Supervised machine learning in multiple sclerosis: applications to clinically isolated syndromes

Multiple sclerosis (MS) is an inflammatory, demyelinating disease that can cause various neurological symptoms. The first episode of this disease is called a clinically isolated syndrome (CIS) and leads to the diagnosis of MS in the majority of patients in the long-term. Fast conversion from CIS to MS is associated with higher disability and more severe disease progression so that it is of high clinical interest to identify risk patients that will convert to MS within a short time. Several risk factors for conversion have been identified but they can only be applied on cohort levels. In this thesis we provide an overview of supervised machine learning approaches that can be used to distinguish individual CIS-stable patients from those who will experience a second attack within one to five years and consequently will be diagnosed with clinically definite MS. This classification is based on information available at baseline derived from routine MRI scans and complemented by clinical information such as lesion masks, age, gender, disability and CIS type of onset. We introduce the classification landscape, an overview of supervised classification studies with respect to their method and task complexity, and show that our experiments cover a large range of feature complexities in this landscape for the rather complex task of outcome prediction in CIS patients. We show that low-level voxel-based information such as tissue density of grey and white matter are not informative and lead to inconclusive results, whereas the introduction of high-level features such as lesion load, age, gender or disability improves accuracies to 71.4 % and 68 % at one- and three-year follow-up respectively in a single-centre data set. Finally, we propose a recursive feature elimination method that is able to identify specific regions that are relevant with respect to disease progression in MS and achieves accuracies of 73.9 % and 74.3 % at one- and three-year follow-up respectively even in a multi-centre setting