Dual-attention temporal graph neural network on resting-state fMRI dynamic functional connectivity identifies risk-related patterns of migraine chronification

Abstract

Background

Migraine chronification remains difficult to characterize at the individual level, and robust baseline imaging markers associated with subsequent conversion from episodic migraine (EM) to chronic migraine (CM) are lacking. Resting-state functional MRI (rs-fMRI) dynamic functional connectivity (dFC) captures transient brain network interactions, but conventional approaches provide limited interpretability for clinical translation.

Methods

We enrolled 195 participants at baseline, including 95 healthy controls (HC) and 100 patients with episodic migraine at baseline. All migraine patients underwent baseline rs-fMRI during the interictal phase and were followed for 12 months; 70 remained EM (non-converters) and 30 converted to CM. Sliding-window dFC matrices (55-TR window, 2-TR step) were constructed from 142 regions of interest and represented as temporal graph sequences. We developed an interpretable temporal graph neural network integrating a two-layer graph isomorphism network, a gated recurrent unit, and dual spatial–temporal attention mechanisms to quantify node-level (“where”) and window-level (“when”) importance. A multilayer perceptron performed three-class discrimination among HC, EM non-converters, and CM converters. Model performance was evaluated using stratified ten-fold cross-validation. Edge-wise ANCOVA controlling for age, sex, and mean framewise displacement with false discovery rate correction was applied to dFC matrices at key temporal-attention windows and within k-means-derived dFC states.

Results

Four recurrent dFC states were identified. State 3 showed the highest fractional occupancy and mean dwell time and was the only state with significant between-group edge-wise differences after correction. Compared with HC, migraine groups exhibited altered connectivity involving sensory, attention, default-mode, and subcortical systems. CM converters showed additional baseline abnormalities involving sensorimotor–visual, default-mode, ventral attention, and limbic-related circuits. Spatial attention shifted from occipital–frontoparietal hubs in HC to frontal–insular prominence in EM non-converters and a frontoparietal–limbic configuration in CM converters. Temporal attention peaked at window 23, where chronification-related abnormalities were most evident. The model achieved AUCs of 0.817(HC), 0.832(EM non-converters), and 0.874(CM converters), with a macro-AUC of 0.841 (95% CI 0.77–0.91). Findings were robust across alternative window lengths and step sizes.

Conclusions

This rs-fMRI dFC–based temporal graph learning framework identified baseline spatiotemporal network patterns associated with subsequent migraine chronification and generated individualized time-resolved importance maps. Although the model showed promising internal discriminative performance, these findings should be interpreted as risk-related pattern identification rather than as evidence of a clinically validated prediction tool. External prospective multicenter validation is required before clinical translation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02567-x.

Introduction

Migraine is a prevalent neurological disorder characterized by recurrent moderate-to-severe headaches accompanied by nausea, photophobia, and phonophobia. Affecting approximately 14% of the global population, migraine imposes substantial personal and societal burdens and is more common among women and working-age adults [1–3]. Beyond pain, patients frequently experience psychiatric and cognitive comorbidities, underscoring widespread functional disturbances across large-scale brain networks.

A major clinical challenge is the transition from episodic migraine (EM) to chronic migraine (CM), defined as headache on ≥ 15 days per month for more than three months, with ≥ 8 days meeting migraine criteria [4]. Approximately 2.5–3% of patients with EM convert to CM annually [5]. Chronification is associated with greater disability, medication overuse, psychiatric comorbidity, and reduced therapeutic responsiveness [6]. Early identification of individuals at high risk of chronification is therefore crucial.

Evidence suggests that central sensitization, involving dysfunctional pain modulation within trigeminal, thalamic, limbic, and brainstem circuits, contributes to chronification [7–11]. However, most insights have been derived from group-level analyses or studies of established chronic migraine. There remains a need to identify baseline imaging patterns associated with subsequent chronification using a single resting-state fMRI scan.

Static functional connectivity (FC) has revealed abnormalities across the default mode network (DMN), salience (SN), frontoparietal (FPN), and sensorimotor (SMN) networks in migraine [12–15]. Dynamic FC (dFC), typically derived using sliding-window correlation and k-means clustering, better captures temporal variability [16–19]. However, conventional dFC approaches face limitations including fixed window parameters, hard state assignment, and limited interpretability at the individual level [20–24].

Deep learning, and particularly graph neural networks (GNNs), offers a flexible framework to model the brain as a dynamic graph with nonlinear dependencies [23, 24]. In this study, we propose a temporal GNN incorporating dual spatial–temporal attention to analyze full-length dFC sequences from a single baseline rs-fMRI. This model preserves fine-grained temporal information, identifies informative windows (“when”) and high-importance brain regions (“where”), and characterizes dynamic connectivity patterns associated with longitudinal chronification outcome among patients with episodic migraine at baseline.

We hypothesized that: (1) converters exhibit features associated with later chronification at baseline; (2) temporal attention identifies brief critical windows with characteristic connectivity; and (3) spatial attention reveals core hubs implicated in central sensitization.

Methods

Participants

The study protocol was approved by Ethics Committee of Shuguang Hospital affiliated to Shanghai University of Traditional Chinese Medicine Clinical Trial Registry Platform (World Health Organization under registration number: 2019-766-121-01) and conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent. Participants were retrospectively selected from a pre-existing large-scale migraine neuroimaging database based on predefined criteria, including complete baseline clinical and MRI data, acceptable imaging quality, and complete 12-month follow-up information. The final analytic sample comprised 195 participants: 95 healthy controls and 100 patients with episodic migraine at baseline. Episodic migraine was diagnosed according to the International Classification of Headache Disorders, 3rd edition (ICHD-3) criteria [4], and no patient met criteria for chronic migraine at baseline. All patients underwent longitudinal follow-up over 12 months with assessments every 3 months, at which headache frequency, monthly migraine days, and medication use were recorded. Conversion to chronic migraine was defined per ICHD-3 as headache on ≥ 15 days/month for more than 3 months, with migraine features on ≥ 8 days/month. All participants were aged 18–55, right-handed, and free of neurological or psychiatric disorders. Exclusion criteria included structural brain lesions, medication overuse, and MRI contraindications.

MRI acquisition

All migraine patients underwent rs-fMRI scanning during the interictal phase. No patient reported an ongoing migraine attack or migraine-related symptoms at the time of scanning.

MRI scans were performed using a United Imaging uMR 790 3.0-T superconducting scanner (United Imaging, Shanghai, China) equipped with a 32-channel rigid head coil. Participants were positioned supine and instructed to remain awake with their eyes closed, relax, and minimize head motion during scanning.

High-resolution structural images were acquired using a three-dimensional T1-weighted imaging sequence with the following parameters: repetition time (TR) = 7.2 ms, echo time (TE) = 3.1 ms, flip angle (FA) = 10°, field of view (FOV) = 256 × 256 mm2, matrix size = 512 × 512, slice thickness = 1 mm, and 192 slices.

Resting-state functional MRI data were acquired using a gradient-echo echo-planar imaging sequence with the following parameters: TR = 2000 ms, TE = 30 ms, FA = 90°, FOV = 224 × 224 mm², matrix size = 64 × 64, slice thickness = 3.5 mm, number of slices = 33, voxel size = 3.5 × 3.5 × 3.5 mm³, and 240 volumes. The total resting-state scan duration was approximately 8 min. Participants with head motion exceeding 2 mm translation or 2° rotation were excluded. Mean framewise displacement was calculated and included as a covariate in subsequent analyses.

Clinical assessment

At baseline, participants completed the Zung Depression Scale (SDS), Headache Impact Test-6 (HIT-6), and Migraine Disability Assessment (MIDAS) questionnaires. Pain intensity was assessed using the visual analog scale (VAS) to quantify the average headache pain intensity during migraine attacks over the preceding month. Medication use was systematically documented for all participants.

Preprocessing

Functional MRI data were preprocessed using DPABI implemented in MATLAB. The first 10 volumes were discarded to allow for signal equilibration and participant adaptation to the scanning environment. The remaining images underwent slice-timing correction and realignment for head-motion correction. Participants with head motion exceeding 2 mm translation or 2° rotation were excluded from further analysis. Mean framewise displacement (FD) was additionally calculated for each participant and included as a covariate in subsequent statistical analyses. No additional scrubbing or censoring procedure was applied.

Structural and functional images were spatially normalized to the Montreal Neurological Institute (MNI) standard space using unified segmentation-based normalization implemented in SPM, and resampled to 3 × 3 × 3 mm³ isotropic resolution. Spatial smoothing was performed using a 6-mm full-width at half-maximum Gaussian kernel. Linear detrending and temporal band-pass filtering (0.01–0.08 Hz) were subsequently applied. Several nuisance covariates, including Friston-24 head-motion parameters, white matter signals, and cerebrospinal fluid signals, were regressed out. Global signal regression was not performed.

Regional mean BOLD time series were extracted from 142 regions of interest (ROIs) using a DPABI-distributed Schaefer-based atlas in MNI space. Specifically, the atlas comprised 100 cortical parcels derived from the Schaefer functional parcellation, supplemented by 42 subcortical regions provided in DPABI [25]. The cortical parcels were assigned to canonical intrinsic functional networks, including the visual network (VN), somatomotor network (SMN), dorsal attention network (DAN), ventral attention/salience network (VAN), frontoparietal network (FPN), and default mode network (DMN). The subcortical component included regions such as the thalamus, basal ganglia, hippocampus, amygdala, and related deep gray-matter structures. Mean BOLD signals were extracted from all ROIs for subsequent dynamic functional connectivity analysis.

Dynamic functional connectivity was estimated using a sliding-window correlation approach. Within each temporal window, pairwise functional connectivity between regional BOLD time series was calculated using Pearson correlation coefficients. Correlation coefficients were subsequently Fisher z-transformed to improve normality and variance stabilization prior to statistical analysis and graph construction. The Fisher z-transformed Pearson correlation coefficient was used as the edge weight in each dynamic functional connectivity matrix.

The window length was set to 55 TRs, corresponding to 110 s given the TR of 2 s. This window length satisfies the criterion proposed by Leonardi and Van De Ville [26] that the window duration should be longer than the reciprocal of the lowest frequency of interest. Because the resting-state fMRI time series were band-pass filtered at 0.01–0.08 Hz, the lowest frequency of interest was 0.01 Hz, corresponding to a minimum window duration of 1/0.01 Hz = 100 s. Thus, the 110-s window was selected to reduce spurious connectivity fluctuations introduced by the sliding-window procedure [27]. The window was shifted with a step size of 2 TRs, corresponding to 4 s; a relatively small step size was used to improve the temporal sampling of brain-state transitions [28].

Dynamic functional connectivity state clustering

To identify recurrent dFC states, all sliding-window dFC matrices across participants and time windows were pooled and submitted to k-means clustering. Prior to clustering, each Fisher z-transformed dFC matrix was vectorized using the upper triangular elements (excluding the diagonal). The number of clusters was set to k = 4 based on clustering validity indices and the interpretability of the resulting state patterns. Each cluster centroid was reconstructed into a 142 × 142 connectivity matrix representing a group-level recurrent dFC state. For each participant, every sliding window was assigned to its nearest centroid, enabling calculation of state-level temporal metrics: fractional occupancy, mean dwell time, and transition probability.

Cluster centroids served as descriptive templates of recurrent dFC configurations and were normalized relative to the mean network activity level for visualization purposes. No edge-wise statistical inference was performed at the state-identification stage; statistical testing was restricted to subsequent between-group edge-wise comparisons with covariate adjustment and false discovery rate correction.

Temporal graph neural network architecture

Each sliding-window dFC matrix was represented as a weighted graph G_t = (V, E_t, A_t), where V denotes the 142 brain regions, E_t the functional connections, and A_t the Fisher z-transformed Pearson correlation matrix at window t. For each participant, the full.

Graph sequence {G₁, G2, …, Gₜ} served as model input.

The temporal GNN was adapted from the dynamic–static fusion framework of Zhao and Zhang [29], retaining the dynamic graph branch to model time-varying rs-fMRI connectivity. The architecture comprised four components: (1) a graph isomorphism network (GIN) for window-wise spatial representation learning via neighborhood aggregation; (2) a gated recurrent unit (GRU) for modeling temporal dependencies across consecutive windows; (3) dual spatial–temporal attention modules for interpretability; and (4) a multilayer perceptron (MLP) classifier with softmax output for three-class discrimination among healthy controls, EM non-converters, and CM converters. The spatial attention module assigned importance weights to brain regions, identifying nodes most discriminative for group classification, while the temporal attention module identified the most informative sliding windows within the resting-state scan.

Converter status was determined by 12-month follow-up outcomes; accordingly, the model characterized baseline imaging patterns associated with subsequent chronification rather than serving as a prospective prognostic tool.

Model training and evaluation used stratified 10-fold cross-validation, preserving class distribution across folds. The Adam optimizer and class-weighted cross-entropy loss were applied to address class imbalance. Final performance metrics were derived from pooled out-of-fold predictions across all ten folds. A schematic overview of the model architecture is shown in Fig. 1.

Fig. 1.

Schematic of the temporal graph neural network pipeline for migraine chronification risk prediction from baseline rs-fMRI. Resting-state fMRI time series were extracted from 142 ROIs using the DPABI-distributed Schaefer-based 142-ROI atlas. Sliding-window dynamic functional connectivity (dFC) matrices were computed (55-TR window, 2-TR step) and represented as weighted graphs . Each participant’s graph sequence was input into a temporal GNN comprising a two-layer graph isomorphism network (GIN), a gated recurrent unit (GRU), dual spatial–temporal attention modules, and an MLP classifier for three-class discrimination. The model was trained with stratified 10-fold cross-validation, the Adam optimizer, and class-weighted loss. Performance was assessed using one-vs-rest AUCs, macro-AUC, sensitivity, specificity, precision, and F1-score (95% CIs via bootstrapping). Dual attention mechanisms generated individualized importance maps. Group-level validation used edge-wise ANCOVA (controlling for age, sex, and mean framewise displacement) with FDR correction on dFC matrices at key temporal-attention windows and the dominant k-means-derived dFC state (State 3). This pipeline enables individualized risk stratification for episodic migraine chronification with temporally and spatially resolved interpretability

Cross-validation workflow and leakage control

Stratified ten-fold cross-validation was performed at the participant level. For each fold, participants in the test set were held out before model training and were not used for parameter updating, class-weight estimation, normalization, hyperparameter selection, or any procedure contributing to the classifier. The temporal graph neural network received each participant’s baseline sliding-window dFC graph sequence as input, and model training was performed only on the training folds. Performance metrics were computed exclusively from out-of-fold predictions and then pooled across the ten folds. No feature selection was conducted using the full dataset before cross-validation. Hyperparameters were predefined and kept fixed across folds; no test-fold information was used for hyperparameter optimization. The k-means dFC state clustering analysis was conducted for descriptive characterization of recurrent connectivity states and was not used as input to the classification model.

Statistical analysis

Demographic and clinical variables were analyzed according to data type and distribution. Normally distributed continuous variables were compared across groups using one-way ANOVA; non-normally distributed variables were compared using the Kruskal-Wallis test. Categorical variables were analyzed with the chi-square test or Fisher’s exact test as appropriate, with post hoc pairwise comparisons corrected for multiplicity.

Edge-wise group differences in dFC—including all-window and temporal-attention key-window analyses-were assessed using ANCOVA with age, sex, and mean framewise displacement as covariates, followed by false discovery rate (FDR) correction across edges. For state-specific analyses, the same ANCOVA framework was applied within each of the four dFC states across three pairwise contrasts (HC vs. EM non-converters, HC vs. CM converters, and EM non-converters vs. CM converters), with FDR correction applied jointly across states, contrasts, and all ROI-to-ROI connections. Cluster centroids were used for descriptive characterization only and were not subjected to formal edge-wise inference. Associations between imaging-derived measures and clinical variables were evaluated using Pearson or Spearman correlation according to data distribution.

Robustness analysis

To assess sensitivity to sliding-window parameter selection, the full analysis pipeline was repeated using window lengths of 45, 55 (primary), and 65 TRs, and step sizes of 1, 2 (primary), and 3 TRs, with all other procedures held constant. Classification performance and spatial–temporal attention patterns were compared across configurations.

Model performance evaluation

Model performance was evaluated using stratified 10-fold cross-validation at the participant level to address class imbalance among healthy controls (HC, n = 95), episodic migraine non-converters (EM, n = 70), and chronic migraine converters (CM, n = 30). For each fold, the temporal graph neural network—comprising GIN layers, GRU, dual spatial–temporal attention mechanisms, and an MLP classifier—was trained on sliding-window dynamic functional connectivity graph sequences derived from baseline rs-fMRI.

Key performance metrics were computed as follows: (1) One-vs-rest AUCs and macro-averaged AUC assessed discriminative ability for each class and overall performance. (2) 95% confidence intervals (CIs) for AUC values were derived via bootstrapping (1000 iterations) to quantify variability. (3) Sensitivity (recall), specificity, precision, and F1-score for each class were calculated from the confusion matrix, using aggregated out-of-fold predictions across all folds. (4) Macro-averages summarized performance across classes with equal weighting. All metrics were implemented in Python with scikit-learn and PyTorch libraries. Class weighting was applied during training with the Adam optimizer and weighted cross-entropy loss to mitigate imbalance.

Results

Participant characteristics

A total of 195 participants were included: 95 healthy controls and 100 patients with episodic migraine at baseline. All migraine patients underwent rs-fMRI at baseline; none met criteria for chronic migraine at the time of scanning. Following 12 months of longitudinal follow-up, 30 patients fulfilled ICHD-3 criteria for chronic migraine (CM converters) and 70 remained episodic throughout (EM non-converters).

No significant between-group differences were observed in age or sex distribution (p > 0.05). Among the migraine subgroups, baseline disease duration, monthly headache days, pain intensity (VAS), HIT-6, MIDAS, and SDS scores were comparable between EM non-converters and CM converters (p > 0.05), indicating well-matched clinical characteristics at baseline (Table 1).

Table 1.

Demographic and clinical characteristics of the study participants

Variable	HC (n = 95)	EM (n = 70)	CM (n = 30)	Statistic	p-value
Age (years)	28.1 ± 6.2	27.9 ± 5.8	28.6 ± 6.0	F = 0.18	0.84
Sex (M/F)	46/49	32/38	14/16	χ²=0.05	0.97
Education (years)	15.8 ± 4.3	16.2 ± 4.5	15.6 ± 4.2	F = 0.26	0.77
Disease duration (years)	—	6.5 ± 4.9	6.7 ± 5.0	t=-0.18	0.86
SDS (score)	48.9 ± 6.1	49.6 ± 6.4	50.2 ± 6.2	F = 0.45	0.64
VAS (score)	—	5.5 ± 1.7	5.7 ± 1.9	t=-0.49	0.63
HIT-6 (score)	—	52.1 ± 9.0	52.8 ± 9.3	t=-0.31	0.76
MIDAS (score)	—	18.6 ± 9.2	19.3 ± 9.5	t=-0.35	0.73
Monthly headache days	—	5.16 ± 2.44	5.77 ± 2.40	t = − 1.14	0.25
Acute medication use within the past month, n (%)	—	58 (82.9%)	25 (83.3%)	χ² = 0.003	0.95
Acute medication days within the past month (days)	—	6.2 ± 1.5	6.5 ± 1.6	t = 0.96	0.33
Preventive medication use, n (%)	—	32 (45.7%)	14 (46.7%)	χ² = 0.009	0.92

Note: Values are presented as mean ± standard deviation (SD) unless otherwise indicated. Categorical variables are presented as number (percentage). Group comparisons among healthy controls (HC), episodic migraine non-converters (EM), and chronic migraine converters (CM) were performed using one-way analysis of variance (ANOVA) or the chi-square test, as appropriate. Variables available only in the migraine groups (disease duration, VAS, HIT-6, MIDAS, monthly headache days, acute medication use, acute medication days, and preventive medication use) were compared between EM and CM using the independent-samples t test or chi-square test, as appropriate. SDS, Zung Self-Rating Depression Scale; VAS, Visual Analog Scale; HIT-6, Headache Impact Test-6; MIDAS, Migraine Disability Assessment. “—” indicates not applicable

Acute medication use in the past month was comparable between EM non-converters and CM converters (82.9% vs. 83.3%, p = 0.953), as were acute medication days (6.2 ± 1.5 vs. 6.5 ± 1.6 days/month, p = 0.339). Preventive medication use was likewise comparable between the two groups (45.7% vs. 46.7%, p = 0.924). Medication overuse was an explicit exclusion criterion; therefore, no included participant met the ICHD-3 criteria for medication overuse. These clinical variables, along with monthly headache days, are summarized in Table 1.

Descriptive characterization of dynamic functional connectivity states and spatial distributions

K-means clustering of pooled sliding-window dFC matrices identified four recurrent group-level dFC states. Cluster centroids were normalized relative to mean network activity and interpreted as descriptive representations of recurrent dFC configurations in the full sample, rather than group-specific statistical contrast maps (Fig. 2).

Fig. 2.

Centroid matrices of the four dynamic functional connectivity (dFC) states. Each panel displays the centroid matrix of one recurrent dynamic functional connectivity state identified by k-means clustering. Rows and columns correspond to the 142 brain regions from the DPABI-distributed Schaefer-based 142-ROI template and are ordered according to the template-defined network labels, including the visual network (VN), somatomotor network (SMN), dorsal attention network (DAN), ventral attention/salience network (VAN), subcortical network (SC), frontoparietal control network (FPN), and default mode network (DMN).The centroid matrices represent descriptive cluster-average dFC patterns and were not statistically thresholded

The four states exhibited distinct network participation profiles. State 1 was dominated by higher frontoparietal–DMN connectivity alongside lower visual and somatomotor coupling, reflecting a higher-order control–DMN configuration. State 2 showed the inverse pattern, with prominent visual–somatomotor coupling and reduced frontoparietal–DMN connectivity, reflecting a sensory–motor dominant configuration. State 3 exhibited higher DMN–VAN connectivity with lower visual, somatomotor, and frontoparietal coupling, representing a DMN–salience dominant configuration. State 4 was characterized by VAN-dominant connectivity with lower visual and frontoparietal coupling. Formal statistical inference was not applied to centroid matrices; edge-wise testing with covariate adjustment and FDR correction was reserved for subsequent group-comparison analyses (Fig. 3).

Fig. 3.

Network-level spatial distribution of the four k-means-derived dynamic functional connectivity (dFC) states. The centroid matrix of each recurrent state was summarized according to seven large-scale brain networks, including visual network (VN), somatomotor (SMN), dorsal attention (DAN), ventral attention/salience (VAN), subcortical network (SC), frontoparietal control (FPN), and default mode (DMN). This figure provides a descriptive visualization of the network-level organization of the recurrent dFC states. a: Short color bar (below each small brain rendering): Normalized strength of each brain region within the state (blue = weak activity, red = strong activity, white = average activity). b: Normalized strength (Fisher z-transformed Pearson correlation). Value 0 represents average baseline activity of the seven networks; positive values indicate stronger coupling (above baseline), and negative values indicate weaker coupling (below baseline). c: Network-level spatial distribution: Four representative brain renderings (one per state) illustrate the spatial pattern of each dFC state, with contributions summarized on the color bar

Descriptive temporal characteristics of dynamic functional connectivity states

Using the sliding-window analysis combined with k-means clustering (k = 4), we identified four recurrent dFC states across participants. Their aggregate properties were as follows:

Fraction of windows All four states were observed at the group level. State 3 occurred most frequently (~ 40% of windows), whereas States 2 and 4 showed relatively lower occupancy (Fig. 4A).

Fig. 4.

Characteristics of the four recurrent dFC states identified by k-means clustering. (A) Mean fractional occupancy of the each state (mean ± SE). (B) Mean dwell time per state (TRs; mean ± SE). (C) Row-normalized state transition probability matrix (diagonal elements indicate self-transition probabilities). The high diagonal values (State 1–1: 0.91; State 2–2: 0.96; State 3–3: 0.92; State 4–4: 0.95) indicate strong within-state stability, while off-diagonal probabilities remain low (< 0.06), confirming infrequent transitions between states

Mean dwell time Consistent with the occupancy results, State 3 exhibited the longest mean dwell time (~ 34 TRs), suggesting greater temporal stability during rest; States 2 and 4 had the shortest dwell times (Fig. 4B).

Transition probability The state transition matrix revealed the highest probabilities along the self-transition (diagonal) elements, indicating strong within-state stability. Cross-state transitions were generally infrequent, with only a subset of state pairs showing notable transitions. The high diagonal values (State 1–1: 0.91; State 2–2: 0.96; State 3–3: 0.92; State 4–4: 0.95) indicate strong within-state stability, while off-diagonal probabilities remain low (< 0.06), confirming infrequent transitions between states (Fig. 4C).

State-specific group differences in dynamic functional connectivity

Edge-wise group comparisons were performed across all four dFC states. Significant between-group differences survived FDR correction exclusively in State 3; no edges in States 1, 2, or 4 reached significance. Results are therefore reported for State 3 only (FDR q < 0.05; Fig. 5).

Fig. 5.

Group differences in functional connectivity within State 3. (A) HC vs. EM non-converters; (B) HC vs. CM converters; (C) EM non-converters vs. CM converters; Colored lines indicate connections showing between-group differences at the State 3; line color and thickness encode the direction (increase/decrease in the first group relative to the second) and the magnitude of the effect. Acronyms: HC = healthy controls; EM = episodic migraine; CM = chronic migraine. All edge-wise comparisons were adjusted for multiple testing (e.g., FDR) and controlled for age, sex, and mean framewise displacement (FD). Acronyms: VN = visual network, SMN = sensorimotor network, VAN = ventral attention/salience, FPN = frontoparietal control, DMN = default mode

HC vs. EM non-converters Relative to HC, EM non-converters exhibited altered connectivity in post-occipital regions within the VN, VN–SMN (post-occipital–precentral gyrus), intra-VAN (frontal region), DMN–DAN (vmPFC–TPJ), DMN–SC (vmPFC-thalamus), and intra-SC (thalamus).

HC vs. CM converters. Relative to HC, CM converters showed altered connectivity in intra-DMN (inferior temporal–mPFC), VN–SC (post-occipital–basal ganglia), and SC–DAN (thalamus-occipital) pathways.

EM non-converters vs. CM converters. CM converters exhibited altered SMN–DMN connectivity between the SMA and posterior cingulate cortex relative to EM non-converters.

Collectively, State 3 group differences comprised region-specific intra- and cross-network connectivity alterations spanning visual, sensorimotor, attentional, default-mode, and subcortical systems.

Key brain regions identified by the spatial-attention mechanism

In the spatial (“where”) readout, the graph neural network (GNN) identified the Top-10 high-importance nodes for each group (Fig. 6).

Fig. 6.

Top-10 high-importance brain regions identified by the spatial-attention mechanism. (A) HC, (B) EM (non-converters), (C) CM (converters). The color bar encodes spatial attention scores (blue → red; red = higher importance). HC shows a visual–frontoparietal hub pattern; EM exhibits up-weighted frontal and insular nodes; CM further shifts toward a frontal–insular configuration, indicating a progressive relocation of network emphasis along the path to chronification

HC group

High-importance nodes were predominantly located in the bilateral superior occipital gyrus (OccSupM), temporal cortex (Temp), precuneus (Precun), middle occipital gyrus (OccMid), right inferior occipital gyrus (OccInfR), and fusiform gyrus (Fusif), with several frontal nodes (dlPFC, vFC) and the insula also ranking among the top ten. Overall, this reflected a VN–frontoparietal hub–dominated spatial pattern (Fig. 6A).

EM group (non-converters)

While occipital and precuneus regions remained among the top nodes, frontal (right dlPFC, vFC) and insular regions showed markedly elevated importance scores, indicating that a VN–frontal–insular pathway played a more prominent role in dynamic discrimination (Fig. 6B).

CM group (converters)

Spatial attention further concentrated on the frontal–insular system. The right dlPFC, vFC, precuneus, right mid-insula (MidInsR), and fusiform gyrus all exhibited substantially increased importance, with right dlPFC and precuneus achieving the highest scores. This constellation forms a “frontoparietal–limbic loop” pattern, highlighting the central involvement of cognitive control and pain-modulation circuits in the progression toward chronification (Fig. 6C).

In sum, spatial attention revealed a gradient from visual hubs (HC) to frontal–insular emphasis (EM) and ultimately to a frontoparietal–limbic configuration (CM), consistent with progressive engagement of control and salience/affective systems along the path of migraine chronification.

Key moments revealed by the temporal-attention mechanism

Temporal attention scores were computed across all 87 sliding windows derived from the 230-volume rs-fMRI sequence (55-TR window, 2-TR step). The group-level attention curve showed sparse weighting across most windows, with the highest peak at window 23-the time segment in which the model detected the strongest discriminative information at the group level. This peak reflects a model-identified informative moment rather than an externally defined physiological event. To characterize the connectivity pattern at this window, dFC matrices at window 23 were extracted from all participants and submitted to edge-wise ANCOVA (covariates: age, sex, mean framewise displacement) with FDR correction (Fig. 7).

Fig. 7.

Group-level temporal-attention profile across 87 sliding windows. Temporal-attention scores were averaged across participants to visualize the relative importance of each sliding window for model discrimination. The rs-fMRI scan contained 230 time points and was divided into 87 windows using a 55-TR window length and 2-TR step size. The highest group-level attention score was observed at window 23, indicating a model-highlighted informative segment of the scan. The dFC matrices from this window were subsequently extracted from all participants for edge-wise group comparison

HC vs. EM non-converters EM non-converters exhibited altered connectivity within the VN (occipital regions) and across VN–DAN and VN–FPN pathways. Additional alterations involved SMN–VAN connections (posterior insula, pre-SMA, mid-insula) and multiple VAN–FPN edges (vPFC, vFC, dACC, mid-insula with dlPFC). FPN–DMN alterations included aPFC–vlPFC, dlPFC–vmPFC, and dlPFC–posterior cingulate/angular gyrus connections, with intra-SC thalamic involvement (Fig. 8A).

Fig. 8.

Group differences in functional connectivity at the temporal-attention key moment. (A) HC vs. EM; (B) HC vs. CM; (C) EM vs. CM. The graphs depict significant edges at each participant’s key window (i.e., the time window with the highest temporal attention weight). Colored lines indicate connections showing between-group differences at the key moment; line color and thickness encode the direction (increase/decrease in the first group relative to the second) and the magnitude of the effect. Acronyms: HC = healthy controls; EM = episodic migraine; CM = chronic migraine

HC vs. CM converters CM converters showed altered intra-VN connectivity (occipital and post-occipital regions), SMN–DMN coupling (SMA–vlPFC), and SMN–VAN connections (parietal–mid-insula, temporal–basal ganglia). DAN–SMN/VAN alterations included precentral gyrus–temporal and precentral gyrus–anterior insula edges. Intra-DMN alterations involved posterior cingulate–precuneus connectivity, with additional DMN-related edges at vFC-occipital and vlPFC-SMA (Fig. 8B).

EM non-converters vs. CM converters CM converters showed altered VN–SMN connectivity (post-occipital–SMA) and VN–DMN connections (post-occipital–posterior cingulate, temporal–posterior cingulate). Additional alterations involved intra-DAN/VN edges and VAN–FPN interactions (aPFC–ventral aPFC), along with VAN–DMN and FPN–DMN alterations including parietal–IPS and ACC–vmPFC connections (Fig. 8C).

Collectively, key-moment group differences comprised specific ROI-to-ROI alterations rather than diffuse network changes. EM non-converters showed predominant VAN–FPN and FPN–DMN involvement, whereas CM converters exhibited broader alterations spanning VN, SMN, VAN, DAN, and DMN circuits, consistent with progressive sensory–salience–default-mode reorganization along the chronification trajectory.

Classification performance (AUC)

The temporal graph neural network achieved robust three-class discrimination among healthy controls (HC), episodic migraine non-converters (EM), and chronic migraine converters (CM) under stratified 10-fold cross-validation. As summarized in Table 2, the model yielded one-vs-rest AUC values of 0.817 (95% CI: 0.76–0.88) for HC, 0.832 (95% CI: 0.77–0.90) for EM, and 0.874 (95% CI: 0.79–0.96) for CM, with a macro-averaged AUC of 0.841 (95% CI: 0.77–0.91) (Supplementary Figure S1).

Table 2.

Classification performance of the temporal graph neural network

Class	AUC	95% CI	Sensitivity	Specificity	Precision	F1-score
HC (Class 0)	0.817	0.76–0.88	72.6% (69/95)	86.0% (86/100)	83.1% (69/83)	0.775
EM (Class 1)	0.832	0.77–0.90	77.1% (54/70)	84.0% (105/125)	73.0% (54/74)	0.750
CM (Class 2)	0.874	0.79–0.96	76.7% (23/30)	90.9% (150/165)	60.5% (23/38)	0.677
Macro-average	0.841	0.77–0.91	75.5%	87.0%	72.2%	0.734

Note: Performance metrics were derived from pooled out-of-fold predictions obtained under stratified 10-fold cross-validation. Class-specific AUC values were calculated in a one-vs-rest manner, and the macro-average represents the unweighted mean across the three classes. 95% confidence intervals (CIs) for AUC values were estimated using bootstrapping with 1000 iterations. Sensitivity, specificity, precision, and F1-score were calculated from the corresponding confusion matrix based on aggregated out-of-fold predictions. HC, healthy controls; EM, episodic migraine non-converters; CM, chronic migraine converters; AUC, area under the receiver operating characteristic curve; CI, confidence interval

Class-specific metrics

Sensitivity ranged from 72.6% (HC) to 77.1% (EM) and 76.7% (CM), while specificity remained consistently high (84.0–90.9%). Precision was 83.1% for HC, 73.0% for EM, and 60.5% for CM, reflecting the impact of class imbalance in the CM converter group (n = 30). Corresponding F1-scores were 0.775 (HC), 0.750 (EM), and 0.677 (CM), with a macro-average of 0.734.

Overall performance

The macro-averaged sensitivity and specificity were 75.5% and 87.0%, respectively, indicating balanced discriminative ability across classes despite the modest sample size in the converter group.

Robustness across sliding-window configurations

Classification performance remained stable across all tested sliding-window configurations, with alternative window lengths and step sizes yielding macro-AUC values comparable to the primary setting (55-TR window, 2-TR step). Key interpretive patterns—including elevated frontal–insular and precuneus spatial attention in CM converters, increased SMN–VN and DMN–VAN coupling, and reduced intra-DMN cohesion—were qualitatively preserved across configurations, indicating that the main results were not dependent on a single arbitrary parameter choice (Supplementary Table S1).

Discussion

This study integrated sliding-window dynamic functional connectivity (dFC) with a temporal graph neural network (GNN) incorporating dual spatial–temporal attention to characterize network reorganization associated with migraine chronification. By directly modeling the full ordered sequence of window-level connectivity graphs, rather than reducing the dFC series to discrete state labels or summary metrics, the framework identified a high-attention temporal segment (window 23) and spatially discriminative regions—including the right dlPFC, precuneus, and mid-insula—that distinguish chronic migraine (CM) converters from episodic migraine (EM) non-converters at baseline. These findings extend prior static FC and state-averaged dFC work [30, 31] by providing temporally resolved, individualized characterization of chronification-related network alterations.

State-based dFC analysis identified four dominant connectivity configurations via k-means clustering [32]. State 3, characterized by prominent DMN–VAN coactivation, exhibited the highest fractional occupancy and longest mean dwell time, and was the only state yielding significant between-group edge-wise differences after FDR correction. These results suggest that DMN–VAN coactivation may represent a recurrent resting-state configuration in which chronification-related alterations are more readily detectable, and further indicate that state-stratified analysis may provide greater sensitivity than whole-scan averaging. Within State 3, both migraine groups showed enhanced VN–SMN coupling and strengthened insula–frontoparietal interactions alongside reduced within-network cohesion in the DMN and FPN. These alterations intensified from EM non-converters to CM converters, with the latter additionally exhibiting amplified SMN–VN coupling and prominent DMN–VAN hyper-coupling, consistent with progressive cross-network dysintegration as the disease advances [33, 34].

The temporal GNN extended this state-based perspective by preserving the temporal order of the dFC sequence [29] and generating interpretable attention weights at both temporal and spatial levels. Temporal attention was sparse, with most windows receiving low weights (approximately 0.03–0.05) and distinct peaks (≥ 0.09) emerging at specific moments, notably window 23. Connectivity analysis at this high-attention window revealed more pronounced DMN–VAN and sensory–attention cross-coupling in CM converters relative to EM non-converters, whereas EM non-converters’ abnormalities were primarily confined to visual–prefrontal pathways. This continuum—from visual–control emphasis in EM non-converters to more widespread DMN–VAN and SMN–VAN hyper-coupling in CM converters—suggests a more extensive and temporally focal large-scale network reorganization associated with subsequent chronification. Compared with k-means clustering, the temporal GNN therefore provided richer mechanistic information by jointly identifying when informative dFC patterns occurred and where the most relevant brain regions were located.

The spatial attention mechanism revealed a progressive topographic gradient across groups. High-importance nodes in healthy controls were concentrated within the visual network and primary frontoparietal regions. In EM non-converters, occipital and precuneus nodes remained prominent, but importance scores of the right dlPFC, ventrolateral frontal cortex, and insula increased markedly. In CM converters, spatial attention shifted further toward a frontoparietal–limbic configuration dominated by the right dlPFC, precuneus, right mid-insula, and fusiform gyrus. This trajectory—from visual–frontoparietal dominance in health, through frontal–insular prominence in episodic migraine, to a frontoparietal–limbic loop in chronic converters—mirrors progressive recruitment of cognitive control, salience, and affective processing circuits as chronification risk increases.

These findings can be interpreted within the framework of central sensitization and impaired descending pain modulation [10, 14, 34, 35]. Elevated importance of the right dlPFC and ventrolateral PFC, together with reduced intra-DMN cohesion and enhanced DMN–VAN coupling, is consistent with insufficient top-down modulation of pain processing. Heightened insula importance and stronger VAN coupling are consistent with overweighting of nociceptive and interoceptive signals and biased attentional allocation toward salient pain-relevant stimuli. Precuneus prominence despite weakened intra-DMN cohesion may reflect a destabilized self-referential template more readily perturbed by external inputs. Greater weighting of occipital and fusiform regions, together with increased SMN–VN coupling, is consistent with facilitated sensory gating and may contribute to the photophobia and multimodal trigger sensitivity commonly reported in patients at risk of chronification. Collectively, these alterations point to a coordinated imbalance of salience bias, intrinsic network decoupling, and sensory pathway facilitation that is transiently expressed at high-attention windows and distinguishes CM converters from EM non-converters at the episodic baseline.

Under ten-fold stratified cross-validation, the temporal GNN achieved a macro-averaged AUC of 0.841 (95% CI: 0.77–0.91), with one-vs-rest AUCs of 0.817, 0.832, and 0.874 for HC, EM non-converters, and CM converters, respectively. Specificity for CM converters reached 90.9%, indicating a low false-positive rate for this high-risk subgroup. The joint spatial–temporal attention readout transforms the classification problem into a three-dimensional map—comprising timestamps, node importance scores, and inter-network pathway signatures—that is more mechanistically interpretable than conventional black-box deep learning models. From a translational standpoint, if validated in independent multicenter cohorts, this framework may support early risk stratification, inform neuromodulation target selection for patients at heightened chronification risk, and provide candidate dynamic biomarkers—such as SMN–VN coupling strength and intra-DMN cohesion—for monitoring treatment response and disease progression.

Several limitations warrant acknowledgment. Although conversion status was determined longitudinally, the present neuroimaging analysis was based on a single baseline rs-fMRI scan and therefore supports associative rather than causal inference. Repeated imaging across the EM-to-CM transition will be necessary to clarify causal directionality. All performance metrics derive from internal cross-validation within a single-center cohort and should be regarded as preliminary; external validation in independent multicenter cohorts, together with calibration curves, decision-curve analyses, and prospective longitudinal testing, is needed before any clinical predictive use can be claimed. The framework relies exclusively on rs-fMRI, and integrating structural connectivity, task-based fMRI, or electrophysiological measures could refine mechanistic interpretation and improve classification. Residual confounding from medication exposure, affective state, hormonal fluctuations, and subtle motion effects cannot be entirely excluded, despite controlling for age, sex, and mean framewise displacement. Finally, although dual attention substantially enhances interpretability relative to standard deep learning, incorporating anatomical priors and established pain neuroscience pathways in future models may yield more biologically grounded and clinically actionable representations.

Conclusion

By integrating sliding-window dFC with a temporal graph neural network incorporating dual spatial–temporal attention, this study identified chronification-related dynamic network patterns from baseline rs-fMRI. Of the four recurrent dFC states, only State 3, characterized by dominant DMN–VAN coactivation, showed significant between-group differences after FDR correction. Both migraine groups exhibited increased sensory-network coupling together with reduced intra-DMN and FPN cohesion, with these abnormalities being more pronounced in CM converters than in EM non-converters. Spatial attention shifted from a visual–frontoparietal pattern in healthy controls to frontal–insular prominence in EM non-converters and a frontoparietal–limbic configuration in CM converters, while temporal attention identified window 23 as the most informative segment, where DMN–VAN and sensory–attention abnormalities were most evident. Under internal ten-fold cross-validation, the model achieved a macro-AUC of 0.841, with the highest specificity for CM converters (90.9%). These findings indicate that combining state-stratified dFC with temporal graph learning may provide complementary spatiotemporal markers of migraine chronification. However, because all results were obtained from single-center internal validation, they should be considered preliminary, and prospective multicenter validation is required before clinical application.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area Under the ROC Curve

CM

Chronic Migraine

CM converters

Patients with episodic migraine at baseline who converted to chronic migraine during follow-up

CSF

Cerebrospinal Fluid

dFC

Dynamic Functional Connectivity

DAN

Dorsal Attention Network

DMN

Default Mode Network

dlPFC

Dorsolateral Prefrontal Cortex

DPABI

Data Processing & Analysis for Brain Imaging

EPI

Echo-Planar Imaging

EM

Episodic Migraine

EM non-converters

Patients with episodic migraine at baseline who remained episodic during follow-up

FA

Flip Angle

FD

Framewise Displacement

FDR

False Discovery Rate

FOV

Field of View

FPN

Frontoparietal Network

GIN

Graph Isomorphism Network

GNN

Graph Neural Network

GRU

Gated Recurrent Unit

HC

Healthy Controls

HIT-6

Headache Impact Test-6

ICHD-3

International Classification of Headache Disorders, 3rd Edition

MIDAS

Migraine Disability Assessment Scale

MLP

Multilayer Perceptron

MNI

Montreal Neurological Institute

MRI

Magnetic Resonance Imaging

ROI

Region of Interest

rs-fMRI

Resting-State Functional Magnetic Resonance Imaging

SD

Standard Deviation

SDS

Zung Self-Rating Depression Scale

SMN

Somatomotor Network

TE

Echo Time

TR

Repetition Time

VAN

Ventral Attention Network

VAS

Visual Analog Scale

VN

Visual Network

WM

White Matter

Author contributions

CYX contributed to study design, analysis, project execution, interpretation, and primary manuscript writing. SHZ and ZG contributed to the conception, design, and data analysis. MZ and YW contributed to data collection and formal statistical analysis. RFG, YNK, and WLT contributed to imaging analysis and project operation. YCH contributed to study design, data collection, and manuscript revision. YCY contributed to study conception and design, study analysis and interpretation, manuscript revisions, and final approval.

Funding

This work was supported by the Shanghai Science and Technology Committee (Grant No. 25TS1411900) and the Science and Technology Development Program of Shanghai University of Traditional Chinese Medicine (2025; Grant No. 25KFL40). The funding bodies had no role in the design of the study; the collection, analysis, and interpretation of data; or in writing the manuscript.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki (1964) and the relevant guidelines and regulations. Approval was granted by the Ethics Committee of Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine (No. 2019-766-121-01). Written informed consent was obtained from all participants before any research relevant procedure started, and all were informed that they could withdraw from the study for no reason.

Consent for publication

All data have been anonymized to adhere to ethical standards, with explicit consent obtained from the participants, and all authors have provided their consent for this manuscript’s publication.

Competing interests

The authors declare no competing interests.