Effect of EOG Artifact Removal on EEG Motor-Imagery Classification

Deep Learning (DL) models have proven to be efficient in Motor Imagery Brain Computer Interface (MI-BCI) task classification from non invasive Electroencephalogram (EEG) signals. However, most neural network models demand high computational power and vast memory to attain satisfactory performance. This poses a major challenge to the edge level implementation of EEG based motor imagery classification. In this paper, the computation and memory limitations of an edge level devices is considered and a compact and temporally aware neural network model with an efficient data pre-processing step is introduced. The model uses a light-weight linear regression model for EOG artifact removal and are parallel-placed Temporal Convoluted Neural Network that can cover small receptive fields along with a conventional exponentially increasing receptive field coverage. Through the model, 4-class MI task classification is performed on the "BCI Competition 2008 - data set 2A", scoring 80.5% in mean accuracy for all 9 subjects, with a standard deviation of 9.54%. For the reduced standard deviation, a single neural network structure with fixed hyper-parameter will be sufficient for resource-limited devices at the edge.