Recurrent Deep Learning for EEG-based Motor Imagination Recognition

Deep Learning has grasped great attention for recognition of Electroencephalography. For the analysis of brain dynamics, non-stationary motor imagery signals are used. Although a number of studies have been carried out for the extraction of hidden patterns and classification of EEG signals, temporal information has rarely been incorporated. In this paper, we propose a spatio-temporal energy maps generation scheme followed by deep learning classification model. Common spatial pattern filters and Fast Fourier Transform Energy Maps are deployed to obtain discriminative and spatio-temporal features. Long-Short-Term-Memory (LSTM) based neural network has been proposed to classify the temporal series of energy maps. This research also investigates preprocessing techniques to obtain optimal parameters which include frequency bands selection and temporal segmentation. The proposed model is evaluated on BCI Competition IV dataset 2a and achieved 0.64 mean kappa for multi-class EEG classification, which is the current state of the art. Furthermore, several empirical findings are also presented, that may be of significant interest to the BCI community.