Abstract

Multiple sclerosis (MS) affects 2.9 million people. Traditional classification of MS into distinct subtypes poorly aligns with its pathobiology, has limited value for prognostication of disease evolution and treatment response, and divergent views on disease classification hamper drug discovery. We have developed a bespoke multivariate hierarchical Bayesian model (probabilistic latent variable followed by hidden Markov model (HMM)) that can handle different data

modalities (binary, count, ordinal and continuous variables) and structured missingness.

We report a new, data-driven classification of MS disease evolution by analysing a large clinical trial database (~8000 patients with 118,000 patient visits, >35,000 MRI scans). Four dimensions define MS disease states: physical disability, brain damage, clinical relapses, and subclinical disease activity. Early/mild/evolving (EME) MS and advanced MS represent two poles of a disease severity spectrum. Transitions to advanced MS occur via accumulation of damage to the central nervous system through inflammatory states, with or without accompanying symptoms. We validated these results in pooled data from three independent clinical trials and in a real-world cohort, totalling >4000 patients with MS.