Discovering Potential Clinical Profiles of Multiple Sclerosis from Clinical and Pathological Free Text Data with Constrained Non-negative Matrix Factorization

Abstract

Abstract. Constrained non-negative matrix factorization (CNMF) is an effective machine learning technique to cluster documents in the presence of class label constraints. In this work, we provide a novel application of this technique in research on neuro-degenerative diseases. Specifically, we consider a dataset of documents from the Netherlands Brain Bank containing free text describing clinical and pathological information about donors affected by Multiple Sclerosis. The goal is to use CNMF for identifying clinical profiles with pathological information as constraints. After pre-processing the documents by means of standard filtering techniques, a feature representation of the documents in terms of bi-grams is constructed. The high dimensional feature space is reduced by applying a trimming procedure. The resulting datasets of clinical and pathological bi-grams are then clustered using non-negative matrix factorization (NMF) and, next, clinical data are clustered using CNMF with constraints induced by the clustering of pathological data. Results indicate the presence of interesting clinical profiles, for instance related to vision or movement problems. In particular, the use of CNMF leads to the identification of a clinical profile related to diabetes mellitus. Pathological characteristics and duration of disease of the identified profiles are analysed. Although highly promising, results of this investigation should be interpreted with care due to the relatively small size of the considered datasets.