Multiclass electrooculography using common spatial pattern

In this paper we apply common spatial pattern (CSP) to the classification of electrooculography (EOG) signals with four distinct classes, namely, eye blinks (EB), eye rotation clockwise (ERC), vertical eye movement (VEM) and horizontal eye movement (HEM). We first describe the CSP and Linear Discriminant Analysis (LDA) algorithms with two classes. We apply the classification method to a database with 9 subjects to evaluate the system performance for long trials (15 s) and short trials (3 s). In both cases the performance was above 86%. We then generalize the method to the multiclass situation (four classes). It is shown that 100% accuracy is obtained (in the scope of the used data set) when the classifier is trained with 8 subjects and tested with another. Even in the extreme situation, when only one subject is used to train the classifier, the system can perform with an accuracy of 84.3%.