segregated, exhibited less communication efficiency, had lower centrality, and exhibited reduced integration between clusters.
Cognitive task
Similar to the resting state, graph metrics values of alpha networks, during the cognitive task, estimated with the PLV method reduced significantly in networks based on aperiodicity-unbiased oscillations as compared with the standard approach (Fig. 12, Table S3). These networks were less segregated, as suggested by clustering coefficient values significantly lower after selection of aperiodicity-unbiased oscillations, regardless of condition (congruent condition: dataset A-HC: t = 11.30, p < 0.0001; dataset A-PD: t = 10.32, p < 0.0001; incongruent condition: dataset A-HC: t = 12.37, p < 0.0001; dataset A-PD: t = 9.75, p < 0.0001). Similarly, characteristic path length values were significantly reduced in these networks based on aperiodicity-unbiased oscillations in both datasets, regardless of condition (congruent condition: dataset A-HC: t = 9.00, p < 0.0001; dataset A-PD: t = 7.07, p < 0.001; incongruent condition: dataset A-HC: t = 9.13, p < 0.0001; dataset A-PD: t = 7.24, p < 0.001). In addition, networks based on aperiodicity-unbiased oscillations had fewer centrality (congruent condition: dataset A-HC: t = 0.08; p < 0.0001; dataset A-PD: t = 19.84; p < 0.0001; incongruent condition: dataset A-HC: t = 10.20; p < 0.0001; dataset A-PD: t = 21.29, p < 0.0001) and less communication efficiency, as suggested by global efficiency values (congruent condition: dataset A-HC: t = 11.75; p < 0.0001; dataset A-PD: t = 21.60; p < 0.0001; incongruent condition: dataset A-HC: t = 12.68; p < 0.0001; dataset A-PD: t = 18.95, p < 0.0001) in both datasets and regardless of condition.
Graph theory metrics of FC alpha networks during the cognitive task estimated with the PLV method according to the approach. The darkest color (left part of the violin plots) indicates values from FC matrices derived from the classical pipeline while the lightest color (right part of the violin plots) corresponds to values from matrices computed on aperiodicity-unbiased oscillations.
Graph metrics of functional networks during the cognitive task, computed from other methods generally showed similar results (Fig. S8, Table S3); however, nonsignificant differences in clustering coefficient can be observed according to the FC method. Overall, alpha-band networks based on aperiodicity-unbiased oscillations, although being the most preserved ones, were less segregated, exhibited less communication efficiency, had lower centrality, and exhibited reduced integration between clusters.
Discussion
Here, we aimed at evaluating the quantitative contribution of aperiodic activity in functional networks estimation. One key finding is that aperiodic activity is a major contributor of resting-state functional networks identified by classical methods. EEG oscillatory networks were drastically sparser than classically estimated: ∼99% of delta, theta, and gamma functional networks, over 90% of beta functional networks, and 23–61% of alpha functional networks were biased by aperiodic activity (not driven by pure oscillatory activity). Both the number of nodes and percentage of kept connections were greatly reduced in all frequency bands, remaining high in the alpha band.
Overestimation of functional networks in resting state
In delta, theta, beta, and gamma bands, considering aperiodicity-unbiased oscillations resulted in dramatically sparser FC matrices. In comparison to the standard approach where connections were often observed across most individuals, considering only aperiodicity-unbiased oscillations significantly reduced the number of connections (<1% of delta, theta, and gamma networks).
A large body of research focused on resting-state FC in these bands, with growing evidence supporting their potential as biomarkers. For example, delta-band FC is a robust signature of conscious states (Bourdillon et al., 2020). Alterations in delta- and theta-band FC were reported in neurological and psychiatric diseases (Koenig et al., 2001; Fingelkurts et al., 2007; Di Lorenzo et al., 2015; Hata et al., 2016; Cai et al., 2021), suggesting delta-band connectivity as a biomarker for large-scale degradation and theta-band connectivity as a marker of cognitive network integrity. Beta-band FC in resting-state networks was shown to support more localized or short-range connections (Samogin et al., 2019) and is associated with motor learning (Sugata et al., 2020), tactile spatial acuity (Sasaki et al., 2023), and Alzheimer’s disease (Stam et al., 2006; Koelewijn et al., 2017). Similarly, EEG gamma FC alterations in resting state were linked to schizophrenia (Di Lorenzo et al., 2015), Alzheimer’s disease (Tao and Tian, 2005), and creative behavior (Chhade et al., 2024).
Overall, resting-state delta, theta, beta, and gamma connectivity were repeatedly presented as meaningful oscillatory markers of cognitive processes. Our results, however, suggest that such associations may have been biased by broadband aperiodic activity rather than reflecting genuine oscillatory coupling, calling for a re-evaluation of prior findings reporting FC in these frequency bands as oscillatory in nature.
Graph theory metrics were systematically lower when aperiodic activity was considered, indicating that oscillatory networks are sparser, less locally clustered, and less globally integrated than previously assumed. This suggests that small-world properties and hub structures previously reported may partly reflect broadband aperiodic comodulation rather than rhythmic coupling, with genuine oscillatory networks being fewer and more specific, particularly in the alpha band. Because graph theory metrics inherently vary with factors such as network size, density, and thresholding, we did not focus on absolute values but rather on relative reductions observed when aperiodic contributions were considered. Importantly, proportional thresholding was applied identically across approaches, confirming that these reductions reflect genuine changes in network structure rather than methodological artifacts.
Overall consistency of alpha functional networks in resting-state
Both the number of nodes and percentage of preserved connections remained highest (but still attenuated) in the alpha band. Such alpha network preservation agrees with previous reports, showing that FC is most prominent in this band (Samogin et al., 2020). Both MEG and EEG studies confirmed that alpha activity in the default mode network is key in supporting functional interactions between nodes (Tang et al., 2017; Samogin et al., 2019). Alpha connectivity is particularly prominent during rest, supporting long-range communication and internal cognition, such as mind-wandering and self-referential thought (Knyazev et al., 2011; Clancy et al., 2022). It may also serve an inhibitory role, suppressing extraneous neural processing, enhancing efficient information processing and cognitive performance (Klimesch et al., 2007; Jensen and Mazaheri, 2010). Beyond the alpha band, our findings reveal an overestimation of connectivity and altered network properties when aperiodic activity is not considered, emphasizing the importance of verifying the presence of aperiodicity-unbiased oscillations in FC studies.
Persistence of results during cognitive control task
One could argue that rest, not requiring specific cognitive processes, would not show widespread oscillations. However, being engaged in a task does not imply that all cortical regions display oscillations in all bands. Assuming widespread oscillations during tasks could also overestimate FC. Applying similarly our approach during a cognitive task revealed results consistent with rest: FC in delta, theta, and gamma bands was dramatically sparse (<1% of networks preserved). In alpha (mostly) and beta bands, more connections were identified, yet connectivity was still reduced with limited overlap across subjects.
These findings challenge a body of literature reporting theta power (Cohen, 2014; Duprez et al., 2020; Kaiser et al., 2022) and theta connectivity as core markers of cognitive control (Cooper et al., 2015; Morales and Buzzell, 2025). Given the near absence of theta FC and limited intersubject consistency when accounting for aperiodic activity, prior findings need re-examination, in line with recent studies supporting that aperiodic activity drives theta effects previously considered oscillatory (Frelih et al., 2024; van Engen et al., 2024).
Furthermore, our results raise an interesting question: if many EEG networks are driven by aperiodic dynamics, rather than oscillatory coupling, what does that mean? The power spectrum aperiodic component is mathematically equivalent to the autocorrelation timescale of the time series (Gao et al., 2020). In fMRI, stronger autocorrelations result in stronger correlations/connectivity, even if underlying time series are “random” (Arbabshirani et al., 2014). For example, two Gaussian distributed white noise signals, which have a flat power spectrum with equal power at all frequencies, have significant correlations at chance level, whereas autocorrelated signals with steeper spectra and longer timescales produce spurious correlations, far more often, up to 60% of the time (Schaworonkow et al., 2015). FC could be a statistical artifact driven by inherent autocorrelations, and any differences in FC between groups or task states are driven by aperiodic/autocorrelation timescale differences, which are also increasingly being demonstrated to be dynamic.
On the importance of a new methodological “good practice”: verifying oscillations’ presence before estimating FC
Our analyses highlight the importance of establishing a new methodological good practice: verifying the presence of aperiodicity-unbiased oscillations in ROIs before estimating oscillation-based FC. If the focus is on oscillations and their implications, verifying such presence is fundamental since FC assumes a facilitation of interregional neural communication through oscillations (Fries, 2015). Consistent with previous studies, our findings show that aperiodic activity biases putative oscillations measurement and highlight the importance of parameterizing spectra as a combination of aperiodic and periodic activity (Donoghue et al., 2020; Ostlund et al., 2022; Donoghue and Watrous, 2023; Del Bianco et al., 2024). Therefore, persisting to assume that EEG signals consist of widespread oscillations without statistically testing this assumption leads to unreliable interpretations about underlying physiological mechanisms.
Methodological limitations
The specparam algorithm has limitations, but its performance is comparable with other methods such as BOSC or IRASA (Donoghue et al., 2020). Test–retest reliability of parameterized activity with specparam is satisfactory in eyes-closed resting state (McKeown et al., 2024). An important question remains regarding the proposed pipeline: to what extent can it detect true oscillatory networks? The proposed pipeline is able to detect oscillations in both (M)EEG, local field potentials (LFP), and simulated data (Gerster et al., 2022), and the robustness of specparam was assessed against violations of model assumptions, including fitting no knee when a knee is present, non-Gaussian peaks, and nonsinusoidal oscillations (Donoghue et al., 2020). However, recovering true oscillatory coupling remains a challenge and still needs further investigations. A simulation framework with a known ground truth would be a possible way to investigate this and would deserve dedicated studies.
Differences between datasets, such as sample size, number of electrodes, and age population, should be considered, but results replication across them strengthens robustness. Despite a different preprocessing, results consistently showed sparse FC networks at rest, further supporting robustness of our findings. Another limitation is the implicit assumption that oscillations are sinusoidal: investigating more realistic waveforms may provide a better understanding.
Our results also show that oscillations are not consistently present at every cortical site at rest, based on oscillatory detection on epoch-averaged spectra. However, studies showed that oscillations are not systematically sustained in time and are often bursty (Jones, 2016; Sherman et al., 2016), having implications on how strongly such transient changes in oscillations are responsible for FC results when averaging signals. Here, we adopted the common assumption of most FC studies that averaged signals contain mostly oscillatory contributions. Considering oscillatory bursts and how strongly these bursts influence FC would be important in further studies and could advocate for further methodological changes in how FC is calculated based on specific periods where oscillations are detected (Cole and Voytek, 2019).
Implications
Aperiodic activity is now widely examined in relation to cognition (Donoghue et al., 2020; Zhang et al., 2023), development (Schaworonkow and Voytek, 2021; Ostlund et al., 2022), and disease (Robertson et al., 2019; Johnston et al., 2023; Monchy et al., 2024). Given its relevance, it is important to consider its potential role in FC. By definition, oscillatory networks reflect narrowband rhythmic interactions, whereas aperiodicity-driven networks may capture broadband modulations of the spectral background (e.g., slope or offset). This distinction suggests that oscillatory networks could support temporally precise rhythmic coordination, while aperiodicity-driven networks may instead reflect global state-dependent fluctuations.
From a functional perspective, oscillatory networks have been linked to specific neurocognitive operations, based on the theory that interareal communication is facilitated by oscillatory dynamics (communication-through-coherence), whereby rhythmic synchronization aligns excitability phases across regions to regulate selective information transfer. In contrast, aperiodic activity covaries with arousal (Lendner et al., 2020), consciousness (Colombo et al., 2019; Maschke et al., 2023), or task performance (Podvalny et al., 2015; Waschke et al., 2021; Preston et al., 2025). By extension, one may infer that aperiodicity-driven networks could also reflect these factors, which remains to be demonstrated. While the physiological mechanisms remain debated and are likely multifactorial, these networks may provide complementary insights into how background cortical states shape the conditions under which oscillatory interactions occur.
Although this concept is gradually emerging, very few studies have so far addressed the notion of “aperiodicity-driven networks,” partly due to the lack of an established methodology for estimating such networks. However, recent work points in this direction. Helfrich et al. (2021) highlighted the need for new approaches to conceptualizing and quantifying such networks during sleep to further understand their relevance and interplay with oscillations in supporting memory consolidation. Complementarily, Chaoul and Siegel (2021) showed that aperiodic parameters are strongly correlated across cortical structures and not redundant with oscillatory correlations, suggesting that part of the network structure commonly attributed to oscillatory FC may in fact depend on aperiodic activity.
More broadly, distinguishing between oscillatory and aperiodicity-driven networks is not only conceptually meaningful but also practically relevant. For research on rhythmic coordination (e.g., frequency-specific neuromodulation), isolating aperiodicity-unbiased networks is essential. Conversely, for investigations into arousal, task engagement, or pharmacological manipulations that alter large-scale cortical excitability, aperiodicity-driven networks may provide particularly informative markers. Importantly, considering both contributions prevents FC metrics misinterpretation, which might otherwise conflate broadband comodulations with genuine rhythmic interactions.
Finally, rather than being a spoiler to oscillatory connectivity, aperiodic activity may play an active role in shaping interregional communication. There is also evidence that aperiodic activity captures transmembrane currents arising within a local population (Gao et al., 2017) and the underlying neuronal timescale (Gao et al., 2020). Dynamic, aperiodic networks could arise from matching timescales and altering population correlation structure (Schaworonkow et al., 2015; Wolff et al., 2022). Ultimately, acknowledging aperiodic activity contribution may require revising major paradigms in cognitive neuroscience, where broadband network dynamics are integrated alongside oscillatory coordination to explain cognition and behavior.
Footnotes
We thank all the volunteers who participated in this study.
This work is dedicated to the memory of our dear colleague and friend Alexandre Legros.
This work was supported by the Region Bretagne (ARED program) and the University of Rennes for the joint PhD fellowship, Bretagne Atlantique Ambition, and the Rennes Clinical Neuroscience Institute.
↵*N.M. and J.D. are the co-first authors.
↵‡B.V. and J.M. are the co-last authors.
The authors declare no competing financial interests.
↵†Deceased, Feb. 10, 2025.
This paper contains supplemental material available at: https://doi.org/10.1523/JNEUROSCI.1041-25.2025
- Correspondence should be addressed to J. Modolo at julien.modolo{at}inserm.fr.
