A new approach to classification of brainwaves

摘要

In past years statistical pattern recognition has been used in the classification of evoked potential waveforms. The statistical approach has more recently been enhanced and often replaced by the structural approach in which patterns are depicted as being built out of subpatterns in various ways of composition. This paper describes an application of a general approach to pattern recognition in which these principal approaches are combined. The result is a two-stage classification algorithm which places evoked potential waveforms into clinically significant classes. Fourier descriptors are used as feature representations of each waveform shape. Euclidean distance is used to determine an optimal number of these features for design and testing of the algorithm. The development of a dissimilarity measure follows and is crucial to the overall success of the algorithm. An interdistance matrix of the training sample which is all that is retained from the original input is projected into a pseudo-Euclidean vector space resulting in a minimal vector representation of the sample. Accuracies of 90% of correctly classified test waveforms have been obtained.