Abstract
Radiologically Isolated Syndrome (RIS) is characterized by incidental MRI findings indicative of multiple sclerosis (MS) in asymptomatic individuals. Factors such as younger age, positive cerebrospinal fluid biomarkers, and specific lesion locations have been previously linked to a higher risk of conversion from RIS to clinical MS. Predicting which individuals will develop clinical MS remains challenging. Based on widely available cross-sectional patient studies, unsupervised machine learning has been proposed to uncover MRI-driven MS phenotypes with distinct temporal progression patterns. We evaluated whether an unsupervised artificial intelligence framework based on generative manifold learning could stratify RIS patients by conversion risk. BrainGML-MS analyzed imaging biomarkers and generated individualized digital twins from MRI data. We studied 152 RIS individuals (32 converters, RIS-C), 152 MS patients, and 152 healthy controls. The model identified four RIS clusters with distinct five-year conversion risks ranging from 10% to 39%. The brain age gap increased progressively from healthy controls to RIS non-converters, RIS-C, and MS. RIS converters showed greater structural atrophy and greater similarity to MS profiles. These findings indicate that MRI-derived brain aging biomarkers and structural deviations measured at the first RIS scan may improve early risk stratification and support clinical decision-making in preclinical MS.