Explainable machine learning reveals threshold associations and distinct phenotypes linking migraine features to white matter hyperintensities in women with episodic migraine.
6 affiliations
- 1School of Nursing, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China.
- 2School of Public Health, Sun Yat-Sen University, Guangzhou, China.
- 3College of Medicine and Veterinary Medicine, The University of Edinburgh, Edinburgh, UK.
- 4School of Nursing, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China. [email protected].
- 5Research Centre for Chinese Medicine Innovation, The Hong Kong Polytechnic University, Hong Kong, China. [email protected].
- 6Joint Research Centre for Primary Health Care, The Hong Kong Polytechnic University, Hong Kong, China. [email protected].
Abstract
Background
White matter hyperintensities (WMHs) are prevalent in migraine, yet their relationship to specific migraine features remains poorly defined. Prior studies have used coarse phenotyping and linear methods, leaving potential threshold effects and the broader migraine phenotype unexplored. We applied an explainable machine-learning framework to characterise associations between migraine features and WMH volume and to identify data-driven phenotypes.Methods
Baseline data from two ongoing randomised controlled trials in Hong Kong Chinese women with episodic migraine were used for analyses. Participants underwent 3T brain magnetic resonance imaging (MRI) and WMH were quantified from manually corrected segmentations by radiologists. Migraine features (attack frequency, duration, pain intensity, aura subtypes, accompanying symptoms, triggers, and acute medication use) were derived from prospective three-month diaries and structured interviews. An XGBoost model was developed with nested five-fold cross-validation and interpreted using SHapley Additive exPlanations (SHAP). Unsupervised K-means clustering on patient-level SHAP values identified phenotypes.Results
The analytic sample comprised 171 women (median age 52.0 years [IQR 38.5-60.0]) with a median WMH volume of 0.90 mL (IQR 0.56-1.31). The model explained modest variance (R² = 0.221) and achieved an area under the curve of 0.801 (95% CI 0.726-0.876) for classifying high WMH burden. SHAP dependence analyses revealed several non-linear patterns of migraine features on WMH. Attack frequency and duration showed apparent saturation, plateauing above approximately 5 attacks/month and 24 h/month, respectively; pain intensity showed an L-shaped threshold pattern above a Numerical Rating Scale score of approximately 5. Somatosensory aura was associated with higher WMH volume. Three phenotypes emerged: Age-Dominant (n = 16; older, low burden, low WMH), Cumulative Risk (n = 59; older with high vascular and migraine burden, highest WMH volume), and Migraine-Dominant (n = 96; younger, migraine as the primary driver, intermediate WMH).Conclusions
Explainable machine learning revealed threshold associations, including saturation patterns for attack frequency and duration and an L-shaped pattern for pain intensity, together with a pattern linking somatosensory aura to higher WMH volume, which may help reconcile longstanding discrepancies in the migraine-WMH literature. Findings from this study are hypothesis-generating and warrant validation in larger, longitudinal, and multi-ethnic cohorts.
Related
- Quantitative MRI Characterization of Fluid-Related Phenotypes in Pediatric Epilepsy: Deep Learning-Assisted Assessment of Enlarged Perivascular Space Burden, Regional Distribution, and Clinical Associations
- Nutritional status, inflammatory burden, and CT-derived muscle phenotypes for predicting chronic postsurgical pain after total hip arthroplasty: an explainable machine learning study
- Diffusion-weighted imaging reversal after reperfusion therapy in acute ischemic stroke: associations with white matter hyperintensities and cerebral microbleeds
- Associations of serum Hcy, SAP, NfL, and carotid plaque Crouse score with white-matter hyperintensities in older patients with cerebral small-vessel disease
- 1078 Mutational signatures as mechanistic biomarkers of tumor immunity: associations with TMB, PD-L1 and immune phenotypes across cancer types
BibTeX
@article{yang2026explainable,
title = {{Explainable machine learning reveals threshold associations and distinct phenotypes linking migraine features to white matter hyperintensities in women with episodic migraine.}},
author = {Qingling Yang and Sen Deng and Jing Qin and Harry Haoxiang Wang and Kin Cheung and Yao Jie Xie},
journal = {The journal of headache and pain},
year = {2026},
month = aug,
doi = {10.1186/s10194-026-02488-8},
url = {https://europepmc.org/article/MED/42963542},
note = {CC BY},
}
RIS
TY - JOUR
TI - Explainable machine learning reveals threshold associations and distinct phenotypes linking migraine features to white matter hyperintensities in women with episodic migraine.
AU - Qingling Yang
AU - Sen Deng
AU - Jing Qin
AU - Harry Haoxiang Wang
AU - Kin Cheung
AU - Yao Jie Xie
PY - 2026
DA - 2026/08/04
JO - The journal of headache and pain
DO - 10.1186/s10194-026-02488-8
UR - https://europepmc.org/article/MED/42963542
AB - Background White matter hyperintensities (WMHs) are prevalent in migraine, yet their relationship to specific migraine features remains poorly defined. Prior studies have used coarse phenotyping and linear methods, leaving potential threshold effects and the broader migraine phenotype unexplored. We applied an explainable machine-learning framework to characterise associations between migraine features and WMH volume and to identify data-driven phenotypes.Methods Baseline data from two ongoing randomised controlled trials in Hong Kong Chinese women with episodic migraine were used for analyses. Participants underwent 3T brain magnetic resonance imaging (MRI) and WMH were quantified from manually corrected segmentations by radiologists. Migraine features (attack frequency, duration, pain intensity, aura subtypes, accompanying symptoms, triggers, and acute medication use) were derived from prospective three-month diaries and structured interviews. An XGBoost model was developed with nested five-fold cross-validation and interpreted using SHapley Additive exPlanations (SHAP). Unsupervised K-means clustering on patient-level SHAP values identified phenotypes.Results The analytic sample comprised 171 women (median age 52.0 years [IQR 38.5-60.0]) with a median WMH volume of 0.90 mL (IQR 0.56-1.31). The model explained modest variance (R² = 0.221) and achieved an area under the curve of 0.801 (95% CI 0.726-0.876) for classifying high WMH burden. SHAP dependence analyses revealed several non-linear patterns of migraine features on WMH. Attack frequency and duration showed apparent saturation, plateauing above approximately 5 attacks/month and 24 h/month, respectively; pain intensity showed an L-shaped threshold pattern above a Numerical Rating Scale score of approximately 5. Somatosensory aura was associated with higher WMH volume. Three phenotypes emerged: Age-Dominant (n = 16; older, low burden, low WMH), Cumulative Risk (n = 59; older with high vascular and migraine burden, highest WMH volume), and Migraine-Dominant (n = 96; younger, migraine as the primary driver, intermediate WMH).Conclusions Explainable machine learning revealed threshold associations, including saturation patterns for attack frequency and duration and an L-shaped pattern for pain intensity, together with a pattern linking somatosensory aura to higher WMH volume, which may help reconcile longstanding discrepancies in the migraine-WMH literature. Findings from this study are hypothesis-generating and warrant validation in larger, longitudinal, and multi-ethnic cohorts.
SN - 1129-2369
SN - 1129-2377
ER -