Combined magnetic resonance imaging and serum analysis reveals distinct multiple sclerosis types

Abstract

Multiple sclerosis (MS) is a highly heterogeneous disease in its clinical manifestation and progression. Predicting individual disease courses is key for aligning treatments with underlying pathobiology. We developed an unsupervised machine learning model integrating MRI-derived measures with serum neurofilament light chain (sNfL) levels to identify biologically informed MS subtypes and stages. Using a training cohort of patients with relapsing-remitting and secondary progressive MS (n = 189), with validation on a newly diagnosed population (n = 445), we discovered two distinct subtypes defined by the timing of sNfL elevation and MRI abnormalities (early- and late-sNfL types). In comparison to MRI-only models, incorporating sNfL with MRI improved correlations of data-derived stages with the Expanded Disability Status Scale in the training (Spearman's ρ = 0.420 versus MRI-only ρ = 0.231, P = 0.001) and external test sets (ρ = 0.163 for MRI-sNfL, versus ρ = 0.067 for MRI-only). The early-sNfL subtype showed elevated sNfL, corpus callosum injury and early lesion accrual, reflecting more active inflammation and neurodegeneration, whereas the late-sNfL group showed early volume loss in the cortical and deep grey matter volumes, with later sNfL elevation. Cross-sectional subtyping predicted longitudinal radiological activity: the early-sNfL group showed a 144% increased risk of new lesion formation (hazard ratio = 2.44, 95% confidence interval 1.38-4.30, P < 0.005) compared with the late-sNfL group. Baseline subtyping, over time, predicted treatment effect on new lesion formation on the external test set (faster lesion accrual in early-sNfL compared with late-sNfL, P = 0.01), in addition to treatment effects on brain atrophy (early sNfL average percentage brain volume change: -0.41, late-sNfL = -0.31, P = 0.04). Integration of sNfL provides an improved framework in comparison to MRI-only subtyping of MS to stage disease progression and inform prognosis. Our model predicted treatment responsiveness in early, more active disease states. This approach offers a powerful alternative to conventional clinical phenotypes and supports future efforts to refine prognostication and guide personalized therapy in MS.

Keywords: disease phenotyping; machine learning; neurodegeneration; neuroinflammation; precision medicine.