Imaging subtypes reveal distinct biological substrates and disability profiles in multiple sclerosis

Multiple sclerosis is characterized by marked biological heterogeneity that is only partly captured by conventional clinical phenotypes and age-at-onset categories. MRI can detect focal lesions, diffuse microstructural damage and atrophy, but these measures are usually considered separately. We therefore aimed to use Subtype and Stage Inference (SuStaIn), an unsupervised machine learning framework, to identify biologically meaningful MRI subtypes of multiple sclerosis and determine their associations with disability, age at onset, cognition, genetic susceptibility to more severe disease, and relapse-independent progression. We applied SuStaIn to multimodal 3T brain MRI data from 1017 multiple sclerosis patients and 548 healthy controls. Imaging features included T2-hyperintense white matter lesion volume, mean diffusivity in white matter lesions and major white matter tracts, and cortical and deep gray matter volumes. Cognitive testing was available in 501 patients, genetic profiling in 650, and longitudinal clinical follow-up in 645 (median follow-up=6.57 years). An independent validation cohort included 247 multiple sclerosis patients and 141 healthy controls. SuStaIn identified four MRI subtypes: lesion-led (44%), cortex-led (23%), tract-led (23%), and deep gray matter/cerebellar-led (10%). Cortex-led and deep gray matter/cerebellar-led subtypes were enriched in progressive disease (47% and 56%), whereas lesion-led and tract-led subtypes were predominantly relapsing-remitting disease (66% and 71%) (p<0.001). Tract-led subtype was overrepresented in pediatric-onset disease (38%) and underrepresented in adult-onset disease (19%); lesion-led subtype predominated in adult-onset disease (48%), and deep gray matter/cerebellar-led subtype was more frequent in late-onset disease (14%) (p<0.001). Expanded Disability Status Scale scores ≥4.0 and ≥6.0 were more frequent in cortex-led and deep gray matter/cerebellar-led disease than in lesion-led or tract-led disease (45.3% and 55.6% versus 34.3% and 31.9%, and 27.8% and 29.3% versus 20.0% and 18.7%; p<0.001). Relapse-independent progression-free survival differed across subtypes (p=0.04) and was 1.67 and 1.93 years shorter in cortex-led and deep gray matter/cerebellar-led disease, respectively, than in lesion-led disease (p<0.001). Cognitive performance and genetic risk scores for multiple sclerosis severity did not differ across subtypes, but within each subtype advancing SuStaIn stage was associated with worse global and domain-specific cognitive performance (all p≤0.030). The four-subtype structure was reproduced in the independent validation cohort. Multimodal MRI analysed with SuStaIn identified reproducible subtypes of multiple sclerosis with clinically meaningful differences in phenotype, age at onset, disability, and relapse-independent progression, while cognitive impairment was related to advancing SuStaIn stage within each subtype. This framework may support biologically informed patient stratification and more mechanism-based prognostic and therapeutic approaches in multiple sclerosis.