Single-Channel Selection for Detecting Steady-State Visual Evoked Potentials in a Brain-Computer Interface Speller

Farzad Saffari; Ali Khadem

doi:10.18502/fbt.v8i3.7109

Farzad Saffari Department of Biomedical Engineering, Faculty of Electrical Engineering, K. N. Toosi University of Technology, Tehran, Iran
Ali Khadem Department of Biomedical Engineering, Faculty of Electrical Engineering, K. N. Toosi University of Technology, Tehran, Iran

DOI: https://doi.org/10.18502/fbt.v8i3.7109

Keywords: Brain-Computer Interface Speller; Steady-State Visual Evoked Potential; Deep Learning; Convolutional Neural Networks; Single-Channel Electroencephalogram

Abstract

Purpose: Brain-Computer Interface (BCI) provides a secondary communication pathway for patients with neuromuscular diseases such as amyotrophic lateral sclerosis (ALS) or brainstem stroke in which they are almost incapacitated to move or talk. BCI enacts neural oscillations to generate a command signal for machines to operate desired tasks instead of patients. Steady-State Visual Evoked Potential (SSVEP) is the brain response to a visual stimulus, with the same frequency as its eliciting signal (or its harmonics), that has been widely used in BCI environments. In order to provide a more convenient situation for BCI users, we aim to find the best single-channel EEG, which results in the highest accuracy for detecting SSVEP.

Materials and Methods: We developed a Deep Convolutional Neural Network with single-channel EEG as input to classify a 40-class SSVEP; each class represents a stimulus, which has been acquired from 35 subjects. We used 3.5 s windows of the data (Trials of 3.5 seconds length for each class) to train our model and leave-one-subject-out cross-validation for the testing.

Results: The proposed method resulted in the average classification accuracy of 74.30%±20.85 and Information Transfer Rate (ITR) of 57.51 bpm which outperforms the previous single-channel SSVEP BCIs in terms of ITR. Also, the O1 channel achieved the best performance criteria among the channels in the occipital and parietal lobes, which seems reasonable according to previous researches for finding the location of neurons, responsible for visual tasks in the brain.

Conclusion: In this study, we dedicated our efforts to reduce the number of EEG channels to a single channel while proposing a deep learning strategy for an SSVEP-based BCI speller to make it more feasible for patients whose lives are dependent on such systems. The overall results, although not ideal, open a new promising window toward a feasible BCI system.