WAVELET-NEURAL SYSTEMS AS APPROXIMATORS OF AN UNKNOWN FUNCTION – A COMPARISON OF BIOMEDICAL SIGNAL CLASSIFIERS

Abstract

Wavelet-neural systems (WNS) presented in this work, inheriting the properties of neural networks, belong to the class of universal approximators of unknown functions, F, describing the relationship between input X ∈IR^N and output Y ∈IR^M of a process or object. Classifier structures described in this work fulfil the role of approximators of functions, which are able to assign the input signal to a particular class with a given accuracy. A performance comparison of elaborated classifier structures with preliminary time-frequency analysis in the wavelet layer has been made for different types of the neural part. A feed forward multi-layer perceptron and a neural net with radial basic functions are analysed theoretically and practically. Results included in this paper present a comparison of the learning and verification stages of a classifier, tested on the basis of non-stationary signals of heart rate variability. Despite the fact that a WNS with the Morlet basic function gives the best results for the learning phase of WNS, the other tested wavelets used in the preliminary layer, Db4, allow us to obtain the best system performance during its verification.