Temporal stability and neural complexity in resting-state MEG predict migraine phenotypes

Abstract

Background: Objective brain signatures for migraine, a leading cause of global disability, have yet to be identified, which limits objective diagnosis and personalised management for the condition. Herein, we introduce a novel magnetoencephalography (MEG)-based approach to capturing dynamic temporal signatures of brain activity in this cross-sectional study. We leveraged MEG’s high temporal and spatial resolution to investigate migraine-related disruptions in neural homeostasis within classification models.

Methods: Resting-state MEG data were collected from 250 right-handed individuals (age: 20–60 years; women: 80%), including 72 healthy control ([HC] group) and 178 migraine patients (diagnosed per the International Classification of Headache Disorders, 3rd edition). Within the migraine group, 98 patients were diagnosed with chronic migraine (CM), and 80 with episodic migraine (EM). MEG data were collected during the interictal period by using a 306-channel system. Time-resolved spectral power and dynamic functional connectivity were analysed across frequency bands ranging from delta to high-frequency oscillations in the default mode, salience, central executive, pain-related, sensorimotor, auditory, and visual networks. Neural stability and complexity were assessed by calculating the temporal standard deviation and entropy of spectral power, region-to-region connectivity, and node strength. Machine learning models were used to differentiate between the migraine and HC groups as well as between the CM and EM groups by using principal component analysis, 5-fold cross-validation, and 10% independent test datasets.

Results: In the spectral power dynamics, the migraine group exhibited elevated entropy in the theta band, particularly in the right insula, along with reduced temporal standard deviation in the delta band within the insula, in the alpha band within the lateral frontal cortex, and in the beta band within the posterior cingulate and lateral frontal regions (corrected p < 0.05). Node strength dynamics of functional connectivity revealed reduced entropy in the migraine group, particularly in the alpha, beta, gamma, and high-frequency oscillation bands across distinct brain networks (corrected p < 0.05). In the classification model between the migraine and HC groups, SVM model achieved a validation accuracy of 79.1% (sensitivity: 82.0%; specificity: 71.9%; AUC value: 0.8507) and a test accuracy of 76%. By contrast, in comparisons of the CM and EM groups, SVM models achieved a validation accuracy of 80.7% (sensitivity: 80.9%; specificity: 80.6%; AUC value: 0.9117) and a test accuracy of 76.5%. Feature importance indicated migraine is linked to disrupted spectral complexity and dynamic coupling across sensory–cognitive networks.

Conclusion: Our MEG-based temporal dynamics approach exhibited good classification accuracy, revealing distinct homeostatic disruptions in individuals with migraine. The findings may guide the development of objective, brain-based signatures, advancing the development of personalised strategies for migraine management. Future research should validate the findings and assess their applicability to scalable modalities such as EEG.

Supplementary Information: The online version contains supplementary material available at 10.1186/s10194-025-02169-y.

Keywords: Brain homeostasis; Diagnostic classification; Entropy; Machine learning; Network dysregulation; Oscillatory dynamics.

Fig. 1

Analytical pipeline for resting-state MEG-based migraine prediction. The figure depicts the methodological workflow for identifying migraine-related biomarkers from temporal stability and neural complexity data derived through resting-state MEG data. (1) MEG data acquisition: 3-minute resting-state recordings were obtained with the participant in an eyes-closed condition. (2) Preprocessing and source reconstruction: The data were notch-filtered, bandpass-filtered, and denoised using signal-space projection.Source localisation was performed using an overlapping spheres head model and minimum norm estimates. (3) Spectral power and functional connectivity analysis: Time-resolved analyses were performed to estimate both absolute and normalised spectral power; in addition, amplitude envelope correlations were explored to capture dynamic oscillatory functional connectivity across key brain regions. (4) Brain signature extraction: Standard deviation and entropy metrics were computed from the spectral power and connectivity time series. Independent t tests and false discovery rate (FDR) corrections were used to identify features with significant between-group differences. (5) Classification modelling: Machine learning models were trained using 5-fold cross-validation and validated on independent datasets. Feature extraction was performed using independent t-tests to identify discriminant features, followed by principal component analysis (PCA) to reduce feature dimensionality. Model performance was evaluated on the basis of classification accuracy, receiver operating characteristic (ROC) curves, and confusion matrices. MEG, magnetoencephalography

Fig. 2

Migraine-related changes in the entropy and standard deviation values of cortical power dynamics. The figure depicts differences in the temporal stability and neural complexity of resting-state MEG data between the migraine and HC groups and between the CM and EM groups. a Differences in the entropy of cortical power dynamics: Differences in the entropy of absolute (Abs.) and normalised (Normd.) spectral power between the migraine and HC groups and the CM and EM groups across the delta, theta, alpha, beta, gamma, high gamma, and high-frequency oscillation (HFO) bands in seven regions of interest (ROIs): default mode network (DMN), salience network (SAN), central executive network (CEN), pain-related network (PN), sensorimotor network (SMN), auditory network (AN), and visual network (VN). The heatmaps denote the t-values between-group differences. * uncorrected p < 0.05. *^+O p < 0.05 with FDR correction. L, left; R, right; T, temporal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; MI, primary motor cortex; PCC, posterior cingulate cortex; ACC, anterior cingulate cortex; SI, primary somatosensory cortex; SII, secondary somatosensory cortex; A1, primary auditory cortex; V1, primary visual cortex. b Differences in the standard deviation of cortical power dynamics: The t value heatmaps depict the between-group differences in standard deviation of spectral power across frequency bands and ROIs. * uncorrected p < 0.05. *^+O p < 0.05 with FDR correction. c Prominent spectral power differences: Box plots illustrate significant between-group differences (p < 0.05) in the entropy and standard deviation of spectral power for specific ROIs and bands. HC, healthy control; CM, chronic migraine; EM, episodic migraine. * p < 0.05 with FDR correction

Fig. 3

Differences in the entropy and standard deviation values of dynamic FC between the migraine and HC groups. The t value heatmaps depict the between-group differences in FC dynamics. The upper triangles represent entropy, and the lower triangles represent standard deviation across the delta, theta, alpha, beta, gamma, high gamma, and high-frequency oscillation (HFO) bands. The ROIs include the default mode network (DMN), salience network (SAN), central executive network (CEN), pain-related network (PN), sensorimotor network (SMN), auditory network (AN), and visual network (VN). Colour intensity (blue to red) represents the t values. The right panel maps the intrinsic network affiliations of each ROI. The separate inset defines the frequency band associated with each heatmap and the measures of dynamic FC (entropy and standard deviation). * uncorrected p < 0.05; FC, functional connectivity; HC, healthy control. L, left; R, right; T, temporal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; PCC, posterior cingulate cortex; ACC, anterior cingulate cortex; SI, primary somatosensory cortex; MI, primary motor cortex; SII, secondary somatosensory cortex; A1, primary auditory cortex; V1, primary visual cortex

Fig. 4

Differences in the entropy and standard deviation values of dynamic FC between the chronic and episodic migraine groups. The t value heatmaps depict between-group differences in FC dynamics. The upper triangles represent entropy and the lower triangles represent standard deviation across the delta, theta, alpha, beta, gamma, high gamma, and high-frequency oscillation (HFO) bands. The ROIs include the default mode network (DMN), salience network (SAN), central executive network (CEN), pain-related network (PN), sensorimotor network (SMN), auditory network (AN), and visual network (VN) networks. Colour intensity represents t values. The right panel maps the intrinsic network affiliations of each ROI. The separate inset defines the frequency band associated with each heatmap and the measures of dynamic FC (entropy and standard deviation). * uncorrected p < 0.05; FC, functional connectivity. L, left; R, right; T, temporal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; PCC, posterior cingulate cortex; ACC, anterior cingulate cortex; SI, primary somatosensory cortex; MI, primary motor cortex; SII, secondary somatosensory cortex; A1, primary auditory cortex; V1, primary visual cortex

Fig. 5

Differences in node strength dynamics between the migraine and HC groups. The left heatmap depicts between-group differences in entropy, whereas the right heatmap depicts between-group differences in standard deviation. The ROIs include the default mode network (DMN), salience network (SAN), central executive network (CEN), pain-related (PN), sensorimotor network (SMN), auditory network (AN), and visual network (VN), with specific subregions. The colour-coded values represent the t values for between-group differences. * uncorrected p < 0.05; *^+O p < 0.05 with FDR correction; HFO, high-frequency oscillation. L, left; R, right; T, temporal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; MI, primary motor cortex; PCC, posterior cingulate cortex; ACC, anterior cingulate cortex; SI, primary somatosensory cortex; SII, secondary somatosensory cortex; A1, primary auditory cortex; V1, primary visual cortex

Fig. 6

Differences in node strength dynamics between the chronic and episodic migraine groups. The left heatmap depicts between-group differences in entropy, whereas the right heatmap depicts between-group differences in standard deviation. The ROIs include the default mode network (DMN), salience network (SAN), central executive network (CEN), pain-related network (PN), sensorimotor network (SMN), auditory network (AN), and visual network (VN), with specific subregions. The colour-coded values represent t values for between-group differences. * uncorrected p < 0.05; HFO, high-frequency oscillation. L, left; R, right; T, temporal cortex; F, frontal cortex; P, parietal cortex; O, occipital cortex; PCC, posterior cingulate cortex; ACC, anterior cingulate cortex; SI, primary somatosensory cortex; MI, primary motor cortex; SII, secondary somatosensory cortex; A1, primary auditory cortex; V1, primary visual cortex

Fig. 7

Model performance and feature importance in the prediction of migraine. Machine learning models incorporating discriminative dynamical features of oscillatory power and functional connectivity were used to differentiate between patients with migraine and healthy control individuals. a Validation accuracy values for 9 models: The bar chart depicts the validation accuracy (%) values for 9 classifiers (e.g., support vector machines [SVM], 79.1%; neural network, 75.5%). b Validation and test confusion matrices for SVM: The matrices present true-positive rates (82.0% and 76.5%, respectively) and true-negative rates (71.9% and 75%, respectively) for validation (left, [L]) and test (right, [R]) data. c Receiver operating characteristic (ROC) curves for 10 models: The plots present ROC curves for each model and corresponding area under the curve (AUC) values for validation (left) and test (right) data. d Weighted precision–recall curves for 9 models: The curves depict precision–recall trade-offs (PR-AUC) for validation (left) and independent test (right) data

Fig. 8

Model performance and feature importance in the classification of migraine phenotypes. Machine learning models were used to differentiate between chronic migraine and episodic migraine. a Validation accuracy values for 9 models: The bar chart depicts the validation accuracy (%) values for 9 classifiers (e.g. support vector machines [SVM], 80.7%; neural network, 83.2%). b Validation and test confusion matrices for the support vector machine model: The matrices present true-positive rates (80.9% and 77.8%) and true-negative rates (80.6% and 75.0%) for validation (left, [L]) and test (right, [R]) data. c Receiver operating characteristic (ROC) curves for 10 models: The plots represent the ROC curves for each model and corresponding area under the curve (AUC) values for validation (left) and test (right) data. d Weighted precision–recall curves for 9 models: The curves depict the precision–recall trade-offs (PR-AUC value) for validation (left) and independent test (right) data