Normative structural MRI deviation factors summarize patient-level neuroanatomical heterogeneity across episodic and chronic migraine

Abstract

Background

Structural MRI findings in migraine are inconsistent across cohorts, suggesting that group contrasts may obscure patient-level heterogeneity. We tested whether individualized normative structural deviations could be summarized into interpretable patient-level factor-weight dimensions related to chronicity and clinical burden.

Methods

In this single-center cross-sectional study, 249 adults with ICHD-3 migraine (198 episodic, 51 chronic) and 57 local healthy controls underwent 3.0-T MRI. Released PCNtoolkit lifespan models were locally adapted with controls to estimate patient-level deviation z scores for 158 FreeSurfer features. Non-negative matrix factorization was applied to the absolute-deviation matrix; directional negative- and positive-deviation models were sensitivity analyses. Covariate-adjusted associations were FDR-corrected within prespecified families.

Result

The primary model identified two stable factor-weight dimensions. Factor 1 captured a subcortical-limbic-brainstem/global-volume deviation pattern and was associated with chronic migraine (OR per 1-SD, 1.50; q = 0.036), MIDAS (beta = 16.46; q = 0.007), and monthly headache days (beta = 9.41; q < 0.001). Factor 2 mainly reflected cortical-thickness deviations, showed a borderline inverse association with chronic migraine (OR = 0.68; q = 0.054), and was associated with high absolute-deviation burden (OR = 2.37; q < 0.001). HIT-6 showed no corrected factor-specific association. Removing cranial-size-sensitive features preserved factor structure and main Factor 1 associations.

Conclusion

Normative structural MRI deviations in migraine can be summarized into non-mutually exclusive deviation-burden factors with patient-level weights. Factor 1 showed the most consistent associations with chronic migraine, headache frequency, and MIDAS; these associations attenuated in restricted sensitivity analyses, and external validation is required before clinical application.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s10194-026-02464-2.

Background

Migraine is a common and disabling neurological disorder with substantial clinical heterogeneity. Global Burden of Disease 2021 estimates identify migraine as one of the leading contributors to neurological disability worldwide [1], and contemporary disease-primer evidence emphasizes its recurrent, multidimensional, and disabling nature [2]. Current clinical categories, including episodic migraine and chronic migraine, are useful for diagnosis and treatment planning, but they are unlikely to capture all biologically relevant variation. This is clinically important because preventive treatment response, progression to chronic migraine, and disability vary markedly across patients even within the same diagnostic category.

Structural MRI studies have reported differences in gray matter volume, cortical thickness, cortical surface area, and subcortical structures in migraine, but the direction and location of findings have varied across cohorts and analytic pipelines [3, 4]. A coordinate-based meta-analysis found no reliable whole-brain gray matter morphological signature for migraine and emphasized heterogeneity, clinical variation, and methodological differences as major barriers to replication [5]. Recent systematic and large-cohort neuroimaging evidence has likewise suggested distributed, clinically heterogeneous alterations rather than a single focal abnormality [6–8]. These findings argue against a simple lesion-like model of migraine anatomy.

Group-level contrasts may therefore be poorly suited to a disorder in which chronicity, aura, medication exposure, medication-overuse headache, disease duration, mood, sleep, and disability differ across individuals [9]. For headache medicine, the relevant question is not only whether people with migraine differ from healthy controls on average, but whether individualized structural deviations can be organized into reproducible patterns that relate to clinically meaningful burden.

Normative modeling addresses this problem by estimating the expected distribution of a brain phenotype for an individual with specified covariates and quantifying that person’s deviation from expectation [10–13]. At the feature level, deviation scores express each observed structural value relative to the normative predicted mean and predictive uncertainty. This approach shifts the inferential target from group mean differences to patient-level departures from normative structure. However, deviation maps remain high dimensional and require principled reduction before they can be interpreted clinically.

We used non-negative matrix factorization (NMF) to summarize individualized structural MRI normative deviations. NMF represents each patient as an additive combination of non-negative latent components and each component as a non-negative loading profile across imaging features [14–16]. For absolute deviations, non-negativity is particularly appropriate because deviation-burden contributions across regions are additive and do not cancel each other. This mixed-membership representation differs from k-means clustering, which forces each patient into one cluster, and from principal-component or independent-component methods, whose signed components can be less directly interpretable as burden profiles. Importantly, NMF does not by itself establish discrete subtypes or clinical classes.

We tested two primary hypotheses: (1) patients with migraine would show more than one latent neuroanatomical deviation pattern; and (2) patient-level factor weights would be differentially associated with chronic migraine status and selected clinical outcomes, with particular attention to MIDAS, HIT-6, and monthly headache days. We did not train a new normative model from scratch. Instead, locally adapted pretrained PCNtoolkit lifespan models were used to generate patient-level deviation scores before factorization. The primary analysis used the absolute-deviation matrix to capture total deviation magnitude irrespective of direction. Negative- and positive-deviation matrices, high-deviation-burden analyses, preventive-medication-free analyses, and incremental model comparisons were prespecified exploratory or sensitivity analyses. The overall aim was to map structural heterogeneity, not to establish causal mechanisms, diagnostic biomarkers, or treatment-selection tools.

Methods

Study design, setting, and participants

This was a cross-sectional analysis of a single-center observational MRI cohort at the Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University, Wenzhou, Zhejiang, China. Patients were consecutively recruited from the Headache Clinic between April 2024 and March 2026. Eligibility was restricted to participants aged 18–65 years. The final migraine sample included 249 patients, comprising 198 patients with episodic migraine and 51 with chronic migraine. Fifty-seven locally recruited healthy controls constituted the local healthy-control calibration set for adaptation of the pretrained normative model and for baseline description (Supplementary Figure S1); they were not included in NMF factor discovery or patient-level association models. Healthy controls were required to have no migraine, no other primary or secondary headache disorder, no chronic pain condition, no major psychiatric or neurologic disorder, no cardiovascular or metabolic disease history, and no headache-related, psychotropic, neurologic, chronic-pain, cardiovascular, or metabolic medication use within 3 months before MRI.

Migraine diagnoses were assigned by board-certified headache neurologists using structured interviews based on the International Classification of Headache Disorders, 3rd edition [17]. Episodic migraine was defined as migraine with or without aura and fewer than 15 headache days per month during the preceding 3 months. Chronic migraine and medication-overuse headache were diagnosed according to ICHD-3; medication-overuse headache was retained as a clinically relevant comorbidity rather than used as an exclusion criterion.

Clinical variables included aura status, disease duration, monthly headache days, medication-overuse headache status, preventive medication exposure, pain intensity, mood, anxiety, sleep, and disability measures. Validated instruments included the Headache Impact Test-6 (HIT-6), Migraine Disability Assessment (MIDAS), Patient Health Questionnaire-9 (PHQ-9), Generalized Anxiety Disorder-7 (GAD-7), and Pittsburgh Sleep Quality Index (PSQI) [18–22]. Pain intensity was assessed using a visual analogue scale (VAS). Monthly headache days referred to average headache days per month over the preceding 3 months. Preventive medication exposure was defined as ongoing migraine-preventive treatment with a stable regimen for at least 3 months before MRI.

Exclusion criteria included other primary or secondary headache disorders except medication-overuse headache, major neurologic, systemic, or psychiatric disorders, substance or alcohol abuse, MRI contraindications, significant structural brain abnormalities, white matter hyperintensities greater than Fazekas grade 1 [23, 24], image-quality failure, or headache attack within 48 h before or 24 h after MRI [25]. Participant flow is summarized in Supplementary Figure S1. Post-MRI exclusions were determined before statistical outcome modeling: white matter hyperintensities greater than Fazekas grade 1 (n = 8), motion artifact or poor image quality (n = 8), and headache attack within 48 h before or 24 h after MRI (n = 12; total n = 28).

The study was approved by the Medical Ethics Committee of the Second Affiliated Hospital of Wenzhou Medical University (approval No. 2024-K-137-01) and was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.

MRI acquisition and structural preprocessing

MRI was performed on the same 3.0-T scanner (Ingenia Elition X, Philips Healthcare, Best, the Netherlands) using a 32-channel head coil during verified interictal sessions. High-resolution T1-weighted images were acquired for morphometric analysis, and axial T2-FLAIR images were acquired for structural review and Fazekas grading. Detailed acquisition parameters are provided in Additional file 1.

T1-weighted images were processed with the standard FreeSurfer version 7.3.2 recon-all stream. Extracted measures included cortical thickness, cortical and global volumes, subcortical volumes, ventricular and cerebrospinal-fluid-related measures, and estimated total intracranial volume (eTIV). Quality control was prespecified and performed without using group labels or clinical outcomes. Detailed preprocessing, quality-control criteria, and the feature inventory are provided in Additional file 1.

Normative deviation scores

The analysis workflow proceeded in four steps: local adaptation of released PCNtoolkit normative models, patient-level deviation z score estimation, NMF-based factor-weight summarization, and covariate-adjusted clinical association testing. A schematic overview of this workflow is provided in Supplementary Figure S2. Patient-level structural deviation scores were generated using released PCNtoolkit lifespan brain-chart normative models rather than models retrained in the present cohort [10, 11, 26]. The pretrained models were locally adapted to the present scanner context using a predefined 57-control local calibration set. This calibration set was selected before outcome modeling as the largest eligible healthy-control subset satisfying prespecified balance thresholds for age, sex, and eTIV against the final migraine cohort. Control selection was blind to episodic or chronic migraine status, disability, headache frequency, medication-overuse headache, preventive medication exposure, NMF results, and all downstream outcomes.

For each FreeSurfer-derived structural feature, the adapted model estimated an individualized expected value and deviation z score. Because eTIV could not be added as a calibration covariate in the released adaptation workflow, the calibration set was balanced for eTIV, eTIV was included as a downstream covariate, and reduced-feature sensitivity analyses excluded global volume, CSF-related, ventricular, choroid plexus, and other cranial-size-sensitive features. This approach is consistent with prior PCNtoolkit transfer-learning work using modest site-specific healthy-control adaptation sets, while residual calibration uncertainty remains possible [12]. Leave-one-out healthy-control calibration diagnostics (Supplementary Figure S3) and full model-provenance, adaptation, feature-compatibility, and calibration-set selection details are provided in Additional file 1.

Deviation matrices and NMF

Three non-negative patient-by-feature deviation matrices were constructed from signed z scores: absolute deviation magnitude, lower-than-expected deviation magnitude, and higher-than-expected deviation magnitude. The absolute-deviation matrix was prespecified as the primary matrix because migraine-related morphometric deviations may be bidirectional and heterogeneous across patients. Directional negative- and positive-deviation matrices were analyzed as sensitivity decompositions. High-deviation burden outcomes were derived from extreme-deviation counts and treated as exploratory internal-organization outcomes of the same deviation space.

Non-negative matrix factorization was applied to summarize patient-level deviation profiles. The resulting patient-by-factor weight matrix was interpreted as cross-sectional mixed-membership descriptors of deviation burden within each patient, whereas the factor-by-feature loading matrix was used to describe anatomical composition. Dominant-factor labels were used only descriptively and were not interpreted as discrete subtypes. Candidate models with one to six factors were compared using reconstruction performance, repeated-run reproducibility, bootstrap stability, split-half recovery, parsimony, and clinical interpretability, with implementation details and methodological references provided in Additional file 1 [27–30]. No independent external validation cohort was available; therefore, factor labels and clinical associations were considered exploratory until replicated.

Statistical analysis and multiplicity control

The primary analysis used the full migraine cohort and the absolute-deviation NMF factors. The primary clinical targets were chronic migraine status, MIDAS, HIT-6, and monthly headache days. Chronic migraine status was tested using Firth penalized logistic regression with standardized factor weights and adjustment for age, sex, and eTIV [31, 32]. MIDAS and HIT-6 were tested using linear regression adjusted for age, sex, eTIV, and chronic migraine status. Monthly headache days were adjusted for age, sex, and eTIV, but not chronic migraine status, because headache-day frequency partly defines chronic migraine. Raw outcomes and rank-based inverse normal transformed outcomes were analyzed in parallel [33].

Disease duration was not included as a primary-model covariate because it may partly reflect accumulated migraine burden or lie on pathways linking chronicity, structural deviation, and clinical outcomes; it was therefore evaluated in sensitivity analysis. False-discovery-rate correction was applied within prespecified hypothesis families. Directional deviation matrices, high-deviation-burden outcomes, disease-duration adjustment, mood- and sleep-adjusted models, medication-overuse-headache restriction, preventive-medication-free restriction, without-aura restriction, factor-weight collinearity diagnostics, reduced-feature NMF, and fixed-weight incremental model comparisons were treated as sensitivity or exploratory analyses [34, 35]. Detailed model specifications, multiplicity families, detectable-effect sensitivity methods, and complete outputs are provided in Additional file 1 and the Supplementary Tables.

Results

Baseline characteristics

Baseline characteristics are shown in Table 1. The analytic sample included 57 healthy controls, 198 patients with episodic migraine, and 51 patients with chronic migraine. The 57-control calibration set was comparable to the migraine groups in sex distribution and eTIV, whereas the chronic migraine group was older than the episodic migraine group and healthy controls. Compared with episodic migraine, chronic migraine was characterized by longer disease duration, more monthly headache days, greater headache impact and disability, and higher frequencies of medication-overuse headache and preventive medication exposure. Preventive medication classes are summarized in Supplementary Table S5G. Migraine with aura and pain intensity did not differ significantly between episodic and chronic migraine.

Table 1.

Demographic and clinical characteristics across healthy controls, episodic migraine, and chronic migraine

Characteristic	HC (n = 57)	EM (n = 198)	CM (n = 51)	P value
Age, years	34.0 +/- 9.7	32.8 +/- 8.2	40.5 +/- 10.2	< 0.001
Female sex, n (%)	47 (82.5%)	166 (83.8%)	43 (84.3%)	0.960
Estimated intracranial volume, ml	1403.4 +/- 164.6	1390.1 +/- 186.4	1380.5 +/- 217.9	0.614
Migraine with aura, n (%)	-	41 (20.7%)	5 (9.8%)	0.104
Medication-overuse headache, n (%)	-	1 (0.5%)	18 (35.3%)	< 0.001
Preventive medication exposure, n (%)	-	17 (8.6%)	12 (23.5%)	0.006
PHQ-9	2.0 [1.0–4.0]	5.0 [3.0–9.0]	7.0 [5.0-10.5]	< 0.001
GAD-7	1.0 [0.0–3.0]	5.0 [2.0–8.0]	4.0 [1.0-8.5]	< 0.001
PSQI	5.0 [3.0–7.0]	7.0 [5.0–10.0]	9.0 [6.0–12.0]	< 0.001
Disease duration, years	-	7.0 [3.0–10.0]	10.0 [6.0–20.0]	0.002
Pain intensity, VAS	-	6.0 [6.0–8.0]	7.0 [6.0–8.0]	0.182
Monthly headache days	-	3.0 [2.0–5.0]	19.0 [15.0-25.5]	< 0.001
HIT-6	-	52.0 [49.0–56.0]	66.0 [62.0–68.0]	< 0.001
MIDAS	-	9.0 [6.0–15.0]	50.0 [45.0-75.5]	< 0.001

Note. Values are n (%), mean +/- SD, or median [interquartile range]. P values compare healthy controls, episodic migraine, and chronic migraine for variables available in all groups; disease-specific variables compare episodic and chronic migraine. Kruskal-Wallis tests were used for three-group continuous variables, chi-square tests for sex, and Mann-Whitney U tests for disease-specific EM-CM comparisons. Monthly headache days refer to average headache days per month over the preceding 3 months. Values below 0.001 are reported as < 0.001. HC, healthy control; EM, episodic migraine; CM, chronic migraine; NMF, non-negative matrix factorization; VAS, visual analogue scale

A stable two-factor solution provided a low-dimensional summary of the absolute-deviation matrix

The primary analysis included 249 patients with migraine. Across one to six factors, reconstruction error decreased as factor number increased. The K = 1 single-factor benchmark had reconstruction error 112.33, explained-variance-like value 0.063, repeated-run reproducibility 1.00, and 1000-bootstrap top-feature Jaccard stability 0.599; silhouette and consensus separation were not applicable because only one dominant factor exists. The primary absolute-deviation model selected two factors (Fig. 1; Supplementary Table S1), reducing reconstruction error to 109.19 and increasing the explained-variance-like value to 0.114, with silhouette 0.426 and 1000-bootstrap top-feature Jaccard stability 0.634. Thus, K = 2 captured approximately 11% of the total deviation-matrix variance and provided stable but low-dimensional summary axes, not a comprehensive explanation of structural deviation variance. The remaining variance likely reflects additional weak dimensions of heterogeneity, individual-specific deviation patterns, measurement noise, and the high dimensionality of FreeSurfer feature space.

Fig. 1.

Model selection for the primary absolute-deviation factor analysis. Model selection metrics for one to six factors in the full migraine cohort. Panels show reconstruction error, explained-variance-like value, silhouette score, and 1000-bootstrap top-feature stability. K = 1 was included as a single-factor benchmark; silhouette was not applicable for K = 1 because no between-factor separation exists. The final primary model selected two factors, balancing reconstruction, stability, separation, and parsimony. EV-like indicates explained-variance-like metric; NMF, non-negative matrix factorization

The K = 1 benchmark also did not show evidence of the main clinical signal: chronic migraine status, Firth odds ratio per 1-SD increase = 1.04 (95% CI, 0.76 to 1.43; p = 0.808; q = 0.808), MIDAS beta = 0.98 (95% CI, -1.02 to 2.99; p = 0.337; q = 0.505), HIT-6 beta = 0.13 (95% CI, -0.44 to 0.69; p = 0.662; q = 0.662), and monthly headache days beta = 0.45 (95% CI, -0.44 to 1.33; p = 0.320; q = 0.505; Supplementary Table S1). Although K = 3 modestly reduced reconstruction error relative to K = 2, it showed lower silhouette (0.381 vs. 0.426), lower 1000-bootstrap factor reproducibility (0.921 vs. 0.982), and lower 1000-bootstrap top-feature stability (0.467 vs. 0.634). K = 3 and K = 4 sensitivity analyses did not identify a more stable or clinically clearer alternative; instead, the main clinical signal was redistributed across factors that matched the two primary dimensions.

A 500-repeat random split-half factor-recovery analysis further supported internal stability of the primary factorization. After independent NMF in two 50/50 halves and factor matching by cosine similarity with Hungarian assignment, split-half loading recovery was high for both factors: Factor 1 mean cosine similarity = 0.930 (median, 0.932; empirical 95% range, 0.898 to 0.952; >=0.90 in 96.4% of splits) and Factor 2 mean cosine similarity = 0.968 (median, 0.970; empirical 95% range, 0.949 to 0.978; >=0.90 in 100% of splits). The direction of the Factor 1 association was also consistent across split halves for MIDAS in 99.0% of splits and for monthly headache days in 99.6% of splits (Supplementary Table S1).

Feature-family balanced and K-sensitivity analyses supported the overall two-factor interpretation, with the main clinical signal remaining concentrated on the subcortical-limbic-brainstem/global volume deviation axis. These analyses are reported as sensitivity checks because they evaluate robustness of the factor structure rather than independent validation.

To address possible confounding by eTIV, global volume, cerebrospinal-fluid-related, and ventricular/choroid features, we refit the K = 2 absolute-deviation NMF after excluding 13 such features, leaving 145 structural features. The reduced factors closely matched the original primary factors (cosine similarity 0.998 for Factor 1 and 0.999 for Factor 2). The Factor 1-matched reduced factor retained associations with chronic migraine status (odds ratio per 1-SD increase, 1.59; 95% CI, 1.13 to 2.25; p = 0.008; q = 0.016), MIDAS (beta per 1-SD increase, 3.07; 95% CI, 1.02 to 5.13; p = 0.004; q = 0.011), and monthly headache days (beta per 1-SD increase, 1.84; 95% CI, 0.96 to 2.71; p < 0.001; q < 0.001). Full sensitivity outputs are provided in the supplementary tables.

Benchmark analyses indicated that the main anatomical axes were not unique to NMF, whereas the patient representation depended on the method. Principal-component analyses recovered broadly similar loading dimensions, but two-cluster k-means did not provide comparably interpretable clinical associations. These findings support the use of NMF factor weights as mixed-membership descriptors rather than a hard-clustering subtype model.

The two primary factors captured distinct deviation-burden profiles. Factor loadings are shown in Fig. 2, and anatomical composition is summarized in Fig. 3. Factor 1 was dominated by subcortical, limbic, brainstem, and global volume-related measures, including subcortical gray matter volume, bilateral ventral diencephalon, hippocampus, pallidum, brainstem, and supratentorial volume measures. Factor 2 was dominated by cortical-thickness features in parietal, frontal, and occipital regions, including supramarginal, precuneus, middle frontal, superior parietal, occipital, calcarine, and angular cortical thickness features.

Fig. 2.

Latent absolute-deviation factor loadings. Cortical surface maps and top-loading structural regions for the primary two-factor absolute-deviation model. Panel a maps cortical-thickness NMF loadings to the fsaverage5 surface using the FreeSurfer Destrieux atlas. Non-cortical volume, subcortical, ventricular, cerebrospinal-fluid-related, and global features are not displayed on the cortical surface and are summarized in the top-feature panel. Panel b shows top-loading features across all structural feature families. Factor 1 was dominated by subcortical, limbic, brainstem, and global volume-related measures, whereas Factor 2 was dominated by cortical-thickness features in parietal, frontal, and occipital regions. L, left; R, right; g., gyrus; s., sulcus; ant, anterior; inf, inferior; sup, superior; CSF, cerebrospinal fluid

Fig. 3.

Anatomical composition of latent deviation factors. Anatomical and feature-family composition of the two primary absolute-deviation factors. Loading proportions were computed from normalized feature loadings and grouped into rule-based anatomical categories

Absolute-deviation factors were associated with chronic migraine status

In the primary absolute-deviation model, Firth logistic regression showed an association between Factor 1 and chronic migraine status (odds ratio per 1-SD increase, 1.50; 95% CI, 1.07 to 2.09; p = 0.018; q = 0.036; Table 2; Supplementary Table S2). Factor 2 showed a directionally consistent but borderline inverse association with chronic migraine status (odds ratio per 1-SD increase, 0.68; 95% CI, 0.46 to 1.01; p = 0.054; q = 0.054). Directional negative- and positive-deviation factors did not show FDR-significant chronic migraine associations.

Table 2.

Primary chronification and clinical-burden results

Analysis	Matrix	Factor or model	Outcome	Effect (95% CI)	P value	FDR q
Chronification	Absolute deviation	Factor 1	Chronic migraine status, Firth model	OR per 1 SD = 1.50 (1.07 to 2.09)	0.018	0.036
Chronification	Absolute deviation	Factor 2	Chronic migraine status, Firth model	OR per 1 SD = 0.68 (0.46 to 1.01)	0.054	0.054
Clinical burden	Absolute deviation	Factor 1	MIDAS	beta = 16.46 (5.95 to 26.97)	0.002	0.007
Clinical burden	Absolute deviation	Factor 1	MIDAS, rank-transformed	beta = 0.84 (0.34 to 1.34)	0.001	0.003
Clinical burden	Absolute deviation	Factor 2	MIDAS, rank-transformed	beta = -0.82 (-1.34 to -0.31)	0.002	0.004
Clinical burden	Absolute deviation	Factor 1	Monthly headache days	beta = 9.41 (4.91 to 13.90)	< 0.001	< 0.001
Clinical burden	Absolute deviation	Factor 2	Monthly headache days	beta = -6.23 (-10.98 to -1.48)	0.010	0.021

Note. Effects are beta coefficients for linear models or odds ratios per 1-SD factor-weight increase for Firth logistic models. P values are raw model p values; q values are family-wise Benjamini-Hochberg FDR values. Abs, absolute deviation; MHD, monthly headache days; INT, rank-based inverse normal transformed outcome

In the non-medication-overuse-headache sensitivity set, 230 participants were analyzable, including 33 patients with chronic migraine. The Factor 1 association with chronic migraine status was directionally consistent but imprecise (odds ratio per 1-SD increase, 1.43; 95% CI, 0.99 to 2.07; p = 0.058; q = 0.058). In the preventive-medication-free sensitivity analysis, the association remained directionally consistent but did not survive FDR correction (odds ratio per 1-SD increase, 1.49; 95% CI, 1.05 to 2.14; p = 0.028; q = 0.055). Disease-duration-adjusted sensitivity models yielded similar full-cohort estimates for Factor 1 and chronic migraine status (odds ratio per 1-SD increase, 1.50; 95% CI, 1.08 to 2.09; p = 0.017; q = 0.033; Supplementary Table S5F). In the without-aura sensitivity analysis, the effect estimate was similar but attenuated after family-wise correction (odds ratio per 1-SD increase, 1.48; 95% CI, 1.02 to 2.14; p = 0.037; q = 0.074). Thus, restricted-subgroup analyses supported the direction of the full-cohort association, but the chronic-migraine association was not fully robust after medication, medication-overuse, or aura-status restriction. This attenuation may reflect either genuine weakening of the association or reduced statistical power in smaller restricted subgroups. Additional exploratory adjustment for PHQ-9, GAD-7, and PSQI did not materially change the full-cohort absolute Factor 1 association with chronic migraine status (odds ratio per 1-SD increase, 1.52; 95% CI, 1.08 to 2.16; p = 0.018; q = 0.036).

Detectable-effect sensitivity analysis indicated that, under this approximation, the full cohort could detect odds ratios of approximately 1.62 or larger per 1-SD factor-weight increase for chronic migraine status. Restricted sets required larger effects: 1.79 in the non-medication-overuse-headache set, 1.72 in the preventive-medication-free set, and 1.68 in the without-aura set. The corresponding observed Factor 1 estimates in these restricted sets were similar in direction but below these approximate detectable-effect ranges, supporting interpretation as imprecise rather than confirmatory negative findings.

MIDAS and monthly headache days, but not HIT-6, showed the clearest factor associations

Primary clinical associations were concentrated in the absolute-deviation matrix and, to a lesser extent, the positive-deviation matrix (Fig. 4; Table 2; Supplementary Table S3). These associations were primarily observed for headache frequency and MIDAS, not consistently for HIT-6. In the full migraine cohort, absolute-deviation Factor 1 was associated with MIDAS in the raw model (beta, 16.46; 95% CI, 5.95 to 26.97; p = 0.002; q = 0.007) and the rank-transformed model (beta, 0.84; 95% CI, 0.34 to 1.34; p = 0.001; q = 0.003). In monthly-headache-days models adjusted for age, sex, and eTIV but not chronic migraine status, Factor 1 was positively associated with monthly headache days (beta, 9.41; 95% CI, 4.91 to 13.90; p < 0.001; q < 0.001). Factor 2 was negatively associated with monthly headache days (beta, -6.23; 95% CI, -10.98 to -1.48; p = 0.010; q = 0.021) and rank-transformed MIDAS (beta, -0.82; 95% CI, -1.34 to -0.31; p = 0.002; q = 0.004).

Fig. 4.

Associations between deviation factors and selected primary clinical outcomes. Forest plot of prespecified rank-transformed primary clinical associations in the full migraine cohort, shown for MIDAS, HIT-6, and monthly headache days across absolute-, negative-, and positive-deviation factor decompositions. Points indicate regression coefficients for factor weights; horizontal lines indicate 95% confidence intervals. Filled points indicate false-discovery-rate-significant associations, and open points indicate nonsignificant associations. P values are raw model p values; q values are Benjamini-Hochberg false discovery rate corrected within the prespecified clinical-outcome family for each matrix. FDR, false discovery rate; HIT-6, Headache Impact Test-6; MIDAS, Migraine Disability Assessment

Positive-deviation Factor 1 was negatively associated with rank-transformed MIDAS (beta, -0.50; 95% CI, -0.85 to -0.14; p = 0.006; q = 0.032) and rank-transformed monthly headache days (beta, -0.55; 95% CI, -0.98 to -0.13; p = 0.011; q = 0.032). HIT-6 did not emerge as a factor-specific signal after FDR correction in the condensed main results.

Distribution-sensitive models supported the direction of the primary MIDAS and monthly-headache-days findings. For MIDAS, robust linear and median quantile regression retained a positive Factor 1 association and a negative Factor 2 association. For monthly headache days, negative-binomial and scaled quasi-Poisson models retained a positive Factor 1 incidence-rate ratio and a negative Factor 2 incidence-rate ratio. These analyses were treated as sensitivity checks rather than replacement primary models.

Exploratory non-medication-overuse-headache clinical sensitivity analyses showed a directionally similar pattern for Factor 1, but with wider uncertainty because the number of chronic migraine cases was smaller. This restricted analysis should be interpreted as a power-limited sensitivity check requiring external replication, not as evidence that the non-medication-overuse-headache subgroup definitively preserves or refutes the primary chronification association. Overall, the clinical signal was primarily associated with headache frequency and MIDAS, not consistently with HIT-6, and was more closely related to total or direction-specific deviation composition than to a unidirectional negative-deviation model alone.

High-deviation burden aligned within specific factor spaces

High-deviation burden aligned with specific latent factor spaces (Fig. 5; Supplementary Table S4). In the primary absolute-deviation model, Factor 2 was associated with high absolute burden (odds ratio per 1-SD increase, 2.37; 95% CI, 1.68 to 3.34; p < 0.001; q < 0.001) and with the subject-level high-deviation profile (odds ratio per 1-SD increase, 2.92; 95% CI, 1.81 to 4.71; p < 0.001; q < 0.001). In the negative-deviation matrix, Factor 1 was associated with high negative burden (odds ratio per 1-SD increase, 3.15; 95% CI, 2.19 to 4.53; p < 0.001; q < 0.001), and Factor 2 was also associated with high negative burden (odds ratio per 1-SD increase, 2.34; 95% CI, 1.68 to 3.26; p < 0.001; q < 0.001).

Fig. 5.

High-deviation burden and latent factor composition. Internal alignment between high-deviation burden phenotypes and latent deviation factors. Panel a shows Firth logistic odds ratios per 1-SD increase in factor weight for FDR-significant high-deviation burden associations. Panel b compares mean primary absolute-deviation factor composition between patients with lower and high absolute-deviation burden. Factor composition was computed by normalizing each subject’s NMF weights so that each subject’s factor proportions summed to 1. These analyses describe internal organization of the deviation space rather than independent clinical validation

Because high-deviation burden variables and NMF factors were derived from the same deviation matrices, these findings do not constitute independent validation. They indicate that the upper tail of structural deviation load was organized along latent factor axes rather than randomly distributed across the cohort.

Fixed factor weights showed small exploratory incremental associations

Supplementary Table S5 summarizes restricted sensitivity analyses, detectable-effect sensitivity, reduced-feature analyses, and exploratory incremental model comparisons. In the primary absolute-deviation analysis, adding NMF factor weights increased cross-validated R-squared for MIDAS from 0.346 to 0.369 (delta + 0.023; permutation p = 0.012; q = 0.036). For HIT-6, adding NMF factors did not improve cross-validated performance (delta cross-validated R-squared − 0.006; q = 0.459). For monthly headache days, adding factor weights increased cross-validated R-squared from 0.030 to 0.067 (delta + 0.037; permutation p = 0.006; q = 0.018). These small, outcome-specific increments were interpreted as exploratory representation-level evidence that mixed-membership deviation-factor weights carry some outcome-relevant information, not as clinically meaningful prediction or evidence for deployable clinical modeling. These results suggest that deviation-factor weights may encode structural heterogeneity not fully captured by diagnostic category and basic covariates, but they do not establish incremental utility over comprehensive clinical prediction models.

The complete 158-feature inventory and full primary absolute-deviation loading matrix are provided in Supplementary Table S6 and Supplementary Table S7.

Discussion

This study mapped individualized structural MRI normative deviations in migraine using NMF. The primary finding was not a single migraine-related anatomical abnormality but two factor-weight dimensions expressed as patient-level structural deviation burden. Factor 1 showed the most consistent associations in the full cohort, primarily with chronic migraine status, headache frequency, and MIDAS rather than HIT-6, although these associations were attenuated in restricted sensitivity analyses. Factor 2 captured predominantly cortical-thickness deviations, showed a directionally consistent but borderline inverse association with chronic migraine status, and aligned strongly with high absolute-deviation burden. These factors were expressed as patient-level weights and should not be interpreted as mutually exclusive subtypes.

The conceptual contribution is to reframe structural MRI heterogeneity in migraine from the search for a universal lesion-like signature toward the mapping of individualized deviation patterns. Conventional group-level MRI contrasts can describe average differences but do not quantify how each patient deviates from an expected neuroanatomical distribution. Normative modeling provides patient-level deviation z scores, but the resulting 158-feature deviation maps remain too high-dimensional for direct clinical or biological interpretation. Brain-age or other scalar summaries compress anatomy into a single value, whereas NMF provides a mixed-membership representation of regional deviation burden. This interpretation is consistent with recent evidence that group-level morphological findings in migraine are distributed, directionally variable, and sensitive to clinical and methodological context [3–8]. The modest explained-variance-like value indicates that K = 2 provided stable but low-dimensional summary axes, not a comprehensive explanation of structural deviation variance. Because most deviation variance remained outside the two-factor reconstruction, factor weights should be interpreted as partial descriptors of structural heterogeneity with limited signal-to-noise for clinical stratification.

Local adaptation reduces, but cannot eliminate, scanner and calibration uncertainty. Prior PCNtoolkit transfer-learning work by Volkmer et al. demonstrated that normative deviations derived from the same large external reference model can be transferred to independent clinical datasets using site- or dataset-specific healthy-control adaptation sets that ranged from 14 to 78 controls, but site-held-out predictive performance remained moderate [12]. This is important for the present study because a calibration error would not simply add random noise: a systematic local shift could alter patient deviation magnitudes, change which features enter the NMF factors, and potentially inflate or attenuate downstream clinical associations. We therefore treated the local healthy-control set as scanner-context calibration rather than disease comparison, selected it using outcome-blind age, sex, and eTIV balance rules, evaluated leave-one-out healthy-control calibration diagnostics, adjusted eTIV downstream, and tested reduced-feature NMF after excluding cranial-size-sensitive features. These steps make major cranial-size or scanner-offset explanations less likely, but they do not replace external validation on an independently calibrated migraine cohort.

Recent headache neuroimaging work has also moved toward network-level and multimodal descriptions of migraine biology. Functional MRI and diffusion imaging reviews emphasize distributed pain, sensory, white-matter, and brainstem-related changes whose interpretation varies across migraine phase, aura status, and chronicity [36, 37]. Li et al. further reported abnormal structural-to-functional connectivity coupling beyond the cortex in chronic migraine, including subcortical and cerebellar findings [38]. The present study addresses a complementary question: whether structural deviations can be organized at the individual-patient level into interpretable factor-weight dimensions. The convergence of subcortical, brainstem, limbic, and cortical themes across methods supports further multimodal validation, but it does not establish a single mechanism.

The benchmark analyses help clarify the role of NMF. Principal-component loadings recovered broadly similar anatomical axes, suggesting that the subcortical/global and cortical-thickness dimensions were not artifacts of the NMF algorithm alone. However, k-means clustering showed little agreement with NMF dominant-factor labels and did not provide similarly interpretable clinical associations. This supports patient-level factor weights rather than hard cluster assignment as the safer representation for this dataset.

The choice of absolute deviation as the primary matrix was deliberate. Migraine structural MRI studies have reported both increases and decreases in cortical and subcortical measures, and the direction of effects may vary by cohort, feature family, clinical phenotype, disease stage, medication exposure, and analytic strategy [3–8]. Recent genetic and Mendelian-randomization work also suggests that cortical macrostructural and microstructural phenotypes may relate to migraine risk and subtypes in bidirectional and region-specific ways, reinforcing the need to avoid a unidirectional interpretation of structural findings [39]. Absolute deviations capture total departure from the normative range without requiring a common direction of change. Directional negative- and positive-deviation matrices were therefore used as complementary exploratory analyses rather than as competing primary models.

The chronic migraine and clinical-outcome findings require cautious interpretation. Absolute-deviation Factor 1 was associated with chronic migraine status and primarily with headache frequency and MIDAS in the full cohort, but not consistently with HIT-6, and the chronic-migraine association was weakened in restricted sensitivity analyses, including the without-aura subgroup. The study was cross-sectional, the chronic migraine group was older than the episodic migraine group, and medication overuse, preventive treatment exposure, mood, anxiety, sleep, and disease duration differed across clinical groups, consistent with contemporary medication-overuse headache and migraine-prevention literature [40–42]. The relatively low prevalence of stable preventive medication exposure is consistent with Chinese real-world migraine data showing limited preventive-treatment uptake and especially low treatment maintenance among prophylaxis-eligible patients [43, 44]. Medication-class-specific analyses were not feasible because class counts were sparse; therefore, medication effects were addressed using preventive-medication-free sensitivity analyses. These variables may be confounders, mediators, or consequences of disease burden. Disease-duration adjustment did not materially change the primary Factor 1 association, and factor-weight collinearity diagnostics were below conventional concern thresholds (Pearson r = -0.697; Spearman rho = -0.698; maximum VIF = 2.01), but residual confounding remains possible. Accordingly, the estimates should be viewed as total clinical-burden associations within an observational cohort, not as direct neuroanatomical effects, aura-independent chronic-migraine signatures, or chronification mechanisms. The mood- and sleep-adjusted sensitivity analysis was reassuring for the Factor 1 association but should not be interpreted as proving an independent structural effect, because mood and sleep disturbance may be intertwined with migraine burden.

Factor 1 is best described as a burden-related structural deviation axis rather than a focal chronic-migraine lesion. Its loading pattern included subcortical, limbic, brainstem, and global volume-related measures, but this should be interpreted as a broad morphometric deviation-burden profile rather than evidence for a specific neuroanatomical circuit. This caution is consistent with recent MRI brain-age evidence suggesting multilevel structural differences in migraine without proving mechanism or directionality [45]. Any biological interpretation remains indirect because functional, vascular, inflammatory, genetic, and longitudinal measures were not acquired. Factor 2 provides an additional caution: cortical-thickness deviation burden was strongly related to high absolute-deviation load but showed a borderline inverse association with chronic migraine status and some burden measures. These inverse associations should not be interpreted as protective structural mechanisms; they may reflect factor competition, nonlinear disease-stage effects, clinical heterogeneity, residual confounding, or the geometry of the decomposition.

From a normative-modeling perspective, the potential added value of this framework lies in separating clinical expression from individualized structural atypicality. Current clinical variables capture symptom frequency, disability, impact, medication overuse, and diagnostic category, but they do not quantify whether a given patient’s brain structure is atypical relative to an expected neuroanatomical distribution for that person’s age, sex, and scanner context. Normative deviation scores provide this patient-level reference frame. The NMF factors then summarize these high-dimensional deviation maps into interpretable mixed-membership axes, allowing patients within the same clinical category to be described by different magnitudes and anatomical compositions of structural-deviation burden.

Accordingly, the present factors should be viewed as candidate research-level stratification variables rather than clinical decision tools. Their potential translational use would be to test, in longitudinal and interventional cohorts, whether baseline structural-deviation burden improves modeling of chronification, remission, persistent disability, medication-overuse headache course, or preventive-treatment response after comprehensive clinical predictors are accounted for. The present cross-sectional data provide only exploratory support for this pathway: factor weights showed small outcome-specific incremental associations for MIDAS and monthly headache days, but these effects were not sufficient to support individual-level prediction or treatment selection. External validation, fully nested model evaluation, calibration assessment, and decision-curve or net-benefit analyses are required before clinical implementation can be considered.

Several limitations should be considered. First, the cross-sectional design cannot establish causal relationships among structural deviation factors, migraine chronicity, medication exposure, medication-overuse headache, and disability. Second, the single-center Chinese cohort and modest local healthy-control adaptation set may limit generalizability to other scanners, populations, and clinical settings; internal repeated-run and 1000-bootstrap analyses support stability, but external replication remains essential. Third, the normative model was transferred rather than retrained, so performance depends on the calibration, representativeness, and feature compatibility of the released pretrained model. This limitation is relevant to lifespan brain-chart and PCNtoolkit workflows, where local calibration and feature compatibility are central to transportability [46, 47]. eTIV also could not be incorporated as a normative calibration covariate; downstream adjustment reduces but does not eliminate possible residual cranial-size confounding, particularly for volume, ventricular, cerebrospinal-fluid-related, and global features. The detectable-effect sensitivity analysis was approximate and intended only to contextualize precision in the chronic-migraine models; it does not replace formal prospective power calculation or exact simulation for Firth logistic regression.

Fourth, feature-family composition may shape NMF loadings despite supportive feature-family balanced analyses, and factor labels remain interpretive summaries of loading patterns rather than direct biological mechanisms. Fifth, high-deviation burden phenotypes and NMF factors were derived from the same patient-level matrices, so these associations describe internal organization rather than independent validation. Sixth, incremental model comparisons were exploratory and not fully nested end-to-end validation analyses; in addition, the modest sample size, smaller restricted subgroups, and absence of formal a priori power calculation mean that cross-validated estimates and subgroup findings should be interpreted cautiously [34, 35].

Conclusions

In this single-center cross-sectional cohort, structural MRI normative deviations in migraine were summarized by two stable but exploratory NMF factors. Factor 1 showed the most consistent associations in the full cohort, primarily with chronic migraine status, headache frequency, and MIDAS rather than HIT-6, although these associations were attenuated in restricted sensitivity analyses. A second predominantly cortical-thickness factor captured high absolute-deviation burden and showed directionally consistent but less robust relationships with chronic migraine status. The two-factor model explained approximately 11% of total deviation variance, indicating that additional dimensions of structural heterogeneity remain uncharacterized. These results support a patient-level framework for studying migraine neuroanatomical heterogeneity, but the factors require external validation and longitudinal testing before clinical translation.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CM

Chronic migraine

CSF

Cerebrospinal fluid

CV

Cross-validation

EM

Episodic migraine

eTIV

estimated total intracranial volume

FDR

False discovery rate

GAD-7

Generalized anxiety disorder-7

HC

Healthy control

HIT-6

Headache impact test-6

ICHD-3

International classification of headache disorders, 3rd edition

MIDAS

Migraine disability assessment

MOH

Medication-overuse headache

MRI

Magnetic resonance imaging

NMF

Non-negative matrix factorization

OR

Odds ratio

PHQ-9

Patient health questionnaire-9

PSQI

Pittsburgh sleep quality index

VAS

Visual analogue scale

Author contributions

Conceptualization: C.Y.D., M.Q.Y., Y.G.C., F.W.F. Methodology: C.Y.D., M.Q.Y., X.Z.L., Y.G.C., F.W.F. Software: C.Y.D., X.H.L., X.Z.L. Validation: Y.L., K.Y.C., B.L.H., Y.G.C., F.W.F. Formal analysis: C.Y.D., M.Q.Y., X.H.L. Investigation: C.Y.D., M.Q.Y., X.H.L., Z.W.Y., L.C.Y., Y.J.M., H.M.W., Y.L. Resources: X.Z.L., Y.G.C., F.W.F. Data curation: C.Y.D., M.Q.Y., X.H.L., Z.W.Y., L.C.Y., Y.J.M., H.M.W., Y.L. Imaging acquisition and quality control: Z.W.Y., L.C.Y., Y.J.M., H.M.W., X.Z.L., K.Y.C., B.L.H. Visualization: C.Y.D., M.Q.Y., X.H.L. Writing - original draft: C.Y.D., M.Q.Y. Writing - review and editing: all authors. Supervision: Y.G.C., F.W.F. Project administration: Y.G.C., F.W.F. Funding acquisition: Y.G.C., F.W.F. All authors reviewed and approved the final manuscript and agree to be accountable for all aspects of the work.

Funding

This study was supported by the Health Commission of Zhejiang Province (2024KY1275) and the Wenzhou Basic Scientific Research Program (YC20250597). The funders had no role in study design, data collection, data analysis, interpretation, manuscript preparation, or the decision to submit the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Code availability

Analysis code supporting the reported tables and figures is available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This single-center, cross-sectional observational study was approved by the Medical Ethics Committee of the Second Affiliated Hospital of Wenzhou Medical University (approval No. 2024-K-137-01) and was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.

Consent for publication

Not applicable; the manuscript contains no identifiable individual-level images, videos, or personal details.

Competing interests

The authors declare no competing interests.