Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 210

et al., 2003] for analysis of long-term multichannel EMG. In the first step of analysis
so-called active segments containing MUAPs were selected by thresholding based
on an estimated signal to noise ratio. The Daubechies wavelets [Daubechies, 1992]
were used and coefficients from lower frequency bands were applied for further steps
of analysis. The high frequency coefficients were considered as more contaminated
by noise and so called time-offset, which depended on the position of MUAP in the
analysis window. The next steps of analysis included supervised classification and
clustering based on the selected wavelet coefficients. The provisions were made to
account for the possible changes of MUAP shape during the experimental session, by
introducing adaptation of MUAPs templates. The achieved accuracy of classification
was reported as 70%.
The problem inherent in application of discrete WT to EMG decomposition is
connected with the fact that the description of the signal shape in terms of wavelet
coefficients depends on the position of MUAP in the data window. This problem is
usually alleviated by alignment of the main peaks of the analyzed potentials. Another
approach which might be used for MUAP identification, which is free of restrictions
imposed by the presence of the measurement window, is matching pursuit. To our
knowledge such an attempt has not yet been undertaken.
4.3.4 Surface EMG
Because of its noninvasiveness, surface EMG (sEMG) has always attracted the
attention of investigators seeking new methods for EMG analysis. One of the first
applications of sEMG was in studying muscle fatigue. With increasing fatigue, the
sEMG spectrum shifts toward lower frequencies. It was suggested that the spectral
shift was caused by the decrease of conduction velocity of action potential along
the muscle fibers [Lindstrom et al., 1970, Sadoyama and Miyano, 1981]. However,
according to [Linssen et al., 1993] despite the definite and strong influence of the
motor fiber conduction velocity on the power spectrum, the frequency shift cannot be
explained by a change in propagation velocity alone. Nevertheless, reported changes
of propagation velocity in some muscle diseases [Zwarts et al., 2000] indicate the
usefulness of sEMG spectral analysis for medical diagnosis.
From sEMG spatial information concerning the MUAPs activity may also be
gained. The analysis of surface EMG maps may provide indirect information on
the spatial recruitment of motor units within a muscle [Falla and Farina, 2008].
The analysis of sEMG signals used for control of prostheses requires fast and effi-
cient signal processing techniques. In this case the application of WT appeared quite
successful [Englehart et al., 1999]. The sEMG signals were recorded from biceps
and triceps during elbow and forearm movements. The aim of the study was the
best distinction of the four classes of EMG patterns. Time-frequency analysis was
performed by means of short-time Fourier transform (STFT), wavelets, and wavelet
packets (WP). Different kinds of wavelets were considered. The best results in case
of WT were obtained for Coiflet-4 and for WP for Symmlet-5. The task of the re-
duction of dimensionality of feature sets was performed by Euclidean distance class
separability (CS) criterion and principal component analysis (PCA). For classifica-