A Profound Analysis on Complex Multiclass BCI Signals Classification by an EEG Device

Abstract

With a major shift in the technology paradigm today in the area of deep learning and artificial intelligence, one of its important applications today is found in brain-computer interface (BCI). Brain computer-interface, also known as neural control interface or mind-machine interface is a device that helps a user to interact with a computer using only his brain-activity, which is measured usually by Electroencephalography (EEG). The usage of such mechanism has also been explored in the field of motion intent recognition, emotion recognition, stress detection, epileptic seizure prediction etc. However, with major focus being given to binary class problems in the EEG domain, it has been observed that the use cases involving complex multi class features have not been tapped into as much. This paper takes a step by step approach by first designing a new methodology for the existing braincomputer interface device which can enhance the performance of EEG signal classifier in a composite multi class problem. A new approach for the prevailing classification algorithm is proposed and a contrast on the differences between the binary and multi class problem in the domain of EEG analysis is also discussed. The key reasons behind the failure of existing classifiers on EEG response resulting from slightly varying multiple stimuli are put forward, followed by an examination of various other approaches of signal classification that rely on coordination between multiple channels and spectral analysis of signal, with their significance being tested upon on in a different problem set. Furthermore, the effect of sampling rate on feature extraction and the effect of multiple channels on accuracy of a complex multiclass EEG signal is investigated with the help of Fast Fourier Transformation and Principle Component Analysis, which are used for a better signal study in frequency domain. 1-D CNN architecture is used to further classify the EEG signals and other algorithms are applied to test their variability. In the end of our implementation stage, 98% accuracy was achieved in binary class problem of classifying digit and non-digit stimuli, and 36% accuracy was observed in classification of signals resulting from stimuli of digit 0 to 9. The outcomes from the above experiment make us believe that more work needs to be done in the field of complex multi class problems with similar stimuli, as it is clear that future brain-computer interface applications will be focused more in the said domain and many other dimensions are still yet to be explored. A different approach of the classification algorithm has been put forward in this paper which yields better results for multi class features as opposed to the existing methodologies and a new hybrid design for the EEG recording device has also been discussed which includes the fusion of two devices resulting in a 4-channelled device consisting of 2 electrode pairs working at different sampling rates so that improved specifics for the same stimuli can be captured at these two rates resulting in improved classification.