A P300 speller enables a user to communicate to a computer without any physical movements. In this chapter, a P300 speller system using Devanagari script (DS) is presented. The variation and large size of DS character set increase the problems for classification. To effectively tackle these problems, the application of binary particle swarm optimization (BPSO) has been proposed for channel selection. The algorithm was applied with three different objectives: to maximize the accuracy, to investigate the optimal trade-off between the numbers of channels and the accuracy, and to achieve the maximum accuracy while selecting fixed number of channels. A modification in BPSO algorithm has also been proposed to achieve the third objective. The dataset was acquired from 10 subjects. The mean accuracy of 93.46% was achieved with BPSO algorithm when maximizing the accuracy was the sole objective. Further, average accuracy of 90.62% was achieved while selecting an optimal subset of eight channels.
TopIntroduction
Brain-computer interface (BCI) is a system that records, analyses and decodes the signals generated by brain to provide a medium of communication, without any muscular activity (Wolpaw, Birbaumer, McFarland, Pfurtscheller, & Vaughan, 2002). This allows patients with severe motor disabilities to communicate and control computer machines. BCI-based applications can improve the productivity and independence of such patients and allow them to have a better quality of life. Electroencephalographic (EEG) signals are most commonly used for BCI (Mason, Bashashati, Fatourechi, Navarro, & Birch, 2007).
One of the most useful applications of BCI is the P300 speller (Mak et al., 2011), which works on the principle of detecting P300 component of event related potential (ERP) in recorded EEGs. P300 wave is a positive component produced after 300 ms of stimulation by the rare event of an odd-ball paradigm-based experiment (Farwell & Donchin, 1988; Wolpaw, et al., 2002). A P300 speller system emulates a keyboard, consisting of various technical components such as visually stimulating a subject (using a P300 display paradigm), acquiring the response EEGs (signal acquisition), signal pre-processing (filtering, artifacts removal), electrode channel selection (optional), feature extraction, and classification for character detection.
The first P300 speller system was developed for communicating English text by Farwell and Donchin in 1988 (Farwell & Donchin, 1988). In this speller, the display paradigm consisted of 6×6 matrix of English alphanumeric characters. The rows/columns (RC) of the matrix were randomly intensified. Because of this method of intensification, the paradigm is commonly known as row-column (RC) paradigm. The subject’s task was to focus on the character that he wanted to communicate (commonly referred to as target character). In the intensification of 12 different rows and columns, only one row and 1one column is supposed to contain target character with the probability of flashing of the target character as 0.167 (2/12). This infrequent intensification of target character elicited P300 ERP. However, the P300 ERP may not be elicited and detected with perfection; this sequence of intensification was repeated multiple times. The signal to noise ratio was improved by averaging over multiple sequences. The classification task was to detect the target character with smallest number of sequences. A step-wise linear discriminant analysis (SWLAD) method was used for classification of data acquired with single EEG channel.
A typical P300 speller today works on almost similar experimentation. However, a tremendous development has been witnessed in the performance of the P300 speller systems in last few years. This is due to several modification and improvements in the display paradigm (R. Fazel-Rezai et al., 2012), signal preprocessing and classification methods (Akcakaya et al., 2014), channel selection algorithms (J. Jin et al., 2010; C.-Y. Kee, Ponnambalam, & Loo, 2015; Rakotomamonjy & Guigue, 2008), applications of various data transformation methods for feature selection (Akcakaya, et al., 2014; H. Cecotti, 2011).