Application of Quantitative Methods of Signal Processing to Automatic Classification of Long-Term EEG Records

The aim of the work described in the paper has been to develop a system for processing long-term EEG recordings, especially of comatose state EEG. However with respect to the signal character, the developed approach is suitable for analysis of sleep and newborn EEG too. EEG signal can be analysed both in time and frequency domains. In time domain the basic descriptive quantities are general and central moments of lower orders, in frequency domain the most frequently used method is Fourier transform. For segmentation, combination of non-adaptive and adaptive segmentation has been used. The approach has been tested on real sleep EEG recording for which the classification has been known. The core of the developed system is the training set on which practically depends the quality of classification. The training set containing 319 segments classified into 10 classes has been used for classification of the 2hour sleep EEG recording. For classification, algorithm of nearest neighbour has been used. In the paper, the issues of development of the training set and experimental results are discussed.