Estimation of depth of anesthesia in canines from electroencephalographic and hemodynamic data using neural networks

The main objectives of this research were: (1) To study some bispectral features of EEG with regard to their ability to predict movement in response to noxious stimuli under halothane and isoflurane anesthesia. (2) To design a monitor for determining depth of isoflurane anesthesia.; The first part of this thesis discusses experiments using halothane as the anesthetic agent. Nine experiments were done on seven different mongrel dogs to collect four-channel electroencephalographic (EEG) data. It is shown using neural networks that the boundary separating the two classes (non-responders and responders to the noxious stimuli) in the bispectral feature space is simpler than the boundary separating the same two classes in the power spectral feature space.; The second part of this thesis discusses experiments using isoflurane as the anesthetic agent. Six mongrel dogs were used in 20 experiments to collect two-channel EEG and hemodynamic data. Using non-parametric bispectral features (like the one used for the halothane experiments) and anesthetic concentration from the first 12 experiments we fail to obtain a clear boundary between the two classes of data. We then look at the parametric approach. Two separate AR models of order ten are derived, one from the third order cumulant sequence and the other from the autocorrelation lags of the EEG. The former models the bispectrum and hence monitors the non-Gaussian components of the EEG while the latter models the power spectrum of the EEG. A multiple neural network approach is conceived which takes in both hemodynamic and EEG derived parameters as inputs. Data from the first eleven experiments were used for training the neural networks. They concentrated on acquiring points with unambiguous labels (label = 0.5 for intermediate or ambiguous state of anesthesia, neither responder nor non-responder). The next nine experiments were used for validating the neural networks. They concentrated on getting points in and around the one MAC (minimum alveolar concentration) region. A fuzzy integral of the individual neural network estimates was used to obtain the final estimate of the depth of anesthesia. The first three of the nine experiments in the validation set were used to obtain the fuzzy densities for the fuzzy integral. The fuzzy integral output was then tested on the remaining six experiments.; The neural network operating on the third order cumulant parameters of EEG classified 79% of the 52 points correctly. The neural network operating on autocorrelation lag parameters of channel one EEG classified 69% of the 52 points correctly and that operating on similar parameters from channel two EEG classified 71% of the 52 points correctly. The last neural network operating on hemodynamic parameters classified 75% of the 52 points correctly. The fuzzy integral output tested on 35 points classified 32 of them correctly (91%).; The multiple neural network scheme with a fuzzy integrated output has been found to be a feasible method in monitoring depth of isoflurane anesthesia. The networks would have to be trained on data from different anesthetic agents to make it more drug independent.