Instrumentation & Measurement Magazine 25-1 - 17

showed that the new solution produced contaminants whose
morphology more closely matched the motion artifact observed
in vivo as compared to more simplistic approaches (i.e.,
autoregressive modelling and Markov chain modelling). The
RNN-based motion artifact model allows the synthesis of a
wide range of motion artifacts, and our group is further exploring
ways to generalize and further validate the model.
Other contaminants, such as powerline interference and
additive white noise are often simulated since their characteristics
are relatively well understood. For instance, powerline
interference consists of contributions from the fundamental
frequency (50 Hz or 60 Hz) as well as associated harmonics.
One challenge with controlled contamination of sEMG
signals is that most studies assume that noise is additive and
linear in nature, i.e., clean sEMG plus contaminants equals
contaminated sEMG. While this simplifies the process, researchers
in [18] suggest that the additive nature may not be
a valid assumption in certain conditions. They report nonlinear
dependencies between contaminants and clean sEMG
signals, particularly at low SNR values in the case of baseline
noise. Further investigation needs to be done in this regard to
ascertain if this is true for all types of contaminants, and how
this can be exploited or accounted for in quality assessment
methods.
Evaluation Metrics
Researchers commonly use accuracy as the key metric to evaluate
the performance of a given algorithm. However, accuracy
alone does not fully characterize the performance of the algorithms,
particularly for highly imbalanced classes. Thus,
other metrics such as precision and recall are often used to gain
further insights into the performance of the algorithms [24].
In the case of quality analysis, the impact of false positives
(i.e., classifying clean signals as contaminated) are arguably
not as substantial as false negatives (i.e., classifying contaminated
signals as clean), since in the former, data can be simply
reacquired as necessary, while the latter may lead to misdiagnosis
during analysis. When evaluating algorithms for quality
assessment, it is imperative to report multiple different metrics
to provide a clearer picture of the characteristics of the
algorithms.
Quality Requirements for Different Applications
Different sEMG based applications have different tolerances
for sEMG signal quality. For instance, diagnostic applications
(e.g., analysis of myopathy) often rely on high resolution and
high quality sEMG signals since the presence of contaminants
may substantially impact the analysis and interpretation.
However, sEMG Pattern Recognition (sEMG-PR) based applications,
such as gesture recognition and myoelectric control
of prostheses, are much less sensitive to signal contamination.
While application-specific tolerance to noise is not a well-studied
area of sEMG, some studies have shown that performance
degradation may not be substantial [19] for sEMG-PR applications
when using devices which only capture part of the sEMG
spectrum and with low ADC resolution.
February 2022
The Myo Armband is a popular wearable sEMG device that
wirelessly acquires sEMG signals from the forearm. It has been
relatively commonly used in the sEMG-PR literature [20], even
though the armband has a sampling rate of 200 Hz and an ADC
resolution of 8-bit, both of which are well below specifications
that are considered 'acceptable' for other clinical applications
(1000 Hz and 14-bit, respectively) [1]. It seems that features for
sEMG-PR can be selected such that they are relatively robust to
contaminants while simultaneously being relatively discernible
between motion classes. Low fidelity systems like the Myo
Armband are inexpensive and portable, making them an attractive
solution for sEMG-PR solutions; however, more work
needs to be done to determine acceptable fidelity limits for
other clinical applications.
Discussion and Conclusions
sEMG signals, like other biomedical signals, are highly susceptible
to contaminants, and there has been work done to
enable assessment of recorded signals to ensure that they are
of sufficient quality for sEMG based applications. We outlined
approaches to quality assessment, and highlighted the contributions
our group has made to this area. Quality assessment of
sEMG signals is a hard problem due to the fact that, like other
biomedical signals, we do not have access to the underlying
true signal for modelling, and creating a clean sEMG dataset
for baseline and/or evaluation can be a formidable task. Although
using simulated sEMG and contaminants is attractive,
establishing a useful model and its associated parameters is a
challenge, and it remains unclear whether the models available
reasonably capture the characteristics and variability of
real signals and contaminants.
Despite the challenges with sEMG quality assessment,
based on our experience, many of the methods being developed
can provide tangible benefits when deployed in real
world applications. While sensitivity to contaminants is currently
low, most of the methods can robustly detect severe
contamination issues. This provides the opportunity to catch
and potentially correct errors in acquisition setup, in early
stages of the data collection, or discard poor quality data from
analysis. At this point in development, various techniques can
be used to supplement manual signal inspection.
While detection itself is useful, because different applications
have different tolerances to noise, quantifying what is
detected can help determine whether or not it will be influential.
From our perspective, an important contribution to better
enabling quality assessment is determining application-specific
tolerances so that thresholds for meaningful detection can
be set. Quality assessment methods which quantify also enable
application developers to explore design choices which
might widen their tolerance to contaminants. For instance, feature
sets and classifiers for sEMG-PR can be selected so that the
classification accuracy does not significantly drop due to contaminants
in the sEMG signal.
Finally, despite the scale and challenges, it is our judgment
that building a comprehensive dataset of clean sEMG signals
and contaminants is central to advancing sEMG quality
IEEE Instrumentation & Measurement Magazine
17
Instrumentation & Measurement Magazine 25-1

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