Modeling individual disability evolution in multiple sclerosis patients based on longitudinal multimodal imaging and clinical data

Abstract

ABSTRACTBackgroundThe individual disease evolution of multiple sclerosis (MS) is very different from one patient to another. Therefore, the prediction of long-term disability evolution is difficult based on only clinical information. Magnetic resonance imaging (MRI) provides a very efficient tool to distinguish between healthy and abnormal brain tissue, monitor disease evolution, and help decision-making for personalized treatment of MS patients.ObjectiveWe aim to develop a patient-specific model to predict individual disease evolution in MS, using demographic, clinical, and imaging data that were collected at study onset.MethodsThe study included 75 patients tracked over 5 years. The latent class linear mixed model was used to consider individual and unobserved subgroup variability. First, the clinical model was established with demographic and clinical variables to predict clinical disease evolution. Second, the imaging model was built using the multimodal imaging variables. Third, the imaging variables were added one by one, two by two, and all three together to investigate their contribution to the clinical model. The clinical disability is measured with the Expanded Disability Status Scale (EDSS). The performances of the clinical, imaging, and the combined models were compared mainly using the Bayesian Information Criterion (BIC). The mean of the posterior probabilities was also given as the secondary performance evaluation criterion.ResultsThe clinical model gave higher BIC value than imaging and any combined models. The means of the posterior probabilities given by the three models were over 0.94. The clinical model clustered the patients into two latent classes: stable evolution class (n=6, 88%) and severe evolution class (n=9, 12%).ConclusionThe latent class linear mixed model may provide a well-fitted prediction for the disability evolution in MS patients, thus giving further information for personalized treatment decisions after thorough validation with a larger and independent dataset.