I. E. Shepelev, B. M. Vladimirskiy

In this paper a neural network classifier for Brain-Computer Interface purposes is proposed. The signal being analyzed is a multichannel human EEG recorded during imagined movement execution. Our feature extraction method, based on the intrinsic time concept in the context of bioelectrical signal processing, is novel. In this approach it is the sequence of the local extrema of EEG signal amplitude, rather than the signal-s spectral characteristics that is the numerical data to be analyzed. We use self-organizing neural maps to assess whether a classifier can be constructed. In order to do that, we analyze how structured the feature space is. We show that the features indicating each imagined movement type form separable clusters in the feature space. The structural organization of the clusters is also shown to exhibit high variability of an individual character. Based on the structural variability of the feature space, we conclude that in terms of the computational complexity of the imagined movement classification procedure the problem at hand is linearly separable in some cases, while in others, strongly nonlinearly separable. As a universal classifier of imagined movement patterns, we use a backpropagation-trained multilayer neural network. To tune the size of the neural network to the individual characteristics of the computational complexity of a given data sample, we utilize a growing neural network following the adaptive resonance theory. Numerical experiments demonstrate that the preliminary tuning allows for a significant decrease in training time and a considerable improvement of the classifier-s generalization properties, which is of practical importance for real-world systems. Our method of feature extraction from the EEG recorded during imaginary movements and the neural network computational tools for processing of the extracted data enable the system to solve the problem within 2 seconds.