Characterization of neural network simulations for optimal classification of intraoperative electroencephalograph data

Abstract

Accurately and consistently determining depth of anesthesia during surgical procedures is a significant problem. A more objective technique than traditional methods is required. Different concentrations of anesthetic drugs have been shown to affect the Electroencephalograph; results, however, are inconsistent when using only visual inspection of the EEG. An automated technique using Neural Networks for classifying anesthetic depth from EEG data is proposed. Neural Networks are reviewed. Reasons for choosing a Backpropagation Network (BPN) are discussed. Ambiguities in previous BPN research are presented. Over 3,000 networks are formed, demonstrating training and classification properties while altering network topologies, parameters and performance criteria. Tests are performed on the Power spectrum and Phase portions of the EEG data. Optimal BPN parameters and topologies are shown. Results are compared with a statistical paradigm.