Application to biomedical signals - Practical Biomedical Signal Analysis Using MATLAB - page 206

force, muscle fatigue, and the condition of unused muscles and muscles in chronic
pain syndromes are considered in [Fuglsang-Frederiksen, 2000].
In the system proposed by Pattichis and coworkers [Pattichis and Elia, 1999] for
differentiation of myogenic and neurogenic signals from normal EMG, both time
domain and frequency domain features of EMG were considered. The following
time-domain parameters were automatically determined: 1) MUAP duration, 2) spike
duration measured from the first to the last positive peak, 3) amplitude: difference
between the minimum positive peak and maximum negative peak, 4) area: sum of
the rectified MUAP integrated over its duration, 5) spike area: sum of the rectified
MUAP integrated over spike duration, 6) phases, 7) turns. The frequency features
were based on the spectra estimated by means of AR model of order 12. They in-
cluded: 1) bandwidth, 2) quality factor—the ratio between the dominant peak fre-
quency and bandwidth, 3) median frequency, 4) spectral moments describing the
shape of the spectrum; additionally for classification AR coefficients were taken into
account.
Univariate analysis was applied to find the coefficients which best separated the
normal from pathological groups and multiple covariance analysis to select stepwise
the best features and find the correlation between the parameters. For classification
into three considered groups artificial neural networks (ANN) were applied. Three
algorithms were used: back-propagation, the radial-basis function network, and self-
organizing feature map. They are all available in MATLAB Neural Network Toolbox.
The univariate and multivariate analysis indicated as the best classifying parame-
ters: among time domain parameters—duration, and among AR spectral measures—
median frequency. Median frequency and central frequency for the neurogenic group
were higher, and for the myogenic group these parameters were lower than for the
normal EMG. In classification by ANN the highest diagnostic yield was obtained for
time domain features and next for frequency domain features; AR coefficients gave
the poorest results.
4.3.3 Decomposition of needle EMG
The problem of automatic decomposition of an EMG signal into its constituent
MUAP trains is important not only for medical diagnosis, but also for basic studies
of the neuromuscular system. One of the first systems performing EMG decomposi-
tion was proposed by LeFever and de Luca [LeFever and De Luca, 1982]. The signal
from the intramuscular needle electrode was sampled at 50 kHz and high-pass fil-
tered in order to reduce the amplitude of slow rise-time MUAP waveforms recorded
from fibers more distant from the electrode. The decomposition program was based
on a template-matching algorithm and included a routine for continuous template
updating, in which every consecutive MUAP could be classified as belonging to the
existing template (then the template was updated), or used as the initial estimate of
a new template, or discarded. The verification of the initial decomposition results
was based on the a priori knowledge of firing statistics, i.e., inter-spike interval dis-
tribution within MUAP trains generated during isometric constant force contraction.
The program might work automatically; however the reliability of the results was