Performance and training strategies in feedforward neural networks: an application to sleep scoring

A comparison is made of the performance of single- and multilayer perceptrons in the scoring of sleep stages under different training conditions. The input to the neural network is a set of feature vectors, and the sleep staging is the output. Performance is the degree of agreement with the human scorer. For this application the single-layer perceptron performed at the same level as the multilayer perceptron. The best strategy for training the network is the use of human a priori knowledge. The neural network performed at the same level as other much more difficult to implement pattern recognition schemes.<>