Comparison of filtering and classification techniques of electroencephalography for brain-computer interface

In this paper several methods are investigated for feature extraction and classification of mu features from electroencephalographic (EEG) readings of subjects engaged in motor tasks. EEG features are extracted by autoregressive (AR) filtering, mu-matched filtering, and wavelet decomposition (WD) methods, and the resulting features are classified by a linear classifier whose weights are set by an expert using a-priori knowledge, as well as support vector machines (SVM) using various kernels. The classification accuracies are compared to each other. SVMs are shown to offer a potential improvement over the simple linear classifier, and wavelets and mu-matched filtering are shown to offer potential improvement over AR filtering.