Lupus Prognosis:a Clinical Study

Abstract

This thesis relies on a partnership between Dr. C. GORDON from the Department of Rheumatology of the Faculty of Medicine and Dentistry and the Neural Computation Research Group. Dr. GORDON works on a disease which is called Lupus. A lot of data has been collected throughout different hospitals from the Midlands on people who suffer from this disease and has been put together in a database. At each visit the patients undergo some clinical tests and answer questionnaires to better follow the evolution of the disease. My goal was to know if, using the database and some neural networks, it is possible to predict the global flares of the disease and moreover in which part of his body the patient is most likely to have complications. In this kind of medical problem, you can usually think of two different ways to treat it: either you adopt graphical models or you use neural networks. In this study, this is the second method which had been used for simple reasons: the first and most important one is that in graphical models, you need the expert. to work with you in order to build the model. This task is quite long and the doctor must give a lot of his/her time, which is not always possible. Moreover, the computations to run on the graphical models for learning are very time-consuming. For these two main reasons, neural networks have been used in this study. Keeping in mind that we wanted to predict the flares of the disease, I have worked on different problems which will be detailed in this thesis. The first one was to find out which variables were the most reliable to help us in the prognosis. I have tried several different, techniques, the first ones based on linear dependencies while the last one is a Bayesian technique called Automatic Relevance Determination.