Classification of EEG signals using fractal dimension features and artificial neural networks

The classification of the brain signals is an issue which has been highly explored to be applied on different fields such as medical diagnostic or in the technological development as can be the Brain-Computer Interfaces (BCI). In order to obtain more feasible BCI applications, there are different issues which must be attended. For example, we can highlight the number of samples used to train the classifier (for avoiding long calibration sessions) and the variability of the brain signals (for avoiding recalibration). In this sense, the feature extraction and the classification stage of a BCI system must be carefully designed to be robust to these issues. In this paper, we propose a methodology based on the use of the fractal dimension as feature and the artificial neural networks as classifier in order to discriminate EEG signals from to different mental tasks. To evaluate the performance of the proposal, we computed the accuracy for two different classification conditions (timeinvariant and reduced number of training samples). After that, the results were compared with those obtained by two classifiers commonly used in BCI applications: Bayesian classifier and linear discriminant. Data set I from the Brain-Computer Interface competition III was used for evaluating and comparing the proposed methodology. The obtained results suggest that the performance of the artificial neural networks are better than those obtained with the other two classifiers.