Stochastic signal classification procedures with reference to electroencephalogram analysis

A number of stochastic models and statistical tests are synthesised to develop a general framework for signal analysis and classification. The particular application which provides a focus for this work is the automatic real-time analysis and classification of human electroencephalograms as a clinical aid to diagnosis, treatment and long term monitoring of epileptic patients. Stochastic modelling and estimation procedures are described; these procedures can be employed together with previously recorded data to determine signal classes which are differentiated on the basis of their first and second order moments. Since nearly all analyses of the electroencephalogram study only these moments, and because these moments have been demonstrated to be reliable indicators of the physiological condition, it is expected that the resulting signal classes will be clinically meaningful. It is shown that standard methods of statical hypothesis testing can be used to classify segments of the electroencephalographic record, various approximations are introduced, including a hierarchical test procedure, to develop suboptimal but computationally efficient classification procedures. The use of expert judgement to relate these stochastically differentiated signal classes to the answers to clinically meaningful questions is also discussed; this relationship then permits the system to provide its classification results in a clinically meaningful form.