A Deep Learning Approach for Steady-State Visual Evoked Potential Classification in a Brain-Computer Interface Speller

Background and Aim: Neuro-muscular diseases such as amyotrophic lateral sclerosis or brainstem stroke would cause a severe effect on these patients. Mostly, these patients lose their ability to communicate with their environment, since they are almost incapacitated to move or talk. In order to provide a secondary communication pathway for them, Brain-Computer Interfaces (BCI) enact neural oscillation in order to generate a command signal for machines to operate desired tasks instead of patients. Consequently, they could communicate well without any requirement to move their facial muscles or whatsoever. In general, most BCIS contains three major parts; firstly, there is a stimulus part of generating the desired pattern in neural activity of the brain, such as Steady-State Visual Evoked Potential (SSVEP). Secondly, these signals should be measured via techniques like EEG, and at last, the expected signals should be detected and classified for each stimulus to function as a command signal for the computer to perform the proper task. Methods: The SSVEP is brought out from visual stimuli that flicker in a specific frequency, and the response of this stimulation has the same frequency as the visual stimulus or its fundamental harmonics. Based on this feature, we developed a Deep Learning method for a ۴۰-class SSVEP; each class represents a stimulus, which has been acquired from ۳۵ subjects. Our model contains three convolution layers following a fully connected and a SoftMax layer to perform the classification task. We used ۴s windows of the data (۴ seconds for each class) to train our model and leave-one-out cross-validation for the testing. Results: The Classification part of these systems is the core factor for evaluating a BCI since the whole performance of the system relies on this crucial part. Despite the considerable progress that these systems have made, there is still a lack of feasibility which mainly relies on the performance of the classification part of the system. In this study, we scrutinize an SSVEP based BCI speller to enhance the classification accuracy of SSVEP detection and consequently, to raise the Information Transfer Rate (ITR) of these systems in order to become more reliable for the actual usage. Conclusion Our three-layered Convolutional Neural Networks reached ۸۶.۵% classification accuracy and ۴۲.۴۵ bit per minute for a nine-channel EEG which outperforms the traditional approach for SSVEP classification. We hope that our contribution would put this technology forward to become more practical for people who require these systems for their everyday use, indeed.

کلیدواژه ها:

Brain-Computer Interface ، Steady-State Visual Evoked Potential ، Deep Learning ، Convolutional Neural Networks