Predicting Multiple Sclerosis Disease Activity from Longitudinal MRI and EDSS Data Using Multimodal Deep Learning: A Pilot Study

Multiple sclerosis (MS) is a chronic inflammatory disease of the central nervous system characterized by high heterogeneity of clinical progression and limited accuracy of traditional prognostic approaches. Predicting the transition to an active disease phase is critical for timely adjustment of disease-modifying therapy and reduction of disability risk. Three deep learning architectures were developed (ResNet3D, PretrainedResNet2D, and ResNeXt) for predicting MS activity within a one-year horizon using longitudinal multimodal data comprising T1-weighted MRI volumes (50 axial slices, 128×128 px) and Expanded Disability Status Scale (EDSS) functional subscores. The dataset included 28 patients (67 annual observations) with a confirmed MS diagnosis and a minimum two-year follow-up. Class imbalance (≈10% active-phase cases) was addressed through weighted cross-entropy loss (1:3.5) and Gaussian noise augmentation. The best performance was achieved by PretrainedResNet2D, which combines a 3D-to-2D MRI embedding with a pretrained ResNet18 backbone and a recurrent classifier, yielding F1 = 0.80. ResNeXt with self-normalizing layers and ReZero connections reached F1 = 0.73, while fully trainable ResNet3D achieved F1 = 0.62, constrained by the limited dataset size. Transfer learning effectively compensates for data scarcity in rare clinical settings. The proposed pipeline demonstrates feasibility as a decision-support tool in neurological practice and establishes a foundation for prospective multicenter validation.