Evaluating spiking neural models in the classification of motor imagery EEG signals using short calibration sessions

Abstract

Since the emergence of brain computer interface (BCI), several methods have been applied to associate an electroencephalographic (EEG) recording with a specific mental task. Particularly, in the classification stage, several techniques such as linear Fisher discriminant (LD), feed-forward artificial neural networks (FNN) and radial basis function neural networks (RBF) have been applied successfully with BCI applications. However, in BCI applications, there is a challenge related to avoid long and tedious calibration session for users, this implies that the classification techniques used during the classification stage, have to be trained with a reduced number of EEG recordings. However, most of the classification techniques require several samples to learn accurately an association with a particular mental task. Since the spiking neural models (SNM) have shown their robustness in pattern recognition problems, this paper is focused on demonstrating that they are potential alternatives to classify EEG recordings when they are trained with a reduced number of data samples. To do that, we computed the coherence from a subset of three electrodes to obtain the feature vector of each EEG recording. Then, this information was classified using the SNM. In order to evaluate the robustness, the SNM was trained varying the number of samples. Furthermore, based on the performance and the confidence interval achieved in the classification, we developed two indexes to evaluate and compare the SNM against LD, FNN and RBF. The experimental results over the IIIa, IVa and V data sets from BCI International Competition III, suggest that the SNM are the best option to avoid long calibration sessions.