UBC Theses and Dissertations
Estimation of the level of anesthesia during surgery by automatic EEG pattern recognition McEwen, James Allen
The feasibility of developing an automatic electroencephalographic (EEG) pattern recognition system for reliably estimating the level of consciousness of surgical patients during general anesthesia is investigated. An effort was made to establish a valid methodology, by identifying and controlling as many extraneous variables as possible and by ensuring that the work would be relevant to current anesthetic practice. The data base that was established for use in all experimental investigations consists of 938 EEG pattern samples from 72 subjects and three types of anesthesia. Each EEG pattern sample corresponds to one of five possible clinical levels of anesthesia. The use of automatic pattern recognition techniques, in conjunction with heuristic techniques of clinical EEG analysis, to develop spectral and time domain EEG pattern recognition systems is described. All of the initially developed systems extract a small number of heuristically derived features from unknown EEG pattern samples. The classifiers in these systems employ Bayes decision rule under the assumption that the extracted features are statistically independent. A rationale concerning the choice of this particular feature extraction scheme and pattern classification algorithm is presented and discussed. Consideration is given to the general problem of how to use a relatively small set of available EEG pattern samples to effectively evaluate the performance of EEG pattern recognition systems. Two non-parametric techniques which provide particularly informative and efficient estimates of the performance of such systems are formulated. Results obtained by employing these techniques to estimate the performance of the initially developed spectral and time domain EEG pattern recognition systems are presented. The results clearly demonstrate the feasibility of estimating the level of anesthesia by means of automatic EEG pattern recognition. However, the results also indicate that the initially developed systems are not sufficiently reliable for immediate and general clinical application. Theoretical techniques are developed to model some relevant statistical properties of spontaneous EEG activity, with a view to improving the performance of the initially developed EEG pattern recognition systems. Results which were obtained by applying the modelling techniques to some specific ensembles of EEG pattern samples are presented. The comparative advantages of employing alternate methods of EEG analysis are then discussed in relation to the estimated statistical characteristics of the particular EEG ensembles under consideration. Several factors which, could adversely affect the reliable performance of EEG pattern recognition systems in general, and the initially developed systems in particular, are identified and discussed. Various schemes for improving the performance of the initially developed systems are suggested and an evaluation of the practicability of each is presented.
Item Citations and Data