Computational Intelligence in Analysis of Electromyography Signals - Computational Intelligence in Biomedical Engineering - page 240

(BEP) and end (EEP) extraction points of the MUAP are identified by sliding
an extraction window of length 3 ms with width 40 µV. The BEP is found

if the voltage to the left of the MUAP waveform after searching for 3 ms
remains within

±

40 µV. The EEP is determined similarly by searching to
the right. A baseline correction is then performed to obtain the area of the
MUAP waveform. Some researchers have extracted MUAP information using
weighted low-pass filters (Xu and Xiao, 2000) and Weiner filters (Zhou et al.,
1986) to obtain the signal decomposition of surface EMGs.

±

5.5.2 Intermediate Processing: Decomposition and
Identification of Motor Unit Action Potential Trains
Decomposition of the EMG signal refers to the separation of the complex
signal, such as MUAPTs, into the individual action potentials. This interme-
diate processing is usually required in surface EMG because of the complexity
of the MUAPTs originating from effects of the adjoining muscles. The objec-
tive of decomposition (Thompson et al., 1996) is to determine the EMG sig-
nals at all levels of contraction, separate reliably and accurately all MUAPs,
and to provide an acceptable level of decomposition in the presence of noise.
An inherent problem in this scheme is that superposition of MUAPs makes
it dicult to determine the true number of MUAPs in the composite wave-
form. Besides that, movements due to surface-electrode placement distorts the
composite MUAPT further, whereas discharge to discharge variability makes
MUAP localization dicult.
Since the 1980s, efforts to extract single MUAP patterns from the composite
MUAPTs have been attempted using basic signal-processing methods
(LeFever and De Luca, 1982a,b). De Luca (1993) described a decomposi-
tion system, which employed a special quadrifilar electrode to measure three
EMG channels simultaneously. The effect of superposition was reduced by
bandlimiting the signals to a window of 1-10 kHz and sampling at 51.2 kHz.
Next, a template-generation process was employed for 150 waveforms using
a nearest-neighbor algorithm. In this process, new data were compared with
templates using a least squares method, which was found to achieve up to
90% classification accuracy when decomposing 4-5 MUAPs. A competing sys-
tem was the automatic decomposition electromyography (ADEMG) system
developed by McGill et al. (1985), which used digital filters to transform the
rising slope of MUAPs into spikes for detection using a zero-crossing detec-
tor. The authors reported an identification rate of 40-70% and argued that
their method outperformed template matchingbecause sampling can be under-
taken at the Nyquist frequency of 10 kHz for maximum information extrac-
tion. Work by Gerber and Studer (1984) involved partitioning the waveforms
into active and inactive segments where the inactive segments were perceived
to contain noise and discarded. The nearest-neighbor algorithm was used
to cluster the segments based on a feature vector and superpositions were
resolved using template matching. Louden (1991) in his PhD thesis studied
the decomposition of needle EMG waveforms using normalization to separate