Motor imagery EEG classification using feedforward neural network

Electroencephalography (EEG) is a complex voltage signal of the brain and its correct interpretation requires years of training. Modern machine- learning methods help us to extract information from EEG recordings and therefore several brain-computer interface (BCI) systems use them in clinical applications. By processing the publicly available PhysioNet EEG dataset, we extracted information that could be used for training feedforward neural network to classify three types of activities performed by 109 volunteers. While volunteers were performing different activities, a BCI2000 system was recording their EEG signals from 64 electrodes. We used motor imagery runs where a target appeared on either the top or the bottom of a screen. The subject was instructed to imagine opening and closing either both his/her fists (if the target is on top) or both his/her feet (if the target is on the bottom) until the target disappears from the screen. We used the EEGLAB Matlab toolbox for EEG signal processing and applied several feature extraction techniques. Then we evaluated the classification performance of feedforward, multilayer perceptron (MLP) networks with different structures (number of layers, number of neurons). Achieved accuracy score for test data was 71.5%