Neural oscillation mechanisms of repetitive subconcussive impacts: a network study of microstate-specific cross-frequency coupling

Abstract

Background

Repetitive subconcussive impacts are linked to headache pathophysiology, yet the role of electroencephalography (EEG) microstates and cross-frequency coupling in repetitive subconcussive (SC) neural alterations remains unclear. This study investigated microstate-specific cross-frequency coupling network (MCFCN) abnormalities in SC-exposed individuals to uncover mechanisms underlying headache vulnerability.

Methods

Resting-state EEG was recorded from 16 experienced male parachuters aged 18–25 years (SC group) and 16 demographically matched healthy male controls of the same age range (HC group). Microstate analysis (four canonical classes: A-D) and cross-frequency coupling (CFC) interactions (n: m) phase synchronization index, PSI) were computed to construct the MCFCNs. The MCFCNs were evaluated using machine learning (LightGBM classifier) to discriminate between the two groups, with key features identified by SHAP values and corrected for multiple comparisons.

Results

SC-exposed individuals exhibited MCFCN disruptions in large-scale networks. Notably, reduced delta/theta to alpha/beta coupling was observed in microstates A, C, and D, except for an increase in delta-band coupling from the default mode network (DMN) to the frontoparietal network (FPN) in microstate A. These alterations involved emotional-motor integration, attentional control, and self-referential processing. LightGBM models achieved significant group discrimination, with SHAP analysis highlighting theta-DMN, beta-SMN, and delta-LIM as critical nodes.

Conclusions

SC exposure induces MCFCNs abnormalities resembling central pain syndromes, even in the absence of overt symptoms. These findings suggest that interactions within MCFCNs may serve as potential early biomarkers for headache vulnerability and chronicity, warranting further exploration in longitudinal studies and targeted neurointerventions.

Introduction

Repetitive subconcussive(SC), often classified as mild traumatic brain insults without overt clinical symptoms, have emerged as an important factor in headache pathophysiology, especially within populations exposed to recurrent head accelerations. Unlike concussions, which are typically associated with acute symptoms, repetitive subconcussive impacts can induce insidious and cumulative alterations in brain structure and function that may underlie chronic headache syndromes and migraine-like phenotypes [1, 2]. Increasing evidence from both human and animal studies indicates that repetitive, non-injurious cranial accelerations result in long-term synaptic dysregulation, axonal injury, and executive dysfunction, which are hallmark features shared with primary headache disorders [3, 4].

Parachuters in contact sports such as boxing and American football are particularly susceptible to SC-related neuropathology. Epidemiological data report a prevalence of closed-head injury as high as 18.4% among military parachutists, who may experience force transmission via axial loading during landings [5, 6]. Importantly, individuals in these cohorts frequently report headache as a dominant post-impact symptom, despite the absence of overt trauma markers, positioning SC as a plausible etiology of post-traumatic headache (PTH) or migraine equivalents [6, 7].

Neuroimaging advances have facilitated the investigation of SC-induced changes in both structural and functional brain integrity. Diffusion tensor imaging (DTI) has shown white matter microstructural degradation, while resting-state functional magnetic resonance imaging(fMRI) studies reveal hyperconnectivity within default mode and visual networks, potentially indicative of compensatory reorganization [8, 9]. Electroencephalography (EEG), due to its millisecond-level temporal resolution, offers a complementary perspective by capturing the fast-scale neurophysiological dynamics underlying SC-related symptoms. For example, increased delta and theta power have been associated with fatigue and attentional deficits, whereas reductions in alpha and beta oscillations reflect diminished cortical inhibition commonly reported in migraineurs [10, 11].

However, traditional frequency-domain EEG analyses often fail to account for the interaction between neural oscillations at different frequencies. Cross-frequency coupling, encompassing phase-amplitude coupling (PAC) and cross-frequency synchronization (CFS), provides a mechanistic framework for the coordination of large-scale brain networks. Recent studies have shown that impaired CFC is not only prevalent in neurodegenerative disorders such as Alzheimer’s disease but also implicated in chronic migraine, where disrupted network integration contributes to symptom burden [12–14].Methodologically, this inter-frequency phase coupling manifests as stable phase coherence between oscillator populations operating at harmonic frequency ratios (n:m). Such synchronization dynamics are quantified through the n:m Phase Synchronization Index (PSI), which evaluates phase relationship stability by accelerating original phase time series to align instantaneous frequencies across distinct spectral components [15].Notably, conventional within-frequency phase synchronization (measured via Phase Locking Value, PLV) represents a specialized case of n:m PSI where n:m = 1:1 [16]. This establishes n:m PSI as a comprehensive metric capable of capturing multiplexed oscillatory coordination across spatiotemporally distributed networks, thereby elucidating the mechanistic foundations of cross-frequency interactions governing cortical network dynamics.Despite successful applications in mapping pathological network dynamics in major depression [13] and disorders of consciousness [17], the implementation of n:m PSI analysis remains conspicuously absent from investigations examining headache pathophysiology following repetitive subconcussive impacts. This methodological gap represents a critical limitation in understanding the neurophysiological substrates of trauma-related cephalalgia.

EEG microstate analysis, which identifies quasi-stable brain activity configurations lasting 60–120 ms, has emerged as a powerful tool for mapping spatiotemporal dynamics of intrinsic brain networks. These microstates are reproducibly categorized into four canonical classes (A–D) and are associated with functional networks such as the default mode, visual, dorsal attention, and salience networks [18]. Notably, abnormalities in microstate duration, occurrence, and transition probabilities have been reported in both migraine patients and those with post-traumatic headache, linking alterations in microstate dynamics to cortical hyperexcitability and network disintegration [14, 19].

Despite their individual utility, CFC and microstate analyses have rarely been combined to provide a holistic view of brain dysfunction in SC populations. Given the interdependence of oscillatory coordination and topographic stability, their integration could illuminate neural mechanisms underpinning post-SC headache phenotypes. Therefore, in this study, the construction of a microstate-specific cross-frequency coupling network (MCFCN) which investigated the combined role of EEG microstates and CFC in characterizing SC-induced neural alterations. This study hypothesized that individuals with repetitive SC exposure would exhibit distinct CFC signatures, reflecting network instability and altered inhibitory control, which are common features in migraine and other headache disorders. This study further examined whether these electrophysiological biomarkers could discriminate SC-exposed individuals from matched controls and correlate with headache burden.

By addressing these questions, our study aims to contribute to the mechanistic understanding of SC-induced headache and provide neurophysiological markers for early identification and stratification of at-risk individuals. Ultimately, this work supports the development of precision diagnostic tools and targeted interventions for headache disorders arising from repetitive, subclinical brain trauma.

Materials and methods

Participants

Traumatic brain injury (TBI) predominantly affects adult males aged 18–65 years [20], with young adults aged 18–23 years facing the highest risk of mild TBI (mTBI) [21]. Guided by this age distribution, the present study focused on male participants aged 18–25 years [7]. The study cohort comprised 16 military paratroopers recruited from specialized airborne units and 16 demographically matched healthy controls selected from Wuhan universities. Inclusion criteria for paratroopers required documented completion of 70 actual parachuting training exercises combined with 3,000 daily simulated platform jumps, ensuring sufficient exposure to subconcussive impacts via spinal-mediated force transmission. Both groups were matched for demographic characteristics and neurological health status. All participants were right-handed with normal or corrected-to-normal visual acuity. The study protocol was approved by the Ethics Committee of the General Hospital of Central Theater Command, and written informed consent was obtained from each participant. Detailed demographic data are presented in Table 1.

Table 1.

Summary of participants’ characteristics

Variable	HC group()	SC group ()	F	p value
Age, years (M±SD)	21.75±1.06	22.00±1.37	0.333	0.568
Education, years (M±SD)	14.50±1.03	14.44±1.21	0.025	0.876
BMI, kg/m2 (M±SD)	23.11±2.91	22.35±1.50	0.878	0.356
MMSE, median (range)	30.0(27-30)	30.0(25-30)	0.076	0.785
Number of parachuting, median (range)	——	84(48-106)
Number of platform jumps, median (range)	——	4500 (1000-4500)
PCS, median (range)	——	2(0-5)

BMI body mass index, MMSE Mini-Mental State Examination, PCS Post-Concussion Scale, HC healthy control, SC group, experienced parachuters with repetitive subconcussive

EEG recording and processing

Raw EEG signals represent a composite of neural activity, physiological artifacts, and non-physiological noise. These signals are particularly vulnerable to interference from power line noise, electromyographic (EMG) signals, and ocular movements. Such contamination may stem from environmental factors, bioelectrical interactions, or limitations inherent to EEG acquisition hardware. Effective noise reduction and artifact removal are essential prerequisites for improving the signal-to-noise ratio (SNR) and ensuring the validity of subsequent analyses [18, 22].

EEG data were recorded using a 64-channel Ag/AgCl electrode cap (eego™ amplifier, ANT Neuro Inc., The Netherlands) arranged according to the international 10–20 system, with CPz as the reference and AFz as the ground. Electrode impedance was maintained below 10 k. A notch filter at 50 Hz and an acquisition band-pass filter of 0.3–100 Hz were applied during recording. Signals were sampled at 1,000 Hz for 6 minutes, yielding 360,000 data points per participant. After artifact rejection, including removal of contaminated epochs, approximately 24,000 clean data points remained. Bad channels were interpolated using spherical spline interpolation to ensure spatial continuity.

In the present study, offline EEG preprocessing was performed using EEGLAB 2022.0 toolbox within MATLAB R2022b (MathWorks Inc., Natick, MA, USA). The initial step involved spatial electrode localization, in which a three-dimensional (3D) head model based on individual cranial morphology was utilized to determine precise electrode positioning. This approach integrated measurements of electrode coordinates and head posture, enabling accurate registration of electrodes on the scalp surface [23]. Subsequently, the raw EEG signals were band-pass filtered between 1 and 30 Hz to remove high-frequency noise and low-frequency drift artifacts. To mitigate reference bias, the signals were re-referenced to the bilateral mastoid electrodes (TP9 and TP10), which is a common method in EEG studies to enhance signal stability and comparability [24].

To further purify the data, independent component analysis (ICA) was applied, decomposing the EEG into temporally independent sources. ICA is widely acknowledged for its ability to isolate and remove artifacts such as eye blinks, muscle activity, and cardiac interference [25]. This study employed the ICLabel plugin to automatically classify and eliminate artifactual components, enhancing neural signal fidelity while preserving meaningful brain dynamics. This preprocessing pipeline is aligned with current best-practice guidelines for artifact minimization and reproducibility in EEG research [26].

The most important network interactions

A wide variety of connectivity methods are available for EEG analysis, including causal/non-causal, linear/non-linear, amplitude- or phase-sensitive, and directed or non-directed approaches. These methods perform differently when capturing brain network dynamics under various psychophysiological and pathophysiological conditions such as ADHD, dementia, and rumination [27–29]. Compared with amplitude-based metrics, phase-based measures are more robust to interindividual differences in signal amplitude and can capture transient network reconfigurations with high temporal precision. For these reasons, the present study focused on phase-based changes in connectivity.

Microstate analysis was conducted on six-minute resting-state EEG recordings from 32 participants—including healthy controls and parachuters —using the MicrostateLab toolbox. Although no universal consensus currently exists on the optimal number of microstate classes [18, 30, 31], and this number may vary depending on the dataset characteristics [32], clinical EEG studies often adopt four canonical microstates (labeled A, B, C, and D), based on early foundational work [33, 34]. These four microstates are consistently observed across diverse participant populations and electrode configurations and collectively account for 65–84% of the global variance in resting-state EEG data [18]. Accordingly, this study adopted this standard four-class solution to facilitate cross-study comparability and employed a k-means clustering algorithm to extract microstate templates that best represent the dominant spatiotemporal patterns in the data [35].

To begin with, clustering was applied at the individual level, using topographic maps extracted from peaks in the Global Field Power (GFP) curve. GFP represents the standard deviation of voltage across all electrodes at a given time point and serves as a measure of the global strength of brain electrical activity. Peaks in the GFP curve correspond to moments of maximal spatial variance and were used to define stable topographic patterns, referred to as “individual topographic maps”.These maps formed the basis for subsequent segmentation.The GFP is defined as:

1

where denotes the voltage at the i electrode, is the average voltage across all electrodes, and N represents the total number of electrodes.

Next, group-level clustering was performed by selecting a random subset of individual maps from both groups and applying k-means clustering to identify shared topographic templates. Spatial correlation was calculated between candidate templates and individual maps, and the clustering procedure was repeated iteratively until both the correlation and Global Explained Variance (GEV) values reached a plateau. GEV quantifies the proportion of total scalp topographic variance explained by a given template and is calculated as:

2

where GFP(t) denotes the Global Field Power (GFP) of single EEG topographic map at time point t; represents the spatial correlation between the topographic map at time t and the template map n; and t is the total number of EEG time points.

Finally, the derived group-level templates were fitted to each participant’s continuous EEG data to generate individualized microstate sequences. Each time point was assigned to the best-matching template based on maximal spatial correlation. To improve temporal continuity and reduce the influence of transient fluctuations, a 20 ms temporal smoothing threshold was applied: microstates shorter than 20 ms were considered noise and reassigned to neighboring states with greater spatial similarity [36].

Source reconstruction

To estimate cortical-level brain activity from scalp EEG recordings, this study employed the weighted minimum norm estimation (wMNE) method implemented in the Brainstorm software platform [37]. A three-layer boundary element method (BEM) head model was constructed using the OpenMEEG plugin (v2.4) and aligned with the ICBM152 anatomical MRI template provided by Brainstorm, without acquisition of individual MRI scans [38, 39]. The electrical conductivity parameters for the model were set as follows: scalp = 0.33 S/m, skull = 0.0066 S/m, and brain = 0.33 S/m, consistent with established modeling practices in EEG source localization [40].

The forward model was solved to compute the lead field matrix, which describes the sensitivity of each EEG electrode to activity in every source location. The inverse solution, representing cortical current density estimates, was computed using the default parameters of the wMNE approach, with dipoles constrained to be perpendicular to the cortical surface and the noise covariance matrix modeled as an identity matrix[29]. This reconstruction yielded current density estimates over a 3D cortical mesh comprising 15,002 vertices.

For regional analysis, the reconstructed cortical activity was parcellated into 100 regions of interest (ROIs) according to the Schafer2018 functional atlas, which enables biologically meaningful large-scale network interpretation [41]. Regional current densities were computed by averaging the voxel-wise estimates within each ROI. These ROIs were then grouped into seven canonical large-scale functional networks: the limbic network (LIM), default mode network (DMN), sensorimotor network (SMN), dorsal attention network (DAN), salience network (SAN), visual network (VSN), and frontoparietal network (FPN). This functional parcellation allowed for the investigation of network-specific neural dynamics and inter-network interactions relevant to cognitive control.

The dynamic network integrating microstate and cross-frequency coupling analyses

To investigate the spectral dynamics within large-scale cortical networks, EEG signals corresponding to the seven predefined functional systems were band-pass filtered into four canonical frequency bands—delta (1-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), and beta (12-30 Hz) [42, 43]—using a two-way, zero-phase-lag finite impulse response filter implemented via the pop_eegfiltnew.m function from EEGLAB [44]. We excluded the gamma band due to its high susceptibility to muscle artifacts in resting-state EEG, which can confound interpretations of neural coupling [45, 46]. Similarly, full-band coupling (e.g., across 0.3–100 Hz) was not used to avoid introducing broad-range interactions that may obscure physiologically meaningful patterns [47]. As described previously [48], these filtered time series were partitioned into temporal windows at five-second intervals.

The resting-state EEG data were acquired at a 1000 Hz sampling rate over a 6-minute recording per participant, yielding a total of 24,000 time points per subject. Each time point was labeled with a specific microstate class based on a four-state clustering template. For each participant, EEG segments corresponding to each microstate were extracted according to their individualized microstate transition sequences. Due to natural variation in microstate duration, the resulting EEG segment lengths and number of temporal windows varied across subjects.

Departing from conventional static-window approaches to network estimation, this study employed a dynamic cross-frequency coupling framework to better resolve the temporal variability of information flow. A sliding-window approach was applied to each segment, with rhythmic coupling averaged across each window to capture transient, recurring patterns of functional interaction. Window length plays a critical role in balancing temporal resolution against estimation reliability: excessively long windows may obscure brief but meaningful events, whereas overly short windows may inflate variance due to insufficient data. Prior research indicates that windows spanning several seconds are optimal for evaluating dynamic functional connectivity during rest. In this study, microstate segmentation was first performed using the Microstate-LAB plugin within EEGLAB to label EEG timepoints as belonging to one of four canonical microstate classes (A–D). Following segmentation, all time segments corresponding to the same microstate class were concatenated to form a continuous time series for each microstate. Dynamic functional connectivity was then calculated within these concatenated microstate-specific time series using a 4-second sliding window with a 2-second step size. This window length was selected based on previous studies demonstrating that 4-second intervals effectively balance temporal sensitivity and connectivity estimation reliability in EEG microstate analyses [49, 50]. Segments shorter than 4 seconds were discarded to ensure consistency across data. Accordingly, a 4-second window with 50% overlap (2-second step) was applied to each microstate time series for sliding-window functional connectivity analysis [48].

To assess the variability of network dynamics across different microstates in resting-state EEG data following repetitive subconcussive impacts, n:m PSI analyses were conducted on the large-scale network signals of each participant. The classification performance of a dynamic network model—MCFCN—was evaluated for distinguishing between two participant groups. Furthermore, critical patterns of information flow were identified under each of the four microstates. To provide a comprehensive understanding of network topology, both global and local graph-theoretical metrics were computed under each microstate for the Dynamic Network Integrating Microstate and Cross-Frequency Coupling framework.

To investigate intra- and inter-frequency coupling within and across the seven networks, phase-based measures were calculated in each time window. Instantaneous phase was computed via the Hilbert transform:

3

where PV represents the Cauchy principal value. Then both the instantaneous amplitude A(t) and the instantaneous phase can be calculated using the following equation:

4

Then n:m PSI was calculated to measure the phase synchronization of two signals and ,with the center frequency and for each frequency band, over a time window T:

5

where the PSI ranges from 0 to 1, and represents the instantaneous phase difference between the two signals, defined as:

6

In the cross-frequency case, n and m are minimum integers that need to meet [51, 52].For example, a delta rhythm at 2.5 Hz and an alpha rhythm at 10 Hz yield a 1:4 ratio, meaning that the phase of the delta oscillation must be time-warped by a factor of 4 to match the alpha rhythm. When a stable phase relationship is maintained under this transformation, the rhythms are said to be phase-synchronized at a 1:4 ratio.

Importantly, same-frequency coupling—such as alpha-alpha or beta-beta—is a special case of this model, where n=m=1. Therefore, the n:m PSI framework provides a unified method to quantify both within- and cross-frequency synchronization across spatially distributed cortical networks, offering insights into the spatiotemporal dynamics of brain rhythms.

Classifications

To investigate whether the connectivity structure of MCFCN under distinct microstate conditions could distinguish individuals exposed to SC from HC, this study extracted PSI- based functional connectivity matrices (28 27/2 = 378 features) for each of the four EEG microstates. The number of features reflects all unique pairwise connections in the 2828 functional connectivity matrix. Because the matrix is symmetric and diagonal elements (self-connections) were excluded, only the upper-triangular elements were retained as features for classification.Classification was conducted separately within each microstate, including only subjects who entered the corresponding state to ensure state-specific analysis.

A 10-fold stratified K-fold cross-validation framework was implemented using the StratifiedKFold function in scikit-learn [53] to preserve the group balance across folds. In each fold, connectivity features were standardized using z-score normalization. A Light Gradient Boosting Machine (LightGBM) classifier [54] —renowned for its high efficiency, robustness to overfitting, and superior performance on structured data—was trained on the training subset and evaluated on the corresponding test set.

Model performance was quantified using true positive rate (TPR), true negative rate (TNR), and balanced accuracy (bACC), defined as:

7

8

9

where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. Balanced accuracy was used to account for potential class imbalance between HC and SC participants.

To interpret how PSI features from MCFCNs contributed to classification decisions, SHapley Additive exPlanations(SHAP) values were calculated for each model. SHAP provides a unified approach to interpreting tree-based machine learning predictions by quantifying the marginal contribution of each feature [55]. For each outer-loop classifier, SHAP values were computed and averaged across the ten folds, enabling identification of PSI feature subsets that consistently drove group discrimination within specific microstates.

Results

Microstate maps

Following topographic clustering of resting-state EEG data from both the skydiver and healthy control groups, four representative scalp maps were identified at the group level (Fig. 1). The resulting microstate templates revealed consistent spatial topographies across both populations, supporting the notion that large-scale resting-state brain dynamics can be captured using a small number of quasi-stable configurations.

Fig. 1.

Comparison of four microstates between SC group and HC group. MsA represents the A microstate

Microstate A exhibited a diagonal orientation extending from the left frontal cortex to the right posterior region. Conversely, microstate B demonstrated an opposite pattern, extending from the right frontal cortex to the left posterior region, forming a bilateral mirror image of microstate A. Microstate C displayed a horizontal gradient from the mid-frontal to occipital areas, while microstate D was characterized by a radially symmetric configuration centered on the central scalp region, forming a circular arc-like pattern.

These topographic configurations closely resemble previously reported canonical microstate classes [33, 34], suggesting strong consistency with established spatial templates derived from HC group EEG data.

Classification performance of MCFCNs

As shown in Table 2, all four MCFCNs achieved favorable classification performance in distinguishing SC participants from healthy controls. Microstate A yielded the highest accuracy at 91.6% ± 3.0%, with a balanced sensitivity (93.1% ± 3.9%) and specificity (90.1% ± 4.5%). Microstate D also performed robustly with an accuracy of 89.5% ± 3.5%. Microstate B achieved the highest sensitivity (90.3% ± 3.9%) but had relatively lower specificity (85.9% ± 5.6%). Microstate C showed slightly lower performance across all metrics, with accuracy of 86.5% ± 3.6%.

Table 2.

Classification performance of the connectivity in MCFCNs (mean ± SD)

Microstate	Accuracy(%)	Sensitivity(%)	Specificity(%)
A	91.6% ± 3.0%	93.1% ± 3.9%	90.1% ± 4.5%
B	88.1% ± 3.2%	90.3% ± 3.9%	85.9% ± 5.6%
C	86.5% ± 3.6%	85.6% ± 5.9%	87.4% ± 4.4%
D	89.5% ± 3.5%	90.0% ± 5.3%	89.1% ± 3.9%

These results indicate that each microstate carries distinctive predictive information, with Microstates A and D providing the most balanced classification outcomes.

The most important network interactions

To further investigate the spatial-rhythmic mechanisms underlying group discrimination, this study performed feature importance ranking for each MCFCN using SHAP, a gradient-boosted decision tree–based interpretable machine learning framework. This method identifies the most influential CFC features that drive classification outcomes under each of the four canonical microstates (Fig. 2).

Fig. 2.

SHAP swarm plots illustrating the feature contributions of MCFCNs under microstates A–D. Each plot ranks the top features based on the mean absolute SHAP value (y-axis), representing their overall importance in the LightGBM classifier. The x-axis indicates the SHAP value, reflecting each feature’s impact on the model output for individual subjects. Each dot corresponds to a single subject, with color encoding the original feature value. SHAP values were computed using the SHAP (SHapley Additive exPlanations) framework

For each participant, STRN feature values—representing averaged PSI of network pairs across all temporal windows—were computed separately within each microstate. The top ten most influential edges were identified from SHAP value rankings for each MCFCN. To examine whether these top-ranked CFC links exhibited group-level statistical differences between the SC and HC groups, this study performed normality assessment using the Kolmogorov–Smirnov test, followed by Levene’s test for homogeneity of variance. Depending on test results, either independent samples t-tests or Wilcoxon rank-sum tests were used for between-group comparisons. Multiple comparisons were corrected using the Benjamini–Hochberg FDR procedure across the ten SHAP-ranked edges per microstate.

As shown in Fig. 3 a total of eleven CFC edges across the four MCFCNs exhibited significant group-level differences following correction. Among these, ten connections showed significantly stronger coupling in the HC group, while one edge showed stronger coupling in the SC group.

Fig. 3.

The statistical analysis of the top ten SHAP values for different microstates. (A–D) Violin plots illustrate feature distributions for microstates A, B, C, and D, respectively, comparing the HC and SC groups. HC group, healthy controls; SC group, experienced parachuters with repetitive subconcussive; *, **

Microstate A yielded four significant edges. Notably, delta DMN to delta FPN coupling was higher in the SC group, while three edges—including delta LIM to alpha DAN, delta DAN to beta SAN, and theta FPN to beta VN—were significantly stronger in the HC group.Microstate B showed no statistically significant MCFCN differences after correction.Microstate C yielded five edges with significantly stronger coupling in the HC group: theta DMN to beta SMN, theta SMN to delta LIM,delta FPN to alpha FPN,delta FPN to alpha LIM, and alpha SMN to beta LIM.Microstate D included two edges with HC-dominant coupling: theta DMN to beta SMN and alpha SMN to beta LIM.

Across all significant edges, theta DMN, beta SMN, beta LIM, and delta LIM emerged repeatedly, suggesting these regions play central roles in the disrupted spatiotemporal-rhythmic interactions associated with subconcussive exposure.

Discussion

After multiple comparison corrections, the characteristic edges that remained significantly different were primarily concentrated in microstates A, C, and D, these features are shown in Fig. 4, which is consistent with the microstates found to be significantly different in previous headache-related studies [14, 19].

Fig. 4.

A,B,C represent significant information interactions under MCFCNs microstate A,C,D, respectively, in both groups. Blue and red solid lines indicate significantly stronger and weaker network connections in the SC group compared to the HC group

Important interactions in the microstate A

Microstate A typically reflects posterior DMN activity and is associated with introspective attention. In subconcussive individuals, delta-band hyperconnectivity between DMN and FPN was observed, suggesting persistent engagement of self-referential processes and impaired disengagement from internally directed cognition, consistent with maladaptive rumination in chronic pain disorders such as migraine [56, 57]. Excessive DMN-FPN synchronization has also been associated with reduced pain inhibition and increased vulnerability to pain chronification [58].

In contrast, the patient group demonstrated significantly reduced delta LIM-alpha DAN and delta DAN-beta SAN couplings, indicative of impaired emotion-attention integration. The LIM-DAN axis is crucial for modulating affective responses to external stimuli, while DAN-SAN coupling is involved in salience detection and attentional shifts to nociceptive cues [59, 60]. Attenuated connectivity within these systems may reflect weakened prioritization of sensory information and diminished top-down emotional regulation, frequently reported in migraine and post-traumatic headache [61].

Additionally, decreased theta FPN-beta VSN coupling implies diminished top-down control of visual attention in pain processing, which may lead to increased distractibility and impaired spatial orientation—factors contributing to visuomotor deficits in persistent pain states [62].

Important Interactions in the microstate C

Microstate C engages fronto-central executive networks and is associated with affect regulation, salience detection, and action initiation. SC group exhibited marked reductions in theta DMN-beta SMN and theta SMN-delta LIM coupling. The former implicates failures in translating self-referential content into motor readiness, while the latter reflects deficient integration of bodily sensation with affective context. These alterations may underpin somatic dissociation and disinhibited motor activity in chronic pain conditions [63, 64].

The disruption of deltaFPN-alphaLIM coupling indicates weakened executive control over emotional reactivity. Such a pattern aligns with neuroimaging evidence linking prefrontal-limbic disconnection to affective instability and pain catastrophizing in chronic headache and fibromyalgia [47, 65]. Similarly, alpha SMN-beta LIM coupling points to impaired top-down modulation of emotional responses via sensorimotor systems [66]. Possibly contributing to flattened affective expression and impaired movement planning observed in patients with persistent pain.

The SC group showed reduced delta FPN–alpha FPN coupling during Microstate C, reflecting impaired executive network coordination. This disruption may weaken affect-cognition integration, contributing to emotional and cognitive dysregulation in chronic pain [67, 68].

Collectively, the Microstate C abnormalities signify a breakdown in the hierarchical flow of emotional and motor information necessary for pain adaptation and behavioral modulation. These findings resonate with network models of pain chronification that propose maladaptive plasticity within the salience-executive-motor triad [68].

Important Interactions in the microstate D

Microstate D, linked to anterior prefrontal and paralimbic areas, exhibited further reductions in theta DMN-beta SMN and alpha SMN-beta LIM connectivity. These repeated patterns across microstates emphasize the state-independent nature of affective-motor disintegration in SC group.

Theta DMN-beta SMN decoupling indicates impaired mapping of self-directed thought onto bodily states, disrupting anticipatory motor control—a critical mechanism in predictive coding frameworks of pain [69]. In parallel, weakened alpha SMN-beta LIM interaction may underlie insufficient emotion-based modulation of motor intent, manifesting as motor hesitancy or disinhibition in chronic pain patients [70].

These deficits may reflect a broader failure in integrative networks to generate coherent embodied representations of pain, particularly in frontal-limbic systems. Abnormalities in these systems have been associated with increased interoceptive noise and misattribution of bodily signals, as seen in migraine aura, fibromyalgia, and functional somatic disorders [71, 72].

Across all microstates, this study observed a pervasive collapse in delta/theta to alpha/beta phase-amplitude coupling between core pain-related networks. These findings point to a failure in coordinating bottom-up sensory input with top-down regulatory mechanisms—an electrophysiological hallmark of central sensitization [73]. Slow-wave phase oscillations are typically involved in large-scale communication and long-range inhibitory control, while alpha and beta amplitude modulations are central to local information encoding and cortical readiness [74]. Disruption in this hierarchical phase-amplitude structure likely compromises temporal coordination of pain processing.

Interestingly, while most couplings were reduced in patients, the deltaDMN-deltaFPN interaction was significantly elevated. This may reflect pathological over-synchronization in slow-wave activity, often interpreted as a compensatory but maladaptive response to impaired higher-frequency modulation. Similar patterns have been described in patients with thalamocortical dysrhythmia and neurogenic pain, where increased delta coherence coexists with alpha suppression and is linked to tonic pain states and disinhibition of pain networks [75, 76].

Moreover, the specificity of disrupted couplings—for instance, thetaFPN-betaVSN and deltaDAN-betaSAN—suggests a targeted breakdown in the executive-visual and attentional-salience axes, which are crucial for detecting and localizing pain stimuli. These results are consistent with behavioral reports of visual hypersensitivity, attentional bias, and impaired threat discrimination in migraine sufferers [77].

This study has several limitations that should be acknowledged. First, the relatively small sample size () limits the statistical power and generalizability of our findings. Future studies with larger and more diverse cohorts are needed to validate the robustness of our results, especially considering the high-dimensional nature of the MCFCN parameters and the machine learning approaches employed.In preliminary analyses, we examined conventional microstate temporal metrics—mean duration, coverage, and occurrence rate—to compare SC and HC groups. Most showed no significant differences and were therefore not emphasized. While scalp- or source-level microstate analyses remain valuable [78], our study focused on microstate-specific cross-frequency coupling to reveal subtle network changes linked to subconcussive impacts.Second, although SHAP analysis provided valuable insights into the most influential features, the biological relevance and specificity of these computationally derived features remain to be confirmed. Integrating multimodal neuroimaging techniques, such as simultaneous fMRI-EEG recordings in symptomatic populations, will be essential to bridge the gap between computational models and underlying neurobiological mechanisms.

Conclusion

Our findings reveal microstate- and network-specific disruptions in cross-frequency coupling among subconcussive individuals that mirror core features of central pain syndromes. The observed alterations span emotional, attentional, motor, and self-referential systems, reinforcing a model of pain as a multisystem network disorder. Importantly, these electrophysiological abnormalities occurred in the absence of overt clinical symptoms, suggesting that MCFCNs may serve as early biomarkers for pain vulnerability and chronicity.

Future longitudinal studies should explore the predictive value of these MCFCNs patterns for symptom development and treatment response. In addition, targeted interventions such as neurofeedback, noninvasive brain stimulation, or cognitive-emotional training may be designed to restore disrupted oscillatory dynamics and improve clinical outcomes in repetitive subconcussive and pain-prone populations.

Authors’ contributions

All authors carried out the studies. And Yida He drafted the manuscript, participated in the design of the study and performed the statistical analysis.

Funding

Supported by the Fundamental Research Funds for the Central Universities of South-Central Minzu University (GrantNumber:CZZ24015,CZZ24014,CZZ25009), in part by the National Natural Science Foundation of China under Grant 81601586.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This work involved human subjects or animals in its research. Approval of all ethical and experimental procedures and protocols was granted by the Ethics Committee of the General Hospital of Central Theater Command (Followed the World Medical Association’s Declaration of Helsinki for medical research involving humans.).

Consent for publication

All authors read and approved the final manuscript.

Competing interests

The authors declare no competing interests.