A study of electroencephalographic descriptors and end-tidal concentration in estimating depth of anesthesia

To study the usefulness of three electro-encephalographic descriptors, the average median frequency, the average 90% spectral edge frequency, and a bispectral variable were used with the anesthetic concentrations in estimating the depth of anesthesia.

Four channels of raw EEG data were collected from seven mongrel dogs in nine separate experiments under different levels of halothane anesthesia and nitrous oxide in oxygen. A tail clamp was used as the stimulus and the dog was labeled as a non-responder or responder based on its response. A bispectral variable of the EEG (just before a tail clamp) and the estimated MAC level of halothane and nitrous oxide combined were the two features used to characterize a single data point. A neural network analysis was done on 48 such data points. A second neural network analysis was done on 47 data points using average 90% spectral edge frequency and the estimated MAC level. The average median frequency of EEG was also evaluated, although a neural network analysis was not done.

The first neural network needed nine weights in order to train and correctly classify all of the 12 points in the training set under a training tolerance of 0.2. It could correctly classify all of the remaining 36 data points as either belonging to responders or non-responders. A cross-validation procedure, which estimated the overall performance of the network against future data points, showed that the network misclassified two out of the 48 data points. The second neural network needed 25 weights in order to train and classify correctly all of the 26 points in the training set under a tolerance of 0.2. It was later able to classify all of the 21 points of the test group correctly.

The bispectral variable seems to reduce the non-linearity in the boundary separating the class of non-responders from the class of responders. Consequently, the neural network based on the bispectral variable is less complex than the neural network that uses a power spectral variable as one of its inputs.