betaFeedback

WorksEngineering

Multidimensional Source-Space MEG Representation of Network Imbalance in ADHD

Abstract

Attention-deficit/hyperactivity disorder (ADHD) is associated with an atypical imbalance between the default mode network (DMN) and task-control systems. However, compressing neural activity into a single macroscopic imbalance measure may obscure individual neurophysiological differences. To address this, we leveraged the source-orientation selectivity and related measurement properties of resting-state magnetoencephalography (MEG) to establish a multidimensional source-space network representation in 21 individuals with ADHD and 21 healthy controls. We defined the macroscopic imbalance in relative Alpha activity between the DMN and task-control systems as the Core, and expanded its constituent information along three dimensions: spatial sources, spectral composition, and temporal network organization: into in-depth measures, assessed for stability, robustness, and complementarity. We found that the ADHD group exhibited a significant Core Alpha imbalance, driven mainly by task-control regions, while the multidimensional expansion revealed complementary ADHD-related neurophysiological information not fully represented by the Core, suggesting that similar macroscopic imbalances may reflect distinct underlying physiological states. Across 40 exploratory within-cohort held-out evaluations with separate development and test subsets, mean ROC-AUC increased from 0.7607 (Core-only) to 0.8995 with multidimensional inputs and to 0.9051 with the final ensemble, and mean accuracy increased from 71.43% to 79.82%. These findings suggest that multidimensional source-space expansion can further clarify the neurophysiological composition of ADHD-related network imbalance and improve individual-level discrimination, offering an interpretable framework potentially applicable to other brain disorders characterized by network imbalance.

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

Introduction

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder affecting approximately 7% of children worldwide [1], [2]. It is characterized by persistent inattention and/or hyperactivity-impulsivity, with substantial individual variability [1]. ADHD identification still relies primarily on clinical interviews, behavioral observations, and standardized rating scales. Although essential for clinical diagnosis, these approaches mainly capture behavioral manifestations and provide limited insight into underlying neurophysiological mechanisms. Identifying stable, reproducible, and physiologically interpretable brain measures to support more objective individual-level assessment therefore remains an important goal in ADHD research [1].

Among the functional abnormalities reported in ADHD, imbalance between the default mode network (DMN) and task-control-related networks is particularly relevant. The DMN primarily supports internally directed processes, whereas task-control networks support externally directed attention and cognitive control [3], [4]. Their coordination is essential for maintaining and shifting attentional states. Previous fMRI and electrophysiological studies have repeatedly identified abnormal activity and connectivity within the DMN and task-related networks in ADHD [5, 6, 7, 8, 9, 10, 11, 12, 13, 14]. These abnormalities have also been linked to genetic variation and treatment response [15, 16, 17], while related network features have been used for individual-level ADHD identification [18, 19, 20, 21]. Thus, DMN–task-control imbalance provides a low-dimensional summary of complex network abnormalities, potentially linking altered brain function to individual-level diagnosis.

However, this low-dimensional summary creates an information bottleneck. Similar imbalance values may occur in both ADHD and control individuals: some with ADHD show little imbalance, whereas some controls exhibit comparable patterns [22], [23]. As later illustrated in Figure 2A, although the DMN–task-control Alpha imbalance measure (hereafter, Core) shows a group-level shift, substantial overlap remains between ADHD and control individuals at the subject level. This overlap suggests that compressing complex network states into a single imbalance measure may obscure neurophysiological differences between individuals, mapping distinct physiological states onto similar Core values and limiting individual-level discrimination. The next step is therefore to analyze the neural activity that gives rise to the observed imbalance, rather than relying on the difference between the two networks alone.

Two-panel plots showing Core Alpha imbalance and decomposition of Core Alpha

Figure 2. Core Alpha imbalance and decomposition of its internal sources. (A) Subject-level distribution of the DMN–task-control Alpha Core in controls and ADHD participants. (B) Separate distributions of task-control-related and DMN-related log Alpha amplitudes. Blue and red indicate controls and ADHD, respectively; overlaid dots represent individual participants. Boxes summarize the distributions, with horizontal black lines indicating medians. Brackets indicate group comparisons, annotated with Hedges’ gg and two-sided permutation pp-values. Task-control-related Alpha showed stronger group separation than the isolated DMN Alpha component.

Such an analysis requires a measurement approach that can resolve the spatial sources of neural activity while preserving its rapid spectral and temporal dynamics. Among fMRI, EEG, and MEG, which are commonly used to investigate brain networks, MEG offers measurement properties particularly suited to this multidimensional analysis. Unlike the slow blood-oxygen-level-dependent response of fMRI, MEG measures magnetic fields generated by neural currents, providing direct access to rapid electrophysiological activity [24, 25, 26, 27]. Compared with EEG, MEG is generally less sensitive to skull conductivity uncertainties in forward modeling and source estimation [28, 29, 30] and exhibits distinct source-orientation selectivity [31], supporting cross-subject source-space anal-

Xiran Guo, Zhihao Liu, and Guanghua Liu are with the School of Cyber Science and Engineering, Huazhong University of Science and Technology, Wuhan 430074, China (e-mail: [email protected]; [email protected]; [email protected]). Corresponding author: Guanghua Liu. ysis [32]. Together, these properties provide a measurement foundation for resolving the spatial contributions of the Core and characterizing complementary spectral and temporal-network information. We further evaluated these measurement properties through matched EEG–MEG simulations to assess their contributions to source-space network recovery.

Building on the source-space measurement foundation provided by MEG, we addressed the information bottleneck of the Core by extracting complementary neurophysiological information from the underlying neural signals. Specifically, we expanded the Core representation along three physiologically connected dimensions. First, similar Core values may arise from different DMN and task-control contributions, requiring spatial decomposition to resolve this ambiguity. Second, because ADHD-related abnormalities extend beyond a single frequency band [10, 33], the Core cannot fully capture their spectral characteristics. Different oscillatory rhythms are associated with distinct neural processes, including Alpha-related information gating and Theta-related cognitive control, while aperiodic measures characterize broadband spectral properties [34, 35, 36, 37]. We therefore combined these measures to examine complementary neurophysiological information beyond the Core. Third, time-averaged spectra cannot capture network reconfiguration or stability, whereas MEG preserves the temporal information needed to characterize dynamic network coordination. These spatial, spectral, and temporal measures were linked by their source-space networks but captured distinct physiological properties. Core-adjusted analyses then tested whether the expanded measures provided additional ADHD-related information rather than merely restating the Alpha contrast.

Overall, this study established a source-space MEG framework that resolves the spatial sources of network imbalance, incorporates complementary spectral and temporal information, and assesses its value for individual-level ADHD discrimination. At the measurement level, we examined MEG properties relevant to rapid neural oscillations and cross-subject source reconstruction. At the representation level, DMN–task-control Alpha imbalance was expanded along spatial-source, spectral, and temporal-network dimensions. At the validation level, temporal reliability, technical-factor adjustment, empirical sensor-noise perturbation, and Core-adjusted complementarity analyses assessed the stability, robustness, spatial-source, spectral, and temporal-network dimensions and evaluated temporal reliability, technical influences, sensor-noise robustness, and Beyond-Core complementarity. Finally, individual-level ADHD discrimination was assessed using repeated stratified train–test splits, with a separately held-out test subset in each split.

Participants and Data Acquisition

This study retrospectively analyzed data from the Open MEG Archive (OMEGA) [38]. The final sample comprised 42 participants (21 ADHD and 21 healthy controls), classified according to OMEGA diagnostic labels. Only participants with complete resting-state MEG, structural MRI, and spatial information required for source reconstruction were included. Selected resting-state recordings were used for the primary analyses.

Ethics and Data Use Statement: This study is a retrospective secondary analysis of previously collected, de-identified resting-state MEG data from the Open MEG Archive (OMEGA) [38], accessed under an approved request and the OMEGA Data Use Agreement. All OMEGA participants were recruited for other MEG research studies and were offered to share their resting-state MEG data in OMEGA, and they signed the OMEGA consent form. The McGill University Health Centre Research Ethics Board reviewed the OMEGA registry and is responsible for its ethical oversight; under the OMEGA Data Use Agreement, no additional authorization was required for this secondary analysis. No new participants were recruited or contacted, and no additional human-subject measurements were performed. In accordance with the Data Use Agreement, only group-level results are reported; no single-participant data and no participant ID codes are presented. The data were kept in a secure environment and were not redistributed; requests for access should be directed to OMEGA.

Mean age was 21.75±4.7221.75 \pm4.72 years in ADHD and 23.62±6.2523.62 \pm6.25 years in controls; the groups comprised 8 males/13 females and 12 males/9 females, respectively (Fisher’s exact p=0.3543p = 0.3543). Age was obtained for the MEG session analyzed. Demographic sensitivity analyses assessed age and sex as covariates.

MEG data were acquired in a magnetically shielded room using a whole-head 275-channel CTF system at 2400 Hz, with a 600-Hz hardware anti-aliasing low-pass filter. ECG and horizontal and vertical EOG were recorded simultaneously. Head shape and anatomical landmarks were digitized using a Polhemus Fastrak system and registered to individual T1-weighted MRI.

MEG Preprocessing and Source-Space Reconstruction

MEG Preprocessing

The 265 MEG channels common to all participants were retained. Preprocessing comprised 1–40 Hz band-limiting, 50-Hz line-noise suppression, and resampling to 200 Hz. From each participant, 250 s of continuous resting-state data were extracted; after 5-s edge trimming, the remaining 240 s were divided into eight nonoverlapping 30-s windows. Theta and Alpha bands were defined as 4–8 Hz and 8–12 Hz, respectively. These windows were used for window-based analyses, whereas Theta network-role stability was derived separately from the first 30 s of each participant’s data.

  1. Source Reconstruction and ROI Time Series: Individual T1-weighted MRI scans were processed with FreeSurfer for cortical surface reconstruction and aparc anatomical parcellation [39]. After MEG–MRI coregistration, individual cortical source spaces and forward models were constructed. Noise covariance was estimated from the corresponding noise recordings, and source activity was reconstructed using dynamic statistical parametric mapping (dSPM) within the minimum-norm framework. The inverse operator used a loose-orientation constraint of 0.2, depth weighting of 0.8, and λ2=1/9\lambda^{2} = 1/9, retaining cortical surface-normal components (pick_ori="normal"). These parameters were fixed before analysis and were not tuned using group differences or classification performance.

For each aparc ROI, vertex-level dSPM time series were combined using the MNE-Python mean_flip method, which corrects signal polarity based on source orientation before averaging to reduce cancellation caused by cortical folding. ROI aggregation preceded all band-limited and spectral analyses.

Leveraging MEG Measurement Properties for Source-Space Representation

Matched EEG–MEG models were used to assess how conductivity uncertainty and source orientation affect source-space network recovery. Ground-truth 10-Hz source activity was generated, with DMN and task-control-related amplitudes varied to simulate different network states. Identical source states were projected through EEG and MEG forward models and reconstructed using the dSPM settings adopted in the main analysis. Simulations lasted 2 s at 200 Hz.

Conductivity analysis involved six participants with complete individual models. Reference skull conductivity was 0.006 S/m, with perturbations of 0.003 and 0.012 S/m. Sensor data generated under the reference model were reconstructed using the reference and both perturbed models. Recovery errors for task-control activity, DMN activity, and Core were measured by normalized mean absolute error (nMAE), using the across-state standard deviation of the reference representation for normalization.

Source-orientation analysis also involved six participants. At nine representative cortical locations, source position and strength were fixed. Local radial directions were defined relative to a reference center fitted to the individual MRI surface, together with two orthogonal tangential directions. EEG and MEG responses were evaluated for all three orientations to derive normalized radial response, tangential preference, and direction selectivity. The reference center was additionally perturbed to assess the stability of orientation measures. Radial and tangential basis vectors were constructed orthogonally.

Core DMN–Task-Control Alpha Imbalance

ROIs were predefined using the Desikan–Killiany parcellation and previous DMN and task-control network studies [4], [39]–[42]. As the resting-state dataset lacked individual functional localizers, these anatomical ROIs approximated the corresponding functional systems. DMN ROIs comprised the bilateral precuneus, posterior cingulate, and medial orbitofrontal cortex. The frontal-control component of the Core comprised the bilateral rostral middle frontal, caudal middle frontal, and superior frontal cortices. Bilateral superior and inferior parietal cortices formed a separate dorsal-attention ROI set and were not included in the Core.

Alpha activity was selected to characterize relative network organization because of its associations with information gating and functional inhibition [36]. The Core was designed as a low-dimensional summary of relative Alpha amplitudes rather than an exclusive measure of network imbalance.

ROI-level dSPM time series were filtered to 8–12 Hz. For each 30-s window ww, Alpha activity in ROI rr was quantified by root-mean-square (RMS) amplitude:

Rr,wα=1Tw∑t∈w[xrα(t)]2.(1)R^{\alpha}_{r,w} = \sqrt{\frac{1}{T_w}\sum_{t\in w}[x^{\alpha}_{r}(t)]^{2}}. \tag*{(1)}

For each system SS, ROI log-RMS amplitudes were averaged within each window:

AS,wα=1NS∑r∈Sln⁡(Rr,wα+ϵ).(2)A^{\alpha}_{S,w} = \frac{1}{N_S}\sum_{r\in S}\ln(R^{\alpha}_{r,w}+\epsilon). \tag*{(2)}

Here NSN_S is the number of ROIs in system SS. These measures are noise-normalized dSPM log amplitudes, not absolute currents in nAm. The window-level Core was defined as:

Cw=AControl,wα−ADMN,wα.(3)C_w = A^{\alpha}_{\mathrm{Control},w} - A^{\alpha}_{\mathrm{DMN},w}. \tag*{(3)}

Here, Control specifically denotes the frontal-control ROI set. This difference reduces sensitivity to common multiplicative scaling. Subject-level Core was the median across eight windows. The two network log-Alpha components were retained separately because distinct component values can yield similar Core values.

Multidimensional Expansion of the Core Imbalance

  1. Spatial-Source Dimension: The frontal-control and DMN Alpha components of the Core were retained separately. Because the Core is a difference, resolving its two terms distinguishes the frontal-control and DMN contributions underlying similar imbalance values. Both were calculated using the same 30-s windows and subject-level median aggregation as the Core, followed by between-group comparisons.

  1. Spectral Dimension: Frontal-control Theta was defined over 4–8 Hz. Source time series from predefined frontal-control ROIs were bandpass-filtered, and ROI-level RMS amplitudes were log-transformed and averaged within the network in each window. The subject-level measure was the median across eight windows. The Theta band was selected based on previous cognitive-control research [37]. Frontal-control Theta therefore probes control-related oscillatory activity outside the Alpha band summarized by the Core.

Aperiodic spectral features were estimated from source-space power spectral density (PSD) using Welch’s method with Hann windows, 4-s segments, and 2-s overlap. For the inter-network gradient, a log–log linear fit was performed over 2–7 and 14–30 Hz to exclude the main Alpha peak:

log⁡10P(f)=a+blog⁡10f,E=−b.(4)\log_{10} P(f) = a + b\log_{10} f,\quad E=-b. \tag*{(4)}

The DMN and frontal-control regions comprised the bilateral precuneus and rostral middle frontal cortex, respectively. The window-level aperiodic gradient was defined as:

Gap,w=EFrontal,w−EDMN,w.(5)G_{\mathrm{ap},w}=E_{\mathrm{Frontal},w}-E_{\mathrm{DMN},w}. \tag*{(5)}

Subject-level gradients were the medians across eight windows. The superior-parietal exponent was estimated separately over 2–30 Hz, excluding 4–13 Hz to reduce Theta- and Alpha-band contributions. Because the gradient fitting range retains 4–7 Hz activity, these exponents were interpreted as empirical spectral-slope measures rather than fully isolated aperiodic components. The gradient probes spectral differences within the two Core systems, whereas the superior-parietal exponent extends spectral assessment to attention-related cortex outside the Core.

3) Temporal-Network Dimension: The Theta network comprised four predefined nodes: bilateral precuneus and rostral middle frontal ROIs, without data-driven node selection. ROI signals were filtered to 4–8 Hz, and pairwise phase-locking values (PLVs) were calculated from Hilbert phases to construct weighted connectivity matrices. Weighted participation coefficients were then calculated using the predefined DMN and frontal-control modules.

For each participant, the first 30 s of data were divided into two 15-s segments, yielding node participation coefficient vectors PC(1)\mathrm{PC}^{(1)} and PC(2)\mathrm{PC}^{(2)}. Theta-band network-role stability was defined as:

Srole=11+mean⁡(∣PC(1)−PC(2)∣).(6)S_{\mathrm{role}}=\frac{1}{1+\operatorname{mean}\left(\left|\mathrm{PC}^{(1)}-\mathrm{PC}^{(2)}\right|\right)}. \tag*{(6)}

Higher values indicate greater consistency of node network roles between the two segments. Unlike the time-averaged Alpha and spectral measures, role stability characterizes short-term consistency of network participation. One stability value was obtained per participant.

The seven measures (Table I) are network-anchored: spatial terms define the window-level Core, while spectral and temporal measures probe the same DMN–frontal systems or related parietal cortex through additional properties. Classification used predefined combinations without assuming equal importance.

DimensionMeasureROIsComputationPhysiological Rationale
CoreCore Alpha imbalanceDMN and frontal-control8–12 Hz ROI log-RMS mean; network difference; eight-window medianLow-dimensional Alpha contrast between networks
SpatialFrontal-control AlphaFrontal-control ROIs8–12 Hz ROI log-RMS mean; eight-window medianFrontal contribution underlying the Core
SpatialDMN AlphaDMN ROIs8–12 Hz ROI log-RMS mean; eight-window medianDMN contribution underlying the Core
SpectralFrontal-control ThetaFrontal-control ROIs4–8 Hz ROI log-RMS mean; eight-window medianControl-related oscillations beyond Alpha
SpectralInter-network aperiodic gradientBilateral precuneus and rostral middle frontal2–7 and 14–30 Hz log–log spectral fitting; frontal minus DMN exponent; eight-window medianInter-network empirical spectral-slope differences
SpectralSuperior-parietal aperiodic exponentSuperior parietal cortexLog–log PSD fitting over 2–30 Hz, excluding 4–13 HzAttention-related regional spectrum beyond the Core
TemporalTheta network-role stabilityBilateral precuneus and rostral middle frontal4–8 Hz PLV; weighted participation coefficients; first 30 s split into two 15-s segmentsShort-term consistency of network participation

Table I. Multidimensional Source-Space Measures and Their Physiological Rationale

Reliability, Robustness, and Complementarity Analyses

1) Temporal Reliability: Temporal reliability was evaluated across eight nonoverlapping 30-s windows using ICC(2,1) and inter-window consistency. Participants, rather than individual windows, were treated as independent units in all group-level analyses.

2) Technical-Factor Adjustment: Technical covariates were selected from MEG–MRI coregistration distance, missing-channel count, available data-channel count, and total MEG-channel count. To reduce collinearity, candidate variables with absolute pairwise Spearman correlation ≥0.90\ge0.90 with an already retained variable were excluded. Coregistration distance (mm) and missing MEG-channel count were retained.

For each major measure, the following model was fitted:

Xi=β0+β1Di+β2Mi+β3Ri+εi.(7)X_i=\beta_0+\beta_1D_i+\beta_2M_i +\beta_3R_i+\varepsilon_i. \tag*{(7)}

Here, DiD_i denotes diagnostic group, MiM_i the number of missing MEG channels, and RiR_i the MEG–MRI coregistration distance in millimeters. For the technical-factor regression, outcomes and continuous covariates were standardized, and ordinary least squares coefficients were tested using conventional residual-variance standard errors. The adjusted diagnostic-group coefficient β1\beta_1 was extracted. Separately, Hedges’ gg was calculated between groups after regressing each measure on the two technical covariates alone and retaining the residuals. This residual-based effect size is distinct from the adjusted regression coefficient.

Separately, demographic sensitivity analyses fitted ordinary least squares models of the Core and each principal spectral measure with diagnostic group, age at the analyzed MEG session, and sex as predictors, using HC3 robust standard errors. The demographic and technical adjustments were evaluated separately; neither changed the classification inputs.

3) Empirical Sensor-Noise Perturbation: Eleven participants with noise recordings available for this additional perturbation analysis (five controls and six ADHD) were included to assess noise propagation into source-space measures, without between-group inference. Empirical MEG noise was assembled using contiguous 10-s block bootstrap sampling to match the baseline recording length and added to baseline sensor-space signals at 20, 10, 5, and 0 dB SNR. Five independent noise realizations were generated per participant and SNR. Source reconstruction and feature extraction were repeated for Core, frontal-control Theta, inter-network aperiodic gradient, and superior-parietal aperiodic exponent.

For each measure and SNR, Spearman correlation and mean absolute error (MAE), normalized by the baseline standard deviation across all 42 participants, were calculated relative to baseline and summarized across noise realizations. Theta-band network-role stability was not recomputed in this experiment.

4) Beyond-Core Complementarity: Spearman and group-adjusted partial Spearman correlations described monotonic associations with Core. Separately, the following Core-adjusted model tested whether each expanded measure kk retained ADHD-related group information:

Xik=β0+β1Ci+β2Di+εi.(8)X_{ik}=\beta_0+\beta_1C_i+\beta_2D_i+\varepsilon_i. \tag*{(8)}

Here, CiC_i is the Core and DiD_i denotes diagnostic group. HC3 robust standard errors were used to test the group coefficient β2\beta_2 after Core adjustment. Multiple tests were corrected using the Benjamini–Hochberg false discovery rate (FDR) procedure.

DMN: default mode network; ROI: region of interest; RMS: root-mean-square; PSD: power spectral density; PLV: phase-locking value. The seven measures characterize the Core and its multidimensional expansion. Candidate feature subsets were specified for exploratory comparison; fusion weights and decision thresholds were selected within each development subset.

Measures showing relatively stable Beyond-Core effects underwent leave-one-subject-out sensitivity analysis. Each participant was excluded in turn, yielding 42 refitted estimates on the original measure scales to assess dependence on influential individuals.

Individual-Level ADHD Diagnostic Feasibility

Core-only classification served as the baseline for an exploratory within-cohort comparison with multidimensional source-space information. Forty repeated stratified splits assigned 28 participants (14 per group) to development and 14 (7 per group) to held-out testing; all comparisons used identical splits.

To assess the contribution of expanded feature inputs under a matched classifier, Core-only and the full seven-measure representation were first compared using the same L2-regularized logistic-regression pipeline with training-only standardization. For each input set, the regularization parameter (C=0.01,0.1,1,C = 0.01, 0.1, 1, or 1010) was selected by three-fold out-of-fold ROC-AUC within the development subset, and the threshold (0.350.35–0.650.65 in steps of 0.050.05) was selected from development-set predictions by balanced accuracy. Each fitted classifier was then evaluated on the unchanged held-out subset.

The final multidimensional ensemble considered three subsets of the seven measures in Table I: all seven measures (All7); Core, frontal theta, aperiodic gradient, and superior-parietal exponent (Spectral4); and these four plus Theta network-role stability (Compact5). L2-regularized logistic regression was used with training-only standardization. The regularization parameter CC for All7 was selected within the development subset in the preceding modeling stage, whereas C=0.3C = 0.3 was fixed for Spectral4 and Compact5. Each subset’s predictions averaged one full-development model and 16 equally weighted models fitted to class-stratified 85% subsamples.

Fusion weights and thresholds were selected using three repetitions of stratified three-fold out-of-fold predictions within the development subset. Seven predefined weight triplets for (All7, Spectral4, Compact5) were considered: (1,0,0)(1, 0, 0), (0.75,0.25,0)(0.75, 0.25, 0), (0.5,0.5,0)(0.5, 0.5, 0), (0.75,0,0.25)(0.75, 0, 0.25), (0.5,0,0.5)(0.5, 0, 0.5), (0.5,0.25,0.25)(0.5, 0.25, 0.25), and (0.4,0.3,0.3)(0.4, 0.3, 0.3). Thresholds ranged from 0.380.38 to 0.620.62 in increments of 0.020.02.

The internal utility combined accuracy, ROC-AUC, and the minimum of sensitivity and specificity with weights of 0.550.55, 0.350.35, and $0.10. The reference configuration was All7 with a threshold of 0.50.5. A candidate was accepted only if its utility improved by at least $0.012, accuracy did not decrease, sensitivity and specificity each decreased by no more than 1/141/14, and ROC-AUC decreased by no more than $0.025 relative to this reference. Otherwise, the reference configuration was retained.

Standardization, model fitting, and configuration selection did not use held-out participants. The selected threshold and fusion weights were applied unchanged to testing. ROC-AUC was the principal discrimination metric; accuracy, sensitivity, and specificity were also computed for each held-out subset and summarized across 40 splits as mean ±\pm SD.

Because the splits reused the same 42 participants and model development followed exploratory comparisons in this cohort, results indicate exploratory within-cohort performance, not independent clinical validation.

Statistical Analysis

Participants were the independent units in all primary statistical analyses. Group differences were reported using means, standard deviations, Cohen’s dd, and Hedges’ gg, with two-sided permutation tests. The primary Core comparison used 10,000 permutations. Multiple parallel tests were corrected using the Benjamini–Hochberg FDR procedure.

Spearman rank correlations were used for continuous variables, with partial Spearman correlations when adjusting for diagnostic group. Core-adjusted and demographic regressions used HC3 robust standard errors; technical-factor regressions used conventional OLS inference. Key Beyond-Core effects were assessed using leave-one-subject-out sensitivity analyses. Matched EEG–MEG experiments used within-subject paired statistics. Sensor-noise analyses reported Spearman correlations and normalized MAE at each SNR.

Unless otherwise specified, two-sided p<0.05p < 0.05 indicated statistical significance. Diagnostic performance was summarized using the mean and SD of 40 repeated held-out evaluations, with ROC-AUC as the principal discrimination metric. Repeated evaluations were not treated as independent participant-level observations. All analyses were performed in Python, primarily using MNE-Python, NumPy, SciPy, pandas, scikit-learn, and statsmodels.

Results

MEG Measurement Advantages Enable Robust Source-Space Network Recovery

We first assessed whether MEG measurement properties supported the source-space network representation used in subsequent analyses. Under the reference head model, both EEG and MEG recovered the predefined task-control, DMN, and Core variations. Differences emerged primarily under modeling uncertainty rather than reference conditions.

When skull conductivity deviated from the reference value, EEG exhibited greater Core recovery errors than MEG. Under low-conductivity mismatch, median normalized MAE was 0.279 for EEG versus 0.065 for MEG; under high-conductivity mismatch, the corresponding values were 0.419 and 0.074. MEG showed lower errors in 11 of 12 subject-condition comparisons across six participants and two mismatch conditions. Thus, MEG better preserved source-space network information under conductivity uncertainty rather than showing universally superior recovery.

Source-orientation analysis revealed a complementary difference. The median direction-selectivity index was 0.686 for MEG versus 0.064 for EEG, with the same direction of difference in all six participants (Wilcoxon p=0.0313p = 0.0313). MEG showed a substantially lower normalized radial response (0.187 vs. 0.983), whereas tangential responses remained relatively strong.

Together, these findings indicate that MEG’s relevance to this framework lies in its lower sensitivity to conductivity mismatch and source-orientation selectivity. These properties support subsequent cross-subject source-space network analysis Figure 1.

Figure 1. MEG measurement properties and source-space network recovery. (A) Spearman correlations between predefined and reconstructed task-control, DMN, and Core information under the reference head model. Blue circles and orange triangles denote EEG and MEG, respectively. (B) Normalized mean absolute error (nMAE) of Core recovery under low- and high-skull-conductivity mismatch. Each point represents one participant, and gray lines connect paired EEG–MEG results. The displayed p-values refer to paired modality comparisons. (C) Normalized EEG and MEG responses to radial and two orthogonal tangential source orientations. Source locations and strengths were held constant across orientation conditions.

Core Alpha Imbalance Captures ADHD Abnormality but Compresses Its Internal Sources

We next examined the relative Alpha state between the DMN and task-control-related systems. Core was substantially lower in ADHD than in controls (−0.7584±0.2542-0.7584 \pm0.2542 vs. −0.4511±0.3308-0.4511 \pm0.3308), with a large group effect (Cohen’s d=1.042d = 1.042, Hedges’ g=1.022g = 1.022) and a significant two-sided permutation test (p=0.0014p = 0.0014). Thus, Core captured a pronounced alteration in the relative Alpha activity of the two systems.

However, a network difference does not reveal the contribution of each system. Decomposition showed markedly lower task-control-related Alpha in ADHD than in controls (−0.5465±0.3635-0.5465 \pm0.3635 vs. −0.0694±0.3940-0.0694 \pm0.3940; d=1.259d = 1.259, g=1.235g = 1.235, p=0.0004p = 0.0004). DMN-related Alpha was also numerically lower in ADHD (0.2159±0.29400.2159 \pm0.2940 vs. 0.3839±0.28580.3839 \pm0.2858), but the effect was smaller and not statistically significant (d=0.580d = 0.580, g=0.569g = 0.569, p=0.0638p = 0.0638). The component-wise results suggest a more pronounced frontal-control Alpha contribution than an isolated DMN Alpha difference.

Different combinations of task-control and DMN activity can produce similar Core values. Although Core summarizes the dominant network imbalance, it cannot fully represent the underlying source-space state. We therefore examined whether multidimensional and Beyond-Core analyses could reveal additional ADHD-related information Figure 2.

Multidimensional Expansion Reveals Information Beyond the Core

Beyond spatial decomposition, four measures extended Core: frontal-control Theta, inter-network aperiodic gradient, superior-parietal exponent, and Theta-band network-role stability. They probe the Core’s DMN–frontal systems or related parietal cortex through distinct spectral and temporal-network properties.

All four measures showed ADHD-related group differences Table II. Frontal-control Theta was lower in ADHD than in controls (−0.532±0.234-0.532 \pm0.234 vs. −0.200±0.277-0.200 \pm0.277; Hedges’ g=1.271g = 1.271, p=0.0001p = 0.0001), as was the inter-network aperiodic gradient (−0.194±0.242-0.194 \pm0.242 vs. −0.038±0.221-0.038 \pm0.221; g=0.660g = 0.660, p=0.0344p = 0.0344). The superior-parietal aperiodic exponent was higher in ADHD (0.993±0.3470.993 \pm0.347 vs. 0.793±0.2170.793 \pm0.217; g=−0.681g = -0.681, p=0.0299p = 0.0299), whereas Theta-band network-role stability was lower (0.967±0.0400.967 \pm0.040 vs. 0.987±0.0120.987 \pm0.012; g=0.671g = 0.671, p=0.0241p = 0.0241).

MeasureControl (Mean ± SD)ADHD (Mean ± SD)Hedges’ ggPermutation pp
Frontal-control Theta−0.200 ± 0.277−0.532 ± 0.2341.2710.0001
Inter-network aperiodic gradient−0.038 ± 0.221−0.194 ± 0.2420.6600.0344
Superior-parietal aperiodic exponent0.793 ± 0.2170.993 ± 0.347−0.6810.0299
Theta-band network-role stability0.987 ± 0.0120.967 ± 0.0400.6710.0241

Table II. Group-level results of the expanded source-space measures.

Diagnostic-group-adjusted partial Spearman correlations with Core were ρ=0.280\rho= 0.280, −0.237-0.237, 0.0760.076, and −0.222-0.222 (FDR-adjusted q=0.206q = 0.206, 0.206, 0.640, and 0.206). These estimates describe monotonic association with Core, distinct from the Beyond-Core group effects tested below.

Beyond-Core analysis tested whether these measures retained group information after Core adjustment. Three of the four expanded measures retained significant ADHD-related group effects: frontal-control Theta (β=−0.710\beta= -0.710, HC3 p=0.0110p = 0.0110, q=0.0326q = 0.0326), inter-network aperiodic gradient (β=−0.904\beta= -0.904, p=0.0163p = 0.0163, q=0.0326q = 0.0326), and Theta-band network-role stability (β=−0.867\beta= -0.867, p=0.0348p = 0.0348, q=0.0465q = 0.0465). The superior-parietal aperiodic exponent did not reach significance after Core adjustment (β=0.463\beta= 0.463, p=0.1286p = 0.1286, q=0.1286q = 0.1286). These results indicate differing levels of statistical support for complementary ADHD-related information, rather than establishing significant differences between the effects of individual measures.

Overall, the expanded measures provided different degrees of ADHD-related information beyond the Core. The Core-adjusted and sensitivity analyses support multidimensional characterization of the underlying neurophysiological state without assuming that all expanded measures have equivalent statistical effects (Figure 3).

Figure 3. Multidimensional expansion and Beyond-Core information. (A) ADHD–Control group effects for four expanded source-space measures, expressed as Hedges’ gg (Control minus ADHD), with permutation pp-values. (B) Partial Spearman correlations between each expanded measure and the Core after adjustment for diagnostic group. Displayed qq-values were corrected using the Benjamini–Hochberg false discovery rate (FDR) procedure. (C) Standardized diagnostic-group coefficients from Core-adjusted regression models, annotated with HC3 robust pp-values and FDR-adjusted qq-values. In panels A–C, horizontal segments extend from zero to the corresponding estimates and do not represent confidence intervals.

Reliability and Sensitivity of the Expanded Representation

To assess whether the observed network measures and principal Beyond-Core effects were driven by short recordings or individual observations, we examined window-level reliability and leave-one-subject-out (LOSO) sensitivity (Figure 4).

Reliability and subject-exclusion sensitivity of source-space MEG measures

Figure 4. Reliability and subject-exclusion sensitivity of source-space MEG measures. (A) ICC(2,1) across eight nonoverlapping 30-s windows for Core Alpha imbalance, frontal-control Theta, inter-network aperiodic gradient, and superior-parietal exponent in 42 participants; the dashed line marks ICC = 0.75. (B) Leave-one-subject-out robustness of the Core-adjusted group associations for frontal-control Theta and inter-network aperiodic gradient. Gray segments indicate the ranges of original-unit group coefficients across 42 exclusions; red diamonds indicate the full-sample estimates. Theta network-role stability was not assessed by window-level ICC.

Across eight nonoverlapping 30-s windows in 42 participants, the Core Alpha imbalance showed high ICC(2,1) reliability (0.906). Frontal-control Theta (0.875), inter-network aperiodic gradient (0.843), and superior-parietal aperiodic exponent (0.825) also exceeded 0.82 (Figure 4A). The superior-parietal reliability estimate was based on window-specific spectral fits. Theta-band network-role stability was not evaluated using window-level ICC because repeated estimates were unavailable.

LOSO analysis further examined whether the Beyond-Core group associations of frontal-control Theta and the aperiodic gradient depended on a single participant. Both effects preserved their direction and statistical significance in all 42 exclusions (Figure 4B); the respective coefficient ranges on the original measure scales were −0.251 to −0.187 and −0.266 to −0.184. Together, the window-level and subject-exclusion analyses support the repeatability of the principal measurements and the stability of these Core-adjusted associations.

Additional sensitivity checks considered technical covariates, age and sex, and empirical sensor-noise perturbations. Technical adjustment retained significant group terms for the Core (β=−1.404\beta= -1.404, p=0.0032p = 0.0032) and frontal-control Theta (β=−1.514\beta= -1.514, p=0.0007p = 0.0007); effects for the other expanded measures were more sensitive to technical covariates. After age and sex adjustment, the Core (p=0.0002p = 0.0002), frontal-control Theta (p=0.0007p = 0.0007), and aperiodic gradient (p=0.0300p = 0.0300) remained significant, whereas the superior-parietal exponent did not (p=0.3212p = 0.3212). Under empirical sensor-noise perturbation, the Core retained a Spearman rank correlation of ρ=0.845\rho= 0.845 even at 0 dB; noise sensitivity varied across other measures. These checks indicate that the strongest results do not depend solely on a single participant or short-window fluctuation, while their robustness is measure-dependent.

Exploratory Within-Cohort ADHD Classification

We assessed the individual-level relevance of the multidimensional source-space representation through exploratory ADHD classification. Core-only and multidimensional approaches were compared using 40 identical stratified held-out splits, each comprising 28 development and 14 test participants. All within-split model fitting and configuration selection used development data only.

In the matched logistic-regression comparison, adding the six expanded measures to Core increased mean ROC-AUC from 0.7607 to 0.8995 (+0.1388) and accuracy from 71.43% to 77.50%, using the same splits and tuning procedure. This comparison supports a discriminative contribution from multidimensional inputs without relying on ensemble fusion.

The final multidimensional approach achieved a mean ROC-AUC of 0.9051±0.05980.9051 \pm0.0598 versus 0.7607±0.10460.7607 \pm0.1046 for Core-only (Table III). Higher AUC occurred in 36 of 40 splits, with two ties and two decreases. Mean accuracy increased from 71.43±10.85%71.43 \pm10.85\% to 79.82±9.13%79.82 \pm9.13\%, sensitivity from 74.64±13.91%74.64 \pm13.91\% to 80.00±17.36%80.00 \pm17.36\%, and specificity from 68.21±22.53%68.21 \pm22.53\% to 79.64±13.29%79.64 \pm13.29\%. These repeated within-cohort results indicate exploratory individual-level discrimination, not independent external validation.

MetricCore-onlyMultidimensional Representation
ROC-AUC0.7607 ± 0.10460.9051 ± 0.0598
Accuracy (%)71.43 ± 10.8579.82 ± 9.13
Sensitivity (%)74.64 ± 13.9180.00 ± 17.36
Specificity (%)68.21 ± 22.5379.64 ± 13.29

Table III. Individual-Level ADHD Identification Performance. Values are mean ± SD across 40 repeated stratified held-out evaluations, each with 28 development and 14 test participants. Models used identical partitions. Repeated test subsets were drawn from the same 42-participant cohort and did not constitute independent external validation.

Discussion

This study addressed information obscured when DMN–task-control activity is reduced to a single Alpha imbalance value. Spatial decomposition resolved the Core’s component contributions, while physiologically related spectral and temporal measures characterized properties not represented by the Alpha contrast. A matched logistic-regression comparison improved mean ROC-AUC from 0.7607 to 0.8995 with expanded inputs, and the final multidimensional approach reached 0.9051. The main contribution is an interpretable link between low-dimensional network imbalance, targeted multidimensional expansion, and exploratory individual-level discrimination.

From MEG Measurement Properties to Source-Space Network Representation

We first examined the physical measurement properties supporting MEG source-space analysis. In matched EEG–MEG simulations, MEG showed lower network-representation errors under skull-conductivity perturbations and greater source-orientation selectivity, consistent with previous studies [28, 29, 30, 31]. These properties are relevant because both Core and its expanded measures depend on cross-regional and cross-subject source-space comparisons, where forward-model uncertainty may propagate through inverse reconstruction into the resulting network measures.

Both modalities recovered network information under the reference model. MEG’s relative robustness to conductivity mismatch and its source-orientation selectivity instead provide measurement-level support for the representation, without implying universally superior source localization or ADHD classification performance.

Neurophysiological Basis and Information Expansion of the Core Imbalance

The principal neurophysiological finding was altered DMN–task-control Alpha balance in ADHD. Previous fMRI and electrophysiological studies have reported abnormalities in the DMN, task-related networks, and their interactions [5, 6, 7, 8, 9, 10, 11, 12]. However, network imbalance alone cannot identify the contribution of each system. Our decomposition showed a stronger task-control-related Alpha alteration, whereas the isolated DMN Alpha difference was weaker. The Core abnormality should therefore not be interpreted simply as increased DMN activity; changes in task-related activity appear to contribute substantially to the relative network state.

Alpha oscillations are associated with information gating, functional inhibition, and network-state regulation [36], but their amplitude does not directly measure a single cognitive function. Accordingly, we interpret the Core as altered relative electrophysiological organization between DMN and task-control systems rather than attributing it to one region or psychological process.

Nevertheless, the Core compresses two Alpha components into a single difference: individuals with similar Core values can have distinct frontal-control and DMN contributions. Spatial decomposition exposes this otherwise obscured information. We then extended the representation to spectral and temporal properties that the Alpha contrast does not measure. Frontal-control Theta captured additional oscillatory information consistent with its association with cognitive control [37].

Spectral expansion complemented band-limited oscillatory measures with empirical spectral-slope descriptors, which may capture distinct aspects of neural activity [34, 35]. Our aperiodic findings indicate that similar Alpha network states may coexist with different background spectral organization. The inter-network gradient examines spectral differences between the two systems comprising the Core, while the superior-parietal exponent extends the analysis to attention-related cortex outside that contrast. These measures were selected to broaden physiological coverage, not to assume that each provides independent ADHD-related information. Thus, spectral expansion adds physiological descriptors that the Alpha contrast does not directly encode, rather than emphasizing any single measure. Importantly, the aperiodic exponent is not a direct measure of excitation–inhibition balance, and our findings do not establish specific changes in that balance.

Theta-band network-role stability characterized temporal consistency of network participation, beyond information available from static spectra. Core-adjusted results showed selective rather than uniform complementarity, supporting a multidimensional representation without treating every expanded measure as an equally informative biomarker. The Core and three principal expanded measures showed high short-window reliability (ICC > 0.82). The Core-adjusted effects of frontal-control Theta and the inter-network aperiodic gradient also persisted across all 42 leave-one-subject-out exclusions, reducing concern that these findings arose from one influential participant. Technical and demographic adjustments provided further support for the Core and frontal Theta associations, although robustness varied across the other descriptors. Together, these checks reinforce the principal expanded representation without assuming uniform robustness across all measures.

Crucially, the matched logistic-regression comparison increased mean ROC-AUC from 0.7607 to 0.8995 and accuracy from 71.43% to 77.50% when expanded features were added. The additional gain to 0.9051 ROC-AUC and 79.82% accuracy with the final ensemble was comparatively modest. Thus, improved discrimination was evident before ensemble fusion, supporting the value of multidimensional physiological inputs rather than attributing the result solely to modeling complexity. These repeated within-cohort evaluations remain exploratory and require independent validation for clinical generalization.

Reliability, Information Gain, and Individual-Level Implications

The cohort was relatively small, with potential age imbalance and limited clinical phenotyping. Demographic and technical adjustments supported the principal Core and frontal Theta findings, although sensitivity varied across expanded measures. Anatomical ROIs approximated functional networks, and medication and comorbidity data were not consistently available. Because the classification comparisons reused one exploratory cohort, independent validation of the locked representation and modeling pipeline is needed.

Limitations and Future Work

Conclusion

A single DMN–task-control Alpha imbalance measure cannot fully characterize its underlying source-space neurophysiological state. Using MEG, we decomposed the core imbalance measure (Core) into its spatial contributions and extracted complementary information along spectral and temporal-network dimensions. Our findings showed that task-control-related regions exhibited more pronounced ADHD-related differences in Alpha activity than the DMN. After adjusting for the Core, Theta activity in task-control-related regions and the inter-network empirical aperiodic gradient still showed significant ADHD-related group differences, revealing neurophysiological information not fully captured by the single imbalance measure. In matched logistic-regression comparisons, multidimensional inputs increased mean ROC-AUC from 0.7607 to 0.8995; the final ensemble achieved a mean ROC-AUC of 0.9051 and a mean accuracy of 79.82%. Overall, MEG-based multidimensional source-space expansion not only provides further insight into the neurophysiological composition of ADHD-related network imbalance but also improves exploratory individual-level discrimination. Independent cohorts are still needed to establish generalizability.

Data Availability Statement

The data were obtained from the Open MEG Archive (OMEGA), subject to its registration and data-use conditions; see https://omega.bic.mni.mcgill.ca and [38].

References

  1. [1]S. V. Faraone et al., “Attention-deficit/hyperactivity disorder,” Nat. Rev. Dis. Primers, vol. 10, 2024, art. no. 11.
  2. [2]R. Thomas, S. Sanders, J. Doust, E. Beller, and P. Glasziou, “Prevalence of attention-deficit/hyperactivity disorder: A systematic review and meta-analysis,” Pediatrics, vol. 135, no. 4, pp. e994–e1001, 2015.
  3. [3]F. X. Castellanos and E. Proal, “Large-scale brain systems in adhd: Beyond the prefrontal–striatal model,” Trends Cogn. Sci., vol. 16, no. 1, pp. 17–26, 2012.DOI
  4. [4]R. L. Buckner, J. R. Andrews-Hanna, and D. L. Schacter, “The brain’s default network: Anatomy, function, and relevance to disease,” Ann. N. Y. Acad. Sci., vol. 1124, pp. 1–38, 2008.DOI
  5. [5]F. X. Castellanos et al., “Cingulate-precuneus interactions: A new locus of dysfunction in adult attention-deficit/hyperactivity disorder,” Biol. Psychiatry, vol. 63, no. 3, pp. 332–337, 2008.pubmed.ncbi.nlm.nih.gov/17888409
  6. [6]D. A. Fair et al., “Atypical default network connectivity in youth with attention-deficit/hyperactivity disorder,” Biol. Psychiatry, vol. 68, no. 12, pp. 1084–1091, 2010.DOI
  7. [7]B. Sutsubasi et al., “Resting-state network dysconnectivity in adhd: A system-neuroscience-based meta-analysis,” World J. Biol. Psychiatry, vol. 21, no. 9, pp. 662–672, 2020.
  8. [8]L. J. Norman et al., “Evidence from ’big data’ for the default-mode hypothesis of adhd: A mega-analysis of multiple large samples,” Neuropsychopharmacology, vol. 48, no. 2, pp. 281–289, 2023.DOI
  9. [9]K. A. Duffy et al., “Increased integration between default mode and task-relevant networks in children with adhd is associated with impaired response control,” Dev. Cogn. Neurosci., vol. 50, 2021, art. no. 100980.DOI
  10. [10]T. W. Wilson et al., “Broadband neurophysiological abnormalities in the medial prefrontal region of the default-mode network in adults with adhd,” Hum. Brain Mapp., vol. 34, no. 3, pp. 566–574, 2013.
  11. [11]E. Heinrichs-Graham et al., “Pharmaco-meg evidence for attention-related hyper-connectivity between auditory and prefrontal cortices in adhd,” Psychiatry Res. Neuroimaging, vol. 221, no. 3, pp. 240–245, 2014.DOI
  12. [12]M. Muthuraman et al., “Multimodal alterations of directed connectivity profiles in patients with attention-deficit/hyperactivity disorders,” Sci. Rep., vol. 9, 2019, art. no. 20028.
  13. [13]G. Li, C. Cao, H. Fu, X. Li, and X. Gao, “Modeling functional brain networks for ADHD via spatial preservation-based neural architecture search,” IEEE Journal of Biomedical and Health Informatics, vol. 28, no. 11, pp. 6854–6864, 2024.DOI
  14. [14]D. Girish, Y. H. Chan, S. Gupta, J. Xia, and J. C. Rajapakse, “Encoding and decoding of brain dynamic functional connectivity for ADHD diagnosis,” IEEE Journal of Biomedical and Health Informatics, 2026, early Access.DOI
  15. [15]L. Liu et al., “Interference of default mode on attention networks in adults with attention-deficit/hyperactivity disorder and its association with genetic variants and treatment outcomes,” CNS Neurosci. Ther., vol. 30, no. 8, 2024, art. no. e4900.DOI
  16. [16]L. J. Norman, G. Sudre, M. Bouyssi-Kobar, W. Sharp, and P. Shaw, “A longitudinal study of resting-state connectivity and response to psychostimulant treatment in adhd,” Am. J. Psychiatry, vol. 178, no. 8, pp. 744–751, 2021.DOI
  17. [17]L. O. Quer et al., “Effects of methylphenidate on default-mode network/task-positive network synchronization in children with adhd,” J. Atten. Disord., vol. 21, no. 14, pp. 1208–1220, 2017.
  18. [18]C. Cao, H. Fu, G. Li, M. Wang, and X. Gao, “Adhd diagnosis guided by functional brain networks combined with domain knowledge,” Comput. Biol. Med., vol. 177, 2024, art. no. 108611.DOI
  19. [19]Z. Wang, B. Jie, C. Feng, T. Wang, W. Bian, X. Ding, W. Zhou, and M. Liu, “Distributed graph neural network for functional connectivity analysis in fMRI-based brain disorder identification,” IEEE Journal of Biomedical and Health Informatics, vol. 26, no. 4, pp. 1602–1613, 2022.DOI
  20. [20]C. Cao, X. Li, M. Wang, X. Chen, and X. Gao, “Function-structural interaction with progressive and multi-level feature fusion for ADHD classification,” IEEE Journal of Biomedical and Health Informatics, vol. 29, no. 10, pp. 7568–7578, 2025.DOI
  21. [21]D. Li, Z. Zhao, J. Hao, X. Dong, Y. Zhang, J. Xiang, and B. Wang, “Attention-enhanced temporal and spatial feature extraction network for ADHD diagnosis based on fMRI,” IEEE Journal of Biomedical and Health Informatics, 2026, early Access.DOI
  22. [22]J. C. Mostert et al., “Characterising resting-state functional connectivity in a large sample of adults with adhd,” Prog. Neuropsychopharmacol. Biol. Psychiatry, vol. 67, pp. 82–91, 2016.DOI
  23. [23]T. Wolfers, C. F. Beckmann, M. Hoogman, J. K. Buitelaar, B. Franke, and A. F. Marquand, “Individual differences in the average patient: Mapping the heterogeneity in adhd using normative models,” Psychol. Med., vol. 50, no. 2, pp. 314–323, 2020.DOI
  24. [24]N. K. Logothetis, J. Pauls, M. Augath, T. Trinath, and A. Oeltermann, “Neurophysiological investigation of the basis of the fmri signal,” Nature, vol. 412, no. 6843, pp. 150–157, 2001.
  25. [25]N. K. Baldwin, “Magnetoencephalography for brain and what we cannot do with fmri,” Nature, vol. 453, no. 7197, pp. 869–878, 2008.
  26. [26]S. Baillet, “Magnetoencephalography for brain electrophysiology and imaging,” Nature Neuroscience, vol. 20, no. 3, pp. 327–339, 2017.DOI
  27. [27]J. Gross, “Magnetoencephalography in cognitive neuroscience: A primer,” Neuron, vol. 104, no. 2, pp. 189–204, 2019.DOI
  28. [28]N. G. Gencer and C. E. Acar, “Sensitivity of eeg and meg measurements to tissue conductivity,” Physics in Medicine and Biology, vol. 49, no. 5, pp. 701–717, 2004.pubmed.ncbi.nlm.nih.gov/15070197
  29. [29]M. C. Piastra, A. Nüssing, I. Vorwerk, M. Clerc, C. Engwer, and C. H. Wolters, “A comprehensive study on electroencephalography and magnetoencephalography sensitivity to cortical and subcortical sources,” Human Brain Mapping, vol. 42, no. 4, pp. 978–992, 2021.DOI
  30. [30]J. Vorwerk, M. B. Höltershinken, C. H. Wolters, and D. Baumgarten, “Global sensitivity of meg source analysis to tissue conductivity uncertainties,” NeuroImage, vol. 323, p. 121618, 2025.DOI
  31. [31]S. P. Ahlfors, J. Han, J. W. Belliveau, and M. S. Hämäläinen, “Sensitivity of meg and eeg to source orientation,” Brain Topography, vol. 23, no. 3, pp. 227–232, 2010.DOI
  32. [32]F. Zhao, I. Kakkos, G. K. Matsopoulos, C. Guan, and Y. Sun, “Explaining E/MEG source imaging and beyond: An updated review,” IEEE Journal of Biomedical and Health Informatics, vol. 29, no. 12, pp. 9271–9286, 2025.DOI
  33. [33]A. Khadmaoui, C. Gómez, J. Poza, A. Bachiller, A. Fernández, J. Quintero, and R. Hornero, “MEG analysis of neural interactions in attention-deficit/hyperactivity disorder,” Computational Intelligence and Neuroscience, vol. 2016, p. 8450241, 2016.DOI
  34. [34]T. Donoghue et al., “Parameterizing neural power spectra into periodic and aperiodic components,” Nature Neuroscience, vol. 23, no. 12, pp. 1655–1665, 2020.DOI
  35. [35]O. Dakwar-Kawar et al., “Aperiodic and periodic components of oscillatory brain activity in relation to cognition and symptoms in pediatric ADHD,” Cerebral Cortex, vol. 34, no. 6, p. b h a e236, 2024.DOI
  36. [36]O. Jensen and A. Mazaheri, “Shaping functional architecture by oscillatory alpha activity: Gating by inhibition,” Frontiers in Human Neuroscience, vol. 4, p. 186, 2010.DOI
  37. [37]J. F. Cavanagh and M. J. Frank, “Frontal theta as a mechanism for cognitive control,” Trends in Cognitive Sciences, vol. 18, no. 8, pp. 414–421, 2014.DOI
  38. [38]G. Niso et al., “OMEGA: The open meg archive,” NeuroImage, vol. 124, pp. 1182–1187, 2016.DOI
  39. [39]R. S. Desikan et al., “An automated labeling system for subdividing the human cerebral cortex on mri scans into gyral based regions of interest,” NeuroImage, vol. 31, no. 3, pp. 968–980, 2006.pubmed.ncbi.nlm.nih.gov/16530430
  40. [42]J. D. Power et al., “Functional network organization of the human brain,” Neuron, vol. 72, no. 4, pp. 665–678, 2011.DOI

Paper details

vlnc:W_qswc22k6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents