Design of a recognition system for monitoring the depth of anesthesia, based on the autoregressive modeling and neural network analysis of the EEG signals

The need for a reliable method of measuring the depth of anesthesia has existed since the introduction of anesthesia. Hemodynamic variables are most commonly used as a measure of anesthetic depth by anesthesiologists in the operating room. It has been found that the hemodynamic variables by themselves may not be providing enough information to predict the depth of anesthesia. Since the system most affected by anesthetic agents is the central nervous system, electroencephalograms (EEG), a record of brain activity, can be used to monitor the anesthetic depth. This thesis establishes the feasibility of using a computer-based EEG recognition system to monitor anesthetic depth during halothane anesthesia. The spectral information contained in the EEG signals was represented using a tenth order autoregressive (AR) model. A four layer perceptron feedforward type of network was used in designing the recognition system. The system was trained and its performance tested on the input-output data pairs collected from animal experiments. The input features to the recognition system were based on the AR parameters and the output of the system was depth of anesthesia. Thirteen experiments were carried out on mongrel dogs at various levels of halothane, which itself was controlled using a closed circuit anesthesia controller. Depth of anesthesia was tested by monitoring the response to tail clamping, which is considered to be a supramaximal stimulus in dogs. The system was able to correctly classify the depth in 94% of the cases. The number of neurons in the hidden layer and the number of training samples required for generalization were obtained through clustering analysis. Addition of noise to the input neurons during training made the network more robust to external disturbances and improved the performance of the network significantly. The performance of the system has been shown to be clinically acceptable and has been shown to be robust with respect to inter-patient variability.