Multiple Sclerosis (MS) impairs the transmission of signals in the nervous system resulting in progressive disability. MS progression is heterogeneous, patients show different evolution trajectories and managing their disease is challenging. Clinicians may benefit from the use of data-driven tools aiming at predicting disease worsening. We develop models to predict MS worsening using data collected in two MS centers. To effectively capture information from the sequences of expanded disability status scale (EDSS) measurements, we explore possible strategies to extract time-dependent variables which well describe trends of this score over time. First, we consider the whole-time span and extract EDSS patterns using a stratification method. Then, we independently consider separate time periods and rely on EDSS first order descriptors. We train a Cox model and a survival support vector machine in three different settings characterized by different subsets of EDSS-derived variables. In the first setting we consider EDSS information by identifying typical EDSS evolution trajectories, in the second one we include first order descriptors of EDSS evolution, and in the third one we include all available information. We trained and tested all models within a framework able to perform optimal feature selection and unbiased estimates of model performance. For the Cox model, the best-performing setting was the first one (C-index = 0.833). However, as the C-index of the second and third settings was comparable according to both considered methodological approaches, one should consider the additional qualitative information that stratification brings to patients and clinicians.