Neural networks for predicting depth of anesthesia from auditory evoked potentials: a comparison of the wavelet transform with autoregressive modeling and power spectrum feature extraction methods

Three neural network (NN) models were used to compare the depth of anesthesia prediction performance of the wavelet transform (WT) with power spectrum and autoregressive parameters of the midlatency auditory evoked potentials. The authors' results show that the NN trained with a combination of the WT parameters and anesthetic concentration correctly classified all of the data belonging to a test set. The size of the network required for complete training was the smallest of the three designs. The better performance of the WT can be attributed to good localization in the time-frequency domain and low sensitivity to signal-to-noise ratio.