An Improved Algorithm For Discrimination Of Motor Imagery Based Brain Computer Interface

Abstract

Brain computer interface (BCI) system is one of the means of communication that allowscontrolling devices or communicating with others by active brain signals only without movement signals. The main application of this system is to help paralyzed people or disabled people to communicate with the outside world. The electroencephalography (EEG) based motor imagery is one of the methods that the BCI system uses to identify the expected behavior through brain signals. This system is considered to be accurate when it can first isolate the irrelevant brain signals from the other sources that mixed with it.These artifact signals are originating from other organs of the body or from the outside medium.Secondly, the system's ability to differentiate between different imagery movements such as to differentiates the movement of the right hand from the left. To achieve this, two algorithms were built through which the noise signals are isolated and then discriminating between four different imagery movements (right hand, left hand, both feet, and tongue). Data set IIa from BCI competition IV was used to test both algorithms. The first algorithm is to use anIIRband pass filter (8 - 30) Hz as preprocessing stage in order to purify the signal and extract brain waves with beta β and mu μ frequencies and then use the combined Wavelet Common Spatial (Wavelet-CSP) method to extract the most efficient features of the signal before it is presented to support vector machine (SVM) classifier. In the second algorithm, Independent component analysis (ICA) method was used in the preprocessing stage, in order to extract the most efficient features in each electrode separately, using two software: SOBI and FastICA to compare them and determine which is faster when applied. It is proved that the method of SOBI is the fastest in the execution of the algorithm. In the second stage, the Wavelet-CSP method is used to extract the most important characteristics of the signal before presenting it to the SVM classifier. The two algorithms proved successful with a Kappa coefficient of 0.53 for the first method and 0.55 for the second method in solving the problem of discrimination of four different imagery movements. Moreover, the second algorithm is more interestingnot because of its high performance only but also the possibility to adopt this system to be used in real time.