Influence of Combination of Criteria for Extraction of Features on the Classification of Biological Signals

Abstract

Expert doctors use the shape of the principal components of the Brain Stem Auditory Evoked Potential (BSAEP) signal to diagnose patients with multiple sclerosis. The diagnosis involves the estimation of the effects of the disease on the form of the waveform components of BSAEPs. Since these components are localized in time and frequency a packet wavelet decomposition of the signal is used to compress it. The information obtained by the packed wavelet can be use to feed artificial neural networks (ANN) with Radial Basis Functions for the same purpose of obtaining a diagnosis. Due to the paucity of data, the signals must be preprocessed. From the hundreds or thousands of wavelet coefficients, only eight are selected using different criteria. Those are used to train an artificial neural network with radial basis functions. We have found that if we combine some of the selection criteria to differentiate sick and healthy people, only one combination of criteria provided better results that using each criterion alone, and other combinations worked better only with some wavelet bases.

Keywords

• Neural networks
• Biological Signals
• Packet Wavelets
• Principal Components
• Multiple Sclerosis
• Radial Basis Functions