Application of artificial neural networks to medical signal processing

The dynamics of human sleep have previously been examined using unsupervised clustering techniques [Roberts and Tarassenko, 1992]. This culminated in the hypothesis that the structure of sleep can be described as a linear combination of three underlying processes. These correspond to the conventional, rule-based stages of wakefulness, REM sleep, and the deepest form of non-REM sleep, stage 4. The mixing fractions, p(W), p(R), and p(S), of these three processes vary as sleep progresses, and to estimate them a system has been developed that comprises an autoregressive (AR) model [Makhoul, 1975, Kay and Marple, 1981] followed by two artificial neural networks: a multi-layer perceptron (MLP) and a radial basis function (RBF) network, operating in parallel. The AR model is used to pre-process the EEG on a second-by-second basis, while the mixing fractions for each second are then estimated using the neural networks. The system is currently undergoing clinical trials, during which time the performance of the MLP and RBF networks will be assessed and a choice made as to which one to retain in the final, commercial system