Development of predictive models for short-term prediction of disability progression in multiple sclerosis

Multiple Sclerosis (MS) is an autoimmune degenerative disease of the central nervous system, in which chronic inflammation leads to demyelination with transient or permanent axon damage. Symptoms of MS include problems with vision, movement, sensation and balance, which can be intermittent or progressively increasing over time until bringing to permanent disability. Predictive models of MS disability progression can be very useful to support the clinician in choosing the best care for each patient. The aim of this work is to develop predictive models of short-term MS disability progression. Data are part of the Multiple Sclerosis Outcome Assessments Consortium (MSOAC) Placebo database, which includes longitudinal demographic and clinical data of 2465 MS patients who were enrolled in the control arm of different MS clinical trials. Variables collected in the first visit were used to predict a binary outcome of disability progression at 6 months and 18 months from the baseline, using a logistic regression model. Disability progression was defined as a 1.5 increase in the Expanded Disability Status Scale (EDSS) value compared to the baseline time. 20 input variables were considered in each model, including demographics, medical history, functional tests, questionnaires, and MS phenotype. Preprocessed data were split into a training and a test set with an 80%-20% proportion. Logistic regression models were trained on the training set, using over-/undersampling techniques for balancing the classes. The identified models were tested on the test set by assessing the area under the receiver operating characteristic curve (AUC). Prediction performance on the test set was satisfactory, although not optimal, with AUC equal to 0.74 at 6 months and 0.71 at 18 months. These prediction performances are comparable with results obtained by other literature studies on smaller cohorts. Future developments of this work include the use of other machine learning techniques for model training, the application of feature selection and variable ranking techniques, the incorporation of new variables (e.g., imaging variables), and the external validation of the models on new populations.