Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 216

Blind source separation method was used by [Holobar et al., 2009] who mea-
sured sEMG from four muscles by means of 13x5 grid during low force contraction.
The approach was based on the convolution kernel compensation (CKC) algorithm
proposed by [Holobar and Zazula, 2007]. The signal was modeled as a convolutive
mixture of sparse pulse trains, which carry information about the rising times of the
detected symbols and the symbols themselves. The spatial and temporal statistics of
symbols, i.e., convolution kernels, was combined with the information about their
overlapping probability, in order to blindly reconstruct their pulse sequences. The
model implied admixture of a white noise, however residual noise is dominated by
the contributions of indecomposable MUAPs and does not meet the criterion of spa-
tial and temporal independence. CKC assumed as well, that action potential of MU
remains constant throughout a contraction. In [Holobar et al., 2009] the accuracy of
the method reported as very good was assessed by the method proposed by [Farina
et al., 2001] (described below). In case of one of the muscles the results give 98%
of agreement in MUAPs identification with the intramuscular recordings. However,
the method is limited to low force contractions (10% of maximal force) and biased
toward the high-threshold MUs, which was admitted by the authors.
The performance of different methods of sEMG analysis is usually evaluated by
comparison with the manual decomposition of a given signal. In view of the multi-
tude of the approaches to the problem of EMG decomposition there emerged the need
for a test tool to evaluate and compare different methods. Such a tool was proposed
in [Farina et al., 2001]. The approach was based on the model for the generation of
synthetic intra-muscular EMG signals. The library of 18 test signals was designed.
The indexes estimating performance for segmentation and classification stages of
decompositions and measure of association between model and detected classes of
MUAPs were proposed. Additionally global indices were introduced, such as the dif-
ference between the number of model classes and the number of classes estimated
by the algorithm, mean firing rate, and activation interval detection.
Yet another way of testing the decomposition algorithm was applied in the recent
work of [Nawab et al., 2010]. It relied on reconstruction of the signal from the identi-
fied MUAPs and decomposition of this signal with some noise added. This procedure
is illustrated in Figure 4.57. In the experiment a five-pin surface sensor was used ( Fig-
ure 4.48) and the pair wise voltages between 5 pins were analyzed. The method of
the decomposition was based on the algorithm proposed by [LeFever and De Luca,
1982]. The operation of the algorithm started by extracting as many MUAPs as pos-
sible from experimental sEMG action potential templates. Then it searched for signal
regions where the extracted templates were in superposition with each other and with
unidentified potentials. The algorithm required that the unidentified action potentials
account for less than 25% of the signal energy. The assumption about inter-pulse
intervals was only that they be less than 0.35 s. The algorithm allowed for changes
in the action potential shape of each MUAP to evolve over the duration of the con-
traction. The authors considered 22 signals recorded from 5 muscles for contractions
reaching 100% force levels. They reported the accuracy estimated by the reconstruct-
and-test procedure (Figure 4.57) ranging from 77-97%. Upon analysis of decompo-
sitions performed on these signals it was found that 92% of over 3000 firings from