Classification of Motor Imagery EEG Signals with Deep Learning Models

Motor imagery (MI) is a mental process of a motor action including preparation for movement, passive observations of action and mental operations of motor representations. Brain computer interfaces can discriminate different status of individuals according to their EEG signals during imagery tasks. Power spectral density and common spatial patterns are both feature extraction methods that are commonly used to in the classification tasks of EEG series. In this paper, we combine recurrent neural networks and convolutional neural networks inspired by speech recognition and natural language processing. Furthermore, we apply deep models consist of stacking random forests to enhance the ability of feature representation and classification abilities for motor imagery EEG signals. Compared with traditional feature extraction methods, our approaches achieve significant improvements both in the MI-EEG dataset of BCI competitions with healthy individuals and the dataset collected from stroke patients.