Instrumentation & Measurement Magazine 25-1 - 15

sensitive to multiple contaminants. For instance, although
SER was crafted to quantify ECG contamination, it has been
reported to also be sensitive to motion artifacts [2]. This is
problematic since multiple sources of contamination can influence
the metrics, making it difficult to guide appropriate
mitigation strategies. The proposed techniques can identify
specific kinds of contaminants accurately only for relatively
low SNR values. For instance, the SVM classifier was able to
identify specific kinds of contaminants at 100% accuracy for
SNR values less than 0 dB. The only contaminant that was
reported to be reliably detected at even low levels of contamination
is ADC clipping, where signals with as few as 0.25%
clipped samples were identified with 90% accuracy [12]. Despite
the relatively low sensitivity, our group regularly uses
one or more of these techniques to identify presence of severe
contamination in newly acquired data, allowing us to flag the
data and take appropriate actions (i.e., reacquire or discard
from analysis).
Challenges with Evaluating Assessment
Techniques
Evaluation of the various automated sEMG signal quality assessment
techniques can be done in a few different ways. One
way is by introducing the technique in the context of a sEMG
application and measuring improvement. For instance, in
sEMG classification for myoelectric control, if the introduction
of an automated quality assessment technique results in
an increase in classification accuracy, then that technique can
be considered to be useful. This approach is limited however,
since it confounds application performance with the quality
assessment evaluation. Another way to evaluate techniques
is to compare one assessment against a consensus from various
other techniques. If the technique being tested strongly
agrees with the consensus, then it is likely that it works as intended.
A weakness of this approach is that deviations from
the consensus may actually be improvements in assessment
performance. The best way to evaluate assessment techniques
is to establish a 'ground truth' about the data being used in
the evaluation. This requires a comprehensive dataset of clean
sEMG signals and a means of contaminating them with known
contaminants. Studies proposing quality assessment techniques
use various combinations of real and/or simulated
sEMG data and real and/or simulated contaminants to construct
the requisite ground truth datasets.
Creating an Evaluation Dataset
Creating Clean sEMG Dataset: The most straightforward, but
perhaps most challenging way of creating a clean sEMG signal
dataset is by recording signals in vivo. Studies that have
used real sEMG data to evaluate an assessment technique use
a wide range of sample sizes, ranging from 5 subjects [9] to
40 subjects [11], showing a lack of consensus. Because of the
wide variability in sEMG (among individuals, muscle groups,
contraction level, etc.), a large dataset is necessary to establish
the full extent of the efficacy of a proposed technique, and
February 2022
constructing a comprehensive dataset with full representation
of the diverse nature of muscles and variation of physiology
across people, is in itself daunting. Moreover, to ensure high
fidelity, these signals have to be recorded using high-quality
equipment, often available only in laboratory conditions,
and by highly-trained personnel, which encumbers distributed
or crowd-sourced efforts for creating the dataset. Finally,
to validate that each signal is clean, the sEMG signals must be
manually inspected by experts. This process can be extremely
time-consuming and error-prone, even for creating a subset
of the dataset (e.g., for a specific muscle). Simulating sEMG
signals is an attractive alternative since the signals can be guaranteed
to be clean, and a large set of signals can be generated
with reasonable computation power.
Several mathematical models have been proposed in the
literature for simulating sEMG, ranging from simple models
with few parameters to complex models with several
hundreds of parameters. A simple model was proposed by
Shwedyk et al. [14] which starts with a parametric model of
the sEMG power spectra and uses that to specify a way to filter
white Gaussian noise to generate sEMG signals. The filtering
process produces a signal with normally distributed amplitudes
and a spectrum shape that matches the general shape for
sEMG. This model requires only two parameters to be specified:
the lower and upper cut-off frequencies of the filter (fl
, respectively).
and
fh
Fig. 3a and Fig. 3b show a real sEMG signal recorded from
the bicep brachii and its PSD, and Fig. 3c and Fig. 3d show an
example signal generated by the Shwedyk model, whose PSD
closely resembles the recorded sEMG. The range of values of
the filter parameters (fl
and fh
) required to represent sEMG
from a wide range of muscles is unknown. In our studies [2],
[6], [12], we used Shwedyk's model to generate clean sEMG
data, with fl
ranging from 30-60 Hz and fh
ranging from 60-
160 Hz.
A more complex model was proposed by González-Cueto
and Parker [15]. It starts with a parametric model used to generate
an SFAP and builds up from there by summing SFAPs
into MUAPTs, and summing MUAPTs to form an sEMG signal.
This model takes into account physiological properties of
the fiber (e.g., conduction velocity, length, and depth), as well
as the MUs (e.g., number of fibers and firing rate) to capture
both signal generation and the effects of the muscle medium
through which it travels. Myosim [4] is a tool developed by
our group that implements this model to generate an sEMG
signal. The tool enables users to specify parameters, such as
the number of motor units, the number of fibers per motor
unit, the length and depth of fibers, the propagation velocity
of fiber impulses, and the starting points of impulses along
the fiber, relative to the measuring electrodes. The tool can
also introduce variability in the model parameters to represent
the physiological variability observed in vivo. As such,
the tool allows the generation of a wide range of sEMG signals.
In addition, the tool provides the ability to simulate instrumentation
characteristics such as ADC quantization, amplifier
saturation, and electrode configuration [4]. An example sEMG
IEEE Instrumentation & Measurement Magazine
15
Instrumentation & Measurement Magazine 25-1

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-1
