A Novel Neural Network Approach to Creating a Brain–Computer Interface Based on the EEG Patterns of Voluntary Muscle Movements

A novel neural network approach to the real-time detection and classification of temporospatial EEG patterns accompanying performance of motor imagery (imaginary movements) based on a local approximation method using radial basis functions and an original algorithm for interpreting the time sequence of the responses of the neural network is proposed. This was used as the basis for creating and testing an asynchronous neural interface whose basic element was a classifier including a committee of five neural networks detecting EEG patterns accompanying four types of motor imagery. We present a comparative assessment of the EEG pattern recognition effectiveness for motor imagery using the method developed here and classical neural network methods, particularly radial basis networks, a multilayer perceptron, and the support vectors method. Experimental studies demonstrated a user learning effect with increases in recognition accuracy and EEG pattern classification, as well as speed of control.