Transfer learning in imagined speech EEG-based BCIs. (April 2019)

• Record Type: Journal Article
• Authors: García-Salinas, Jesús S.; Villaseñor-Pineda, Luis; Reyes-García, Carlos A.; Torres-García, Alejandro A.
• Is Part Of: Biomedical signal processing and control. Volume 50(2019)
• Page Start: 151
• Page End: 157
• Publication Date: 2019-04
• Subjects: 00-01 -- 99-00; Imagined speech -- Bag of Features -- EEG -- Brain Computer Interfaces -- Transfer learning; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.01.006
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Bag of Features is able the discrimination of imagined speech in electroencephalograms. Bag of Features aids the transfer learning of imagined speech to recognize new imagined words. Using time domain features of the electroencephalogram for a Bag of Features model achieves a competitive classification performance. Abstract: The Brain–Computer Interfaces (BCI) based on electroencephalograms (EEG) are systems which aim is to provide a communication channel to any person with a computer, initially it was proposed to aid people with disabilities, but actually wider applications have been proposed. These devices allow to send messages or to control devices using the brain signals. There are different neuro-paradigms which evoke brain signals of interest for such purposes. Imagined speech is one of the most recent paradigms, and it is explored in this work, it consists of the internal pronunciation of a word, i.e. a subject imagines the utterance of a word without emitting sounds or articulating facial movements. Under this neuro-paradigm, to increase the initial vocabulary reducing drastically the training time using few or none new data is an open challenge. The proposed method extracts characteristic units (i.e. codewords ) of the EEGs associated with the words of an initial vocabulary. Subsequently, a new imagined word is represented with these codewords, and then a classification algorithm is applied. The method was tested both, with and without calibration examples, in a 27 subjects dataset. An initial vocabulary of 4 words, with 33 epochs for each word was considered. The results were obtained by averaging the accuracies of every subject, without calibration data a 65.65% accuracy was achieved. In comparison to the baseline method, which obtained an average accuracy of 68.9%, the proposed method showed no statistical difference.