Unmasking the noise: aberrant cortical oscillations in visual snow syndrome

Abstract

Background

Visual snow syndrome (VSS) is characterised by persistent visual static and disabling perceptual disturbances, yet its underlying neural mechanisms remain poorly understood and are frequently confounded by migraine comorbidity.

Methods

We examined resting-state cortical oscillatory activity in VSS (n = 30) using EEG with distributed source modelling. Absolute and normalised spectral power was quantified across cortical regions, and machine-learning classifiers were trained to identify diagnostic spectral signatures.

Results

Relative to healthy controls (HC, n = 47), VSS showed a frequency-dependent redistribution of power, with increased low-frequency (delta–alpha) activity across parieto-occipital and frontal cortices, and reduced high-frequency (beta–high frequency oscillation) activity. When contrasted with migraine-only controls (n = 45), VSS with migraine (n = 25) exhibited additional enhancements of low-frequency activity within the parietal cortex. Machine-learning models reliably discriminated VSS from both HC and migraine-only patients using distinct sets of spectral features. Among the acceptable models, classification of VSS versus HC achieved an accuracy > 0.80 and an area under the curve (AUC) > 0.81, while classification of VSS with migraine versus migraine-only achieved an accuracy > 0.75 and an AUC > 0.81. Across models, alpha-band activity in the frontal and parietal regions showed the strongest predictive contribution.

Conclusion

These findings indicate that VSS is characterised by a distinct and spatially distributed pattern of cortical dysrhythmia that is not attributable to migraine, and identify parietal low-frequency oscillatory activity as a potential physiological signature and target for neuromodulation.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s10194-026-02355-6.

Introduction

According to the 11th revision of the WHO International Classification of Diseases (ICD-11), visual snow syndrome (VSS) is diagnosed when a patient experiences continuous visual disturbances for more than three months, characterized by persistent, dynamic tiny dots across the visual field. Diagnosis also requires at least two of the following accompanying symptoms: photophobia, palinopsia (either stationary or moving), blue field entoptic phenomena, or nyctalopia [1, 2]. Comorbidities further exacerbate the burden: migraine, tinnitus, paraesthesia, dizziness, neck pain, irritability, concentration problems, and oculomotor deficits such as convergence insufficiency and accommodative insufficiency, which contribute to reading difficulties and are highly prevalent [3, 4]. Symptoms are chronic, with fluctuations but no full remission. Triggers such as stress, fatigue, insufficient sleep, alcohol, caffeine, and screen exposure often worsen symptoms, while improved sleep, stress management, and tinted lenses may provide relief [3]. Epidemiological data suggest VSS affects approximately 2% of the population, based in part on its association with tinnitus and migraine [5]. Overall, the burden of VSS is multifaceted, involving neurological, visual, psychiatric, and functional impairments that diminish independence and quality of life.

Neurophysiological investigations employing electrophysiology, neuroimaging and psychophysical paradigms indicate cortical hyperexcitability, thalamocortical dysrhythmia and large-scale network disruption across visual, attentional, salience and limbic systems [6–17]. These observations support the view that VSS reflects a disorder of sensory processing rather than a focal lesion, characterised by aberrant excitability and connectivity within visual pathways. Nevertheless, the neural signatures that represent its underlying mechanisms, as well as candidate targets for intervention, remain to be defined. Interpretation of existing findings is further constrained by methodological limitations, most notably confounding by comorbidity. Migraine is present in over 50% of VSS cohorts [1, 2, 5, 18] and shares several reported features, including alterations in cortical dysexcitability, multiple network interactions, and psychiatric comorbidities such as depression and anxiety [1–4].

Nonetheless, substantial evidence indicates that visual snow condition, akin to migraine, is classified under the thalamocortical dysrhythmia disorders. To ascertain the electroencephalographic anomalies and their anatomical localisation associated with dysrhythmia in patients with VSS, we investigated the resting EEG power spectrum of these patients and their subgroups with migraine comorbidity, compared with healthy subjects and a cohort of migraine patients. Spectral characteristics in VSS were quantified using absolute and normalised power across predefined cortical regions. Machine learning with Shapley analysis, transcending the limitations of univariate statistics, was used to uncover nonlinear, multivariate, and interpretable relationships within spectral power activity, enabling the discovery of robust, biologically meaningful signatures. Finally, model interpretability was assessed to evaluate the translational potential of these approaches for clinical application.

Materials and methods

Participants

All participants were aged between 20 and 60 years, and had no history of systemic or major neurological disorders. They exhibited normal findings on both physical and neurological examinations and were recruited from the Headache Clinic at Taipei Veterans General Hospital. Patients were diagnosed with VSS by Dr. SJ Wang according to the criteria [18]. Notably, in the patients with VSS, symptoms are not consistent with typical migraine visual aura and are not better accounted for by another medical condition. Patients with migraine were diagnosed according to the criteria of the International Classification of Headache Disorders, Third Edition (ICHD-3) [19]. None of the migraine patients had received preventive migraine treatment or reported overuse of acute headache medications. Healthy controls (HC) had no personal or family history of pain-related disorders and had not experienced any significant visual disturbance. Individuals who were taking prophylactic medications, hormonal therapies, or other daily pharmacologic treatments were excluded. The health status of the control group was verified through structured screening at enrollment. Specifically, healthy controls were confirmed to have no history of major neurological or psychiatric disorders, no current significant medical illness, and no diagnosis of migraine or VSS.

Study design

At the initial visit, all participants completed semistructured questionnaires capturing demographic variables and psychometric measures. Anxiety, depression, and perceived stress were assessed using the Hospital Anxiety and Depression Scale (HADS) [20], Beck Depression Inventory (BDI) [21], and the Perceived Stress Scale (PSS), respectively [22]. Moreover, sleep quality and efficiency were evaluated using Pittsburgh Sleep Quality Index (PSQI). Participants with VSS or migraine additionally completed headache-related questionnaires, and their clinical profiles were verified in person by neurologists. Headache characteristics included monthly headache days (MHDs) and the presence or absence of visual aura. The Migraine Disability Assessment (MIDAS) [23] and Headache Impact Test-6 [24] were further administered to evaluate headache-related disability and its impact on daily functioning. Following questionnaire completion, all participants underwent resting-state EEG acquisition. All the participants were asked not to consume caffeine or take analgesics 48 h before the EEG assessment.

Resting-state EEG recording

Figure 1 outlines the data acquisition and analysis pipeline. Scalp EEG data were recorded using a 64-electrode BrainVision actiCAP system (Brain Products GmbH, Munich, Germany), which adheres to the extended International 10–20 system. In this system, for impedance conversion, active circuits are integrated into the streamlined actiCAP electrodes, providing superior signal quality even at higher impedance, compared with conventional passive electrodes. Crucially, an electronic circuit is integrated into each active electrode for performing impedance conversion directly at the scalp, offering the acquisition of high-quality EEG signals that remain unaffected by extraneous noise or movements possibly affecting the electrode cables. The electrodes were referenced online to an electrode positioned at the Fz plane, with a common ground connection established at the FPz site. EEG signals were amplified and digitised at a sampling rate of 1,000 Hz by using a BrainAmp DC amplifier (Brain Products GmbH) that was interfaced with Brain Vision Recorder software (version 2.1, Brain Products GmbH).

Fig. 1.

Pipeline of EEG acquisition, processing, analysis, and classification. (1) Resting-state EEG was recorded with eyes closed. (2) Preprocessing and source modelling included removal of EOG/ECG artifacts, notch and band-pass filtering, forward modelling, and minimum-norm source reconstruction, producing time-resolved source activity in cortical regions of interest. (3) Source-level spectral power was obtained by segmenting the time-resolved source activity into overlapping windows, applying fast Fourier transforms using the Welch method, and averaging power within canonical frequency bands; both absolute and normalised spectral power were derived and statistically compared between groups. (4) Group-discriminative spectral features (ranked by ANOVA) were then used to train machine-learning classifiers, which were optimised using cross-validation and evaluated on held-out test data. Model performance was quantified using confusion matrices and receiver-operating characteristic curves

During the 5-min resting-state EEG recording, participants were instructed to keep their eyes closed while remaining awake, relaxed, and task-free. They were specifically asked to relax the jaw, forehead, and neck muscles, minimize talking and swallowing, and maintain a stable posture throughout the recording. These procedures were intended both to capture resting-state spontaneous brain activity and to reduce contamination from temporal and facial muscle activity. If a participant fell asleep or had excessive within-run head movement, the recording was stopped and then rerun. Electrooculography (EOG) and electrocardiography (ECG) activities were also acquired simultaneously for offline artefact elimination.

Data preprocessing and analysis

To minimise the influence of non-neural and environmental artefacts on spontaneous resting-state EEG recordings (Fig. 1), preprocessing was performed in sequential stages. First, raw data were visually inspected to identify channels affected by environmental noise or electrode malfunction. In addition, segments contaminated by muscle artifacts were manually rejected based on their characteristic features, including high-frequency activity with focal scalp distributions, typically over the peripheral, temporal, or frontal regions, and irregular, burst-like time courses. Second, artefacts arising from ocular and cardiac activity were removed using signal-space projection based on principal component analysis, with EOG and ECG traces serving as reference signals [25]. This supervised approach removes only components that correlate with the reference channels, thereby limiting overfitting and reducing computational demand. Third, a notch filter was applied at 60 Hz and its harmonics to attenuate powerline interference. Finally, a 1–200 Hz bandpass filter was used to eliminate slow drifts and high-frequency noise.

To derive source-level neural activity and minimise the confounding effects of volume conduction, we applied distributed current source modelling using minimum norm estimates (MNE). For each cortical vertex, the forward model computed the characteristic scalp signal pattern generated by a unit dipole. A realistic head model was constructed using the symmetric boundary element method (BEM) [26], which provides greater anatomical fidelity than spherical approximations. The inverse operator was then applied to estimate the distributed current sources from the electrode data at each time point, yielding dynamic cortical activation maps projected onto each participant’s surface model.

Individual cortical anatomy was registered to the ICBM152 template. Regions of interest (ROIs) were defined using the Mindboggle cortical parcellation of the T1 template [27] and comprised bilateral primary visual cortex (V1), lateral occipital cortex, inferior parietal lobule, precuneus, superior parietal cortex, posterior cingulate cortex (PCC), superior temporal cortex, insula, medial temporal cortex, primary motor (MI) and somatosensory (SI) cortices, lateral orbital frontal cortex, anterior cingulate cortex (ACC), middle frontal gyrus, and medial orbital frontal cortex. These ROIs primarily encompass regions within the visual, sensorimotor, attentional, salience, and limbic networks. Owing to the limited spatial resolution and signal-to-noise ratio of EEG, deeper structures within these circuits were not included in the analysis.

To characterise source-level oscillatory activity, dynamic current density estimates were first extracted from each ROI. Absolute spectral power was computed using the Welch method (3-s windows with 50% overlap). Normalised spectral power was then obtained by dividing the power in each frequency band by the total power of the corresponding ROI. Spectral power was quantified across the following frequency ranges: delta (2–4 Hz), theta (5–7 Hz), alpha (8–13 Hz), beta (14–25 Hz), gamma (26–40 Hz), high gamma (41–90 Hz), and high-frequency oscillations (HFO; 91–200 Hz). All analyses—including preprocessing, source modelling, amplitude estimation, and spectral quantification—were performed in Brainstorm [28], consistent with our previous work [29–34].

Statistical analysis and classification using machine learning

Group differences in demographic variables, psychometric measures, and spectral power were assessed using independent t-tests. Categorical variables were compared with the χ² test. In addition, comparisons were conducted between patients with VSS comorbid with migraine (VSS + MIG) and those with migraine-only. All tests were two-tailed with a significance threshold of P < 0.05. Multiple comparisons were controlled using false discovery rate (FDR) correction [35].

To construct the classification models, the dataset was randomly partitioned into a training set (80%) and an independent test set (20%) (Fig. 1). Within the training set, absolute and normalised spectral power features across frequency bands were ranked using F values of one-way ANOVA tests. The top N features (N = 50, 40, 30, 20, 10, or 5) were then used to train the classifiers. To mitigate class imbalance, a cost matrix was applied to impose higher penalties on misclassification of the minority class, thereby reducing bias toward the majority group. Seven supervised learning algorithms were trained: decision tree, discriminant analysis, naïve Bayes, support vector machine (SVM), k-nearest neighbours (KNN), ensemble methods, and neural networks. Each classifier was applied in pairwise settings to distinguish (i) VSS from HC and (ii) VSS + MIG from migraine-only. A 5-fold cross-validation strategy was used to minimise overfitting and improve generalisability [36]. Model performance was assessed using standard metrics: overall accuracy, weighted F1 score (harmonic mean of precision and recall accounting for class proportions), and weighted area under (AUC) the receiver operating characteristic (ROC) curve. These metrics were computed both during validation and on the independent test datasets. To interpret model decisions, Shapley values were calculated to quantify the contribution of each feature to prediction outputs [37]. Comparative Shapley Impact was then used to contrast the relative influence of features across models. All analyses were implemented using the Machine Learning Toolbox in MATLAB (R2024b).

Results

Clinicodemographic characteristics of participants

The VSS (N = 30) and HC (N = 47) groups did not differ significantly in age (t(75) = -0.8697, p = 0.39) or sex distribution (c² = 0.93, p = 0.33)(Table 1). By contrast, patients with VSS exhibited substantially higher levels of affective and stress-related symptoms. Mean scores were markedly elevated for anxiety (t(75) = 6.31, p < 0.001), depression (t(75) = 6.11, p < 0.001), and depressive symptoms measured by the BDI (t(75) = 7.37, p < 0.001). Perceived stress was also higher in the VSS group (t(75) = 3.44, p < 0.001). Sleep disturbance was prominent in VSS. Patients reported poorer sleep quality (t(75) = 6.31, p < 0.001) and reduced sleep efficiency (t(75) = -2.18, p = 0.03) compared with healthy controls. Taken together, the cohorts were comparable in demographic variables, but individuals with VSS demonstrated substantially greater psychological distress and sleep disruption.

Table 1.

Demographics and clinical profiles in patients with visual snow syndrome and healthy control

	VSS	HC	p-value
N	30	47
Demographics
Age (years)	29.0 ± 9.7	30.7 ± 7.0	p = 0.39
Sex	20 F/10 M	26 F/21 M	p = 0.33
Psychometrics
HADS_A	9.5 ± 4.6	3.8 ± 3.2	p < 0.001
HADS_D	7.3 ± 4.8	2.4 ± 2.2	p < 0.001
BDI	16.0 ± 9.8	3.9 ± 4.3	p < 0.001
Sleep Quality (0–21)	8.7 ± 3.6	4.5 ± 2.1	p < 0.001
Sleep Efficiency (0-100)	83.7 ± 15.1	90.1 ± 10.3	p = 0.03
PSS	28.2 ± 12.0	20.8 ± 6.9	p < 0.001

HC: Healthy control; VSS: Visual snow syndrome; F: female; M: male; HADS: Hospital anxiety and depression score; A: Anxiety; D: Depression; BDI: Beck’s depression inventory; PSS: perceived stress scale

The VSS with migraine (VSS + MIG, N = 25) and migraine-only (N = 45) groups were comparable in migraine subtype distribution (Table 2), with similar proportions of chronic migraine (x² = 0.04, p = 0.84) and episodic migraine (x² = 0.05, p = 0.82). Age showed a non-significant trend toward younger in the VSS + MIG group (t(68) = -1.96, p = 0.054), and sex ratios did not differ. Psychometric indices were broadly similar between groups. Anxiety, depression, depressive symptom severity, and PSS did not differ significantly. Headache characteristics were likewise comparable. Mean monthly headache days were similar, and the proportion reporting visual aura did not differ. Headache-related impact showed a trend toward lower scores in the VSS + MIG group. Moreover, migraine-related disability (MIDAS) was also a trend toward lower scores in the VSS + MIG cohort. Overall, patients with migraine and comorbid VSS exhibited clinical and affective profiles largely comparable to migraine-only patients.

Table 2.

Demographics and clinical profiles in visual snow syndrome with migraine and migraine

	VSS་MIG	Migraine	p-value
N	25	45
%CM	9/25	15/45	p = 0.84
%EM	16/25	30/45	p = 0.82
Demographics
Age (years)	29.6 ± 8.8	33.6 ± 7.7	p = 0.054
Sex	18 F/7 M	32 F/13 M	p = 0.94
Psychometrics
HADS_A	9.4 ± 4.5	8.0 ± 3.7	p = 0.15
HADS_D	7.3 ± 4.6	6.5 ± 3.7	p = 0.42
BDI	16.8 ± 10.0	12.9 ± 6.4	p = 0.058
PSS	27.7 ± 11.6	26.8 ± 8.3	p = 0.70
Migraine profile
Headache days (/month)	13.5 ± 10.8	12.8 ± 9.1	p = 0.78
Visual aura	6/25	17/45	p = 0.24
HIT-6	56.6 ± 7.5	58.9 ± 5.3	p = 0.13
MIDAS	22.9 ± 47.0	39.0 ± 44.3	p = 0.16

VSS + MIG: Visual snow syndrome with migraine; CM: chronic migraine; EM: episodic migraine; F: female; M: male; HADS: Hospital anxiety and depression score; A: Anxiety; D: Depression; BDI: Beck’s depression inventory; PSS: Perceived stress scale; HIT-6: Headache impact test; MIDAS, migraine disability assessment

Psychiatric and sleep-related problems were common comorbidities in the VSS and migraine groups, as reflected by their higher psychometric scores in Table 1. According to inclusion criteria, migraine patients with other systemic or major neurological disorders were excluded. In the VSS group, the symptoms and associated comorbidities reported by patients are summarized in Supplementary Table 1, including tinnitus, vertigo, and astigmatism.

Absolute and normalized spectral power differences between VSS and HC

Absolute spectral power activities of cortical ROIs revealed marked frequency-dependent alterations in patients with VSS relative to HC (Fig. 2). Across the cortical surface, t-value maps indicated that regions showing increased power in VSS were widespread in the delta, theta, and alpha bands. In the delta and theta ranges, VSS patients exhibited significantly elevated power bilaterally in the inferior parietal, superior parietal, middle frontal, MI, SI, and precuneus cortices. These increases extended anteriorly into the ACC and posteriorly along the midline. Alpha-band enhancement showed a similar topography but further involved lateral occipital and V1 areas, consistent with exaggerated posterior network activity. Beta and higher frequency activity (gamma, high gamma and HFO) did not differ significantly between groups.

Fig. 2.

Differences of absolute spectral power between VSS and HC. Top: Group-level t-value cortical maps for distinct bands show widespread low-frequency power increases in VSS, predominantly over parieto-occipital, sensorimotor, and frontal cortices. Bottom: Boxplots depict absolute power differences across prominent cortical regions. VSS, Visual Snow Syndrome; HC, Healthy Controls; LH, left hemisphere; RH, right hemisphere. MF, middle frontal gyrus; V1, primary visual cortex; LO, lateral occipital cortex; IP, inferior parietal lobule; PCu, precuneus; SP, superior parietal cortex; PCC, posterior cingulate cortex; ST, superior temporal cortex; MI, primary motor cortex; SI, somatosensory cortex; ACC, anterior cingulate cortex; L, left; R, right. *p < 0.05, **p < 0.01, ***p < 0.001 (FDR-corrected)

Box-plot analyses demonstrated widespread increases in low-frequency absolute spectral power in patients with VSS compared with HC. Significant elevations were observed in the delta and theta bands across bilateral parietal cortex (inferior and superior parietal lobules and precuneus; t(75) = 3.0–4.3), extending to sensorimotor cortices (t(75) = 3.2–4.7), frontal regions including the middle frontal cortex and anterior cingulate cortex (t(75) = 3.2–4.6), as well as temporal and occipital areas (all corrected p < 0.01). Group differences were most pronounced in the alpha band, with robust bilateral power increases involving occipital (V1 and lateral occipital), parietal, precuneus, and sensorimotor cortices (t(75) = 3.9–5.3; all corrected p < 0.005).

To further characterise relative oscillatory redistribution, normalised cortical power was compared across seven frequency bands (Fig. 3). VSS patients exhibited increased normalised delta power within frontal regions, including lateral and medial orbitofrontal and middle frontal cortices (t(75) = 2.8–4.1, p < 0.05), and increased normalised alpha power within bilateral primary visual cortex (t(75) = 2.9–3.1, p < 0.05). In contrast, normalised beta, gamma, high-gamma, and HFO power was significantly reduced across widespread frontal, parietal, precuneus, sensorimotor, and occipital regions (t(75) = − 2.6 to − 3.6; all p < 0.05).

Fig. 3.

Differences of normalized spectral power between VSS and HC. Top: Group-level t-value cortical maps for distinct bands show high-frequency power decreases in VSS, predominantly over sensorimotor, parietal, frontal, and visual cortices. Bottom: Boxplots depict normalized power differences across prominent cortical regions. VSS, Visual Snow Syndrome; HC, Healthy Controls; LH, left hemisphere; RH, right hemisphere; MF, middle frontal gyrus; V1, primary visual cortex; LO, lateral occipital cortex; IP, inferior parietal lobule; PCu, precuneus; SP, superior parietal cortex; INS, insula; ST, superior temporal cortex; MI, primary motor cortex; SI, somatosensory cortex; ACC, anterior cingulate cortex; LOF, lateral orbital frontal cortex; MOF, medial orbital frontal cortex; L, left; R, right. *p < 0.05 (FDR-corrected)

Differences of spectral power between VSS with migraine and migraine

Relative to migraine-only participants, patients with VSS + MIG showed widespread increases in absolute low-frequency spectral power, as indicated by t-value cortical maps (Fig. 4). Power elevations were most prominent in the delta, theta, and alpha bands and were spatially overlapping across frontal (middle frontal and lateral orbitofrontal), insular, temporal, parietal (superior and inferior parietal), and precuneus cortices. Additional alpha-band increases extended into occipital regions. In contrast, group differences in higher-frequency bands were spatially limited, with focal increases in gamma, high-gamma, and HFO activity confined mainly to medial temporal and lateral occipital cortices. Notably, normalised spectral power did not show significant group differences after correction for multiple comparisons.

Fig. 4.

Spectral power differences between VSS with migraine and migraine-only. (a) Absolute power: t-value cortical maps (VSS + MIG > Migraine) reveal increased low-frequency (delta–alpha) power across parieto-occipital, insular, sensorimotor, and frontal cortices. Corresponding boxplots illustrate the cortical regions with significantly elevated low-frequency power in VSS + MIG. Boxplots illustrate the cortical regions with the most pronounced power differences between groups. VSS + MIG, visual snow syndrome with migraine; LH, left hemisphere; RH, right hemisphere; MF, middle frontal gyrus; V1, primary visual cortex; LO, lateral occipital cortex; IP, inferior parietal lobule; PCu, precuneus; SP, superior parietal cortex; INS, insula; PCC, posterior cingulate cortex; ST, superior temporal cortex; MT, medial temporal cortex; MI, primary motor cortex; SI, somatosensory cortex; ACC, anterior cingulate cortex; LOF, lateral orbital frontal cortex; MOF, medial orbital frontal cortex; L, left; R, right. *p < 0.05, **p < 0.01, ***p < 0.001 (FDR-corrected)

Box-plot analyses confirmed these findings quantitatively, demonstrating significantly higher absolute delta–alpha power in VSS + MIG patients across multiple cortical regions (all corrected p < 0.05). Consistent low-frequency increases were observed in frontal (middle frontal and lateral orbitofrontal; t(68) = 2.6–3.5), insular (t(68) = 2.7–3.9), temporal (t(68) = 2.7–4.0), parietal (t(68) = 2.6–4.4), and precuneus regions (t(68) = 2.8–4.7). Alpha-band power was also significantly elevated in occipital cortex, including lateral occipital and primary visual areas (t(68) = 2.8–3.4). In addition, focal increases in high-frequency activity were detected within medial temporal cortex and lateral occipital regions (t(68) = 2.5–3.4, p < 0.05).

Identification of VSS patients using machine learning algorithms

Machine-learning approaches were applied to identify optimal classifiers and neurophysiological features distinguishing patients with visual snow syndrome (VSS) from healthy controls, using absolute and normalised spectral power features ranked by ANOVA F-values. As shown in Fig. 5a, model performance was systematically evaluated across increasing numbers of top-ranked features (5–50), revealing algorithm-specific performance profiles. Models achieving validation and test accuracies, as well as weighted F1 scores, of at least 80% were considered optimal. Based on these criteria, eight models met the predefined threshold, including decision tree classifiers using 20–50 features, support vector machines (SVM) using 30 and 40 features, and neural network models using 30 and 40 features. Performance was consistent between validation and independent test datasets, indicating limited overfitting and good generalisability. Receiver operating characteristic analyses (Fig. 5b) further confirmed robust classification performance, with area-under-the-curve (AUC) values ranging from 0.8052 to 0.9388 in validation and from 0.8104 to 0.8859 in test sets. On the basis of their superior and stable performance, SVM and neural network models incorporating 30 or 40 features were selected for subsequent analyses.

Fig. 5.

Machine-learning classification of visual snow syndrome versus healthy controls. (a) Classification performance across algorithms and feature-set sizes, showing accuracy and weighted F1 scores for validation and test datasets. Support vector machines (SVM), decision tree, and neural network models exhibited consistently good performance. Asterisks (*) indicate models that met the predefined criteria for optimal classification. (b) Receiver operating characteristic (ROC) curves and corresponding area under curve (AUC) values demonstrate robust discrimination between patients with VSS and HC across multiple algorithms. Asterisks (*) indicate models were selected for optimal classification. (c) Shapley value summary plots illustrating the top 15 predictive spectral-regional features across SVM and neural network models. (d) Ranked top-10 features based on normalised Shapley impact across models, highlighting the dominant spectral power predictors driving classification. (e) Node-level relative impact derived from aggregated feature contributions, identifying cortical regions with the highest influence as candidate neurophysiological signatures of VSS. (f) Frequency-band contributions showing the total relative impact of each frequency band across models in distinguishing VSS from HC. KNN, k-nearest neighbours. aPSD, absolute power spectral density; nPSD, normalised power spectral density; L, left hemisphere; R, right hemisphere; Hgamma, high gamma; HFO, high-frequency oscillations; MF, middle frontal gyrus; V1, primary visual cortex; LO, lateral occipital cortex; IP, inferior parietal lobule; PCu, precuneus; MI, primary motor cortex; SI, somatosensory cortex; ACC, anterior cingulate cortex; LOF, lateral orbital frontal cortex; MOF, medial orbital frontal cortex

Shapley summary plots (Fig. 5c) illustrate the relative contribution of individual features to model predictions across participants, with Shapley values indicating both the magnitude and direction of feature influence. To identify robust predictors across classifiers, Shapley values were first normalised within each model such that total feature contribution equalled one, and the mean normalised impact of each feature was then computed across the four selected models. As shown in Fig. 5d, the ten most influential features accounted for a substantial proportion of total predictive contribution. The strongest predictor was theta-band absolute power in the right anterior cingulate cortex (11.01%), followed by high-frequency oscillatory activity in the right lateral orbitofrontal cortex (8.93%) and alpha-band power in the left precuneus (5.84%). Additional high-impact features included beta- and delta-band activity in frontal regions and features arising from inferior parietal, visual, and sensorimotor cortices. Both absolute and normalised power measures were represented among the top predictors, indicating complementary contributions of signal magnitude and relative spectral redistribution.

Node-level relative impact was subsequently computed by summing normalised feature contributions within each cortical region (Fig. 5e). This analysis identified a restricted set of regions with disproportionately high influence, led by the right lateral orbitofrontal cortex (13.58%), right medial orbitofrontal cortex (12.77%), and right anterior cingulate cortex (12.45%), followed by the left precuneus (10.01%) and right middle frontal cortex (9.25%). Beyond these leading nodes, relative impact declined across visual, parietal, and sensorimotor cortices, indicating that predictive information was not uniformly distributed. When feature contributions were aggregated across frequency bands, low-frequency oscillatory activity predominated, with alpha-band features accounting for the largest proportion of total model impact (33.93%), followed by theta-band features (20.36%). Collectively, these findings indicate that classification performance was driven primarily by alterations in alpha–theta oscillatory activity within a limited set of frontal and parieto-occipital regions.

To distinguish patients with visual snow syndrome and comorbid migraine (VSS + MIG) from those with migraine only, classification performance was evaluated across multiple machine-learning algorithms using prominent spectral power features. As shown in Fig. 6a, performance was systematically assessed across decreasing numbers of ANOVA-ranked features (50–5), revealing algorithm-specific performance profiles. Models achieving validation and test accuracies, as well as weighted F1 scores, of at least 75% were considered reliable. Based on these criteria, twelve models met the predefined threshold, including a discriminant classifier (5 features), naïve Bayes models (20, 10, and 5 features), support vector machines (SVM; 50 and 5 features), ensemble classifiers (40, 30, 20, and 5 features), and neural network models (20 and 10 features). Receiver operating characteristic analyses (Fig. 6b) further confirmed performance, with all selected models showing area-under-the-curve (AUC) values ≥ 0.75 in both validation (0.7625–0.8481) and independent test sets (0.7696–0.8557), indicating robust and generalisable discrimination between groups.

Fig. 6.

Machine-learning classification of visual snow syndrome with migraine versus migraine-only. (a) Classification performance across algorithms and feature-set sizes showing accuracy and weighted F1 scores for the validation and test data. Discriminant, naïve Bayes, support vector machines (SVM), ensemble, and neural network classifiers exhibited consistently good performance. Asterisks (*) indicate models that met the predefined criteria for optimal classification. (b) Receiver operating characteristic (ROC) curves and corresponding area under curve (AUC) values demonstrating robust discrimination between VSS + MIG and migraine-only patients across these selected models. Asterisks (*) indicate models were selected for optimal classification. (c) Shapley value summary plots illustrating the top five predictive spectral-regional features from the twelve optimal models. (d) Ranked top-10 features based on normalised Shapley impact across models, highlighting the dominant predictors driving classification performance. (e) Node-level relative impact derived from aggregated feature contributions, identifying cortical regions with the highest influence as candidate neurophysiological signatures of VSS. (f) Frequency-band contributions showing the total relative impact of each frequency band across models in distinguishing VSS + MIG from migraine patients. aPSD, absolute power spectral density; nPSD, normalised power spectral density; L, left hemisphere; R, right hemisphere; Hgamma, high gamma; HFO, high-frequency oscillations; IP, inferior parietal lobule; PCu, precuneus; SP, superior parietal cortex; INS, insula; PCC, posterior cingulate cortex; ST, superior temporal cortex; MT, medial temporal cortex; MI, primary motor cortex; MOF, medial orbital frontal cortex

Shapley summary plots illustrate the contribution of the top five features within each model to classification predictions (Fig. 6c). To identify robust predictors across classifiers, the mean normalised Shapley impact of each feature was calculated across the twelve selected models. As shown in Fig. 6d, the ten most influential predictors were dominated by alpha-band absolute power, predominantly localised to posterior cortical regions. The strongest predictor was alpha-band absolute power in the right precuneus (12.12%), followed by alpha-band power in the left precuneus (11.18%) and left superior temporal cortex (10.54%). Additional high-impact predictors included alpha-band activity in bilateral superior parietal cortices, as well as contributions from gamma-band activity in bilateral middle temporal cortex, high-frequency activity in the right precuneus, theta-band power in the left superior temporal cortex, and delta-band power in the left insula. Absolute power measures predominated among the top predictors, indicating that spectral amplitude rather than relative normalisation primarily drove feature-level importance.

When feature contributions were aggregated at the regional level (Fig. 6e), the right precuneus emerged as the most influential cortical node, accounting for 21.29% of the total model impact, followed by the left superior temporal cortex (17.88%) and left precuneus (12.18%). Additional contributions were observed in the bilateral superior parietal cortices and left middle temporal cortex, with progressively smaller effects across medial temporal, insular, inferior parietal, and somatosensory regions, indicating that predictive information was concentrated within a restricted set of posterior and temporal cortical areas. Aggregation across frequency bands (Fig. 6f) revealed a marked predominance of alpha-band activity, which accounted for more than half of the total model impact (56.48%). Theta-band features contributed 15.30%, followed by gamma-band activity (12.09%), whereas high-frequency, delta, and high-gamma bands each accounted for smaller proportions (2.5–8.5%). Collectively, these results demonstrate that classification of VSS + MIG versus migraine-only patients was driven primarily by alpha-band oscillatory alterations, with secondary contributions from theta and gamma activity.

Discussion

Compared with HC, patients with VSS exhibited a broadband shift in absolute spectral power toward enhanced low-frequency cortical activity, most prominently within parieto-occipital and sensorimotor networks. In contrast, normalised spectral power demonstrated a concomitant reduction in high-frequency activity, predominantly involving parieto-occipital and frontal regions. Relative to migraine-only patients, VSS + MIG patients showed increased low-frequency (delta–alpha) activity in absolute spectral power measures. The spatial distribution of these alterations—encompassing frontal, sensorimotor–parietal, and occipital regions—highlights a frequency-dependent redistribution of cortical oscillatory activity in VSS. Using machine learning approaches, VSS was reliably distinguished from HC using SVM and neural network models, achieving classification accuracies exceeding 0.80 and ROC–AUC values ranging from 0.81 to 0.93 with optimized spectral feature sets. The most discriminative features were dominated by low-frequency oscillatory activity, particularly theta- and alpha-band power within frontal, parietal, and sensorimotor cortices. Furthermore, patients with VSS + MIG were successfully differentiated from migraine-only participants using discriminant, naïve Bayes, SVM, ensemble, and neural network classifiers. In this comparison, classification performance was driven primarily by alpha-band features localised to parietal regions, notably the precuneus and superior parietal cortex.

Neurophysiological findings in visual snow syndrome

Neurophysiological investigations employing visual evoked potentials (VEPs), EEG, magnetoencephalography (MEG), and transcranial magnetic stimulation (TMS) have examined cortical function within the occipital and association visual cortices in VSS. These studies provided converging evidence for cortical hyperexcitability, including impaired habituation [38], increased gamma activity [8], and reduced phosphene thresholds [38]. Further findings indicate dysfunction within visual association cortex, characterized by delayed N145 latencies, reduced N75–P100 amplitudes [7], and altered inhibitory network dynamics beyond V1, reflected by reduced alpha power over temporo-parietal regions [9]. Together, these results suggest that aberrant excitability and impaired inhibition within the visual association cortex represent core neurophysiological signatures of VSS.

Neuroimaging studies of VSS, employing structural MRI, functional MRI (resting-state and task-based), and positron emission tomography (PET), have revealed evidences of structural, functional, and metabolic abnormalities across distributed brain networks. Structural MRI investigations demonstrated increased grey matter volume within the visual and lingual cortices, cerebellum, insula, ACC, frontotemporal cortex, middle temporal gyrus, and parahippocampal regions [6, 10, 15, 17]. Functional MRI studies identified hyperconnectivity between visual, attentional, and salience networks [6, 10, 13, 16], while magnetic resonance spectroscopy revealed elevated lactate and glutamate levels in the right lingual gyrus [10], suggesting localized metabolic stress and cortical hyperexcitability. Complementary perfusion imaging using pulsed arterial spin labelling further demonstrated increased baseline cerebral blood flow within visual, parietal, motor, and attention-related cortices [13]. In addition, PET findings confirmed hypermetabolism within the lingual and visual cortices [15, 17]. Collectively, these multimodal findings indicate that VSS is not confined to abnormalities of the visual cortex alone, but rather involves distributed alterations encompassing attentional, limbic, and salience networks—highlighting the widespread cortical dysregulation underlying its neuropathophysiology.

In this study, compared with HC, absolute spectral power in VSS exhibited a broadband shift toward enhanced low-frequency power, most prominently within parieto-occipital and sensorimotor cortices. Such resting-state absolute power elevations likely represent increased baseline cortical activity, indicating an altered neurophysiological state with distributed cortical hyperexcitability. In addition, normalized spectral power demonstrated a concomitant reduction in high-frequency (beta–HFO) activity, extending from frontal regions along the midline to parietal and occipital cortices. This indicates that VSS is not merely characterized as overall increased activity, but involves a spectral redistribution reflecting altered cortical excitatory–inhibitory dynamics, especially within parietal, occipital and frontal networks. These findings support the presence of distributed cortical dysexcitability and disinhibition, reinforcing the notion of impaired thalamo-cortical balance and inefficient sensory modulation. They align with converging structural, metabolic and functional imaging evidence demonstrating altered cortical volume, hypermetabolism, and aberrant network function in VSS [6, 10, 13, 15–17].

Altered cortical spectral activity in visual snow syndrome independent of migraine

Although converging evidences indicates cortical alterations in the neuronal excitability, inhibitory modulation, and network integrity in VSS, potential confounders–such as psychiatric comorbidities (e.g., depression and anxiety) and the high prevalence of migraine overlap–have not been fully disentangled. More importantly, shared cortical abnormalities between VSS and migraine–including cortical dysexcitability, habituation deficit, central sensitization within primary visual, somatosensory, and auditory cortices, and aberrant neural coupling [2, 6–10, 12–17] across the visual, default mode, salience, attention, and sensorimotor networks–further obscure explicit delineation of the neuropathological mechanisms underlying VSS. To address these critical limitations, the present study examined the differences between VSS patients with migraine (n = 25) and a migraine-only cohort (n = 45) carefully matched for CM/EM ratio, sex, age, psychometric scores, and headache frequency. The findings indicate that, independent of migraine comorbidity, VSS is characterized by a spectral redistribution marked by enhanced low-frequency synchronization in absolute power and diminished high-frequency oscillations in normalized power. This altered spectral profile in VSS + MIG patients may reflect disrupted cortical excitability and impaired inhibitory control, extending beyond visual perception to encompass attentional and salience processing. Such alterations likely contribute the perceptual, emotional, and cognitive symptoms characteristic of VSS. The pattern aligns with the framework of thalamocortical dysrhythmia [39], wherein disruption of thalamic relay activity induces excessive low-frequency entrainment and attenuated gamma-band coupling [8], ultimately leading to perceptual instability.

The altered spatial pattern–characterized by widespread cortical increases in absolute power–suggests dysfunction across network. The present study further confirmed these widespread cortical alterations even when comparing VSS with migraine to migraine-only participants. Such abnormalities may underlie visual overload in VSS, whereby early visual areas such as V1 and the lateral occipital cortex amplify spontaneous visual noise into the percept of flickering dots. In addition, impaired higher-order visual processing within the lateral occipital cortex may hinder normal suppression mechanisms, contributing to persistent visual afterimages and palinopsia [8, 16]. Apart from the visual system, the dysregulation of attentional and salience networks in top-down processing for VSS reflected the impair spatial gating, salience misassignment, and functional impairment within parietal, precuneus, and frontal regions [10, 13]. Taken together, spectral power differences of VSS from both healthy and migraine populations underscore the existence of a unique, distributed cortical dysrhythmia underlying VSS, marked by frequency-dependent imbalance and network-level disinhibition.

Cortical signatures underlying VSS

Using machine learning approaches, classifiers reliably distinguished VSS from HC, indicating that VSS is not merely a subjective visual complaint, but a measurable brain disorder with distinct cortical spectral features. In the classification models distinguishing VSS from HC, spectral features within orbitofrontal, ACC, precuneus, and sensorimotor regions emerged as the most predictive contributors, with theta power in the right ACC exerting the dominant influence on model performance. The ACC is critically involved in detecting and prioritizing salient stimuli, with theta-band oscillations facilitating error detection, conflict monitoring, and the enhancement of perceptual salience for otherwise irrelevant visual inputs [40, 41]. Orbitofrontal cortex integrates sensory inputs with emotional and reward processing [42]. In VSS, altered power activity here might reflect emotional distress from visual symptoms, or it could indicate disrupted top-down modulation of visual perception. Studies suggest orbitofrontal cortex hyperactivity in VSS, potentially amplifying perceptual anomalies [11]. The elevated power activity observed within the sensorimotor cortex among VSS patients may reflect impaired cross-modal inhibition, wherein reduced GABAergic modulation weakens inhibitory control over visual pathways [43, 44]. Such disinhibition could contribute to cortical hyperexcitability and aberrant visual–somatosensory coupling. Furthermore, in migraine, both the ACC and sensorimotor cortices have been implicated in cortical dyexcitability, sensory disinhibition, and disrupted pain-related network integration [32, 34, 45, 46]. Notably, given the high comorbidity rate of migraine among individual with VSS, therefore, these overlapping alterations likely contribute to the predictive features identified by the models in distinguishing VSS from HC.

In the classification models distinguishing VSS with migraine from migraine-only patients, alpha power within the precuneus, superior temporal, and superior parietal cortices emerged as the most predictive features, delineating a core cortical signature of VSS that is independent of migraine pathology. The parietal cortex plays a critical role in visual-spatial attention, multisensory integration, and top-down suppression of irrelevant visual input, functions disrupted in VSS patients. Puledda et al. [12, 14] demonstrated aberrant activation in inferior and superior parietal regions during visual snow simulation tasks, reflecting impaired visuospatial attention that normally filters irrelevant visual input and maintains perceptual stability. This dysfunction may lead to over-attention to aberrant visual noise, perpetuating the continuous percept of visual snow. Strik et al. [16] identified reduced local efficiency in the superior parietal region accompanied by hyperconnectivity with the occipital cortex, suggesting deficient multisensory integration—a failure to merge visual and proprioceptive information—manifesting clinically as photophobia and heightened sensory sensitivity. Meanwhile, Latini et al. [47] reported white matter abnormalities in the sagittal stratum, a tract linking parietal and occipital regions, indicating defective top-down suppression of primary visual cortex activity, which may underlie palinopsia and entoptic phenomena such as afterimages and floaters. Furthermore, superior temporal cortex involved in auditory processing, language, and multisensory integration, the altered alpha features highlight polymodal nature in VSS—many patients report tinnitus or auditory hypersensitivity alongside visual symptoms [2, 15]. Together, these findings highlight the parietal cortex as a critical hub in the neuropathophysiology of VSS, where disrupted visuospatial attention, impaired multisensory integration, and weakened top-down inhibitory control converge to produce the core perceptual disturbances characteristic of the syndrome.

The enhancement of low-frequency alpha and theta activity in the parietal cortex in VSS may reflect the interplay of four interrelated mechanisms. First, top-down attentional overdrive may occur, wherein the parietal cortex enhances theta–alpha oscillations to maintain visual focus and suppress irrelevant input—its normal role in attentional gating [48, 49]. In VSS, however, the persistent visual noise drives sustained but ineffective theta–alpha engagement, resulting in compensatory overactivation. Second, failed inhibition due to excitatory–inhibitory imbalance may contribute, as reduced GABAergic inhibition within the parietal cortex diminishes alpha suppressive control over irrelevant visual activity [50–52], while theta power increases as a default low-frequency mode, leading to overall enhancement of low-frequency synchronization. Third, network dysrhythmia may emerge from thalamocortical loop entrainment in theta frequencies [39, 53], secondary to occipital hyperexcitability, thereby disrupting normal cortical resetting and sustaining pathological theta–alpha synchronization. Fourth, compensatory hyperperfusion in the parietal cortex, as shown by increased cerebral blood flow in VSS [14], may support the elevated metabolic demand associated with persistent oscillatory overactivity, further amplifying low-frequency power.

Low-frequency parietal activity may constitute a pivotal neurophysiological signature underlying the pathophysiology of VSS. This finding suggests potential therapeutic targets aimed at restoring cortical excitability balance and improving sensory–attentional regulation. First, alpha-regulating neurofeedback could be employed to enhance parieto-occipital alpha oscillatory control, thereby reinforcing inhibitory gating mechanisms and suppressing the intrusion of irrelevant visual percepts [48, 49]. Second, parietal transcranial magnetic stimulation (TMS) protocols designed to downregulate pathological theta activity may modulate top-down attentional control and recalibrate thalamocortical rhythmicity, potentially normalizing dysrhythmic network activity [54, 55]. Third, GABAergic modulation, either through pharmacological enhancement of inhibitory neurotransmission or noninvasive interventions that increase cortical inhibition, may help rebalance the excitatory–inhibitory dynamics [56, 57] and attenuate visual hyperexcitability. Collectively, these approaches offer a rational framework for developing frequency-specific neuromodulation strategies that directly target the cortical and thalamocortical dysrhythmia underlying VSS, warranting systematic investigation in future clinical trials.

Limitations

This study has several limitations. First, it was conducted at a single center and used a cross-sectional design, which limits generalisability for relatively small cohort size and precludes inferences regarding symptom fluctuation (due to the phase of migraine) or longitudinal progression. Larger, multi-site investigations using standardised acquisition and analysis pipelines and follow-up assessments will be needed to translate aberrant cortical oscillatory signatures into clinically reliable biomarkers. Second, migraine is a cyclic disorder [31], yet the relevance of dynamic sensory cortex excitability to VSS symptom variation remains uncertain. Third, comprehensive interrogation of thalamocortical dysrhythmia — for example using phase–amplitude coupling — is required to strengthen mechanistic interpretation in VSS. Fourth, although potential neuromodulatory targets are implicated here, optimal stimulation parameters for parietal-targeted TMS remain undetermined, and systematic dose–response experiments will be required to confirm neuromodulatory efficacy. Finally, although Shapley analyses were employed to highlight prominent predictors, how these spectral features relate causally to the underlying pathophysiology remains to be established.

Conclusion

These findings demonstrate that VSS is characterised by neurophysiological substrates that extend beyond the migraine spectrum, supporting the view that VSS and migraine, although frequently comorbid, are pathophysiologically dissociable. The machine-learning results further indicate that spectral signatures — particularly theta–alpha dynamics within parietal cortices — are both robust and discriminative for VSS classification, underscoring their potential translational value as physiological signatures. Taken together, these observations suggest that modulation of parietal hyperexcitability, inhibitory control, and associated neurochemical processes warrants further investigation as a candidate treatment target in VSS.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

F.J. Hsiao, L.L.H. Pan, S.J. Wang contributed to the study design. F.J. Hsiao, W.T. Chen, S.P. Chen, Y.F. Wang, K.L. Lai, and S.J. Wang recruited patients and collected data. F.J. Hsiao, F. Puledda, W.T. Chen, S.P. Chen, L.L.H. Pan, Y.F. Wang, K.L. Lai, and S.J. Wang performed data analysis and interpretation. F.J. Hsiao, F. Puledda, and G. Coppola wrote the manuscript. F.J. Hsiao, F. Puledda, W.T. Chen, G. Coppola, and S.J. Wang critically reviewed the article. All authors interpreted the data, reviewed the manuscript, and approved the final version.

Funding

This study was founded by the National Science and Technology Council (112-2321-B-075-007 and 114-2321-B-A49-016 to SJ Wang, and 112-2221-E-A49 -012 -MY2 and 114-2221-E-A49-080 to FJ Hsiao). The funders had no role in the study design, data collection and analysis, decision to publish, or manuscript preparation.

Data availability

Derived data supporting the findings of this study are available on request from the corresponding authors.

Declarations

Ethics approval and consent to participate

The study protocol was conducted in accordance with Declaration of Helsinki and was approved by the Institutional Review Board of Taipei Veterans General Hospital (IRB number: 2022-02-010 A). Written informed consent was obtained from all participants prior to enrolment.

Competing interests

FJ Hsiao, F Puledda, WT Chen, YT Wu, LLH Pan, YF Wang, SP Chen, KL Lai, and G Coppola declare no potential conflicts of interest. SJ Wang reports grants and personal fees from Norvatis Taiwan, personal fees from Daiichi-Sankyo, grants and personal fees from Eli-Lilly, personal fees from AbbVie/Allergan, personal fees from Pfizer Taiwan, personal fees from Biogen, Taiwan, outside the submitted work.