Classification of EEG data

Abstract: "The objective of the work presented in this report is to explore prerequisites of a system currently developed at the University of Technology in Graz, a Brain-Computer Interface (BCI). This system is designed to recognise a limited set of 'brain states' based on a number of EEG recordings from the subject's scalp and to issue the classification result as a command to a computer. The aim of such a system is to provide handicapped persons with an additional means of communication. There are several preconditions which have to be fulfilled before a BCI can be put to operation: (i) the system must be quick enough to process and interpret several EEG channels in parallel; current hardware and software fulfil this criterion; (ii) classification of EEG must be quick and accurate, the classification method Learning Vector Quanitzation [sic] (LVQ) fulfils this requirement; (iii) the features presented to the classifier must contain maximum information about the differentiation of the specified brain states. This report is part of a PhD thesis (chapters 1,2,5,7,8 and Appendix of 'Feature Selection and Classification of EEG Data - Prerequisites for a Brain-Computer Interface', Graz University of Technology 1994) which focuses on this last point, i.e. on the investigation of the length and position of the time interval used for classification as well as on the question of which electrode positions should be incorporated into the classification process. Furthermore, the thesis describes possibilities to use artificial intelligence approaches for feature selection (Genetic Algorithms and Numeric Decision Trees), i.e. to find a subset of a given set of features which is minimal in size and at the same time maximal in informational content, to minimise the computational cost during operation of a BCI. Two experiments, one designed to test the differentiability of two kinds of movement and one designed for four kinds of movement, are analysed to find suitable 'brain s tates' for the construction of a BCI prototype. Based on these results two prototypes, Graz BCI I and II, are presented, whereby the former is based on the two brain states 'left and right hand movement planning' and the latter includes the additional brain state 'right foot movement planning'. On-line results of both prototypes are discussed and several off-line analyses are presented which allow some guidelines for future BCI prototypes."