computers
Article
Implementation of an Intelligent EMG Signal Classifier Using
Open-Source Hardware
Nelson Cárdenas-Bolaño, Aura Polo * and Carlos Robles-Algarín
Facultad de Ingeniería, Universidad del Magdalena, Santa Marta 470004, Colombia;
nelsoncardenas28@gmail.com (N.C.-B.); croblesa@unimagdalena.edu.co (C.R.-A.)
* Correspondence: apolol@unimagdalena.edu.co; Tel.: +57-5-4217940
Abstract: This paper presents the implementation of an intelligent real-time single-channel electromyography (EMG) signal classifier based on open-source hardware. The article shows the experimental design, analysis, and implementation of a solution to identify four muscle movements from
the forearm (extension, pronation, supination, and flexion), for future applications in transradial
active prostheses. An EMG signal acquisition instrument was developed, with a 20–450 Hz bandwidth and 2 kHz sampling rate. The signals were stored in a Database, as a multidimensional array,
using a desktop application. Numerical and graphic analysis approaches for discriminative capacity
were proposed for feature analysis and four feature sets were used to feed the classifier. Artificial
Neural Networks (ANN) were implemented for time-domain EMG pattern recognition (PR). The
system obtained a classification accuracy of 98.44% and response times per signal of 8.522 ms. Results
suggest these methods allow us to understand, intuitively, the behavior of user information.
Keywords: myoelectric prosthesis; electromyography (EMG); artificial neural networks (ANN);
pattern recognition
1. Introduction
Citation: Cárdenas-Bolaño, N.;
Polo, A.; Robles-Algarín, C.
Implementation of an Intelligent
EMG Signal Classifier Using
Open-Source Hardware. Computers
2023, 12, 263. https://doi.org/
10.3390/computers12120263
Academic Editor: Hersh Sagreiya
Sagreiya
Received: 14 July 2023
Revised: 22 August 2023
Accepted: 25 August 2023
Published: 18 December 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Individuals with disabilities constitute one of the populations that are most affected
in terms of accessing services such as education, employment, and healthcare [1]. Governments and organizations around the world are promoting policies and strategies to
improve the quality of life for this population. In 2014, the World Health Organization
(WHO) established action plans to support people with disabilities by strengthening rehabilitation, assistive technology, assistance and support services, and community-based
rehabilitation [2]. In the matter of motor disabilities, research and development regarding
assistive technology, such as active prostheses, can play an important role in reaching these
aims. Surface EMG signal processing is crucial in building active prostheses that give users
a more realistic experience [3].
EMG research is being used in different applications, such as in sports medicine for
fatigue monitoring [4–6], education [7,8], human-machine interfaces (HMI) [9], and clinical
applications like diagnosis of neuromuscular disorders, among others [10,11]. In the field
of prosthetics, pioneering investigations, such as those implemented in references [12–14],
explored the use of EMG signals and served as a support for many later works, among
which the most notable are the Myo Armband [15], an EMG band of eight channels [16],
prostheses with electro-mechanical designs based on EMG signals, and the EMG band
designed by Microsoft with a wireless system based on Zigbee with six EMG channels [17].
For the classification of EMG signals, a process of the capture and subsequent extraction of features is performed, which can be represented in the domains of time, frequency, time-frequency or space [9]. Much of the recent research focuses on the development of EMG prostheses based on multi-channel acquisition or high-density measurements [9,18–20]. However, these approaches came with specific challenges: having more
Computers 2023, 12, 263. https://doi.org/10.3390/computers12120263
https://www.mdpi.com/journal/computers
Computers 2023, 12, 263
2 of 18
channels will result in larger datasets [21]. Additionally, they will generate more complex
configurations, incur higher manufacturing costs, lead to increased energy consumption,
reduced efficiency, and pose greater difficulties for users to adapt [22,23]. Due to this, the
process of signal feature extraction can potentially generate numerous representations.
Consequently, this could lead to a decline in the performance of the recognition system [24].
To cater to these requirements, employing analytical tools for evaluating the extracted
information is the appropriate approach. Such tools assist researchers in selecting various
features and channels, while also assessing the performance of the algorithms [25].
Therefore, it is necessary to know a methodological approach that allows a better
understanding of features’ behavior. For this, the analysis of the features can be performed
using several approaches [26]. One of these is the filter approach, which allows the creation
of information quality coefficients according to their separation, correlation or consistency.
In the field of EMG signals, some of these have been the Mahalanobis distance [27,28],
Bhattacharyya distance [27,29], Davies–Bouldin index [28], Fisher–Markov Selector, and
Minimum Redundancy Maximum Relevance [25].
The second approach is known as Wrapper, and it involves applying an evaluation
criterion, a search strategy (exponential, sequential or random), and a stop criterion to
discover an optimal subset of features. In the study of biomedical signals, this approach
has been utilized alongside techniques like Particle Swarm Optimization [24,30], Sequential
Forward Searching [28,29], and Bacterial Memetic Algorithm [31], among other methods.
In line with the previously mentioned requirements and methods, this paper combines the simplicity approach with information analysis strategies to develop a real-time
intelligent classifier for single-channel EMG signals, employing open-source hardware. In
the context of EMG pattern recognition, time-domain signals were employed alongside
a feed-forward neural network. Testing was conducted using forearm muscle signals
corresponding to the movements of extension, pronation, supination, and flexion.
Machine Learning (ML) techniques have several advantages in various applications,
making them ideal for automating complex processes and large-scale data processing.
These methods improve accuracy and efficiency in decision-making and are ideal for
classification and prediction tasks [32], as they learn from data and adapt over time. Among
the ML methods widely used in the literature are supervised learning, unsupervised
learning, reinforcement learning, decision trees, and artificial neural networks (ANN) [33].
Therefore, in this research we implement ANN because of their usefulness for modeling linear and nonlinear systems, as confirmed by the results obtained in the consulted
literature [11,13,20,34–36]. Additionally, neural networks have also demonstrated their
utility in processing time series for health-related applications. The authors of reference [37]
implemented an ANN for the early prediction of COVID-19 outcomes. The preliminary
findings revealed that the ANN accurately predicts intensive care unit hospitalization
using only five laboratory indices. Similarly, the researchers in reference [38] utilized deep
learning and ANNs to predict the health of football players, obtaining simulation results
that serve as the foundation for data-driven monitoring and training.
Finally, signals were used in time-domain due to the low computational cost that
facilitates implementation in low-performance platforms [10,13,19,39]. In this research, the
implementation of the classifier on the Arduino Uno board using neural networks stands
out as a significant contribution, with real-time processing supported by execution time
calculations. Additionally, the developed user interface is highlighted, enabling the easy
training of the ANN.
This paper is structured as follows: Section 2 outlines the design of the EMG signal
acquisition and conditioning device, the database creation process, the proposed analysis
strategies, the design of the pattern detection system, and the hardware restrictions for
real-time response. The results are detailed in Section 3. Section 4 focuses on the discussion
of the experiment. Section 5 describes the conclusions of the investigation.
Computers 2023, 12, x FOR PEER REVIEW
Computers 2023, 12, x FOR PEER REVIEW
Computers 2023, 12, 263
3 of 18
3 of 18
real-time response. The results are detailed in Section 3. Section 4 focuses on the discus3 of 18
real-time
response.
The results
detailed in
3. Section
4 investigation.
focuses on the discussion of the
experiment.
Sectionare
5 describes
theSection
conclusions
of the
sion of the experiment. Section 5 describes the conclusions of the investigation.
2. Materials and Methods
2.2.Materials
and Methods
Materials
Methods
Figure 1and
shows
the block diagram of the prototype implemented for the EMG signal
Figure
1
shows
the
block
diagram
of the
prototype
implemented
foracquisition,
the
signal
Figure
1 shows
the
block
of the
prototype
implemented
for EMG
the EMG
sigclassifier,
which
is composed
ofdiagram
the general
blocks
responsible
for the
digiticlassifier,
which
is
composed
of
the
general
blocks
responsible
for
the
acquisition,
digitinal
classifier,
which is composed
of the general blocks responsible for the acquisition,
zation,
and transmission
of variables.
zation,
and transmission
of variables.
digitization,
and transmission
of variables.
EMG conditioning
EMG conditioning
EMG digitalization
and
transmission
EMG
digitalization
and transmission
Desktop app
Desktop app
Database
Database
Figure 1. General configuration of the acquisition system.
Figure
General
configuration
the
acquisitionsystem.
system.
Figure
1.1.
General
configuration
ofof
the
acquisition
2.1. Signal
Acquisition
2.1.
Signal
Acquisition
2.1.
Signal
Acquisition
A
prototype
for the
the acquisition
acquisition of
of EMG
EMG signals
signals was
was designed,
designed, which
which is
is composed
composed of
of
A prototype for
Apre-amplification,
prototype for the
acquisition
of EMG
signals
was
designed,
whichstages
is composed
of2).
the
filtering,
amplification,
and
DC
offset
summing
(Figure
the pre-amplification, filtering, amplification, and DC offset summing stages (Figure 2).
the
pre-amplification,
filtering,
amplification,
and to
DC
offset summing
stages (Figurehave
2).
The
input is
is aa differential
signal
that is
is connected
connected
terminals
and 2.
2. The
The
input
differential signal
that
to terminals
11 and
The electrodes
electrodes have
The
input
is
a
differential
signal
that
is
connected
to
terminals
1
and
2.
The
electrodes
have
been placed
placed across
across two
two different
differentmuscle
musclegroups.
groups.Although
AlthoughSENIAM
SENIAM(Surface
(SurfaceElectroMyoElectroMybeen
been
placedfor
across
two different Assessment
muscle groups.
Although
SENIAM (Surface
ElectroMyoGraphy
the
Non-Invasive
of
Muscles)
recommendations
are
not strictly
strictly
Graphy for the Non-Invasive Assessment of Muscles) recommendations are not
oGraphy
for[40],
the some
Non-Invasive
Assessment
of Muscles)
recommendations
are
not strictly
followed
investigations
indicate
that
electrode
placement
is
possible
followed [40], some investigations indicate that electrode placement is possible using using
more
followed
[40],
some
investigations
indicate
that
placement
is possible
using
moreone
than
one
innervation
zone
per
channel
[9].electrode
a single
channel
can contain
the
than
innervation
zone per
channel
[9]. Thus,
aThus,
single
channel
can contain
the group
more
than
one
innervation
zone
per
channel
[9].
Thus,
a
single
channel
can
contain
the
group muscle
information
sufficient
for recognition
pattern recognition
without
themodel
need atoparticular
model a
muscle
information
sufficient
for pattern
without the
need to
group
muscle
information
sufficient
forwere
pattern
recognition
without thethe
need
to model
a
particular
muscle
[9,41,42].
filters
designed
considering
frequency
specmuscle
[9,41,42].
The
filtersThe
were
designed
considering
that the that
frequency
spectrum
of
particular
muscle
[9,41,42].
The
filters
were
designed
considering
that
the
frequency
spectrum of
EMGhas
signal
has the
mostinpower
in the
of 20
to[4,43–45].
500 Hz [4,43–45].
EMG
signal
the most
power
the range
of range
20 to 500
Hz
trum of EMG signal has the most power in the range of 20 to 500 Hz [4,43–45].
1
2
1
2
Pre-amplification
Pre-amplification
Gain = 10
Gain
= 10
IC=INA128
IC=INA128
Notch Filter
Notch
60Filter
Hz
60
Hz
Q=6.05
Q=6.05
IC=UAF42
IC=UAF42
Sub-ckt PP4
Sub-ckt PP4
High-pass Filter
High-pass
Filter
20Hz
20Hz
IC=UAF42
IC=UAF42
Sub-ckt PP3
Sub-ckt PP3
Low-pass Filter
Low-pass
Filter
450Hz
450Hz
IC=UAF42
IC=UAF42
Sub-ckt PP3
Sub-ckt PP3
Amplifier
Amplifier
Gain = 50
Gain
= 50
IC=LF353N
IC=LF353N
DC Adder
DC Adder
2.5V
2.5V
IC=LF353N
IC=LF353N
Figure
Figure 2.
2. Block
Block diagram
diagram of
of the
the electromyograph
electromyograph implemented.
implemented.
Figure 2. Block diagram of the electromyograph implemented.
Samples were taken from the right forearm of two test subjects, using a non-invasive
Samples were taken from the right forearm of two test subjects, using a non-invasive
technique
from
electrodes
in
differential
configuration
(1 and 2) and aa non-invasive
reference elecSamples
were2
fromin
theaa right
forearm
of two test subjects,
technique from
2 taken
electrodes
differential
configuration
(1 and 2)using
and a reference electrode
(Figure
3)
[5,9,46].
Electrode
positions
were
strategically
chosen
based
on
the muscles
technique
from 3)
2 electrodes
in a differential
(1 andchosen
2) andbased
a reference
trode (Figure
[5,9,46]. Electrode
positionsconfiguration
were strategically
on theelecmusrelevant
to
the
study
and
within
areas
feasible
for
an
individual
with
transradial
trode
(Figure 3)to[5,9,46].
Electrode
positions
were strategically
chosen with
basedtransradial
on theamputamuscles relevant
the study
and within
areas feasible
for an individual
amtion
[47]. The
M foam monitoring
electrodes
2228
were
used.
cles
relevant
to 3the
and within
areas
feasible
for2228
an individual
putation
[47].
Thestudy
3 M foam
monitoring
electrodes
were used.with transradial amputation [47]. The 3 M foam monitoring electrodes 2228 were used.
Computers 2023, 12, 263
Computers 2023, 12, x FOR PEER REVIEW
4 of 18
4 of 18
Figure 3. Position of electrodes located on the right arm of the test subjects.
Figure 3. Position of electrodes located on the right arm of the test subjects.
Oncethe
thesignal
signalwas
wasacquired,
acquired,the
thesamples
sampleswere
weredigitized
digitizedusing
usingan
anArduino
ArduinoUno
Uno
Once
board
(Single.board
microcontroller,
Arduino
LL,
Torino,
PIE,
Italy)
with
a
sampling
rate
board (Single.board microcontroller, Arduino LL, Torino, PIE, Italy) with a sampling
of
2
kHz
and
quantization
of
10-bit.
Serial
communication
was
used
between
the
Arduino
rate of 2 kHz and quantization of 10-bit. Serial communication was used between the
and the and
computer.
Based on
the on
literature,
it is observed
that the
length
can have
Arduino
the computer.
Based
the literature,
it is observed
thatsignal
the signal
length
can
values
between
50 and
the sampling
frequency
can be
500 and
2000
have
values
between
50 400
and ms
400and
ms and
the sampling
frequency
canbetween
be between
500 and
Hz [9,28,45,48].
2000
Hz [9,28,45,48].
2.1.1.
2.1.1.Pre-Amplification
Pre-Amplification
The
level
ofof
1010
mV
and
high
levels
of noise
from
different
TheEMG
EMGsignal
signalhas
hasa amaximum
maximum
level
mV
and
high
levels
of noise
from
differsources.
To reduce
thisthis
noise
andand
increase
thethe
amplitude
ofof
the
ent sources.
To reduce
noise
increase
amplitude
thesignal,
signal,the
theINA128P
INA128P
(Instrumentation
Dallas,
TX,
USA,
EE.UU.)
used,
which
(Instrumentationamplifier,
amplifier,Texas
TexasInstruments,
Instruments,
Dallas,
TX,
EE.UU.)
waswas
used,
which
has
has
a
Common
Mode
Rejection
Ratio
(CMRR)
of
120
dB.
Equation
(1)
was
used
to
a
a Common Mode Rejection Ratio (CMRR) of 120 dB. Equation (1) was used to set aset
gain
gain
of
10.8
(Rg
=
5.1
kΩ).
of 10.8 (Rg = 5.1 kΩ).
50 kΩ
G = 1+
(1)
50 𝑘Ω
𝐺 = 1 Rg
(1)
𝑅𝑔
2.1.2. Filters
ToFilters
keep only the spectrum components with more information, several studies high2.1.2.
light the
remove
noise from
the electrical
network
using a band-rejection
filter,highand
To need
keep to
only
the spectrum
components
with
more information,
several studies
that
of thetopower
of an
EMG
signal
lies between
the range
Hz [4,39–41].
lightmost
the need
remove
noise
from
the electrical
network
usingofa20–500
band-rejection
filter,
Therefore,
it
was
decided
to
suppress
signals
below
20
Hz,
above
450
Hz,
and
noise
from
and that most of the power of an EMG signal lies between the range of 20–500
Hz [4,39–
the
grid (60
Hz).
The designs
were performed
with
circuit
41].power
Therefore,
it was
decided
to suppress
signals below
20the
Hz,UAF42
above integrated
450 Hz, and
noise
and
Filter42
software
[49].
from the power grid (60 Hz). The designs were performed with the UAF42 integrated
For the first stage, a 60 Hz notch filter was implemented using a UAF42 configured
circuit and Filter42 software [49].
with the PP4 sub-circuit [49], using the components described in Table 1.
For the first stage, a 60 Hz notch filter was implemented using a UAF42 configured
with the PP4 sub-circuit [49], using the components described in Table 1.
Table 1. Characteristics of the 60 Hz Notch filter.
TableComponent
1. Characteristics of the 60 Hz
Notch filter.
Value
Component
Value
f0
60
Hz
fz
60.05
Hz
Component
Value
Component
Value
Qf0
6.05Hz
Gain
1.026Hz
60
fz
60.05
RF1
2.65 MΩ
Rz1
2 kΩ
Q
6.05
Gain
1.026
RF2
2.65 MΩ
Rz2
2 kΩ
R
F1
2.65
MΩ
R
z1
2 kΩ
RQ
4.99 kΩ
Rz3
12.1
kΩ
RF2
2.65 MΩ
Rz2
2 kΩ
RQ
4.99 kΩ
Rz3
12.1 kΩ
For the second stage, a 20 Hz high pass filter with a Butterworth response was implemented, which makes it possible to attenuate motion artifacts [4]. The filter was impleFor
the asecond
a 20 Hzwith
highthe
pass
with a Butterworth
implemented
from
UAF42stage,
configured
PP3filter
sub-circuit
[49]. Table 2 response
shows thewas
values
of
mented,
which
makes
it
possible
to
attenuate
motion
artifacts
[4].
The
filter
was
implethe components used.
mented from a UAF42 configured with the PP3 sub-circuit [49]. Table 2 shows the values
of the components used.
Computers2023,
2023,12,
12,xxFOR
FORPEER
PEERREVIEW
REVIEW
Computers
Computers 2023, 12, 263
18
55 ofof 18
5 of 18
Table2.2.Characteristics
Characteristicsofofthe
the20
20Hz
Hzhigh
highpass
passfilter
filterwith
withButterworth
Butterworthresponse.
response.
Table
Table 2. Characteristics of the 20 Hz high pass filter with Butterworth response.
Component
Component
Component
f0f0
f0
QQ
Q
RF1F1
R
RF1
R
RF2
F2F2
Value
Value
Value
20.10
Hz
20.10 Hz
20.10
Hz
0.707
0.707
0.707
4.21MΩ
MΩ
4.21
4.21 MΩ
4.21
MΩ
4.21
4.21 MΩ
MΩ
Component
Component
Component
RRQQ
R G
RRQ
G
RG
Gain
Gain
Gain
2A
RR2A
2A
Value
Value
Value
4.7
kΩ
4.7 kΩ
4.7
kΩ
50
kΩ
50 kΩ
50 kΩ
0.9894
0.9894
0.9894
5.49
kΩ
5.49
5.49 kΩ
In thethird
third stage,aa450
450 Hzlow
low passfilter
filter withaaButterworth
Butterworth responsewas
was impleIn
In the
the third stage,
stage, a 450 Hz
Hz low pass
pass filter with
with a Butterworth response
response was impleimplemented
[9,22,42].
The
components
used
for
the
PP3
sub-circuit
of
the
UAF42
can
beseen
seen
mented
[9,22,42].
The
components
used
for
the
PP3
sub-circuit
of
the
UAF42
can
be
mented [9,22,42]. The components used for the PP3 sub-circuit of the UAF42 can be seen in
in
Table
3.
in
Table
Table
3. 3.
Table3.3.Characteristics
Characteristics ofthe
the 450Hz
Hzlow
lowpass
passfilter
filterwith
withButterworth
Butterworthresponse.
response.
Table
Table 3.
Characteristics of
of the 450
450 Hz low
pass filter
with Butterworth
response.
Component
Component
Component
f0f0
f0
Q
Q
Q
R
RF1
F1F1
RRF2F2
R
Value
Value
Value
445.8Hz
Hz
445.8
445.8 Hz
0.707
0.707
0.707
357kΩ
kΩ
357
357
357kΩ
kΩ
357
357
kΩ
Component
Component
Component
RRQQ
RQ
RGG
R
RG
Gain
Gain
Gain
Value
Value
Value
47kΩ
kΩ
47
47 kΩ
51
kΩ
51
51 kΩ
kΩ
1
1
2.1.3.Amplification
Amplification
Amplification
2.1.3.
Inthis
thisstage,
stage,the
theLF353
LF353integrated
integratedcircuit
circuitconfigured
configuredas
asan
aninverting
invertingamplifier
amplifierwas
was
In
used.
A
linear
potentiometer
was
implemented
to
have
a
variable
gain
between
0
and
200.
A linear
linear potentiometer
potentiometer was
was implemented
implemented to
to have
have aa variable
variable gain
gain between
between00and
and200.
200.
used. A
gainof
of50
50was
wasused
usedfor
forthe
theexperiments
experiments(Figure
(Figure4).
4).
AAgain
Figure4.4.Amplifier
Amplifierbased
basedon
onthe
theLF353.
LF353.
Figure
Figure 4.
Amplifier based
on the
LF353.
2.1.4.DC
DCOffset
OffsetSumming
Summing
2.1.4.
DC
Offset
Summing
2.1.4.
For
the
digitization
of
the
signal,
the
Arduino
board
uses
an
analog–digital
converter
Forthe
thedigitization
digitizationof
ofthe
thesignal,
signal,the
theArduino
Arduinoboard
boarduses
usesan
ananalog–digital
analog–digitalconverter
converter
For
from
0
to
5
V,
therefore,
it
is
necessary
to
add
2.5
V
to
the
signal.
Figure
5
shows
the
from
0
5
V,
therefore,
it
is
necessary
to
add
2.5
V
to
the
signal.
Figure
5
shows
the
circuit
from 0 to 5 V, therefore, it is necessary to add 2.5 V to the signal. Figure 5 shows the circuit
circuit
implemented
with
the LF353N.
In addition,
adiode
Zener
diode
was implemented
for
implemented
withthe
the
LF353N.
Inaddition,
addition,
Zenerdiode
was
implemented
forstabilizastabilizaimplemented
with
LF353N.
In
aaZener
was
implemented
for
stabilization
of the reference
tionof
ofthe
thereference
reference
voltage.voltage.
tion
voltage.
Figure5.5.Summing
Summingamplifier
amplifierbased
basedon
onthe
theLF353.
LF353.
Figure
Summing
amplifier
Computers 2023, 12, x FOR PEER REVIEW
Computers
2023,
12,
263
Computers
2023,
12,
FORPEER
PEERREVIEW
REVIEW
Computers
2023,
12,
xxFOR
Computers 2023, 12, x FOR PEER REVIEW
6 of 18
of18
18
666 of
of
18
6 of 18
2.2. Experimental
Experimental Procedure
Procedure
2.2.
2.2.Experimental
ExperimentalProcedure
Procedure
2.2.
2.2.
Experimental
Procedure
The
experiments
were performed
performed with
with the
the participation
participation of
of 22 non-amputated
non-amputated subjects,
subjects,
The
Theexperiments
experimentswere
wereperformed
performedwith
withthe
theparticipation
participationof
of22non-amputated
non-amputatedsubjects,
subjects,
The
experiments
were
The
experiments
were
performed
with
the
participation
of
2
non-amputated
subjects,
who
previously
signed
an
informed
consent.
The
movements
of
extension,
pronation,
suwho
previously
signed
an
informed
consent.
The
movements
of
extension,
pronation,
who
previously
signed
an
informed
consent.
The
movements
of
extension,
pronation,
suwho previously
previously
signed
an informed
informed
consent.6).
The movements
movements of
of extension,
extension, pronation,
pronation, susuwho
signed
an
consent.
The
pination,
and
flexion
were
recorded
(Figure
supination,
and
flexion
were
recorded
(Figure
6).
pination,and
andflexion
flexionwere
wererecorded
recorded(Figure
(Figure6).
6).
pination,
pination,
and flexion
were recorded
(Figure 6).
(a)
(a)
(a)
(a)
(b)
(b)
(b)
(b)
(c)
(c)
(c)
(c)
(d)
(d)
(d)
(d)
Figure 6. Hand movements to acquire the EMG signals. (a) extension; (b) pronation; (c) supination;
Figure6.6.Hand
Handmovements
movementsto
toacquire
acquirethe
theEMG
EMGsignals.
signals.(a)
(a)extension;
extension;(b)
(b)pronation;
pronation;(c)
(c)supination;
supination;
Figure
Figure
6.
acquire the
(a) extension;
(b) pronation;
(c) supination;
(d)
flexion.
Figure
6. Hand
Hand movements
movements to
the EMG
EMG signals.
(d)flexion.
flexion.
(d)
(d) flexion.
flexion.
(d)
The session begins with a skin preparation procedure, involving cleansing with alcoThesession
sessionbegins
begins
with
askin
skin
preparation
procedure,
involving
cleansing
with
alcosession
begins
with
preparation
procedure,
involving
cleansing
with
alcoThe
with
aaaskin
preparation
procedure,
involving
cleansing
with
alcohol
The
session
begins
with
skin
preparation
procedure,
involving
cleansing
with
alcohol toThe
remove
moisture
from
the
skin.
A desktop
application
that communicates
with
the
hol
to
remove
moisture
from
the
skin.
A
desktop
application
that
communicates
with
the
hol
to remove
remove
moisture
from
the
skin.
A
desktop
application
that
communicates
withthe
the
to
remove
moisture
from
thethe
skin.
AA
desktop
application
that
communicates
with
hol
to
moisture
from
skin.
desktop
application
that
communicates
with
the
acquisition
system
was
developed
with
MATLAB
R2023a. The
application
is responsible
acquisition
system
was
developed
with
MATLAB
R2023a.
The
application
is
responsible
acquisition
system
was
developed
with
MATLABR2023a.
R2023a.
The
application
isresponsible
responsible
acquisition
system
developed
with
MATLAB
The
application
is
acquisition
system
was
developed
with
MATLAB
R2023a.
The
application
is
responsible
for
managing
datawas
collection
during
a measurement
session
with
a subject.
The
session
for
managing
data
collection
during
a
measurement
session
with
a
subject.
The session
session
for
managing
data
collection
during
a
measurement
session
with
a
subject.
The
for
data
collection
measurement
session
with
aa subject.
The
for
managing
data
collection
during
measurement
session
with
subject. among
The session
session
wasmanaging
performed
by taking
160 during
trials
ofaasteady-state
signals
evenly
distributed
the 4
was
performed
by
taking
160
trials
of
steady-state
signals
evenly
distributed
among
the444
was
performed
by
taking
160
trials
of
steady-state
signals
evenly
distributed
among
the
was
performed
by
taking
160
trials
evenly
distributed
among
was
performed
bytrials
taking
160
trials of
of steady-state
steady-state
signals
evenly
distributed
among the
the
4
study
classes (40
per
movement
class). Eachsignals
movement
is requested
randomly
and
study
classes
(40
trials
per
movement
class).
Each
movement
is
requested
randomly
and
study
classes
(40
trials
per
movement
class).
Each
movement
is
requested
randomly
and
study
classes
(40
trials
per
movement
class).
Each
movement
is
requested
randomly
and
study
(40 trials per
movement
Each movement
is requested
randomly
and
storedclasses
in the database,
resulting
in an class).
information
structure that
is represented
in Figure
storedinin
inthe
thedatabase,
database,resulting
resultingin
in
an
information
structure
that
isrepresented
represented
inFigure
Figure
stored
an
information
structure
that
is
in
stored
an
information
structure
that
isis
represented
inin
Figure
7.
stored
inthe
thedatabase,
database,resulting
resultinginin
an
information
structure
that
represented
Figure
7.
7.
7.
7.
Signal length (256)
Signallength
length(256)
(256)
Signal
Signal
length (256)
Samples by
Samplesby
by
Samples
Samples
by
movement
Flexion
movement
movement
Flexion
movement
Flexion
Flexion
Supination
Supination
Supination
Supination
Pronation
Pronation
Pronation
Pronation
Extension
Extension
Extension
Extension
Figure7.
7. Scheme of
of database
database information.
information.
Figure
Figure7.7.Scheme
Schemeof
ofdatabase
databaseinformation.
information.
Figure
Scheme
Figure
7. Scheme
of database
information.
Thedata
dataacquisition
acquisitionwas
wasperformed
performedaccording
accordingtoto
the
session
presented
in Figure
8,
The
the
session
presented
in Figure
8, in
The data
data acquisition
acquisition was
was performed
performed according
according to
to the
the
session
presented
in Figure
Figure
The
session
presented
in
8,8,
The
data
acquisition
was
performed
according
to
the
session
presented
in
Figure
8,
in which
a rest
period
of ~1.5
interleaved
with
a runofof~6~6s.s.Each
Eachrun
runfollows
followsthe
thesteps
steps
which
a rest
period
of ~1.5
s isssis
interleaved
with
a run
in
which
a
rest
period
of
~1.5
is
interleaved
with
a
run
of
~6
s.
Each
run
follows
the
steps
in which
which
aFigure
rest period
period
of ~1.5
~1.5 ss is
is interleaved
interleaved with
with aa run
run of
of ~6
~6 s.
s. Each
Each run
run follows
follows the
the steps
steps
in
rest
of
shown
inaFigure
9.
shown
shownin
inFigure
Figure9.9.9.
shown
in
shown in Figure 9.
Rest
Rest
Rest
Rest
Run1
Run1
Run1
Run1
Rest
Rest
Rest
Rest
Run 2
Run22
Run
Run
2
Rest
Rest
Rest
Rest
Run 160
Run160
160
Run
Run
160
1.5 s
1.5ss
1.5
1.5
s
6s
6s
66 ss
1.5 s
1.5ss
1.5
1.5
s
6s
6s
66 ss
1.5 s
1.5ss
1.5
1.5
s
6s
6s
66 ss
Figure 8. Outline of the sampling session.
Figure8.8.8.Outline
Outlineof
ofthe
thesampling
samplingsession.
session.
Figure
Outline
of
the
sampling
session.
Figure
Figure
8. Outline
of
the
sampling
session.
Speech
Speech
Speech
Speech
Cue
Cue
Cue
Cue
Subject
Subject
Subject
Feedback
Subject
Feedback
Feedback
Feedback
1.2 s
1.2ss
1.2
1.2
s
2,8 s
2,8ss
2,8
2.8
s
Figure 9. Scheme of a run.
Figure9.9.Scheme
Schemeof
of arun.
run.
Figure
Figure
9. Scheme
Scheme of
of aaarun.
run.
Figure 9.
Recording Speech
Speech
Speech
Recording Speech
Cue
Recording
Recording
Cue
Cue
Cue
1s
1s
11 ss
1s
1s
11 ss
Computers 2023, 12, 263
7 of 18
Speech Cue: a random instruction is presented to the test subject with the 4 possible
movements (pronation, supination, flexion, and extension) through an audible signal of
~1.2 s.
Subject Feedback: after the audible signal has been reproduced, the subject performs
the movement, maintains it, and instructs the operator to begin the process of recording
the trial. This process takes approximately ~2.8 s.
Recording: the signal is recorded with a total duration of 128 milliseconds.
After the acquisition of all the trials, a randomization process was performed in order
to eliminate the effects of the capture order in the session.
2.3. Feature Extraction and Analysis
The feature extraction of EMG signals was performed with a time-domain analysis.
Features with low computational complexity, and widely used in the state of the art, were
preferred [9]. Eleven features were selected: Integrated EMG (IEMG), Mean absolute
value (MAV), Modified Mean Absolute Value 1 (MMAV1), Modified Mean Absolute Value
2 (MMAV2), Simple Square Integral (SSI), Variance of EMG (VAR), Root Mean Square (RMS),
Waveform Length (WL) [13], Zero Crossing (ZC), Slope Signal Change (SSC), and Willison
Amplitude (WAMP). The equations of each algorithm are described in reference [35].
These algorithms assume that the signal is centered at 0 and varies between positive and
negative values.
To evaluate the features, two methods are proposed, one graphic and one numerical,
which are specified in the following sections.
2.3.1. Numerical Analysis of Discriminative Capacity
An approach is presented to establish the quality of the representation provided by
each feature, based on the minimum distance classification method [50], for which a class
separability coefficient was proposed.
Let {xi (h,k) , (i = 1, . . ., nk )} be a set of nk elements of one dimension, generated by the h
feature from the xi signal, which belongs to the k class, where k = 1, . . ., K for each of the
K classes and where h = 1, . . ., H for each of the H features. Based on this data, the class
separability coefficient S(h) for an h feature is:
S(h) =
K
1
∑
(h)
(h)
m 0 k =1
(h)
m k − m0
, (h = 1, . . . , H )
(2)
where mk (h) is the mean of the k class for an h feature to evaluate and m0 (h) is the mean of
the classes for an h feature to evaluate:
(h)
mk
(h)
m0
=
=
nk
1
nk
i =1
1
K
∑ mk
∑ xi
K
(h,k)
(h)
, (h = 1, . . . , H )
(3)
, (h = 1, . . . , H )
(4)
k =1
2.3.2. Graphic Analysis of Discriminative Capacity
Anscombe’s Quartet [51] and Datasaurus [52], among other data sets, have shown
that graphs are a fundamental element in understanding a problem. In order to have an
intuitive understanding of the differentiation capacity of each feature extraction algorithm,
the data resulting from the algorithms were plotted. For each feature, a chart is made,
where the x-axis represents the movement, and the y-axis represents the value obtained
from the feature extracted. The movements and their associated colors are presented in
Figure 10.
Computers2023,
2023,12,
12,263
x FOR PEER REVIEW
Computers
88 of
18
of 18
Computers 2023, 12, x FOR PEER REVIEW
8 of 18
Figure 10.
10. Movements and their
their associated
associated colors
colors for
for graphical
graphical analysis.
analysis.
Figure
The
to to
represent
thethe
statistical
performance
of the
The violin
violinplot
plotmethod
methodwas
wasused
used
represent
statistical
performance
offeatures.
the feaIf
thereIfisthere
no overlap,
with these
charts
is possible
intuit the
algorithm
effectiveness
to
tures.
is no overlap,
with
theseit charts
it is to
possible
to intuit
the algorithm
effecFigure 10. Movements
and their associated
for graphical
analysis.
provide
representation
on thecolors
pattern
recognition
of one or more
classes.
tivenessuseful
to provide
useful representation
on the
pattern recognition
of one
or more classes.
The violin
plot
method was
used Configuration
to represent the statistical performance of the fea2.4.
EMG
Pattern
Recognition
System
2.4.
EMG
Pattern
Recognition
System
tures.
If there
is no
overlap, with
theseConfiguration
charts it is possible to intuit the algorithm effecThis
isisresponsible
for for
the the
configuration
theofElectromyographic
PatternPattern
Recogtiveness
provide
useful
representation
onconfiguration
the patternof
recognition
one or more classes.
Thistostage
stage
responsible
the ofElectromyographic
nition
System.
From
the
analysis
of
the
features,
four
sets
were
proposed
to
make
the
Recognition System. From the analysis of the features, four sets were proposed to make
classification.
2.4.classification.
EMG Pattern Recognition System Configuration
the
Full
== IEMG,
MAV,
MMAV2,
RMS,
and
Thisset
stage
is responsible
for the configuration
ofVAR,
the Electromyographic
Pattern
Full
set
IEMG,
MAV, MMAV1,
MMAV1,
MMAV2, SSI,
SSI,
VAR,
RMS,WL,
WL,ZC,
ZC,SSC
SSC
andWAMP.
WAMP.
Recognition System. From the analysis of the features, four sets were proposed to make
Hudgins
setset
= =MAV,
ZC,
SSC
and
WL.
Hudgins
MAV,
ZC,
SSC
and
WL.
the classification.
Full set = IEMG, MAV, MMAV1, MMAV2, SSI, VAR, RMS, WL, ZC, SSC and WAMP.
Subset1
= IEMG,
SSI, RMS,
WL,
SSC and WAMP.
Hudgins
set =SSI,
MAV,
ZC, WL,
SSC
and
Subset1
= IEMG,
RMS,
SSCWL.
and WAMP.
Subset2
= SSI,WL,
RMS
and
Subset1 = IEMG,
SSI, RMS,
SSC
andWL.
WAMP.
Subset2 = SSI, RMS and WL.
Subset2 = SSI, RMS and WL.
2.4.1. Offline Training
2.4.1. Offline Training
A three-layer feed-forward ANN, with a sigmoid activation function for their hidden
three-layer
feed-forward ANN, with a sigmoid activation function for their hid2.4.1.AOffline
Training
and output layers, was employed for machine learning classification, as depicted in Figure
den and
output layers,
was employed
machine
learning
classification,
as depicted in
A three-layer
feed-forward
ANN, withfor
a sigmoid
activation
function
for their hidden
11 [34]. The input layer’s size is the number of features on which the model is being trained
Figure
11 [34].
The
layer’s
size is the
number
of features
on which
the model is
and output
layers,
wasinput
employed
for machine
learning
classification,
as depicted
in Figure
(11
input
neurons
for the Full
set, 4 for of
the
Hudgins
set, 6the
formodel
Subset1,
and 3 for Subset2.
11 [34].
The input
is the number
features
is being
being
trained
(11layer’s
inputsize
neurons
for the Full
set, 4on
forwhich
the Hudgins
set,
6 fortrained
Subset1, and
hidden
layer
consists
oflayer
6 neurons,
output
layer
has
neurons,
1 correspond(11
neurons
for
the Full
set,
4 for
the and
Hudgins
set, 6 for
Subset1,
3 forlayer
Subset2.
3The
forinput
Subset2.
The
hidden
consists
ofthe
6 neurons,
and
the 4and
output
has 4 neuing
to
each
movement
to
be
classified.
The
model’s
predicted
class
was
determined
by
The hidden
layer consiststoofeach
6 neurons,
and theto
output
layer has 4 The
neurons,
1 correspondrons,
1 corresponding
movement
be classified.
model’s
predicted class
identifying
the
corresponding
neuron
with
the
highest
activation.
The
system
is
auging determined
to each movement
to be classified.
The model’s predicted
classthe
washighest
determined
by
was
by identifying
the corresponding
neuron with
activation.
The
identifying
the
corresponding
neuron
with
the
highest
activation.
The
system
is
augmented
by
the
Majority
Voting
(MV)
decision
technique
to
avoid
errors
in
the
transient
system is augmented by the Majority Voting (MV) decision technique to avoid errors in the
mented
by the
Majority Voting The
(MV)ANN
decision
technique
to
avoidusing
errors MATLAB
in the transient
state
of the
signal
trained
offline
R2023a
and
transient
state
of the[10,18,20].
signal [10,18,20].
Thewas
ANN
was trained
offline using
MATLAB
R2023a
state
of
the
signal
[10,18,20].
The
ANN
was
trained
offline
using
MATLAB
R2023a and
subsequently
implemented
on
the
Arduino
Uno
board
(Single
board
microcontroller,
Arand subsequently implemented on the Arduino Uno board (Single board microcontroller,
subsequently
implemented
on the
Arduino Uno board (Single board microcontroller, Arduino
LL,
Torino,
PIE,
Italy)
[53].
Arduino LL, Torino, PIE, Italy) [53].
duino LL, Torino, PIE, Italy) [53].
Figure 11. Three-layer feed-forward ANN with n input neurons, 6 hidden neurons, and 4 output
neurons for movement classification.
Computers 2023, 12, 263
9 of 18
For training the classifier, the information in the database was organized in three
segments: Training (60%), Validation (20%), and Test sets (20%) [54] (Table 4). Data were
normalized to have a variance equal to 1 and a mean equal to 0. The averages of the features
and standard deviations were estimated using only the training data.
Table 4. Confusion matrix for test set data.
Movement
Training
CrossValidation
Test
Totals
1. Extension
2. Pronation
3. Supination
4. Flexion
Totals
24
24
24
24
96
8
8
8
8
32
8
8
8
8
32
40
40
40
40
160
In the selection of an appropriate structure for a network, an iterative trial and error
process is often required. The possible suitable networks are described by parameters
such as number of layers, number of neurons per layer, regularization value, number of
iterations, input data, activation function, and more. Given this complexity, and the fact that
the same structure can produce high error rates, this led us to propose a selection strategy
based on programmed batch training. In this method, the training can be performed
several times for the same structure and the performance outcomes can then be graphically
represented for subsequent selection. The optimal number of neurons in the hidden layer
was chosen based on the best results obtained on the validation set while also aiming
to minimize hardware resource utilization. The fmincg training algorithm included an
automatic adjustment of the learning rate and stopping criteria. Moreover, we fine-tuned
the regularization parameter, lambda, to enhance the model’s generalization capabilities.
This structure is validated through the computation of accuracy performance and FValue [55]. These metrics are widely adopted in the evaluation of classification systems [11,36].
2.4.2. Implementation for Real-Time Processing
An adjacent windowing technique was used. It is emphasized that the identification
and generation of a response by the classification system must be less than the time required
for the acquisition of a new signal. The following statements are defined as a necessary and
sufficient condition for the proposed solution to be executed in real time [56]:
1.
2.
The system must perform the acquisition of a signal while discriminating another
immediately preceding it.
Considering that the digitalization of 256 samples takes 29.18 ms (empirical value
measured by means of an oscilloscope) and that the classification window is 128 ms,
the system must classify a signal in a time less than 99.6 ms.
To validate the operation, 8 random signals (2 for each class) were analyzed with the
Arduino board. A comparison between the average execution time for all stages with the
maximum time available was performed. It should be noted that the evaluations performed
verify the execution time with EMG signals from the dataset obtained in this work (See
Supplementary Material).
2.5. Desktop App and Prototype
The desktop app was designed to perform the deployment of a complete experiment for each user, from the creation of the database to the hardware implementation.
The application is divided into 4 functions that can be accessed from the main interface
(Figure 12).
2.5. Desktop App and Prototype
Computers 2023, 12, 263
The desktop app was designed to perform the deployment of a complete experiment
for each user, from the creation of the database to the hardware implementation. The
10 ofap18
plication is divided into 4 functions that can be accessed from the main interface (Figure
12).
Figure12.
12.(a)
(a)Main
Maininterface
interfaceof
ofthe
theapplication;
application;(b)
(b)Interface
Interfacefor
forthe
theanalysis
analysisof
offeatures;
features;(c)
(c)Interface
Interface
Figure
for
the
creation
of
the
database;
(d)
Interface
for
the
training
of
the
neural
network.
for the creation of the database; (d) Interface for the training of the neural network.
CreateDatabase:
Database:configures
configuresthe
theEMG
EMGsignal
signalcapture
captureexperiment
experimentfor
foraauser.
user.Generates
Generates
Create
the
database
file
in
mat
format.
the database file in mat format.
InformationAnalysis:
Analysis:
this
section
allows
the application
of the Numerical
and
Information
this
section
allows
the application
of the Numerical
and Graphic
Analysis
Discriminative
Capacity proposed
in Sections
and 2.3.2.
It will
represent
Graphic of
Analysis
of Discriminative
Capacity proposed
in2.3.1
Sections
2.3.1 and
2.3.2.
It will
the
scores inthe
descending
order to inform
thetoresearcher
about
the quality
of each
characteristic.
represent
scores in descending
order
inform the
researcher
about
the quality
of each
ANN training: contains options to configure the artificial neural network to be trained
characteristic.
offlineANN
(Section
2.4.1). A
project options
file is obtained
with the
trained
neural
networks
and their
training:
contains
to configure
the
artificial
neural
network
to be
performance.
trained offline (Section 2.4.1). A project file is obtained with the trained neural networks
Implementation: starting from the ANN project file, this stage writes a code
and Real-time
their performance.
for Arduino
thatImplementation:
can then be copied
and pasted
into
theproject
sketch file,
ino this
file to
loadwrites
the project
Real-time
starting
from the
ANN
stage
a code
with
all the information
thatbethe
network
process
the signals
online.
for Arduino
that can then
copied
and needs
pastedtointo
the sketch
ino file
to load the project
12, some that
of the
implemented
forthethe
application
withInallFigure
the information
theinterfaces
network needs
to process
signals
online. are shown.
FigureIn12a
shows
the
main
interface
of
the
desktop
application.
Figure
shows the
Figure 12, some of the interfaces implemented for the application 12b
are shown.
Figstage
responsible
for
the
analysis
of
features.
In
Figure
12c,
the
interface
for
creating
the
ure 12a shows the main interface of the desktop application. Figure 12b shows the stage
database
is observed.
Figure 12d
shows the
interface
developed
for ANN
training.
responsible
for the analysis
of features.
In Figure
12c,
the interface
for creating
the dataFinally,
Figure
13
shows
the
prototype
implemented
for
the
EMG
signal
classifier.
base is observed. Figure 12d shows the interface developed for ANN training.
Finally, Figure 13 shows the prototype implemented for the EMG signal classifier.
Computers 2023, 12, x FOR PEER REVIEW
11 of 18
Computers
2023,2023,
12, x FOR
PEER REVIEW
Computers
12, 263
11 of 11
18 of 18
Figure 13. Prototype implemented.
Figure
13. Prototype
implemented.
Figure
Prototype
3. 13.
Results
andimplemented.
Discussion
3.
Results
and Discussion
3.1.
Data
3. Results
andAcquisition
Discussion
3.1. Data
Acquisition
Before
performing the integration of all the elements of the prototype, the circuits
3.1. Data Acquisition
Before
performing
the integration
of all
thetested
elements
of the
theNI
prototype,
the circuits
designed for the acquisition
of the signals
were
using
Educational
LaboraBefore performing
the integration
of all were
the elements
of the
prototype,
the circuits
designed
for
the
acquisition
of
the
signals
tested
using
the
NI
Educational
Laboratory
tory Virtual Instrumentation Suite (NI ELVIS, National Instruments, Austin, TX,
EE.UU.)
designed
for the acquisition Suite
of the(NI
signals
were
testedInstruments,
using the NI Educational
LaboraVirtual
ELVIS,
National
TX, USA,
EE.UU.)
and theInstrumentation
NI myDAQ device (Data
acquisition
device, NationalAustin,
Instruments,
Austin,
TX,
tory and
Virtual
Instrumentation
Suite
(NI
ELVIS,
National
Instruments,
Austin,
TX,
EE.UU.)
the NI myDAQ device (Data acquisition device, National Instruments, Austin, TX,
EE.UU.).
and EE.UU.).
the NI myDAQ device (Data acquisition device, National Instruments, Austin, TX,
Using a 1 Vpp test sinusoidal signal, the following results were obtained: low cutoff
EE.UU.). Using a 1 Vpp test sinusoidal signal, the following results were obtained: low cutoff
frequency of 21.4 Hz, high cutoff frequency of 469 Hz, and attenuation of −23.61 dB at the
Using a 1 Vpp21.4
testHz,
sinusoidal
signal,
the following
results
were obtained:
cutoff
frequency
high cutoff
frequency
of 469
Hz,
and attenuation
of −low
23.61
dB at the
frequencyof
of 60 Hz.
These
results
show good
performance
obtained with
the high-pass,
frequency
of 21.4
Hz,
high
cutoffresults
frequency
ofgood
469 Hz,
and attenuation
of −23.61
dB high-pass,
at the
frequency
of
60
Hz.
These
show
performance
obtained
with
the
low-pass and Notch filters.
frequency
of 60 Hz.
These
results show good performance obtained with the high-pass,
low-pass
Notch
filters.
The and
frequency
spectrum
obtained for extension and flexion movements is shown in
low-pass The
and frequency
Notch filters.
spectrum
for extension
and flexion
movements
is shown
in
Figure 14. As expected, these obtained
signals showed
low amplitude
levels
at frequencies
below
The
frequency
spectrum
obtained
for
extension
and
flexion
movements
is
shown
in
Figure
As 450
expected,
these
signals
showed low amplitude levels at frequencies below
20 Hz, 14.
above
Hz, and
close
to 60 Hz.
Figure
14. As
expected,
20 Hz,
above
450 Hz,these
and signals
close toshowed
60 Hz. low amplitude levels at frequencies below
20 Hz, above 450 Hz, and close to 60 Hz.
(a)
(b)
(b) corresponding totothe
(a)
Figure
Figure14.
14. Frequency
Frequency spectrum
spectrum obtained
obtained for
for four
four signals
signals corresponding
the four
fourpositions
positions of
ofthe
the
database.
(a)
Spectrum
up
to
1000
Hz;
(b)
Spectrum
up
to
100
Hz.
database. (a) Spectrum up to 1000 Hz; (b) Spectrum up to 100 Hz.
Figure 14. Frequency spectrum obtained for four signals corresponding to the four positions of the
database. (a) Spectrum up to 1000 Hz; (b) Spectrum up to 100 Hz.
Computers 2023, 12, x FOR PEER REVIEW
Computers 2023, 12, x FOR PEER REVIEW
Computers 2023, 12, 263
12 of 18
12 of 18
12 of 18
3.2.
3.2.Numerical
Numericaland
andGraphic
GraphicAnalysis
AnalysisofofDiscriminative
DiscriminativeCapacity
Capacity
3.2. Numerical
and
Graphic
Analysis
of
Discriminative
Capacity obtained for all the features
Figure
15
shows
the
results
of
the
numerical
analysis
Figure 15 shows the results of the numerical analysis obtained for all the features
Figure 15
shows thestudy
results of
the numerical
analysis
obtained MMAV1,
for all the features
addressed
addressedin
inthe
thepresent
present studyfor
foreach
eachsubject.
subject.With
Withthe
theIEMG,
IEMG,MAV,
MAV, MMAV1,MMAV2,
MMAV2,
addressed
in the present
study for
each
subject. This
Withisthe
IEMG,
MAV,
MMAV1,
MMAV2,
and
RMS
methods,
good
results
were
obtained.
the
same
case
for
and RMS methods, good results were obtained. This is the same case forthe
theSSI
SSIand
andRMS
RMS
and RMS methods, good results were obtained. This is the same case for the SSI and
methods.
methods.
RMS methods.
(a)
(a)
(b)
(b)
Figure
15. Results
of the
numerical analysis
of discriminative
capacity. (a)
subject 1;
(b) subject
2.
Figure
Figure15.
15. Results
Results of
of the
the numerical
numericalanalysis
analysisof
ofdiscriminative
discriminativecapacity.
capacity.(a)
(a)subject
subject1;1;(b)
(b)subject
subject2.2.
In
In
addition,
Figure
16
shows
violin
plot
representation
of
each
feature
extracted
Inaddition,
addition,Figure
Figure16
16shows
showsaaaviolin
violinplot
plotrepresentation
representationof
ofeach
eachfeature
featureextracted
extracted
from the
the EMG
EMG signals
signals of
of the
the test
test subjects,
subjects,with
withrespect
respectto
tothe
thetype
typeof
ofmovement
movementexecuted.
executed.
from
from the EMG signals of the test subjects, with respect to the type of movement executed.
The results
results obtained with
with four of the features are displayed.
displayed. It can be
be observed that
that the
The
The resultsobtained
obtained withfour
four of the features are displayed.ItItcan
can beobserved
observed thatthe
the
IEMG, MAV,
MAV,
features have very similar distributions. The
IEMG,
IEMG, MAV,MMAV1,
MMAV1,MMAV2,
MMAV2,and
andRMS
RMS features have very similar distributions.The
The
same happens
happens with the
the SSI and
and
same
same happenswith
with theSSI
SSI andVAR
VARfeatures.
features.
Figure 16.
16.Graphical
Graphicalrepresentation
representation
information
distribution
EMG
signals
of four
Figure
ofof
thethe
information
distribution
for for
the the
EMG
signals
of four
feaFigure 16. Graphical representation of the information distribution for the EMG signals of four feafeatures:
(a) IEMG;
RMS;
(c) SSI;
VAR.
tures:
(a) IEMG;
(b) (b)
RMS;
(c) SSI;
(d) (d)
VAR.
tures: (a) IEMG; (b) RMS; (c) SSI; (d) VAR.
The numerical
numerical analysisshowed
showed a better
performance, terms
in termsthe
of the separability
The
The numericalanalysis
analysis showedaabetter
betterperformance,
performance,in
in termsof
of theseparability
separabilitycococoefficient,
for
energy-based
features
(IEMG,MAV,
MAV,MMAV1,
MMAV1,MMAV2,
MMAV2,SSI,
SSI,VAR,
VAR,RMS,
RMS, and
efficient,
for
energy-based
features
(IEMG,
efficient, for energy-based features (IEMG, MAV, MMAV1, MMAV2, SSI, VAR, RMS,and
and
WL), and
and lower
lower performance
performance for
for frequency-based
frequency-basedfeatures
features(WAMP,
(WAMP, SSC, and
and ZC), which
which
WL),
WL), andthe
lower
performance
forPhinyomark
frequency-based
features (WAMP,SSC,
SSC, andZC),
ZC), which
matches
results
reported
by
[28].
matches
the results
reported by
[28].
matches
results
byPhinyomark
Phinyomark
[28].16), it is worth highlighting the cases of
Withthe
respect
to reported
the graphic
graphic
method (Figure
(Figure
With
respect
to
the
method
16),
it isisworth
highlighting the
of
With
respect
to
the
graphic
method
(Figure
16),
worth
thecases
cases
of
pronation and
and supination,
supination, since
since all
all the
the features
features have
have
anitoverlap
overlap
forhighlighting
these movements.
movements.
This
pronation
an
for
these
This
pronationshows
and supination,
sincethe
allsystem
the features
an overlap for
these
movements.
This
situation
the difficulties
difficulties
has in
inhave
discriminating
these
signals.
Flexion and
and
situation
shows
the
the
system
has
discriminating
these
signals.
Flexion
situation
shows
the
difficulties
the
system
has
in
discriminating
these
signals.
Flexion
and
extension
are
the
simplest
movements
to
classify.
Flexion
can
be
easily
differentiated
from
extension
extensionare
arethe
thesimplest
simplestmovements
movementsto
toclassify.
classify.Flexion
Flexioncan
canbe
beeasily
easilydifferentiated
differentiatedfrom
from
Computers 2023, 12, 263
13 of 18
the IEMG, SSI, and WL features, while the extension can be detected from the IEMG, WL,
WAMP, SSC, and ZC features.
The worst performances were presented in the SSC and ZC features, in which the
flexion, pronation, and supination have a significant overlap. These observations are
consistent with the measure of discriminative capacity used in this work, indicating that
SSC and ZC are the features that have the lowest quality. The above suggests that the
numerical method is consistent with an intuitive interpretation of the dataset information.
3.3. Classification System
Tables 5–12 show the results obtained with subjects 1 and 2 for the four sets proposed
in this work (Full Set, Hudgins Set, Subset1, and Subset2).
Table 5. Confusion matrix for test set data (subject 1—full set).
Predicted
Full Set
Actual
Extension
Pronation
Supination
Flexion
Extension
8
0
0
0
Pronation
0
8
0
0
Supination
0
0
8
0
Flexion
0
0
0
8
Table 6. Confusion matrix for test set data (subject 1—Hudgins set).
Predicted
Hudgins Set
Actual
Extension
Pronation
Supination
Flexion
Extension
8
0
0
0
Pronation
0
8
0
0
Supination
0
0
8
0
Flexion
0
0
0
8
Table 7. Confusion matrix for test set data (subject 1—subset1).
Predicted
Subset1
Actual
Extension
Pronation
Supination
Flexion
Extension
8
0
0
0
Pronation
0
8
0
0
Supination
0
0
8
0
Flexion
0
0
0
8
Table 8. Confusion matrix for test set data (subject 1—subset2).
Predicted
Subset2
Actual
Extension
Pronation
Supination
Flexion
Extension
8
0
0
0
Pronation
0
8
0
0
Supination
0
0
8
0
Flexion
0
0
0
8
Table 9. Confusion matrix for test set data (subject 2—full set).
Predicted
Full Set
Actual
Extension
Pronation
Supination
Flexion
Extension
8
0
0
0
Pronation
0
8
2
0
Supination
0
0
6
0
Flexion
0
0
0
8
Computers 2023, 12, 263
14 of 18
Table 10. Confusion matrix for test set data (subject 2—Hudgins set).
Predicted
Hudgins Set
Extension
Pronation
Supination
Flexion
Actual
Extension
8
0
0
0
Pronation
0
8
2
1
Supination
0
0
6
0
Flexion
0
0
0
7
Table 11. Confusion matrix for test set data (subject 2—subset1).
Predicted
Subset1
Extension
Pronation
Supination
Flexion
Actual
Extension
8
0
0
0
Pronation
0
8
2
1
Supination
0
0
6
0
Flexion
0
0
0
7
Table 12. Confusion matrix for test set data (subject 2—subset2).
Predicted
Subset2
Actual
Extension
Pronation
Supination
Flexion
Extension
8
0
0
0
Pronation
0
8
1
0
Supination
0
0
7
0
Flexion
0
0
0
8
The results obtained with the performance measures for the two test subjects are
shown in Tables 13–15. The confusion matrix shows that the classification system has high
accuracy, although there are some difficulties with subject 2 in the detection of pronations
and supinations (see Tables 9–12), which is confirmed by the values obtained for the
accuracy and the F-values (Tables 13 and 14). These errors may be due to the supination
signals of the test subjects having very low values for IEMG, SSI, and RMS, which are
quite close to the extension and supination movements. It is worth highlighting that the
best results in terms of the test data came from Subset2, which has the fewest features (see
Tables 12 and 14). An accuracy of 98.44% was obtained (100% for subject 1 and 96.88% for
subject 2 using Subset2), which meets the accuracy percentage greater than 90% necessary
for the system to be used by a user [23]. The accuracy results are comparable with those
obtained in previous research, both with single channel and multichannel [9,10,23], but
these results cannot be generalized to amputated subjects [57].
Table 13. Performance measures applied to the test set (subject 1).
Set
Accuracy
F-Extension
F-Pronation
F-Supination
F-Flexion
Full set
Hudgins Set
Subset1
Subset2
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
Table 14. Performance measures applied to the test set (subject 2).
Set
Accuracy
F-Extension
F-Pronation
F-Supination
F-Flexion
Full set
Hudgins Set
Subset1
Subset2
0.9375
0.9063
0.9063
0.9688
1.0000
1.0000
1.0000
1.0000
0.8889
0.8421
0.8421
0.9412
0.8571
0.8571
0.8571
0.9333
1.0000
0.9333
0.9333
1.0000
Computers 2023, 12, 263
15 of 18
Table 15. Performance measures applied to the test set (subject 1 and subject 2).
Subject 1
Set
Training
Subject 2
Cross-Validation
Test
Training
Cross-Validation
Test
1.0000
1.0000
1.0000
1.0000
0.9375
0.9063
0.9063
0.9688
Accuracy
Full Set
Hudgins Set
Subset1
Subset2
1.0000
0.9896
0.9896
0.9896
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
1.0000
0.9896
0.9896
0.9896
0.9896
3.4. Execution Time
Execution times were obtained for eight EMG signals. All these signals were randomly
chosen from the database, two for each class analyzed (Extension, Pronation, Supination,
and Flexion) with an average time of 8.522 ms (Table 16).
Table 16. Execution times of the EMG signal processor implemented in the Arduino uno board
(2 signals were recorded for each class used in the study for a total of 8 signals).
Signal
Array
Loading
(ms)
Feature
Extraction
(ms)
ANN (ms)
Majority
Voting (ms)
Total (ms)
1
2
3
4
5
6
7
8
Mean
0.372
0.372
0.372
0.372
0.372
0.372
0.372
0.372
0.372
4.460
4.464
4.460
4.476
4.460
4.460
4.428
4.448
4.457
3.652
3.664
3.696
3.680
3.668
3.676
3.668
3.712
3.677
0.016
0.016
0.015
0.016
0.017
0.016
0.016
0.015
0.016
8.500
8.516
8.543
8.544
8.517
8.524
8.484
8.547
8.522
With the processor incorporated in the Arduino Uno board, all calculations were
performed with an average execution time of 8.522 ms. In this way, it was demonstrated
that the system can run in real time, since an execution time much lower than the 99.6 ms
limit value discussed in Section 2.4.2 was obtained. The processing capacity of a low-cost
tool for the execution of a neural network and for the processing of EMG signals in real time
was demonstrated. The processor was programmed to automatically copy the synaptic
weights and general structure of the neural network from MATLAB R2023a.
4. Limitations and Future Works
Though this study employed a 60%/20%/20% split, it is important to note that, in
situations with limited data, using k-fold cross-validation is recommended for optimal
evaluation with less biased or less optimistic estimates. Future research will aim to improve validation estimates using this approach. Furthermore, there is the possibility of
experimenting with alternative hardware options to implement a system that is even more
efficient and flexible. It is also advisable to work on datasets with a greater number of trials
per subject and that have records of amputee subjects.
5. Conclusions
The techniques described in this work have the potential to develop simple devices
for the acquisition, conditioning, and recognition of EMG patterns in order to facilitate
future implementations of prostheses. Numerical methods are a valuable tool for quickly
assessing the quality of the information generated by feature extraction algorithms. Both
the graphical and numerical methods proposed in this research align with previous studies
and exhibit a consistent level of agreement between them. Reducing redundancy is a
Computers 2023, 12, 263
16 of 18
constructive approach to enhancing generalization performance. The Arduino Uno board,
being a relatively inexpensive product, has showcased adequate hardware capabilities for
the development of efficient real-time systems. This work is conceived as a proof of concept,
with its subsequent stage involving the creation of a more resilient database containing
signals from both healthy and amputated subjects.
Supplementary Materials: The following supporting information can be downloaded at: https:
//github.com/nelsoncardenas/emg_dataset.
Author Contributions: Formal analysis, A.P. and N.C.-B.; Investigation, A.P. and N.C.-B.; Methodology, A.P., N.C.-B. and C.R.-A.; Software, N.C.-B.; Validation, A.P. and C.R.-A.; Visualization, N.C.-B.;
Writing—review and editing, A.P. and C.R.-A. All authors have read and agreed to the published
version of the manuscript.
Funding: This research was funded by the Fondo General de Regalías of Colombia, as part of the
project titled “Desarrollo de un sistema inteligente multiparamétrico para el reconocimiento de patrones asociados a disfunciones neurocognitivas en jóvenes en conflicto con la ley en el departamento
del Atlántico”, under code BPIN 20200000100006. Collaborating institutions in this project include the
Institución Universitaria de Barranquilla (IUB), the Universidad del Magdalena, and the Universidad
Autónoma del Caribe.
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement: The data presented in this study are available in Supplementary
Material.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
Disability and Health. Available online: https://www.who.int/news-room/fact-sheets/detail/disability-and-health (accessed
on 23 May 2023).
WHO Global Disability Action Plan, 2014–2021. Available online: https://www.who.int/publications/i/item/who-globaldisability-action-plan-2014-2021 (accessed on 24 May 2023).
Cordella, F.; Ciancio, A.L.; Sacchetti, R.; Davalli, A.; Cutti, A.G.; Guglielmelli, E.; Zollo, L. Literature review on needs of upper
limb prosthesis users. Front. Neurosci. 2016, 10, 209. [CrossRef]
De la Peña, S.; Polo, A.; Robles-Algarín, C. Implementation of a Portable Electromyographic Prototype for the Detection of Muscle
Fatigue. Electronics 2019, 8, 619. [CrossRef]
Jiménez, C.R.; Bennett, P.; García, A.O.; Cuesta Vargas, A.I. Fatigue Detection during Sit-To-Stand Test Based on Surface
Electromyography and Acceleration: A Case Study. Sensors 2019, 19, 4202. [CrossRef] [PubMed]
Moreno-Barriga, O.S.; Romero-Morales, C.; Becerro-de-Bengoa-Vallejo, R.; Losa-Iglesias, M.E.; Gómez-Salgado, J.; CaballeroLópez, J.; Vidal-Valverde, L.C.; López-López, D. Effects of Foot Structure Type on Core Stability in University Athletes. Life 2023,
13, 1487. [CrossRef]
Polo, A.; Narvaez, P.; Robles Algarín, C. Implementation of a Cost-Effective Didactic Prototype for the Acquisition of Biomedical
Signals. Electronics 2018, 7, 77. [CrossRef]
Staderini, E.M.; Mugnaini, S.; Kambampati, H.; Magrini, A.; Gentili, S. Improved Multichannel Electromyograph Using Off-theShelf Components for Education and Research. Sensors 2022, 22, 3616. [CrossRef]
Hakonen, M.; Piitulainen, H.; Visala, A. Current state of digital signal processing in myoelectric interfaces and related applications.
Biomed. Signal Process. Control 2015, 18, 334–359. [CrossRef]
Naik, G.R.; Selvan, S.E.; Nguyen, H.T. Single-channel EMG classification with ensemble-empirical-mode-decomposition-based
ICA for diagnosing neuromuscular disorders. IEEE Trans. Neural Syst. Rehabil. Eng. 2015, 24, 734–743. [CrossRef]
Gokgoz, E.; Subasi, A. Comparison of decision tree algorithms for EMG signal classification using DWT. Biomed. Signal Process.
Control 2015, 18, 138–144. [CrossRef]
Battye, C.K.; Nightingale, A.; Whillis, J. The use of myo-electric currents in the operation of prostheses. J. Bone Jt. Surg. 1955, 37,
506–510. [CrossRef]
Hudgins, B.; Parker, P. A New Strategy for Multifunction Myoelectric Control. IEEE Trans. Biomed. Eng. 1993, 40, 82–94. [CrossRef]
[PubMed]
Childress, D.S. Myoelectric control of Powered Prostheses. IEEE Eng. Med. Biol. Mag. 1982, 1, 23–25. [CrossRef]
Hassan, H.F.; Abou-Loukh, S.J.; Kasim Ibraheem, I. Teleoperated robotic arm movement using electromyography signal with
wearable Myo armband. J. King Saud Univ. Eng. Sci. 2020, 32, 378–387. [CrossRef]
Bebionic. Available online: https://www.ottobockus.com/prosthetics/upper-limb-prosthetics/solution-overview/bebionichand/ (accessed on 30 May 2023).
Computers 2023, 12, 263
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
17 of 18
Saponas, T.S.; Tan, D.S.; Morris, D.; Turner, J.; Landay, J.A. Making muscle-computer interfaces more practical. In Proceedings of
the SIGCHI Conference on Human Factors in Computing Systems, Atlanta, GA, USA, 10–15 April 2010. [CrossRef]
Huang, H.; Li, T.; Bruschini, C.; Enz, C.; Koch, V.M.; Justiz, J.; Antfolk, C. EMG pattern recognition using decomposition
techniques for constructing multiclass classifiers. In Proceedings of the 6th IEEE International Conference on Biomedical Robotics
and Biomechatronics (BioRob), Singapore, 26–29 June 2016. [CrossRef]
Samuel, O.W.; Zhou, H.; Li, X.; Wang, H.; Zhang, H.; Sangaiah, A.K.; Li, G. Pattern recognition of electromyography signals based
on novel time domain features for amputees’ limb motion classification. Comput. Electr. Eng. 2018, 67, 646–655. [CrossRef]
Wei, W.; Wong, Y.; Du, Y.; Hu, Y.; Kankanhalli, M.; Geng, W. A multi-stream convolutional neural network for sEMG-based
gesture recognition in muscle-computer interface. Pattern Recognit. Lett. 2019, 119, 131–138. [CrossRef]
Phinyomark, A.; Scheme, E. EMG Pattern Recognition in the Era of Big Data and Deep Learning. Big Data Cogn. Comput. 2018,
2, 21. [CrossRef]
Mane, S.M.; Kambli, R.A.; Kazi, F.S.; Singh, N.M. Hand motion recognition from single channel surface EMG using wavelet &
artificial neural network. Procedia Comput. Sci. 2015, 49, 58–65. [CrossRef]
Tavakoli, M.; Benussi, C.; Lourenco, J.L. Single channel surface EMG control of advanced prosthetic hands: A simple, low cost
and efficient approach. Expert Syst. Appl. 2017, 79, 322–332. [CrossRef]
Too, J.; Abdullah, A.; Mohd Saad, N.; Tee, W. EMG Feature Selection and Classification Using a Pbest-Guide Binary Particle
Swarm Optimization. Computation 2019, 7, 12. [CrossRef]
Liu, J.; Li, X.; Li, G.; Zhou, P. EMG feature assessment for myoelectric pattern recognition and channel selection: A study with
incomplete spinal cord injury. Med. Eng. Phys. 2014, 36, 975–980. [CrossRef]
De Silva, A.M.; Leong, P.H.W. Grammar-Based Feature Generation for Time-Series Prediction, 2015a ed.; Springer-Verlag:
Berlin/Heidelberg, Germany, 2015; pp. 13–24.
Al-Angari, H.M.; Kanitz, G.; Tarantino, S.; Cipriani, C. Distance and mutual information methods for EMG feature and channel
subset selection for classification of hand movements. Biomed. Signal Process. Control 2016, 27, 24–31. [CrossRef]
Phinyomark, A.; Scheme, E. A feature extraction issue for myoelectric control based on wearable EMG sensors. In Proceedings of
the 2018 IEEE Sensors Applications Symposium (SAS), Seoul, Republic of Korea, 12–14 March 2018. [CrossRef]
Adewuyi, A.A.; Hargrove, L.J.; Kuiken, T.A. Evaluating EMG Feature and Classifier Selection for Application to Partial-Hand
Prosthesis Control. Front. Neurorobotics 2016, 10, 15. [CrossRef] [PubMed]
Udhaya Kumar, S.; Hannah Inbarani, H. PSO-based feature selection and neighborhood rough set-based classification for BCI
multiclass motor imagery task. Neural Comput. Appl. 2016, 28, 3239–3258. [CrossRef]
Zhou, D.; Fang, Y.; Ju, Z.; Liu, H. Multi-length Windowed Feature Selection for Surface EMG Based Hand Motion Recognition. In
Proceedings of the International Conference on Intelligent Robotics and Applications, Newcastle, Australia, 9–11 August 2018.
[CrossRef]
Rahimi, I.; Gandomi, A.H.; Asteris, P.G.; Chen, F. Analysis and Prediction of COVID-19 Using SIR, SEIQR, and Machine Learning
Models: Australia, Italy, and UK Cases. Information 2021, 12, 109. [CrossRef]
Anbalagan, T.; Nath, M.K.; Vijayalakshmi, D.; Anbalagan, A. Analysis of various techniques for ECG signal in healthcare, past,
present, and future. Biomed. Eng. Adv. 2023, 6, 100089. [CrossRef]
Hagan, M.T.; Demuth, H.B.; Hudson, M.; De Jesús, O. Neural Network Design, 2nd ed.; Oklahoma State University: Stillwater, OK,
USA, 2014; pp. 38–48.
Chowdhury, R.; Reaz, M.; Ali, M.; Bakar, A.; Chellappan, K.; Chang, T. Surface Electromyography Signal Processing and
Classification Techniques. Sensors 2013, 13, 12431–12466. [CrossRef]
Gandolla, M.; Ferrante, S.; Ferrigno, G.; Baldassini, D.; Molteni, F.; Guanziroli, E.; Cotti Cottini, M.; Seneci, C.; Pedrocchi, A.
Artificial neural network EMG classifier for functional hand grasp movements prediction. J. Int. Med. Res. 2017, 45, 1831–1847.
[CrossRef]
Asteris, P.G.; Kokoris, S.; Gavriilaki, E.; Tsoukalas, M.Z.; Houpas, P.; Paneta, M.; Koutzas, A.; Argyropoulos, T.; Alkayem, N.F.;
Armaghani, D.J.; et al. Early prediction of COVID-19 outcome using artificial intelligence techniques and only five laboratory
indices. Clin. Immunol. 2023, 246, 109218. [CrossRef]
Alghamdi, W.Y. A novel deep learning method for predicting athletes’ health using wearable sensors and recurrent neural
networks. Decis. Anal. J. 2023, 7, 100213. [CrossRef]
Parajuli, N.; Sreenivasan, N.; Bifulco, P.; Cesarelli, M.; Savino, S.; Niola, V.; Esposito, D.; Hamilton, T.J.; Naik, G.R.; Gunawardana,
U.; et al. Real-Time EMG Based Pattern Recognition Control for Hand Prostheses: A Review on Existing Methods, Challenges
and Future Implementation. Sensors 2019, 19, 4596. [CrossRef]
Seniam. Available online: http://www.seniam.org/ (accessed on 30 May 2023).
Shenoy, P.; Miller, K.J.; Crawford, B.; Rao, R.P.N. Online Electromyographic Control of a Robotic Prosthesis. IEEE Trans. Biomed.
Eng. 2008, 55, 1128–1135. [CrossRef] [PubMed]
Simao, M.; Mendes, N.; Gibaru, O.; Neto, P. A Review on Electromyography Decoding and Pattern Recognition for HumanMachine Interaction. IEEE Access 2019, 7, 39564–39582. [CrossRef]
Lajante, M.M.P.; Droulers, O.; Amarantini, D. How reliable are “state-of-the-art” facial EMG processing methods? Guidelines for
improving the assessment of emotional valence in advertising research. J. Advert. Res. 2017, 57, 28–37. [CrossRef]
Computers 2023, 12, 263
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
18 of 18
Ehrmann, G.; Blachowicz, T.; Homburg, S.V.; Ehrmann, A. Measuring Biosignals with Single Circuit Boards. Bioengineering 2022,
9, 84. [CrossRef] [PubMed]
Miften, F.S.; Diykh, M.; Abdulla, S.; Siuly, S.; Green, J.H.; Deo, R.C. A new framework for classification of multi-category hand
grasps using EMG signals. Artif. Intell. Med. 2021, 112, 102005. [CrossRef]
Fang, Y.; Hettiarachchi, N.; Zhou, D.; Liu, H. Multi-modal sensing techniques for interfacing hand prostheses: A Review. IEEE
Sens. J. 2015, 15, 6065–6076. [CrossRef]
Farina, D.; Jiang, N.; Rehbaum, H.; Holobar, A.; Graimann, B.; Dietl, H.; Aszmann, O.C. The extraction of neural information from
the surface EMG for the control of upper-limb prostheses: Emerging avenues and challenges. IEEE Trans. Neural Syst. Rehabil.
Eng. 2014, 22, 797–809. [CrossRef]
Kaur, A.; Kumar, A. SEMG based classification using wavelet function for around shoulder muscles. J. Eng. Sci. Technol. Rev.
2017, 10, 109–114. [CrossRef]
Filter Design Program for the UAF42. Available online: https://www.ti.com/lit/pdf/sbfa002 (accessed on 29 May 2023).
Wang, R. Introduction to Orthogonal Transforms with Applications in Data Processing and Analysis, 1st ed.; Cambridge University
Press: Cambridge, UK, 2012; pp. 450–470.
Parish, C.M.; Edmondson, P.D. Data visualization heuristics for the physical sciences. Mater. Des. 2019, 179, 107868. [CrossRef]
Matejka, J.; Fitzmaurice, G. Same Stats, Different Graphs: Generating Datasets with Varied Appearance and Identical Statistics
through Simulated Annealing. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, Denver,
CO, USA, 6–11 May 2017. [CrossRef]
Setiawan, J.; Ariyanto, M.; Nugroho, S.; Munadi, M.; Ismail, R. A Soft Exoskeleton Glove Incorporating Motor-Tendon Actuator
for Hand Movements Assistance. IREACO 2020, 13, 1–11. [CrossRef]
James, G.; Witten, D.; Hastie, T.; Tibshirani, R. An Introduction to Statistical Learning, 1st ed.; Springer: Berlin/Heidelberg, Germany,
2013; pp. 127–168.
Labatut, V.; Cherifi, H. Accuracy measures for the comparison of classifiers. In Proceedings of the 5th International Conference on
Information Technology, Amman, Jordan, 11 May 2011.
Kuo, S.M.; Lee, B.H.; Tian, W. Real-Time Digital Signal Processing: Fundamentals, Implementations and Applications, 3rd ed.; Wiley:
Hoboken, NJ, USA, 2013; pp. 1–42.
Kumar, D.K.; Jelfs, B.; Sui, X.; Arjunan, S.P. Prosthetic hand control: A multidisciplinary review to identify strengths, shortcomings,
and the future. Biomed. Signal Process. Control 2019, 53, 101588. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.