Wavelet and Unsupervised Learning Techniques for Experimental Biomedical Data Processing

Abstract

Learning theories and algorithms for both supervised and unsupervised Neural Networks (NNs) have already been accepted as relevant tools to cope with difficult problems based on the processing of experimental electromagnetic data. These kinds of problems are typically formulated as inverse problems. In this paper, in particular, the electrical signals under investigations derive from experimental electromyogram interference patterns measured on human subjects by means of non-invasive sensors (ElectroMyoGraphic, sEMS surface data). The monitoring and the analysis of dynamic sEMG data reveal important information on muscles activity and can be used by clinicians for both preventing dramatic illness evolution and improving athletes performances. The paper proposes the use of the Independent Component Analysis (ICA), an unsupervised learning technique, in order to process raw sEMG data by reducing the typical “cross-talk” effect on the electric interference pattern measured by the surface sensors. The ICA is implemented by means of a multi-layer NN scheme. Since the IC extraction is based on the assumption of stationarity of the involved sEMG recording, which is often inappropriate in the case of biomedical data, we also propose a technique for dealing with non-stationary recordings. The basic tool is the wavelet (time-frequency) decomposition, that allows us to detect and analyze time-varying signals. An auto-associative NN that exploits wavelet coefficients as an input vector is also used as simple detector of non-stationarity based on a measure of reconstruction error. The proposed approach not only yields encouraging results to the problem at hand, but suggests a general approach to solve similar relevant problems in some other experimental electromagnetics applications.