Time-Invariant Eeg Classification Based On The Fractal Dimension

Several computational techniques have been proposed in the last years to classify brain signals in order to increase the performance of Brain-Computer Interfaces. However, there are several issues that should be attended to be more friendly with the users during the calibration stage and to achieve more reliable BCI applications. One of these issues is related to the BCI's time-invariant robustness where the goal is to keep the performance using the information recorded in previous sessions to classify the data recorded in future sessions, avoiding recalibration. In order to do that, we have to carefully select the feature extraction techniques and classification algorithms. In this paper, we propose to compute the feature vector in terms of the fractal dimension. To evaluate the feasibility of the proposal, we compare the performance achieved with the fractal dimension against the coefficients of an autoregressive model using a linear discriminant classifier. To asses the time-invariant robustness of the fractal dimension, we train and evaluate the classifier using the data recorded during one day; after that, the trained classifier is evaluated using the data recorded in a different day. These experiments were done using the data set I from Brain-Computer Interface Competition III. The results show that the performance achieved with fractal dimension is better than the autoregressive model (which is one of the most common method used in BCI applications).