Analysis of Effect of RSVP Speller BCI Paradigm Along with CNN to Analysis P300 Signals

Abstract

People suffering from neurological disorders, including spinal cord injury (S.C.I.), stroke, Parkinson’s disease may be severely paralyzed and incapable of performing any motor functions. However, they may still have some cognitive abilities, and that can be accessed through brain-computer interaction (BCI). Electroencephalography (E.E.G.) pattern classification is attractive for many researchers in the field of BCI. P300 may be a style of ERP signal which is employed by the BCI system. P300, well known as a prominent component of event-related potential (ERP) from E.E.G. signal. It’s also elicited in an oddball paradigm. In some cases, patients get completely locked in until losing control of their ocular movements. Researchers have shifted toward the BCI system because of these people, which will work without the eye movement. Hence the proposed approach is attempting to implement the BCI system without using oculomotor movements. The traditional methods/algorithms such as the hidden Markov model, support vector machine (SVM), and Linear discriminant analysis (LDA) may not perform well for cross-subject variations with extreme variations. Hence, in recent years, deep neural networks, particularly the conventional neural networks (CNN) widely used, have shown high performance compared to the traditional approach for various applications. CNN can extract data from raw data as well as give us unknown information about the data. Most of the research work related to P300 and various speller system involved the gaze of the subject. However, these systems are not suitable for the patient having problems in oculomotor control. The proposed convolution neural network (CNN) has been used for high-level feature extraction and improves the performance of P300 classification to predict target and non-target character. Along with CNN to analyze P300 signals, this study used gaze independent BCI speller paradigm called rapid serial visual presentation (RSVP). We have selected/formed the letters intuitively by attending target letters in the stream of visual stimuli. A vocabulary of 30 symbols was presented one by one in a pseudo-random sequence at the same display location. We applied the CNN on the RSVP Dataset which gave us an average accuracy of 97% which is better than the previously implemented on the BCI competition dataset II without channel selection before the classification, i.e. 95.5%.