Temporal neural dynamics patterns in episodic and chronic migraine: a magnetoencephalography study

Abstract

Background

Chronic migraine (CM) severely affects patients’ work and daily life, imposing a significant economic burden. However, the underlying neural mechanisms of migraine chronification remain unclear. This study aimed to characterize temporal neural dynamics patterns of migraine via magnetoencephalography (MEG) combined with dynamic network mode (DyNeMo), providing further neuroimaging evidence for migraine chronification.

Methods

This cross-sectional study recruited patients with episodic migraine (EM), CM, and healthy controls (HC). MEG data were acquired during resting and somatosensory stimulation states. The DyNeMo model was applied to source-reconstructed MEG data to quantify temporal neural dynamics metrics, including mean lifetime, mean interval, switching rate and fractional occupancy. Permutation-based analysis of covariance (ANCOVA) with Bonferroni correction and Spearman correlation with false discovery rate (FDR) correction were applied.

Results

Six distinct brain modes were identified: visual network (VN), anterior and posterior default mode networks (aDMN/pDMN), right and left sensorimotor networks (rSMN/lSMN), and auditory network (AN). During resting state, EM showed prolonged mean lifetime and decreased switching rate of the VN vs. HC; CM showed increased switching rate of the VN vs. EM; CM showed prolonged mean lifetime and decreased switching rate of the AN vs. HC. During somatosensory stimulation state, both EM and CM showed prolonged mean lifetime and decreased switching rate of the AN vs. HC. In CM patients, longer duration of disease was correlated with shorter mean lifetime and higher switching rate of the AN during somatosensory stimulation state.

Conclusions

This study identifies distinct alterations in the temporal dynamics of VN and AN in EM and CM. Abnormalities in AN were observed during both resting and somatosensory stimulation states, while disease duration in CM was associated with altered temporal metrics of the AN. These cross-sectional findings support the involvement of altered sensory and cross-modal network processing in migraine and warrant longitudinal validation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s10194-026-02466-0.

Introduction

Migraine is a common primary headache disorder characterized by highly disabling attacks [1]. It affects 14.1% of the global population and ranks second among all neurological disorders in terms of disease burden according to disability-adjusted life years (DALYs) [2]. Migraine is classified into episodic migraine (EM) and chronic migraine (CM) based on monthly headache frequency (≥ 15 headache days per month for more than three months) according to the International Classification of Headache Disorders, 3rd edition (ICHD-3) [3]. Although CM patients develop from only a small fraction of the EM population (approximately 3%), its unremitting and highly disabling nature profoundly disrupts patients’ daily functioning, imposing substantial economic and systemic healthcare burdens globally [4, 5].

The transition from EM to CM is not merely a chronological increase in headache frequency, but rather a complex functional disturbance of the central nervous system, driven by central sensitization and maladaptive neural plasticity involved in multisensory processing and descending pain modulation pathway [4, 6]. However, the precise neural mechanisms underlying migraine chronification remain incompletely understood [4, 7].

Neuroimaging studies have revealed migraine chronification is accompanied by progressive flattening of the cortical hierarchy, involving the visual network (VN) and somatomotor network (SMN) [8]. Graph theory analyses have demonstrated that CM patients exhibit reduced global and local network efficiency compared with EM patients and healthy controls (HC), with disrupted topological organization in frontal and temporal regions [9]. Dynamic functional connectivity (FC) studies further indicate that CM patients exhibit aberrant cross-network interactions, including altered FC among default mode network (DMN), auditory network (AN), executive control network (ECN) and VN [10]. Collectively, these findings suggest that the transition from EM to CM involves a progressive loss of network stability across multiple coexisting networks, a process that depends on long range temporal dependencies. However, conventional static or sliding-window approaches are ill-equipped to capture such temporal neural dynamic patterns, as they assume stationary activity or mutually exclusive brain states [11]. Consequently, how alterations of brain networks at millisecond timescales contribute to migraine chronification remain unknown.

To address this gap, this study applied the time-delay embedded dynamic network mode (DyNeMo), a generative approach that captures the continuous mixing of coexisting networks and their long-range temporal dependencies at the native temporal resolution of magnetoencephalography (MEG) [12–14]. We hypothesized that migraine patients would exhibit altered temporal neural dynamics of sensory networks, particularly the mean lifetime and switching rate, and that these alterations would be associated with disease progression, especially in the chronic stage. Accordingly, this study aimed to characterize millisecond-scale temporal neural dynamics patterns across the migraine spectrum and to provide novel neurophysiological evidence for the pathophysiology of migraine chronification.

Materials and methods

Participants

Between May 2020 and February 2024, a total of 319 participants were consecutively recruited from the Headache Center at the Department of Neurology, Beijing Tiantan Hospital, Capital Medical University. The initial cohort comprised 258 patients diagnosed with migraine (89 presenting with EM and 169 with CM) and 61 HC who possessed no prior history of psychiatric or neurological disorders (see Fig. 1).

Fig. 1.

Flowchart of the process for participant inclusion. Note: HC: healthy controls; EM: episodic migraine; CM: chronic migraine

Diagnoses for all recruited patients were established independently by two experienced, board-certified neurologists strictly adhering to the diagnostic criteria outlined in the ICHD-3 [3]. Specifically, the diagnosis of CM required the presence of headache on fifteen or more days per month for over three months, with at least eight days per month fulfilling criteria for migraines with or without aura or responding to migraine-specific treatment. Participants were required to be between 14 and 75 years of age and to have abstained from any prophylactic migraine treatments for a minimum of three months prior to neuroimaging. General exclusion criteria encompassed the presence of other types of primary or secondary headache disorders, major systemic or psychiatric diseases, an inability to tolerate the MEG or MRI scanner environments due to severe claustrophobia, the presence of incompatible metallic implants, and current pregnancy or lactation. Comprehensive clinical histories detailing duration of disease and headache day, neuroimaging data and validated psychometric clinical scales were collected from all participants during the interictal period (defined as at least 72 h before or after a migraine attack). The clinical scales included Migraine Disability Assessment Scale (MIDAS), Headache Impact Test-6 (HIT-6), Generalized Anxiety Disorder-7 (GAD-7), Patient Health Questionnaire-9 (PHQ-9), Pittsburgh Sleep Quality Index (PSQI), Montreal Cognitive Assessment (MoCA) and Visual Analogue Scale (VAS), which assessed headache impact, anxiety and depression symptoms, sleep quality, cognitive function and pain severity, respectively. MoCA scores were adjusted for years of education according to established norms.

This study, a sub-study of the China HeadAche DIsorders RegiStry Study, was approved by the Institutional Review Committee of Beijing Tiantan Hospital of Capital Medical University (KY2022-044) and was registered at https://www.clinicaltrials.gov (Clinical trial number: not applicable). All participants provided informed written consent in accordance with the Declaration of Helsinki.

MRI data acquisition

All participants were imaged with a 3.0 Tesla MR scanner (GE Healthcare, Milwaukee, WI, USA) at the Nuclear Medicine Department of Beijing Tiantan Hospital. Examinations were performed by a neuroradiologist who was unaware of the participant diagnoses. Participants were asked to keep their heads and necks still, remain awake, and keep their eyes closed, with measures taken to reduce noise and head movement. T1-weighted volumetric images were obtained using the 3D BRAVO sequence (coronal acquisition, field of view = 256 mm, acquisition matrix = 256, slice number = 192, flip angle = 15^°, repetition time (TR) = 8.5 ms, echo time (TE) = 3.2 ms, voxel size = 1 × 1 × 1.5 mm³).

MEG data acquisition

Neural activity during resting state and somatosensory stimulation was recorded using an Elekta Neuroimage scanner with 306 channels (102 magnetometers and 204 gradiometers). Head position of participants was measured by four head position indicator (HPI) coils, and recordings were restarted if excessive head movement occurred during the scan. Head shape was digitally marked (Polhemus Fastrak^®). Approximately 300 points were marked on the nasion, anterior ear points, and scalp for MRI co-registration. Data acquisition was performed at a sampling rate of 2000 Hz using a low-pass filter with a cut-off of 660 Hz, while the participants’ electrooculogram and electrocardiogram were also recorded. Participants were instructed to keep their heads and necks still, remain awake and keep their eyes closed during the scan. Resting-state MEG data were collected for five minutes. Following the resting-state recording, somatosensory stimulation was administered by delivering electrical pulses to the median nerve at the wrist. Each pulse had a duration of 10 ms, and the interstimulus interval (ISI) was 1 s. Stimulation was first applied to the left wrist (200 trials) and then to the right wrist (200 trials). Stimulation intensity was adjusted to exceed the motor threshold (eliciting thumb movement) without causing pain.

MEG preprocessing

In the preprocessing pipeline, spatiotemporal signal space separation (tSSS) was employed offline using MaxFilter software to mitigate artifacts and interpolate malfunctioning channels [15]. Motion compensation was conducted using the continuous recordings of the HPI coils. MEG data were preprocessed using the OHBA software library (OSL, https://github.com/OHBA-analysis). The time series was downsampled to 500 Hz before a notch filter at 50 Hz (and harmonics) was used to remove power line noise. Data were then bandpass filtered between 0.5 and 250 Hz to reduce computational demands. Finally, an automated bad segment and bad channel detection algorithm in OSL was used to remove particularly noisy segments and channels. Independent Component Analysis (ICA) was subsequently performed to estimate 60 independent components, among which noise components related to ocular movements and cardiac activity were removed based on correlations between ICA time courses and the electrooculograms and electrocardiogram, followed by further visual inspection.

After filtering into the frequency band of interest (1–200 Hz), structural MRI images were coregistered with the MEG data using an iterative close-point algorithm. Digitised head points were matched to individual participants’ scalp surfaces extracted with FSL’s BET tool [16]. Using a single-shell head model [17], preprocessed MEG sensor data were reconstructed onto an 8 mm isotropic grid with a linearly constrained minimum variance beamformer. Voxels were parcellated into 38 regions using the Giles38 parcellation distributed with osl-dynamics, an ICA-derived fMRI parcellation originally developed from 100 subjects. The full list of parcels and MNI coordinates is provided in Supplementary Table S1. A time series for each parcel was extracted by applying the spatial basis method to each region of interest.

MEG source reconstruction can lead to artefactual correlations between parcel time courses, referred to as source leakage. We minimised source leakage using the symmetric multivariate leakage reduction technique implemented in OSL [18]. As the signs of source-reconstructed time courses are arbitrary, data from multiple participants cannot be concatenated directly. This problem was addressed by applying a search approach to adjust the sign of the source-reconstructed time courses on the assumption that the covariance matrix between each pair of brain areas has the same sign across participants [19]. This algorithm maximized the “sign agreement” across covariance matrices of all participants by flipping the sign of the parcel time series of certain participants.

Data preparation and DyNeMo model training

Prior to training the DyNeMo model, the leakage-corrected, sign-aligned source data underwent additional preparation. First, time-delay embedding was implemented to encode critical phase and spectral information directly into the observation model’s covariance matrix. We utilized 15 temporal embeddings, specifically applying time lags ranging from − 7 to + 7 samples, which expanded the dimensionality of our data to 570 channels per participant. To optimize computational efficiency while retaining the vast majority of variance, Principal Component Analysis (PCA) was subsequently applied to reduce the dimensionality to 80 principal channels. Finally, the embedded and reduced data underwent standardization via z-transformation across the temporal dimension to facilitate stable gradient descent during model training.

The fully prepared time-delay embedded data were concatenated along the time dimension across all participants and entered the time-delay embedded DyNeMo using the OSL-dynamics toolbox [12]. Compared to the Hidden Markov Model (HMM), DyNeMo could model data as a linear mixture of patterns rather than mutually exclusive states, allowing the modelling of brain dynamics by patterns that can flexibly combine and mix over time. Additionally, for traditional HMMs, the state at any time point was only conditionally dependent on the state at the previous point (i.e., the model is Markovian), which limited its ability to model long-range temporal dependencies [20]. In contrast, DyNeMo models long-range time dependencies in data rather than relying solely on the previous time point.

Here, we trained on time-delay embedded data to ensure that all spectral information was contained within the mode covariance matrices. DyNeMo was used to learn the mixing coefficients and mode covariances. The mixing coefficients described latent dynamics in the training data and the reweighted mixing coefficients were calculated by multiplying the raw mixing coefficients by the trace of each mode covariance to explain each mode’s contribution to the time-varying covariance. We used the raw inferred mode mixing coefficients to estimate quantities characterizing the training data, including summary statistics for each mode’s temporal properties (i.e., fractional occupancy, mean lifetime, mean interval, switching rate). The fractional occupancy represents the fraction of total time that a given brain network mode is activated, reflecting its relative prevalence over the recording period. The mean lifetime represents the average duration for which a brain network mode remains continuously activated before deactivating, reflecting its temporal stability. The mean interval represents the average duration between successive activations of the same brain network mode, reflecting the recurrence interval of that mode. The switching rate represents the number of times a brain network mode is activated per second, reflecting its activation frequency.

For the primary analysis, we fitted a six-mode time-delay embedded DyNeMo model separately to the resting-state and somatosensory stimulation recordings. The six-mode solution was selected because it yielded reproducible and anatomically interpretable networks across both conditions (resting-state and somatosensory stimulation), corresponded to VN, anterior/posterior default mode networks (aDMN/pDMN), right/left sensorimotor networks (rSMN/lSMN) and AN. Modes 2 and 3 were labelled as aDMN and pDMN based on their dominant parcel-level topography. The aDMN mode showed prominent relative power in medial prefrontal/anterior cingulate and frontal midline parcels, whereas the pDMN mode showed prominent relative power in posterior cingulate/midline precuneus, bilateral precuneus/superior medial parietal and inferior parietal/angular-supramarginal parcels. Except for six modes, we also fitted an eight (see Supplementary Fig. S1 and S2) and ten (see Supplementary Fig. S3 and S4) modes time-delay embedded DyNeMo model separately to the resting-state and somatosensory stimulation recordings, while higher model orders examined during model development primarily split these networks into subcomponents. For each condition, the model was trained 10 times and the run with the lowest final loss after KL annealing was retained. Each run used random subset initialization (3 initializations, 1 epoch each, 20% of the data). The DyNeMo configuration was: n_modes = 6, n_channels = 80, sequence_length = 100, inference_n_units = 64, model_n_units = 64, layer normalization in both networks, learnable alpha temperature initialized at 1.0, learn_means = False, learn_covariances = True, batch_size = 64, learning_rate = 1e-3, n_epochs = 60, and tanh KL annealing for 30 epochs. A schematic overview of the DyNeMo pipeline is provided in Fig. 2, illustrating the processing and training steps, including MEG and MRI preprocessing, data preparation, model training and inference, and post-hoc analysis for resting state and somatosensory stimulation.

Fig. 2.

Illustration of the time-delay embedded DyNeMo approach applied to MEG data. Whole-brain MEG data from 306 channels were preprocessed using MaxFilter and OSL for motion compensation and denoising. After filtering into the frequency band of interest, MRI images were co-registered with the MEG data, and voxels were parcellated into 38 regions of interest. Before training, time-delay embedding and PCA were applied to capture spectral properties and improve computational efficiency. During DyNeMo training, individual neural activity was decomposed into six distinct brain modes, each reflecting unique functional states or patterns of network dynamics. Post-hoc analyses were employed to characterize these dynamic processes of brain activity, represented as mixtures of different brain modes. Mode time courses are shown at three stages: raw DyNeMo mixing coefficients (alpha), covariance-trace-reweighted/normalized mixing coefficients, and binary mode activation time courses. Binary time courses were obtained by fitting a two-component Gaussian mixture model to the normalized coefficient of each mode; time points assigned to the high-probability component were classified as active. These binary time courses were used to calculate fractional occupancy, mean lifetime, mean interval and switching rate. DyNeMo: dynamic network mode; PCA: principal component analysis; ROI: region of interest

Statistical analysis

Statistical analyses were performed using R software (Version 4.5.1; R Foundation for Statistical Computing, Vienna, Austria). Demographic and clinical data were analyzed using the base R statistical environment. Normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed variables are presented as mean ± standard deviation (SD) and were compared using one-way analysis of variance (ANOVA); non-normally distributed variables are presented as median (interquartile range, IQR) and were compared using the Kruskal-Wallis H test for three group comparisons and the Mann-Whitney U test for two group comparisons. Differences in gender were tested using the chi-square test.

To examine the mean lifetime, mean interval, switching rate, and fractional occupancy of DyNeMo modes, a permutation-based analysis of covariance (ANCOVA) was employed. For each metric, the dependent variable was first regressed on age and gender using a linear model, and the residuals were used as adjusted values. A total of 5000 permutations were performed to establish the null distribution of the F statistic under the Freedman-Lane procedure. For metrics showing a significant overall effect, post-hoc pairwise permutation tests were conducted. Bonferroni correction was applied for multiple comparisons, with adjusted P < 0.05 considered statistically significant. Effect sizes for pairwise comparisons were reported as Cohen’s d with 95% confidence intervals (CI) estimated via bootstrap resampling with 2000 iterations (resampling within each group separately).

Sensitivity analysis was performed to evaluate whether the observed differences between EM and CM were robust to additional confounding variables. Specifically, for metrics that showed significant group differences in the primary analysis, a pairwise permutation test (5000 permutations) was conducted between EM and CM groups with extended covariates including age, gender, sleep quality (PSQI), depressive symptoms (PHQ-9), anxiety symptoms (GAD-7), photophobia, phonophobia, and duration of disease. Bonferroni correction was applied to the resulting P values.

Spearman’s correlation analysis was used to explore the relationship between altered network metrics and clinical variables, separately within the EM and CM groups. Correlation coefficients with 95% CIs were reported. The resulting P values were corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) method, with q < 0.05 considered statistically significant.

Results

Demographic and general clinical characteristics

The initial recruitment yielded 319 participants. Following stringent quality control, 23 migraine patients and 9 HC were excluded from the final analysis due to excessive motion artifacts resulting in poor data quality. Finally, a highly robust analytical cohort consisted of 78 EM patients, 157 CM patients and 52 HC (see Fig. 1). There were no significant differences in age and gender among the three groups. Compared with EM patients, CM patients had higher headache days per month, MIDAS, HIT-6, PHQ-9, GAD-7 and PSQI scores, and lower MoCA scores (P < 0.05; see Table 1). No significant differences were found in duration of disease, VAS, photophobia, phonophobia between the EM and CM groups.

Table 1.

Demographic and clinical characteristics of participants

	HC (n = 52)	EM (n = 78)	CM (n = 157)	H /χ² /Z	P value
Age	37.50 (27.75, 50.25)	38.50 (26.25, 49.75)	46.00 (31.00, 53.00)	3.26	0.20a
Gender (Female/Male)	27/25	55/23	106/51	5.39	0.07b
Duration of disease (year)	NA	15.00 (7.00, 24.00)	19.00 (8.00, 24.00)	-1.26	0.21c
Photophobia, n (%)	NA	58 (74.36%)	118 (75.16%)	0.02	0.89b
Phonophobia, n (%)	NA	58 (74.36%)	124 (78.98%)	0.64	0.42b
VAS	NA	7.00 (6.00, 8.00)	8.00 (6.00, 9.00)	-1.55	0.12c
Headache day (per month)	NA	8.00 (4.00, 11.75)	20.00 (17.00, 30.00)	-10.58	< 0.01c
MIDAS	NA	50.00 (23.00, 80.00)	153.00 (90.00, 200.00)	-6.85	< 0.01c
HIT-6	NA	63.50 (42.00, 69.75)	66.00 (60.00, 70.00)	-2.52	0.01c
PHQ-9	NA	5.00 (2.00, 9.00)	8.00 (4.00, 12.00)	-3.73	< 0.01c
GAD-7	NA	4.00 (1.25, 9.00)	7.00 (3.00, 12.00)	-2.98	< 0.01c
PSQI	NA	5.50 (3.00, 8.00)	9.00 (4.00, 12.00)	-3.87	< 0.01c
MoCA	NA	28.00 (27.00, 29.00)	27.00 (24.00, 29.00)	3.26	< 0.01c

HC = healthy controls; EM = episodic migraine; CM = chronic migraine; VAS = Visual Analogue Scale; MIDAS = Migraine Disability Assessment Scale; HIT-6 = Headache Impact Test-6; PHQ-9 = Patient Health Questionnaire-9; GAD-7 = Generalized Anxiety Disorder-7; PSQI = Pittsburgh Sleep Quality Index. MoCA: Montreal Cognitive Assessment; a: Kruskal-Wallis H test; b: chi-square test; c: Mann-Whitney U test

Identifying distinct brain network dynamics under resting and somatosensory stimulation

To investigate abnormal temporal neural dynamics patterns associated with migraine, we applied DyNeMo to continuous MEG data from the combined cohort of 235 migraine patients and 52 HC during resting state and somatosensory stimulation. The preprocessed MEG data were source-localized using beamforming and mapped to a 38-region parcellation with reduced spatial source leakage [12, 13]. DyNeMo was then applied to the source-localized time-series from these 38 regions, concatenated across all participants. This data-driven temporal modelling, using a recurrent neural network, identified six spatially distributed brain modes with distinct temporal profiles and cortical power distributions. Remarkably, these brain modes were highly consistent across resting state and somatosensory stimulation corresponding to well-established functional brain networks reported in previous literature (Figs. 3A and 4A). Specifically, mode 1 aligns with the VN [21], modes 2 and 3 are identified as the anterior and posterior default mode networks (aDMN and pDMN) [22], mode 4 and mode 5 correspond to the right and left sensorimotor network (rSMN and lSMN) [23] and mode 6 represents the AN [24].

Fig. 3.

Brain modes estimated by DyNeMo during resting state and their temporal neural dynamics patterns. (A) Power maps show mode-specific power after subtracting the mean power across all modes. Red indicates higher relative power, and blue indicates lower relative power for a given mode. No statistical threshold was applied to these descriptive mode maps. (B) Temporal neural dynamics patterns differed among the three groups during the resting state. Significant group differences were found in the mean lifetime and switching rate of the VN and AN. For the VN, EM patients showed prolonged mean lifetime compared with HC (P = 0.005, Bonferroni correction); EM patients showed a decreased switching rate compared with HC (P = 0.036, Bonferroni correction), whereas CM patients showed an increased switching rate compared with EM patients (P = 0.043, Bonferroni correction). For the AN, CM patients showed prolonged mean lifetime (P = 0.019, Bonferroni correction) and decreased switching rate compared with HC (P = 0.013, Bonferroni correction). HC: healthy control; EM: episodic migraine; CM: chronic migraine; VN: visual network.

Fig. 4.

Brain modes estimated by DyNeMo during somatosensory stimulation and their temporal neural dynamics patterns. (A) Power maps show mode-specific power after subtracting the mean power across all modes. Red indicates higher relative power, and blue indicates lower relative power for a given mode. No statistical threshold was applied to these descriptive mode maps. (B) Temporal neural dynamics patterns differ among groups during somatosensory stimulation. Significant group differences were found in the mean lifetime and switching rate of the AN. Both EM and CM patients showed prolonged mean lifetime (EM vs. HC: P < 0.001; CM vs. HC: P < 0.001, Bonferroni correction) and decreased switching rate of the AN compared with HC (EM vs. HC: P < 0.001; CM vs. HC: P < 0.001, Bonferroni correction). HC: healthy control; EM: episodic migraine; CM: chronic migraine; AN: auditory network

Abnormal brain network dynamics in migraine patients under resting state

During resting state, group-averaged power maps (subtracted from mean across modes) derived from all migraines and healthy participants are shown in Fig. 3A. To examine the abnormal temporal dynamics of brain modes in migraine patients, we tested several brain mode metrics, including fractional occupancy, mean lifetime, mean interval, and switching rate. Among the four metrics, significant group differences were observed exclusively in mean lifetime and switching rate, whereas fractional occupancy and mean interval did not differ among the three groups. For the VN, permutation-based ANCOVA revealed significant overall group effects for mean lifetime and switching rate. Post-hoc pairwise comparisons showed that EM patients had prolonged mean lifetime compared with HC (P = 0.005, Cohen’s d = 0.638, Bonferroni correction), whereas the comparison between CM and HC did not reach statistical significance. Regarding switching rate, EM patients showed a decreased switching rate compared with HC (P = 0.036, Cohen’s d = 0.556, Bonferroni correction) and CM patients showed an increased switching rate compared with EM patients (P = 0.043, Cohen’s d = 0.451, Bonferroni correction), see Table 2; Fig. 3B. However, this difference between EM and CM did not remain significant after additionally controlling for sleep quality, depressive and anxiety symptoms, photophobia, phonophobia, and duration of disease in the sensitivity analysis (P = 0.078, Bonferroni correction, see Supplementary Table S2).

Table 2.

Group differences in temporal neural dynamics metrics under rest and somatosensory stimulation states

Mode	A Permutation ANCOVA		Post-hoc Pairwise Comparisons (Bonferroni-corrected P value)
	F	P value	HC vs. EM	HC vs. CM	EM vs. CM	Cohen_d
Mean lifetime (rest)
Mode 1 (VN)	6.251	0.002	0.005	0.074	0.150	HC vs. EM = 0.638
Mode 6 (AN)	3.943	0.020	1.000	0.019	0.328	HC vs. CM = 0.323
Switching rate (rest)
Mode 1 (VN)	4.194	0.020	0.036	1.000	0.043	EM vs. CM = 0.451 HC vs. EM = 0.556
Mode 6 (AN)	4.280	0.014	0.576	0.013	0.632	HC vs. CM = 0.381
Mean lifetime (somatosensory)
Mode 6 (AN)	12.595	P < 0.001	P < 0.001	P < 0.001	1.000	HC vs. EM = 0.694 HC vs. CM = 0.609
Switching rate (somatosensory)
Mode 6 (AN)	13.309	P < 0.001	P < 0.001	P < 0.001	1.000	HC vs. EM = 0.753 HC vs. CM = 0.673

HC = healthy controls; EM = episodic migraine; CM = chronic migraine

For the AN, significant overall group effects were found for mean lifetime and switching rate. CM patients exhibited prolonged mean lifetime (P = 0.019, Cohen’s d = 0.329, Bonferroni correction) and decreased switching rate (P = 0.013, Cohen’s d = 0.381, Bonferroni correction) compared with HC, whereas no significant differences were observed between EM and HC (see Table 2; Fig. 3B). The fractional occupancy and mean interval did not differ significantly among the three groups.

Abnormal brain network dynamics in migraine patients under somatosensory stimulation

During somatosensory stimulation, brief electrical pulses were delivered to the median nerve at the wrist bilaterally. The brain power maps obtained during somatosensory stimulation were consistent with those identified during the resting state (Fig. 4A), indicating that the six brain modes were highly reproducible across conditions.

Significant group differences were observed in the AN. Permutation-based ANCOVA revealed significant overall group effects for both switching rate and mean lifetime. Post-hoc pairwise comparisons showed that both EM and CM patients exhibited prolonged mean lifetime (EM vs. HC: P < 0.001, Cohen’s d = 0.694; CM vs. HC: P < 0.001, Cohen’s d = 0.609, Bonferroni correction) and decreased switching rate (EM vs. HC: P < 0.001, Cohen’s d = 0.753; CM vs. HC: P < 0.001, Cohen’s d = 0.673, Bonferroni correction) of the AN compared with HC (see Table 2; Fig. 4B). No significant differences were observed between EM and CM in either metric. The fractional occupancy and mean interval did not differ significantly among the three groups.

Correlations between brain dynamics and clinical metrics

To investigate the clinical significance of the altered temporal neural dynamics patterns, Spearman’s correlation analysis was performed between the abnormal network metrics and clinical variables, separately within the EM and CM groups. Correlation coefficients with 95% CI and FDR-corrected q-values were reported.

In the CM group, longer duration of disease was significantly correlated with shorter mean lifetime [r = -0.329, 95% CI (-0.464, -0.179), FDR-corrected q < 0.001, see Fig. 5A] and higher switching rate [r = 0.255, 95% CI (0.100, 0.398), FDR-corrected q = 0.012, see Fig. 5B] of the AN during somatosensory stimulation. No other significant correlations were found between the altered network metrics and clinical variables in either the EM or CM group after FDR correction.

Fig. 5.

Correlation analysis between temporal neural dynamics patterns and clinical characteristics. In the CM group, longer duration of disease was significantly correlated with shorter mean lifetime (r = -0.329, FDR-corrected q < 0.001, Fig. 5A) and higher switching rate (r = 0.255, FDR-corrected q = 0.012, Fig. 5B) of the AN during somatosensory stimulation. CM, chronic migraine

Discussion

This study employed the DyNeMo model on MEG data to investigate millisecond-scale temporal neural dynamics patterns underlying migraine chronification during both resting state and somatosensory stimulation. We identified six distinct brain modes predominantly corresponding to the VN, aDMN, pDMN, rSMN, lSMN and AN. Among the four temporal metrics examined, significant group differences were observed exclusively in mean lifetime and switching rate, whereas mean interval and fractional occupancy did not differ among the three groups in either brain mode.

Visual network alterations during resting state

During resting state, the VN showed a distinct pattern of temporal dynamics alterations across the migraine spectrum. Specifically, EM patients exhibited prolonged mean lifetime and decreased switching rate of the VN compared with HC, while CM patients showed increased switching rate of the VN compared with EM patients. The mean lifetime reflects the average duration for which a given brain network mode remains continuously activated, indexing the temporal stability of coherent neural activity, whereas the switching rate measures the frequency of mode activations per unit time and reflects the network’s propensity for state reconfiguration [12, 13, 25, 26]. In electrophysiological terms, a prolonged mean lifetime may indicate sustained local field potential synchronization, while an increased switching rate implies more frequent transitions between network configurations, consistent with elevated cortical excitability or reduced inhibitory control [26, 27]. The pattern observed in our VN results is consistent with previous studies demonstrating that the visual cortex in migraine patients is characterized by altered excitability and habituation deficits [28–32]. Visual cortex hyperexcitability is one of the most consistently reported neurophysiological features of migraine, even during the interictal period [28, 33]. The prolonged mean lifetime observed in EM patients may reflect sustained activation of the VN, consistent with altered visual processing in the context of cortical hyperexcitability. However, in CM patients, this pattern shifted toward increased switching rate, suggesting a cross-sectional difference in VN temporal stability between EM and CM subgroups. Recent MEG studies have similarly reported that migraine chronification is associated with aberrant beta-band connectivity within pain-related cortical regions [34], providing convergent evidence for network-level differences between migraine subgroups. Importantly, our sensitivity analysis, which additionally controlled for sleep quality, depressive and anxiety symptoms, photophobia, phonophobia, and duration of disease, revealed that the differences between EM and CM in the VN were no longer significant. This suggests that the differences between EM and CM in VN dynamics are partly attributable to these clinical confounders, consistent with previous research documenting the impact of psychiatric comorbidities and sleep disturbances on brain network organization in migraine [35, 36].

Auditory network alterations across resting and somatosensory stimulation states

Our findings also revealed significant alterations in the AN, a network that has received less attention in migraine research compared with the VN. During resting state, CM patients showed a prolonged mean lifetime and decreased switching rate compared with HC. Under somatosensory stimulation, alterations in the AN were even more pronounced: both EM and CM patients showed prolonged mean lifetime and decreased switching rate relative to HC. However, no significant differences were observed for mean interval or fractional occupancy among the three groups in either the resting or somatosensory stimulation state. The mean interval represents the average duration between successive activations of the same network mode, indexing the recovery period before a network can be re-engaged, while the fractional occupancy represents the proportion of total time a mode is active, which integrates both frequency and duration [12, 13]. This selective sensitivity of mean lifetime and switching rate suggests that migraine primarily affects the intrinsic stability and reconfiguration frequency of individual network activations rather than the overall temporal budget or recovery interval.

A particularly noteworthy observation is that under somatosensory stimulation, the alterations of VN that were present during resting state were no longer significant, whereas the AN alteration persisted and even intensified. This differential pattern warrants careful consideration. Previous electrophysiological studies have suggested that visual cortex excitability in migraine patients exhibits state-dependent normalization during non-visual sensory tasks, possibly due to cross-modal inhibitory interactions [37, 38]. In contrast, the auditory cortex in migraine patients may be inherently more susceptible to aberrant activation across sensory modalities. Phonophobia, the heightened sensitivity to sound, is one of the most prevalent and disabling symptoms during migraine attacks and persists interictally in many patients [39, 40]. This suggests that the baseline excitability threshold of the AN is already lowered in migraine, rendering it particularly vulnerable to non-specific sensory input. When a salient but non-auditory stimulus (i.e., median nerve electrical stimulation) is presented, the hyper-responsive AN may be preferentially engaged, leading to prolonged activation and reduced switching flexibility. This interpretation is consistent with the concept of multisensory integration abnormalities in migraine, wherein defective cross-modal sensory gating allows the AN to be aberrantly recruited by non-auditory stimuli [41, 42].

Supporting this view, Masson et al.. demonstrated using MEG and EEG that migraine patients exhibit altered auditory attentional processing, including reduced auditory evoked field amplitudes and aberrant oscillatory responses, indicative of fundamental dysfunction in the AN even at resting state [43]. More recently, Liu et al.. reported aberrant static and dynamic functional connectivity specifically within auditory processing networks during auditory stimulation in migraine, further confirming that AN dysfunction is a core feature of migraine pathophysiology beyond mere symptom-related modulation [44]. The cross-modal engagement of the AN observed in our study extends these findings by showing that the dysfunction of AN can be elicited by somatosensory rather than auditory input, suggesting altered cross-modal sensory processing in the migraine brain [41, 42].

Furthermore, traditional resting-state fMRI studies, which capture slower fluctuations on the order of seconds, have reported aberrant cross-network interactions among the DMN, AN, ECN, and VN in chronic migraine [10, 45]. However, these methods are limited in their ability to resolve the millisecond-scale temporal structure of brain network dynamics [12, 13]. The DyNeMo model employed in our study overcomes this limitation by capturing the continuous mixing of coexisting networks and their long-range temporal dependencies at the native temporal resolution of MEG [12, 13]. This approach enabled us to identify that the temporal dynamics of AN, specifically its mean lifetime and switching rate, are disrupted during somatosensory stimulation, a finding that would be inaccessible to conventional sliding-window connectivity analyses that assume mutually exclusive brain states or stationary activity over extended time windows [12, 13]. Our results thus complement earlier fMRI findings by revealing the fine-grained temporal architecture underlying the network-level alterations previously described in migraine chronification [8].

Progressive changes with disease duration in CM

A key finding of the present study is the significant correlation between temporal dynamics of AN and disease duration specifically in CM patients under somatosensory stimulation. Longer duration of disease was associated with shorter mean lifetime and higher switching rate of the AN. These correlations suggest that with increasing disease chronicity, the AN showed reduced temporal stability, spending shorter periods in coherent activation and switching more frequently between activation states. This pattern may reflect a loss of neural efficiency or a maladaptive increase in cortical excitability over the course of the disease [46, 47]. Previous neuroimaging studies, including voxel-based morphometry meta-analyses, cortical thickness studies, and diffusion imaging, have shown that migraine is associated with cumulative brain changes that correlate with disease duration [48–51]. Our findings extend these observations to the temporal domain of brain network dynamics, demonstrating that the AN’s millisecond-scale temporal organization is associated with longer disease duration in CM patients. Intriguingly, this correlation was observed only in the CM group and only during somatosensory stimulation, suggesting that alterations in the AN dynamics are a feature specific to the chronic stage and is unmasked under sensory challenge conditions [52]. The absence of such correlations in the EM group may reflect either a smaller sample size or a genuinely distinct pathophysiological trajectory in which network alterations have not yet reached a threshold sufficient to correlate with clinical measures.

Clinical implications

The significant correlation between dynamics of AN and disease duration in CM patients raises the possibility that temporal dynamics metrics may serve as candidate neuroimaging markers warranting longitudinal investigation for disease monitoring. Current clinical assessment of migraine chronification relies primarily on headache frequency according to the ICHD-3 criteria [3] and patient-reported outcomes [53]; however, our findings suggest that the temporal organization of brain network activity may provide complementary information about the underlying neurophysiological state [54]. Furthermore, the identification of a brain state (somatosensory stimulation) in which the alterations of AN are most prominent may inform the design of future provocation paradigms for diagnostic or prognostic purposes. The fact that the AN was affected during median nerve stimulation rather than during an auditory task is particularly noteworthy, as it suggests that AN dysfunction in migraine represents a generalized sensory hyper-responsivity rather than a modality-specific deficit [44].

Limitations

Several limitations should be considered when interpreting these findings. First, the cross-sectional design precludes causal inferences regarding the temporal sequence of network alterations and migraine chronification. Longitudinal studies tracking the same individuals from EM to CM would be necessary to establish whether these dynamic changes precede or follow clinical progression. Second, although our sensitivity analysis controlled several clinical confounders, residual confounding from unmeasured variables (e.g., medication overuse, aura state, acute medication intake, time from last migraine attack, allodynia, headache laterality) cannot be excluded. Third, the absence of PSQI, PHQ-9, and GAD-7 data in the HC group prevented us from including these covariates in the three-group comparisons; future studies should collect these data in HC to enable full adjustment. Fourth, the somatosensory stimulation was delivered in a fixed order (left wrist followed by right wrist), which may introduce potential order effects on neural responses, although this fixed sequence was applied consistently across all participants and is therefore unlikely to have systematically biased the group comparisons, future studies should consider a randomized or counterbalanced stimulation order to exclude this potential confound. Finally, replication in independent cohorts is necessary to confirm the generalizability of these findings.

In conclusion, this study identifies distinct alterations in the temporal dynamics of VN and AN in EM and CM. Abnormalities in AN were observed during both resting state and somatosensory stimulation, while disease duration in CM was associated with altered temporal metrics of the AN. These cross-sectional findings support the involvement of altered sensory and cross-modal network processing in migraine and warrant longitudinal validation.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

aDMN

Anterior default mode network

AN

Auditory network

ANOVA

Analysis of variance

CM

Chronic migraine

DALYs

Disability-adjusted life years

DyNeMo

Dynamic network modes

ECN

Executive control network

EM

Episodic migraine

FDR

False discovery rate

GAD-7

Generalized anxiety disorder-7

HC

Healthy controls

HIT-6

Headache impact test-6

HPI

Head position indicator

ICA

Independent component analysis

ICHD-3

International Classification of Headache Disorders, 3rd edition criteria

lSMN

Left sensorimotor network

lSMN

Left sensorimotor network

MoCA

Montreal cognitive assessment

PCA

Principal component analysis

pDMN

Posterior default mode network

PHQ-9

Patient health questionnaire-9

PSQI

Pittsburgh sleep quality index

rSMN

Right sensorimotor network

SD

Standard deviation

tSSS

Spatiotemporal signal space separation

VAS

Visual analogue scale

VN

Visual network

Author contributions

CXL, DQ and PL supported the conception design of this study and analyzed the data. CXL wrote the original draft. GYL, JL, ZHX, XSL prepared figures and tables. CYD, ZW, ZG, YXZ, ZLG, XL, YXL, TSG, MTZ, JHG, BJL and YGW reviewed and edited the final draft. All authors contributed to the article and approved the submitted version.

Funding

This study was supported by Joint Funds of the National Natural Science Foundation of China (U24A20683), Beijing Natural Science Foundation (F252058) and Shenzhen Medical Research Fund (C2501004) and Changping Laboratory.

Data availability

The datasets used and analyzed are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study, a sub-study of the China HeadAche DIsorders RegiStry Study, was approved by the Institutional Review Committee of Beijing Tiantan Hospital of Capital Medical University (KY2022-044) and was registered at https://www.clinicaltrials.gov (Clinical trial number: not applicable). All participants provided informed written consent in accordance with the Declaration of Helsinki.

Consent for publication

All authors consent for the publication.

Competing interests

The authors declare no competing interests.