Unsupervised learning on advanced diffusion MRI metrics for clinical characterization of migraine

Abstract Current migraine classification does not fully capture its clinical heterogeneity, and no objective biomarkers are available for diagnosis or progression. Neuroimaging, especially Magnetic Resonance Imaging (MRI), offers promising biomarkers; however, diffusion tensor imaging (DTI) oversimplifies diffusion properties, motivating the development of advanced approaches compatible with clinical routine. We applied unsupervised machine learning to diffusion MRI data to explore patient stratification and evaluate whether sophisticated diffusion techniques, such as AMURA, provide information complementary to DTI. A total of 107 patients with migraine (56 chronic [CM], 51 episodic [EM]) were included. Diffusion metrics from 48 white matter regions were extracted using standard DTI parameters and advanced AMURA descriptors, alone or combined with DTI. After dimensionality reduction and feature selection, unsupervised clustering was performed using K-means, Agglomerative, and Spectral algorithms. Two clusters consistently emerged, each characterized by distinct clinical features. DTI mainly separated patients based on chronicity-related features, particularly time since onset of CM, and showed an incremental longitudinal association with topiramate response. AMURA-based clusters were associated with migraine frequency, disease duration, and medication overuse headache. These results position the combination of advanced diffusion descriptors and unsupervised learning as an exploratory framework for migraine characterization requiring independent validation.