EEG Based Brain Computer Interface System for Decoding Covert Speech using Deep Neural Networks

TL;DR: In this paper , a two-dimensional convolutional neural network and LSTM was used to learn deep features from EEG signals corresponding to the imagery of words to design a BCI(Brain computer interface) system based on human thoughts.

Abstract: The study aims to learn deep features from EEG signals corresponding to the imagery of words to design a BCI(Brain computer interface) system based on human thoughts. Topological plots of time-averaged EEG signals across all the trials corresponding to the imagery of a particular word are fed as input to the designed deep neural network algorithm. Designed deep learning architecture has an amalgamation of the Two-dimensional convolutional neural network and LSTM that takes the assistance of the capabilities and assets of both neural network architectures. The proposed neural network architecture achieves admirable accuracy in identifying imagined words from the EEG-based Kara one dataset. The accuracy of the designed CNN-LSTM model with topological plots as input features was 20-25% more than chance level accuracy and comparable to the state of arts.