A Channel Selection Approach Based on Convolutional Neural Network for Multi-channel EEG Motor Imagery Decoding

For many disabled people, brain computer interface (BCI) may be the only way to communicate with others and to control things around them. Using motor imagery paradigm, one can decode an individual's intention by using their brainwaves to help them interact with their environment without having to make any physical movement. For decades, machine learning models, trained on features extracted from acquired electroencephalogram (EEG) signals have been used to decode motor imagery activities. This method has several limitations and constraints especially during feature extraction. Large number of channels on the current EEG devices make them hard to use in real-life as they are bulky, uncomfortable to wear, and takes lot of time in preparation. In this paper, we introduce a technique to perform channel selection using convolutional neural network (CNN) and to decode multiple classes of motor imagery intentions from four participants who are amputees. A CNN model trained on EEG data of 64 channels achieved a mean classification accuracy of 99.7% with five classes. Channel selection based on weights extracted from the trained model has been performed with subsequent models trained on eight selected channels achieved a reasonable accuracy of 91.5%. Training the model in time domain and frequency domain was also compared, different window sizes were experimented to test the possibilities of realtime application. Our method of channel selection was then evaluated on a publicly available motor imagery EEG dataset.