Classification of imagined spoken word-pairs using convolutional neural networks

Imagined speech is gaining traction as a communicative paradigm for brain-computer-interfaces (BCI), as a growing body of research indicates the potential for decoding speech processes directly from the brain. The development of this type of direct-speech BCI has primarily considered feature extraction and machine learning approaches typical to BCI decoding. Here, we consider the potential of deep learning as a possible alternative to traditional BCI methodologies in relation to imagined speech EEG decoding. Two different convolutional neural networks (CNN) were trained on multiple imagined speech word-pairs, and their performance compared to a baseline linear discriminant analysis (LDA) classifier trained using filterbank common spatial patterns (FBCSP) features. Classifiers were trained using nested cross-validation to enable hyper-parameter optimization. Results obtained showed that the CNNs outperformed the FBCSP with average accuracies of 62.37% and 60.88% vs. 57.80% (p