Prediction of Second Neurological Attack in Patients with Clinically Isolated Syndrome Using Support Vector Machines

Abstract

The aim of this study is to predict the conversion from clinically isolated syndrome to clinically definite multiple sclerosis using support vector machines. The two groups of converters and non-converters are classified using features that were calculated from baseline data of 73 patients. The data consists of standard magnetic resonance images, binary lesion masks, and clinical and demographic information. 15 features were calculated and all combinations of them were iteratively tested for their predictive capacity using polynomial kernels and radial basis functions with leave-one-out cross-validation. The accuracy of this prediction is up to 86.4% with a sensitivity and specificity in the same range indicating that this is a feasible approach for the prediction of a second clinical attack in patients with clinically isolated syndromes, and that the chosen features are appropriate. The two features gender and location of onset lesions have been used in all feature combinations leading to a high accuracy suggesting that they are highly predictive. However, it is necessary to add supporting features to maximise the accuracy.