Analysis of the human EEG using self-organising neural nets
Stephen Roberts, Lionel Tarassenko · 1992
The authors report on a neural network approach to a problem for which an existing rule-based system gives not entirely satisfactory results. The precise formal rules of such a system do not lend themselves to the analysis of a signal as variable and complex as the sleep EEG and the authors have sought, instead, to utilise the pattern processing capabilities of artificial neural networks to tackle the problem from a radically different perspective. Starting with no a priori assumptions about the number of states which exist within the EEG during sleep, they have used a version of Kohonen's self-organising algorithm for unsupervised clustering of the parameterised EEG signal. This showed the existence of eight clusters in the feature map, but, more importantly, revealed that there were three types of trajectories between these cluster sites. This observation led the authors to develop a quantitative analysis system by constructing a multilayer network, the intermediate, or hidden, layer being the spatially-organised feature map. >