Unveiling state-specific neural dynamics in migraine with and without depressive symptom: a hidden Markov model and interpretable machine learning approach

Migraine with depressive symptom (dMIG) constitutes a more severe clinical condition than migraine without depressive symptom (ndMIG), and is likely underpinned by distinct neuropathophysiology. While traditional neuroimaging has linked these clinical differences to static functional connectivity (FC) alterations, the role of dynamic brain network interactions which may more directly reflect the fluctuating nature of symptoms remains poorly understood. In this cross-sectional study, we examined the spatiotemporal brain dynamics from resting-state functional Magnetic Resonance Imaging (rs-fMRI) data of 204 migraine patients (100 dMIG and 104 ndMIG), and 90 healthy controls (HCs) using Hidden Markov Model (HMM). By integrating multiple machine learning algorithms: Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Random Forest, with Shapley additive explanations (SHAP)-based interpretability analysis, we aimed to identify and elucidate potential dynamic neuroimaging biomarkers capable of distinguishing these two subtypes. Six HMM states were identified in this study. Significant differences in brain network dynamics were observed among the groups. The dMIG group showed higher transition probabilities between states 4, 5, 6 and enhanced activity in sensorimotor, dorsolateral prefrontal, and temporal regions. Conversely, the ndMIG group exhibited prolonged dwell time in state 3, a reduced global transition rate, and heightened sensorimotor activity. The XGBoost model achieved superior classification performance (test set AUC = 0.86, accuracy = 75.81%). SHAP analysis identified the fractional occupancy of states 3, 5 and 2 as the top three discriminative features. Migraine with depressive symptom is characterized by brain state instability and co-activation of pain and mood network, while migraine without depressive symptom exhibits functional inflexibility, persistently engaging pain-related regions. These distinct spatiotemporal patterns offer biomarkers with the potential to inform subtype-specific diagnosis and therapeutic strategies.