A Comparative Study of Time and Frequency Features for EEG Classification

Despite the different nature of sounds and electrophysiological potentials, there are certain similarities between these two kinds of signals. Therefore, feature extraction methods used in both cases are based on similar signal processing concepts. In this work, we used a set of audio features based on time, and frequency domains, including Mel-Frequency Cepstral Coefficients (MFCC), which were applied to electroencephalographic signals (EEG) recorded for Motor Imagery (MI) tasks. A classification stage using Multilayer Perceptron (MLP) was implemented, to evaluate the accuracy to differentiate two task of Motor Imagery, for each proposed descriptor, as well as for groups of descriptors. Moreover, the influence of the length size windows was studied with this approach. The results suggest that MFCC are more robust than other descriptors through different window lengths, achieving high classification accuracy both individually and in the group of MFFC filters.