www.nature.

com/scientificreports

OPEN A smart approach to EMG envelope

extraction and powerful denoising
for human–machine interfaces
Daniele Esposito , Jessica Centracchio *
, Paolo Bifulco & Emilio Andreozzi

Electromyography (EMG) is widely used in human–machine interfaces (HMIs) to measure muscle

contraction by computing the EMG envelope. However, EMG is largely affected by powerline
interference and motion artifacts. Boards that directly provide EMG envelope, without denoising
the raw signal, are often unreliable and hinder HMIs performance. Sophisticated filtering provides
high performance but is not viable when power and computational resources must be optimized.
This study investigates the application of feed-forward comb (FFC) filters to remove both powerline
interferences and motion artifacts from raw EMG. FFC filter and EMG envelope extractor can be
implemented without computing any multiplication. This approach is particularly suitable for very
low-cost, low-power platforms. The performance of the FFC filter was first demonstrated offline by
corrupting clean EMG signals with powerline noise and motion artifacts. The correlation coefficients
of the filtered signals envelopes and the true envelopes were greater than 0.98 and 0.94 for EMG
corrupted by powerline noise and motion artifacts, respectively. Further tests on real, highly noisy
EMG signals confirmed these achievements. Finally, the real-time operation of the proposed approach
was successfully tested by implementation on a simple Arduino Uno board.

Electromyography (EMG)1 is one of the most used techniques to realize human–machine interfaces (HMIs)2,3
and assistive device control. EMG allows measuring the contraction of muscles by recording their electrical
activity. This is usually performed via skin electrodes, so called surface EMG (sEMG). sEMG signal amplitude
is a few millivolts and its power spectrum is mainly concentrated in 10–250 Hz b ­ and1,4. The EMG signal is rich
in information, but only concise information on whether the muscle is active and/or the level of muscular con-
traction is usually of interest for HMI applications. To this aim, the EMG signal envelope is commonly used,
rather than the raw signal. The combination of raw EMG signal rectification and low pass filtering is referred to
as the EMG Linear Envelope (EMG-LE)1. The envelope of an EMG signal can also be estimated by computing
the local root mean square (RMS) value of the raw EMG via a moving window. EMG-LE gives a measure of the
local signal power and can be computed either via analog or digital ­processing5.
To date, many electronic boards are available for EMG signal acquisition, which can be interfaced with micro-
controller boards to easily implement H ­ MIs2. These EMG boards usually provide the EMG-LE to substantially
reduce the required sampling rate (by at least one order of magnitude) compared to that required for acquisition
of raw EMG signals (at least 1 kHz). However, sEMG signals are affected by various noises, the main ones being
powerline interference (PLI) and motion artifacts, which can seriously impair the quality of the EMG-LE signals
provided by such boards.
PLI is an electromagnetic interference that is caused by the capacitive or inductive coupling between EMG
electrodes and the powerline in the surrounding ­environment6–10. In both cases, PLI makes a 50 Hz or 60 Hz
sinewave (depending on the national regulations) and higher harmonics appear superimposed to the actual EMG
signal. The PLI signal in EMG recordings can be much larger than the actual EMG signal, and its amplitude can
vary over time.
Motion artifacts in EMG recordings are unavoidable, because they come from body motion, increase of mus-
cle cross-section, and the resulting slippage of various tissues and skin l­ ayers11. The two main sources of motion
artifacts in sEMG are the mechanical disturbance of the electrode charge layer, due to relative motion between
the electrode and the underlying skin, and the deformation of the skin under the electrodes, which makes the
potentials between skin layers change when skin is deformed or s­ tretched8,12. The latter is much more difficult
to attenuate, except by strongly abrading or puncturing the skin, which is not feasible in most practical cases.

Department of Electrical Engineering and Information Technologies, University of Naples Federico II, Via Claudio,
21, 80125 Naples, Italy. *email: jessica.centracchio@unina.it

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 1

Vol.:(0123456789)
 www.nature.com/scientificreports/

According to the International Society of Electrophysiology and Kinesiology standards, the frequency band
of the diagnostic EMG signal ranges from 5–10 Hz up to 400–500 H ­ z7. Motion artifacts are generally confined
1,13,14
to very low frequencies, typically below 20–30 ­Hz . The spectrum of powerline interference should be con-
centrated at the powerline frequency (e.g., 60 Hz in Japan and almost all American continent, 50 Hz in Europe,
Asia, Africa, Australia and some South American countries) but, due to non-linear loads increasingly used
in modern equipment, higher harmonics of the fundamental powerline frequency are also ­present8. Figure 1
shows a qualitative sketch of sEMG and noise signals in the time domain, along with their power spectra in the
frequency domain.
Most of the above-mentioned EMG boards providing EMG-LE signals directly, usually do not allow reducing
the powerline interference nor the motion artifacts, which must be filtered before extracting the envelope, to
obtain suitable EMG-LE signals. Consequently, these EMG boards often end-up providing noisy and uncertain
envelope signals that cause malfunctioning of HMI applications, especially when the users are in environments
with strong powerline electromagnetic fields, and when they perform large ­movements15. For this reason, some
EMG boards also provide the raw EMG signal, in addition to the envelope signal.
In literature, many studies describe rather sophisticated methods to reduce EMG noises, which are usu-
ally based on digital ­filters16,17, neural networks (NN)18, ­wavelets19, ensemble empirical mode decomposition
(EEMD)20, canonical correlation analysis (CCA)21, among many others. Many of these methods are currently not
suitable for real-time operation. This is often due to heavy computational burdens that pose strong limitations,
especially in the case of wearable or portable HMI applications, which are based on battery-powered devices. In
fact, these devices are usually equipped with resource-limited computing platforms to achieve very low-power
operation and thus cannot support demanding calculations. However, there has recently been an increasing
interest in developing small neural networks, and in general processing algorithms that operate in real time,
which are particularly optimized for embedded systems based on microcontrollers or programmable hardware
(e.g., Field Programmable Gate Arrays, System-on-Chip), within the paradigm of edge computing.
This study presents a simple and elegant solution for real-time, noise free EMG-LE extraction with extremely
low computational complexity. A feed-forward comb (FFC) ­filter22 removes both powerline interference and
motion artifacts, while subsequent rectification and averaging provide the EMG-LE, without using a single
multiplication.
Although the FFC filter is well-known in literature, its use for EMG denoising has never been investigated
thoroughly. Probably, this has been due to the fact that the frequency response of the FFC filter causes substantial
distortion to the EMG spectrum. Nevertheless, this study demonstrates that the FFC filter is effective for PLI
and motion artifacts removal and preserves very well the morphology of the EMG-LE. Considering that this
performance can be obtained at the cost of a single sum computation, as compared to other filtering approaches
for noise removal, it follows that this solution offers great computational advantages for HMI applications. In
fact, it can be easily implemented on very resource-constrained platforms developed for low-power applications,
or used to save computational resources for more demanding tasks.

Methods
Feed‑forward comb filter. The FFC filter operation is based on adding a delayed version of the input sig-
nal to itself, so as to produce constructive and destructive interference. The destructive interference mechanism
is precisely that used to cut off a specific frequency component, along with its harmonics. The general difference
equation of the FFC filter can be expressed as follows:
y(k) = x(k) + αx(k − N), (1)
where x is the input signal, N is the delay expressed in number of samples, and α is a parameter that regulates
some aspects of the filter behaviour. In particular, when α is equal to ± 1, the minima of the FFC amplitude

Figure 1.  Qualitative sketch of sEMG signal, powerline interference and motion artifacts in time and frequency
domains.

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 2

Vol:.(1234567890)
www.nature.com/scientificreports/

response are equal to zero, so they become nulls of the amplitude response. This choice is used to cancel out
specific frequency components, such as the fundamental frequency of powerline interference and its harmonics.
The following equations report the expressions of the FFC amplitude responses for α = ± 1:
 
(2)
 
H fn = 2sin(πNfn ), α = −1,

 
(3)
 
H fn = 2cos(πNfn ), α = 1,

where fn is the normalized frequency, i.e. the absolute frequency f divided by the sampling frequency fs. The
nulls of the amplitude responses in Eqs. (2) and (3) are reported in the following Eqs. (4) and (5), respectively:
fs
fnull = k , k ∈ Z, (4)
N

fs fs fs
fnull = + k = (1 + 2k) , k ∈ Z. (5)
2N N 2N
The amplitude response in Eq. (2) has nulls for frequency values that are integer multiples of fs/N, while the
amplitude response in Eq. (3) has nulls for odd integer multiples of fs/(2N). For this reason, the choice of α = − 1
is the one that allows removing the fundamental frequency of powerline noise interference and all its harmon-
ics (both odd and even). Indeed, by choosing N equal to fs/fpli, with fpli being the fundamental frequency to be
cancelled, the nulls of the amplitude response would correspond to all integer multiples of fpli. On the contrary,
by choosing α = 1, one would need to set N equal to fs/(2fpli) to make the first null to correspond to fpli and all
other nulls to odd multiples of fpli, which leaves even multiples of fpli practically unaltered, thus resulting in poor
PLI removal performances.
The difference equation of the specific FFC filter that was used in this study is reported below:
y(k) = x(k) − x(k − N). (6)
Considering 50 Hz as the fundamental powerline interference frequency and assuming a sampling frequency
of 1000 Hz, N must be set equal to 20. The amplitude response of the FFC filter with N = 20 is depicted in Fig. 2,
where it could be observed that the nulls correspond to 0 Hz, 50 Hz and its integer multiples.
It is worth noting that the FFC filter has a zero DC gain with a high-pass behaviour at low frequencies. The
− 3 dB cut-off frequencies can be computed as follows:
 1
H(fn3 dB ) 1    1 + Nk
= √ ⇒ sin πNfn3 dB  = √ ⇒ fn3 dB = 4N 3 k ,
Hmax 2 2 4N + N

1
f (7)
+ k Ns = 41 + k fpli = 12.5 Hz + k · 50 Hz
 
f3 dB =  43  fs 3  .
4 + k N = 4 + k fpli = 37.5 Hz + k · 50 Hz

The FFC filter has two main effects: it operates as a comb filter for the removal of powerline interference and
filters out the DC and near-DC components, where the power spectra of motion artifacts and baseline oscillations
are mainly concentrated (below 20–30 ­Hz4). These aspects are very important for the extraction of the EMG-LE,

Figure 2.  Amplitude response of the FFC filter with N = 20 and fs = 1000 Hz. The − 3 dB line is depicted as a red
horizontal dashed line, while the black vertical dashed lines correspond to the cut-off frequencies.

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 3

Vol.:(0123456789)
 www.nature.com/scientificreports/

which is usually computed by cascading a full-wave rectification (absolute value or power functions with even
exponents) and a low-pass filtering (usually at 5-10 Hz for HMI applications)3,7,23. Indeed, both powerline inter-
ference and motion artifacts may severely corrupt the EMG-LE signal, and must be filtered out upstream of the
actual envelope computation. Figure 3 shows an example of these issues (the EMG-LE signals were computed
by cascading a full-wave rectification and a low-pass filtering at 5 Hz).
In most low-cost, low-power platforms for HMI implementation, a unipolar A/D converter is commonly used,
so the digitized EMG signal exhibits only positive values and cannot be directly fed to the full-wave rectification
­stage24,25. In fact, the true reference voltage must first be determined and subtracted from the just digitized EMG,
in order to recover the positive and negative values of the EMG. This is not always a straightforward operation, as
unpredictable offset voltages may appear at the output of EMG measurement devices, which result in improper
compensation of reference voltage. The zero DC gain of the FFC filter offers a simple and elegant solution to this
problem, by providing a zero-mean EMG signal that is suitable for EMG-LE extraction.
It is worth noticing that the FFC filter rejects frequency bands of fpli/2 width, centered around the powerline
interference components k·fpli, thus distorting the original EMG signal. However, it will be shown below that for
the sole purpose of extracting the EMG-LE, such distortions are not of concern in practice.
In summary, the FFC filter has two main advantages: (a) it ensures the joint removal of powerline interfer-
ence (with harmonics), offset voltages and motion artifacts/baseline oscillations, thus providing a pre-processed
EMG signal that is suitable for EMG-LE extraction; (b) it has an extremely simple computation, that, in fact,
requires only a single subtraction, thus resulting in an almost negligible computational burden even for resource-
constrained platforms.

Offline tests on noisy EMG signals. A quantitative assessment of the FFC filter performance was carried
out by corrupting a noiseless real EMG signal with a simulated powerline interference, both with flat ampli-
tude and with amplitude modulation. All signal processing operations were performed in MATLAB® R2017b
(The MathWorks, Inc., 1 Apple Hill Drive, Natick, Massachusetts, 01760, USA). The noiseless EMG signal was
extracted from a dataset acquired in a previous s­ tudy26. No measurements were carried out during this study.
The sEMG signals had been recorded at sampling frequency of 1 kHz from forearm muscles.
The powerline interference was simulated by using amplitudes and phases of the 50 Hz component and its
harmonics up to 500 Hz, which had been extracted from a measurement of real powerline interference obtained
with the same experimental set-up. An amplitude modulated powerline interference was also designed by using
a slow sinewave as the modulating signal. The simulated noise was then added with increasing amplitudes to
the noiseless EMG signal, so as to obtain SNR values between 0.05 and 10. The noisy EMG signals thus obtained
were processed with the FFC filter. The linear envelopes of the noiseless, noisy and filtered EMG signals were
computed by means of a full-wave rectification (absolute value operation) and a low-pass filtering at 5 Hz via a
moving average (MA) on 88 samples (with a sampling frequency of 1000 Hz, a MA filter with N = 88 samples
has a cut-off frequency of 5 Hz). The block diagram illustrating the various filtering stages is presented in Fig. 4.
Moreover, the cross-correlation between the noiseless EMG-LE signal and each noisy and filtered EMG-LE
signal was computed by considering the whole length. Then, the maximum of the cross-correlation was located,
the corresponding lag was used to re-align the signals, and the Pearson’s correlation coefficient was determined.
The trends of the correlation coefficients with the SNR were finally compared.
A further quantitative assessment of the FFC filter performances was carried out by corrupting the same
noiseless EMG signal with real motion artifacts. Signals chunks including only motion artifacts were extracted
from the ­dataset26. The noiseless EMG signal was then corrupted by motion artifacts signals with increasing
amplitudes, so as to obtain SNR values between 0.05 and 10. The corrupted EMG signals were processed with the
FFC filter. Finally, the linear envelopes of the noiseless, corrupted and filtered EMG signals and the Pearson’s cor-
relation coefficients between the noiseless EMG-LE signal and each corrupted and filtered EMG-LE signal were
computed via the same procedures. The trends of the correlation coefficients with the SNR were finally compared.
Noisy EMG signals with powerline interference and motion artifacts were extracted from the same dataset
to further test the performance of the FFC filter on actual noisy data.

Performance comparison with other denoising algorithms. The performance of the FFC filter was
compared with two denoising approaches proposed in ­literature27,28. The first approach operates in the frequency
domain and mainly consists of Fast Fourier Transform (FFT) computation, cancellation of spectral coefficients
corresponding to the harmonics of the powerline frequency and to the main spectral content of motion artifacts
at low frequencies, and finally inverse FFT computation. This FFT-based approach was designed with the aim of
reducing as much as possible the number of samples for FFT computation, to reduce the computational burden
per sample in a sliding window application, or the computational delay in a block processing application. How-
ever, since the number of samples determines the frequency resolution of the FFT, a maximum resolution value
of 25 Hz was set, which for EMG signals sampled at 1 kHz was achieved by computing the FFT on 40 samples,
in order to be able to precisely cancel the spectral coefficients corresponding to the harmonics of the powerline
frequency (50 Hz) and those corresponding to the 0–25 Hz range for motion artifacts removal. The second
approach, referred to as FilterDxN, operates in the time domain and is essentially a high quality factor comb
filter with integer coefficients, properly tuned to exhibit zeros at the harmonics of the powerline frequency. The
FilterDxN features two parameters, namely N and D. As per indication of the original publication, N was set to
17, while D was set to the ratio of the sampling frequency (1000 Hz) and the powerline frequency (50 Hz), which
in this study was equal to 20. Therefore, a Filter20x17 was actually considered for the tests. The two algorithms
were tested both for PLI and motion artifacts removal, via the same methodologies adopted for the performance
assessment of the FFC filter, previously described in the “Offline tests on noisy EMG signals” section. In particu-

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 4

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 3.  (a) Noiseless EMG signal with superimposed EMG-LE; EMG signal corrupted by (b) variable
amplitude powerline interference, (c) baseline oscillations, (d) both, with corrupted EMG-LE and true EMG-LE;
(e) EMG signal processed with FFC filter, recovered EMG-LE, true EMG-LE.

lar, the tests on PLI with constant and modulated amplitude were performed by considering only the lowest SNR
(0.05), because all algorithms showed stable performance at varying SNR. The performance on motion artifacts

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 5

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 4.  Block flow diagram of the various filtering stages applied on the raw EMG signal.

removal, instead, were assessed for all SNR values. Moreover, the computational and memory resources require-
ments were evaluated and compared for all the considered algorithms.

Real‑time tests on noisy EMG signals. The real time performance of the proposed approach was
assessed via the following test scene. An arbitrary waveform generator (GW Instek AFG-2005, Good Will
Instrument Co., Ltd., No.7-1, Jhongsing Road., Tucheng Dist., New Taipei City 236, Taiwan) was used to output
voltage signals from real recordings of noisy EMG signals (previously acquired in Ref.26), so as to resemble a real
scenario. The output voltage of the generator was fed to an off-the-shelf EMG sensor board “MyoWare Muscle
Sensor”29, in order to amplify the EMG signals. The amplified signals, provided as an analog output on the “Raw
EMG signal” pin of the “Myoware Muscle Sensor” board (pin 7), were fed to an analog input of an Arduino UNO
board. At this stage, the raw EMG signals were first digitized at 1 kHz, using the on-board analog-to-digital
converter (10-bit resolution), and then processed in real time via the proposed methodology. The open-source
library ­TimerOne30 was used to ensure a precise sampling time. The EMG processing performed by the Arduino
UNO board consisted of the FFC filtering for PLI and motion artifacts removal, signal rectification, and low-pass
filtering at about 4 Hz, by using a MA filter with a window size of 128 samples to eventually obtain the EMG-
LE. Each new sample of the EMG-LE was released at a rate of about 8 Hz and sent to a personal computer via a
Bluetooth serial interface (HC-0531), connected to the Arduino UNO board. The full Arduino code is available
as Supplementary Material.
It is worth noticing that other commercial EMG sensors, such as Seed EMG d ­ etector24 or Gravity EMG
­sensor25, do not offer a specific port for raw EMG output, rather providing only the EMG envelope. For this
reason, such sensors are not compatible with the proposed methodology. The correct execution of the algo-
rithm strictly requires a precise uniform sampling, ensured by the TimerOne library. Indeed, adding simple idle
cycles to sample the raw EMG signal would result in nonuniform sampling and impair correct filter operation.
In addition, by only changing the sampling frequency of the raw EMG signal (e.g., from 1000 to 1200 Hz), the
algorithm could be easily adjusted for the removal of 60 Hz (and higher harmonics) powerline interference, as
well as motion artifacts. Other parameters, such as the number of FFC filter coefficients (N = 20) and the width
of the MA filter window (128 samples) for linear envelope computation, can remain unchanged. The use of a
window size that is a power of 2 allows replacing the division required in the MA filter by a simple bit shifting,
which has a negligible computational cost. Table 1 summarizes the algorithm specifications for 50 Hz or 60 Hz
powerline frequencies.

Results
Offline tests on noisy EMG signals. Figure 5 shows the noiseless EMG signal (top panels), the noisy
signals (central panels) with SNR equal to 2 (both with flat amplitude and amplitude modulated noise) and the
related filtered signals (bottom panels). It could be observed that the original EMG signals were efficiently recov-
ered from both the very noisy EMG signals.
This result is confirmed by the Pearson’s correlation coefficients scored by the filtered EMG-LE signals with
respect to the corresponding noiseless EMG-LE signals, which turned out to be in excess of 0.98 for all values of
the SNR (ranging from 0.05 to 10), as reported in Fig. 6. The same figure also depicts the correlation coefficients
of the noisy EMG-LE with respect to the corresponding noiseless EMG-LE signals. It is worth noticing that for
a SNR of 7, the correlation coefficients of the noisy signals already dropped below 0.9, and for a SNR of 2 (the

Parameters for 50 Hz PLI Parameters for 60 Hz PLI

Raw EMG sampling frequency 1000 Hz 1200 Hz
Number of FFC filter coefficients 20 20
Lower cut-off frequency of FFC filter ≈ 12 Hz ≈ 15 Hz
Number of samples of MA filter window for EMG-LE computing 128 128
Upper cut-off frequency of MA filter for EMG-LE computing ≈ 4 Hz ≈ 5 Hz
Data output frequency from the Arduino Bluetooth serial port ≈ 8 Hz ≈ 9 Hz

Table 1.  Parameters setting for 50 Hz or 60 Hz powerline interference.

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 6

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 5.  Top panels: noiseless EMG signal; central panels: noisy EMG signals obtained by corrupting the
noiseless EMG signal with flat amplitude powerline noise (left) and amplitude modulated power-line noise
(right); bottom panels: signals obtained by applying the FFC filter to the noisy EMG signal with flat amplitude
(left) and amplitude modulated (right) powerline noise.

Figure 6.  Pearson’s correlation coefficients of noisy and filtered EMG-LE with respect to noiseless EMG-LE.
Solid lines refer to the EMG signal corrupted with flat amplitude powerline interference (yellow line) and the
related filtered signal (blue line); dashed lines refer to the EMG signal corrupted with amplitude modulated
powerline interference (purple line) and the related filtered signal (orange line).

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 7

Vol.:(0123456789)
 www.nature.com/scientificreports/

same SNR value of the signals depicted in Fig. 5) correlation coefficients as low as 0.51 and 0.36 were obtained,
respectively, for EMG signals corrupted by flat amplitude and amplitude modulated powerline interference.
Figure 7 shows the noiseless EMG signal in the top panel, the EMG signal corrupted by motion artifacts
with SNR equal to 2 in the central panel, and the related filtered signal in the bottom panel. Once again, the FFC
filter successfully recovered the original EMG signal from its corrupted version. Indeed, the FFC filter provided
Pearson’s correlation coefficients in excess of 0.94 for SNR values ranging from 1 to 10 (see Fig. 8). For lower
SNR values, the correlation coefficients exhibited a rather brisk drop, while still maintaining a correlation gain
in excess of 2 for SNR values down to 0.2. Regarding the corrupted EMG signals, it could be observed that for a
SNR of 7, the correlation coefficients already dropped below 0.8, and for a SNR of 2 (the same SNR value of the
signals depicted in Fig. 7) a correlation coefficient as low as 0.22 was obtained.
Figure 9 shows the results achieved by the FFC filter on actual noisy EMG signals affected by amplitude
modulated powerline interference (top panel) and motion artifacts (bottom panel). In Fig. 9a it could be observed
how the amplitude variations of the powerline interference in the noisy EMG (in blue) resulted in spurious
contraction peaks in the related EMG-LE (in red), and how they were effectively removed by the FFC filter in
the processed EMG (in yellow) and the related EMG-LE (in green). Figure 9b demonstrates the capability of
the FFC filter to remove the baseline oscillations due to motion artifacts, which likewise resulted in spurious
contraction peaks in the noisy EMG-LE.

Performance comparison with other denoising algorithms. The results obtained for the EMG sig-
nals corrupted by PLI with SNR equal to 0.05, both with constant and modulated amplitude, are outlined in
Table 2. The Filter20x17 provided roughly a 1% increase in Pearson’s correlation coefficient between the filtered
and the noiseless EMG-LE as compared to the FFC filter, while the FFT-based approach provided a slightly lower
performance.
The results obtained for EMG signals corrupted by motion artifacts (SNR ranging from 0.05 to 10) and pro-
cessed via the compared algorithms are shown in Fig. 10. In particular, Fig. 10a depicts the Pearson’s correlation
coefficients computed between the filtered EMG-LE signals and the noiseless EMG-LE reference, while Fig. 10b
depicts the differences between the correlation coefficients obtained via the FFC algorithm and those obtained
via the Filter20x17 and the FFT-based approach. The results show that for lower SNR values (SNR < 2) the FFC
substantially outperforms the Filter20x17, by providing up to 35% increases in correlation coefficients, while for
higher SNR values (SNR > 2) the Filter20x17 provides slight increases (< 1 %) in correlation coefficients with

Figure 7.  Top panel: noiseless EMG signal; central panel: EMG signal corrupted with motion artifacts
(SNR = 2); bottom panel: filtered signal obtained by applying the FFC filter to the corrupted EMG signal
depicted in the central panel.

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 8

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 8.  Pearson’s correlation coefficients of EMG-LE of signals corrupted by motion artifacts (red line) and
related filtered signals (blue line) with respect to noiseless EMG-LE.

respect to the FFC filter. The FFT-based approach provides slightly lower performance as compared to the FFC
for all SNR values.
The computational and memory resource requirements of the analyzed algorithms were outlined in Table 3,
from which one can observe that the FFC filter features both the lowest computational burden, requiring the
computation of only one sum and no multiplications, and the lowest memory storage of 20 input signal samples.
The Filter20x17 requires the highest memory storage of 321 input signal samples, while the FFT-based filter
presents the highest computational burden of at least 106 additions and 106 multiplications.

Discussion
The algorithm presented in this study has the significant advantage of removing both powerline interference
and motion artifacts on sEMG recordings in real time, as well as providing the EMG linear envelope with an
extremely low computational load. It can be used to remove both 50 or 60 Hz PLI (and higher harmonics) by
changing only the sampling frequency of the raw EMG signal. In addition, once the raw EMG signal is digitized
and the FFC filter is applied, the EMG-LE is calculated using a rectification and a moving average low-pass filter.
Different tests were performed on EMG signals, artificially corrupted by powerline interference and motion
artifacts, and proved the capacity of the FFC filter to effectively remove such noises. The proposed filter was
capable of providing extremely accurate estimates of the noiseless EMG envelopes. Indeed, the Pearson’s cor-
relation coefficients computed between the filtered and the true linear envelopes were always greater than 0.98,
even for signals corrupted by PLI with SNR values as low as 0.05. Similar results were obtained for EMG signals
corrupted by motion artifacts, with correlation coefficients in excess of 0.94, even for SNR values low as 1.
The FFC filter was further compared with two approaches previously proposed in literature, namely FilterDxN
and an FFT-based approach, both in terms of quality of the filtered EMG-LE signals obtained, and in terms of
computational and memory resource requirements. The results showed that both the FilterDxN and the FFT-
based approach have substantially higher computational and memory resource requirements as compared to
the FFC filter, which, in addition, always outperforms the FFT-based approach in terms of filtering efficacy, and
also outperforms the FilterDxN on EMG signals that are heavily corrupted by motion artifacts while providing
comparable performances for higher SNRs and for PLI removal. For these reasons, the FFC filter appears to be
preferrable over the compared approaches for the purpose of reliable EMG-LE extraction, as it ensures, at the
same time, substantially improved filtering efficacy and computational efficiency.
The proposed method provided excellent results also on real noisy EMG signals. Considering that the
MyoWare board does not filter out the noises, its analog EMG-LE output would have provided very inaccurate
estimates of muscle activity (see Fig. 3). This largely explains the practical problems encountered by many users
of EMG sensor boards. The real-time performance of the FFC filter was confirmed by the effective implementa-
tion on the Arduino UNO board. Indeed, the design of a processing scheme with no multiplications and a rather
small number of sums allowed its implementation on a simple and low-cost microcontroller like the Atmel
ATmega328 equipped on the Arduino UNO board.
It is worth mentioning that the proposed filter cannot be used as a noise suppressor for other biopotentials
such as electroencephalogram (EEG) and electrocardiogram (ECG). Indeed, the fundamental components of
these signals would be unavoidably cancelled or distorted due to the high-pass cut-off, which is at about 12 Hz for
50 Hz PLI (about 15 Hz for 60 Hz PLI). The use of the proposed filter is also not recommended for the acquisition
of EMG signals for diagnostic purposes. Although the high-pass cut-off at 12 Hz is recommended in many EMG
applications, the FFC filter exhibits stop-bands at multiples of the powerline frequency that are too wide and
cause distortions of the EMG s­ ignal14. This behaviour does not ensure that the FFC filter sufficiently preserves

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 9

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 9.  Results obtained on EMG corrupted by: (a) powerline interference; (b) motion artifacts. The noisy
EMG is reported in blue, its linear envelope in red. The filtered EMG in yellow and its linear envelope in green.

Algorithm PLI (constant amplitude) PLI (modulated amplitude)

FFC 0.9857 0.9855
Filter20 × 17 0.9948 0.9946
FFT-based filter 0.9819 0.9817

Table 2.  Pearson’s correlation coefficients between the noiseless EMG-LE and the EMG-LE obtained by
filtering noisy EMG signals with SNR = 0.05 via the considered algorithms.

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 10

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 10.  Results obtained on EMG corrupted by motion artifacts with SNR ranging from 0.05 to 10, and then
processed via the FFC filter, the Filter20x17, and the FFT-based approach. (a) Pearson’s correlation coefficient
between the noiseless and the filtered EMG-LE for the three compared algorithms; (b) differences between the
correlation coefficients achieved via the FFC filter and those achieve via the Filter20x17 and the FFT-based
approach.

Algorithm Additions Multiplications Divisions Memory storage (samples)

FFC 1 0 0 20
Filter20 × 17 17 0 1 321
FFT-based filter > 106 > 106 0 40

Table 3.  Computational and memory resource requirements of the compared algorithms.

morphology of EMG signal components (i.e. motor unit action potential trains); nonetheless, this study showed
that the FFC filter almost perfectly preserves the envelope of EMG signals. Moreover, the FFC filter is expected
to provide lower performances in the removal of particularly brisk motion artifacts, as they would present with
significant spectral power densities within the pass-bands of the filter. The performance of the FFC filter should
be assessed on a larger cohort of cases, which may include EMG signals acquired from more subjects, in different
conditions (e.g., fatigue), and performing different physical activities (e.g., running, playing sports).
As shown in a recent s­ urvey2, EMG-based HMIs are the most popular among those that exploit biopotentials
as control signals. However, EMG recording presents well-known issues, such as need for stable electrode place-
ment and skin preparation, degradation of wet electrodes contact when conductive gel dries out, degraded SNR
that results from the use of dry ­electrodes12, susceptibility to electromagnetic interferences (e.g., PLI), motion
artifacts, crosstalk with other biopotentials (e.g., EMG signals from near muscles, ECG, etc.). A further common
issue of EMG is that electrodes applied to the skin expose the subjects to a higher risk of electrical s­ hock6,32–34.
This risk could be removed/reduced by using battery-operated devices and wireless c­ onnections35.
It is useful to highlight that, in order to overcome many of the EMG limitations, alternative methods such as
Force-myography (FMG) are becoming increasingly popular to monitor muscle ­activity2,36 and control HMIs
(e.g., hand ­prostheses37–40, ­exoskeletons41, gesture recognition d
­ evices42). FMG is currently the best alternative
to EMG linear e­ nvelope43 exploits morphological changes in muscles during contraction by using force/pressure
sensors, which have also been useful to detect respiration and mechanical heart ­activity44–46. Moreover, the FMG
presents various advantages over the EMG: no need for electrodes, i.e. no electrical risk; much simpler process-
ing, since the FMG signal can be used as it is and has a strong similarity with EMG-LE47; less susceptibility to
electromagnetic interference. However, while well-established standards for EMG measurements are available
(e.g., ­SENIAM7), clear guidelines to achieve accurate FMG measurements on specific muscles have not been
provided yet.
It is interesting to note that integration of EMG and FMG improves classification of human intent and con-
trol of powered prosthetic devices, as compared to EMG a­ lone48. Such integration would require an increased
computational demand, because both EMG and FMG signals would need to be processed. The extremely low
computational burden of the FFC filter would certainly help to reduce the computational burden of reliable
EMG-LE extraction, thus favoring the development of these new kind of HMIs on very resource-constrained
platforms developed for low-power applications.
In conclusion, although new emerging techniques are becoming more and more widespread and show com-
parable or even superior performances compared to EMG, this technique is still the current gold standard to

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 11

Vol.:(0123456789)
 www.nature.com/scientificreports/

monitor muscle a­ ctivity4,14. Consequently, it is still useful to investigate new EMG signal filtering techniques that
can increase its performance, especially in HMI ­applications14,49–51.

Data availability
The data presented in this study are available on request from the corresponding author.

Received: 8 July 2022; Accepted: 11 April 2023

References
1. Basmajian, J. V. & de Luca, C. J. Muscles Alive: Their Functions Revealed by Electromyography (Williams & Wilkins, 1985).
2. Esposito, D. et al. Biosignal-based human–machine interfaces for assistance and rehabilitation: A survey. Sensors 21, 6863. https://​
doi.​org/​10.​3390/​s2120​6863 (2021).
3. Parajuli, N. et al. Real-time EMG based pattern recognition control for hand prostheses: A review on existing methods, challenges
and future implementation. Sensors 19, 4596. https://​doi.​org/​10.​3390/​s1920​4596 (2019).
4. Konrad, P. The ABC of EMG: A Practical Introduction to Kinesiological Electromyography (Noraxon Inc., 2005).
5. Esposito, D. et al. Measurement of muscle contraction timing for prosthesis control: A comparison between electromyography and
force-myography. In Proc. 2020 IEEE International Symposium on Medical Measurements and Applications (MeMeA) 1–6 (2020).
6. Webster, J. G. Medical Instrumentation Application and Design 4th edn. (Wiley, 2009).
7. Stegeman, D. & Hermens, H. Standards for Surface Electromyography: The European Project Surface EMG for Non-Invasive Assess-
ment of Muscles (SENIAM) (2007).
8. Clancy, E. A., Morin, E. L. & Merletti, R. Sampling, noise-reduction and amplitude estimation issues in surface electromyography.
J. Electromyogr. Kinesiol. 12, 1–16. https://​doi.​org/​10.​1016/​s1050-​6411(01)​00033-5 (2002).
9. Jiang, Y. et al. Effective biopotential signal acquisition: Comparison of different shielded drive technologies. Appl. Sci. 8, 276.
https://​doi.​org/​10.​3390/​app80​20276 (2018).
10. Andreozzi, E., Gargiulo, G. D., Fratini, A., Esposito, D. & Bifulco, P. A contactless sensor for pacemaker pulse detection: Design
hints and performance assessment. Sensors 18, 2715. https://​doi.​org/​10.​3390/​s1808​2715 (2018).
11. Fratini, A., Cesarelli, M., Bifulco, P. & Romano, M. Relevance of motion artifact in electromyography recordings during vibration
treatment. J. Electromyogr. Kinesiol. 19, 710–718. https://​doi.​org/​10.​1016/j.​jelek​in.​2008.​04.​005 (2009).
12. Gargiulo, G. et al. Dry electrode bio-potential recordings. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2010 6493–6496. https://​doi.​
org/​10.​1109/​IEMBS.​2010.​56273​59 (2010).
13 Merletti, R. & Cerone, G. L. Tutorial. Surface EMG detection, conditioning and pre-processing: Best practices. J. Electromyogr.
Kinesiol. 54, 102440. https://​doi.​org/​10.​1016/j.​jelek​in.​2020.​102440 (2020).
14. De Luca, C. J., Gilmore, L. D., Kuznetsov, M. & Roy, S. H. Filtering the surface EMG signal: Movement artifact and baseline noise
contamination. J. Biomech. 43, 1573–1579. https://​doi.​org/​10.​1016/j.​jbiom​ech.​2010.​01.​027 (2010).
15. Naik, G. R. Applications, Challenges, and Advancements in Electromyography Signal Processing (IGI Global, 2014).
16. Strzecha, K. et al. Processing of EMG signals with high impact of power line and cardiac interferences. Appl. Sci. 11, 4625. https://​
doi.​org/​10.​3390/​app11​104625 (2021).
17. Xu, L. et al. Comparative review of the algorithms for removal of electrocardiographic interference from trunk electromyography.
Sensors 20, 4890. https://​doi.​org/​10.​3390/​s2017​4890 (2020).
18. Kale, S. N. & Dudul, S. V. Intelligent noise removal from EMG signal using focused time-lagged recurrent neural network. Appl.
Comput. Intell. Soft Comput. 2009, e129761. https://​doi.​org/​10.​1155/​2009/​129761 (2009).
19. Jiang, C.-F. & Kuo, S.-L. A comparative study of wavelet denoising of surface electromyographic signals. Annu. Int. Conf. IEEE
Eng. Med. Biol. Soc. 2007 1868–1871. https://​doi.​org/​10.​1109/​IEMBS.​2007.​43526​79 (2007).
20. Sun, Z. et al. Surface electromyography signal denoising via EEMD and improved wavelet thresholds. MBE 17, 6945–6962. https://​
doi.​org/​10.​3934/​mbe.​20203​59 (2020).
21. Al Harrach, M. et al. Denoising of HD-SEMG signals using canonical correlation analysis. Med. Biol. Eng. Comput. 55, 375–388.
https://​doi.​org/​10.​1007/​s11517-​016-​1521-x (2017).
22. Proakis, J. G. & Manolakis, D. G. Digital Signal Processing (Pearson Prentice Hall, 2007).
23. Medved, V. Standards for Reporting EMG Data (2000).
24. Grove—EMG Detector—Seeed Wiki. https://​wiki.​seeed​studio.​com/​Grove-​EMG_​Detec​tor/ (Accessed 27 January 2022).
25. Gravity: Analog EMG Sensor by OYMotion—DFRobot. https://​www.​dfrob​ot.​com/​produ​ct-​1661.​html (Accessed 27 January 2022).
26. Polisiero, M. et al. Design and assessment of a low-cost, electromyographically controlled, prosthetic hand. Med. Devices (Auckl.)
6, 97–104. https://​doi.​org/​10.​2147/​MDER.​S39604 (2013).
27. Simao, M., Mendes, N., Gibaru, O. & Neto, P. A review on electromyography decoding and pattern recognition for human-machine
interaction. IEEE Access 7, 39564–39582. https://​doi.​org/​10.​1109/​access.​2019.​29065​84 (2019).
28. Tabakov, S., Iliev, I. & Krasteva, V. Online digital filter and QRS detector applicable in low resource ECG monitoring systems. Ann.
Biomed. Eng. 36(11), 1805–1815 (2008).
29. MyoWare Muscle Sensor Kit—Learn.Sparkfun.Com. https://​learn.​spark​fun.​com/​tutor​ials/​myowa​re-​muscle-​sensor-​kit#​myowa​re-​
muscle-​sensor (Accessed 27 January 2022).
30. TimerOne—Arduino Reference. https://​www.​ardui​no.​cc/​refer​ence/​en/​libra​ries/​timer​one/ (Accessed 27 January 2022).
31. HC-05—Bluetooth Module. https://​compo​nents​101.​com/​wirel​ess/​hc-​05-​bluet​ooth-​module (Accessed 16 February 2022).
32. Webster, J. G. & Eren, H. Measurement, Instrumentation, and Sensors Handbook: Electromagnetic, Optical, Radiation, Chemical,
and Biomedical Measurement (CRC Press, 2017).
33. Burgess, R. C. Electrical safety. Handb. Clin. Neurol. 160, 67–81. https://​doi.​org/​10.​1016/​B978-0-​444-​64032-1.​00005-9 (2019).
34. Lee, W. R. The hazard of electrocution during patient monitoring. Postgrad. Med. J. 46(536), 355–359. https://​doi.​org/​10.​1136/​
pgmj.​46.​536.​355 (1970).
35. Esposito, D., Gargiulo, G.D., Polley, C., D’Addio, G. & Bifulco, P. Improvements of a simple piezoresistive array armband for gesture
recognition. In Proc. 2020 International Conference on e-Health and Bioengineering (EHB) 1–5 (2020).
36. Xiao, Z. G. & Menon, C. A review of force myography research and development. Sensors 19, 4557. https://​doi.​org/​10.​3390/​s1920​
4557 (2019).
37. Esposito, D. et al. The “Federica” hand. Bioengineering 8, 128. https://​doi.​org/​10.​3390/​bioen​ginee​ring8​090128 (2021).
38. Esposito, D. et al. Experimental study to improve “Federica” prosthetic hand and its control system. In Proceedings of the XV Medi-
terranean Conference on Medical and Biological Engineering and Computing—MEDICON 2019 (eds Henriques, J. et al.) 586–593
(Springer, 2020).
39. Esposito, D. et al. Study on the activation speed and the energy consumption of “Federica” prosthetic hand. In Proceedings of the
XV Mediterranean Conference on Medical and Biological Engineering and Computing—MEDICON 2019 (eds Henriques, J. et al.)
594–603 (Springer, 2020).

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 12

Vol:.(1234567890)
www.nature.com/scientificreports/

40. Esposito, D. et al. Evaluation of grip force and energy efficiency of the “Federica” hand. Machines 9, 25. https://​doi.​org/​10.​3390/​
machi​nes90​20025 (2021).
41. Esposito, D. et al. Design of a 3D-printed hand exoskeleton based on force-myography control for assistance and rehabilitation.
Machines 10, 57. https://​doi.​org/​10.​3390/​machi​nes10​010057 (2022).
42 Esposito, D. et al. A piezoresistive array armband with reduced number of sensors for hand gesture recognition. Front. Neurorobot.
13, 114. https://​doi.​org/​10.​3389/​fnbot.​2019.​00114 (2020).
43. Belyea, A., Englehart, K. & Scheme, E. FMG versus EMG: A comparison of usability for real-time pattern recognition based control.
IEEE Trans. Biomed. Eng. 66(11), 3098–3104. https://​doi.​org/​10.​1109/​TBME.​2019.​29004​15 (2019).
44. Andreozzi, E. et al. Respiration monitoring via forcecardiography sensors. Sensors 21, 3996. https://​doi.​org/​10.​3390/​s2112​3996
(2021).
45. Andreozzi, E. et al. Forcecardiography: A novel technique to measure heart mechanical vibrations onto the chest wall. Sensors 20,
3885. https://​doi.​org/​10.​3390/​s2014​3885 (2020).
46. Andreozzi, E., Gargiulo, G. D., Esposito, D. & Bifulco, P. A novel broadband forcecardiography sensor for simultaneous monitoring
of respiration, infrasonic cardiac vibrations and heart sounds. Front. Physiol. 12, 1988. https://​doi.​org/​10.​3389/​fphys.​2021.​725716
(2021).
47 Esposito, D. et al. A piezoresistive sensor to measure muscle contraction and mechanomyography. Sensors 18, 2553. https://​doi.​
org/​10.​3390/​s1808​2553 (2018).
48. Sanford, J., Patterson, R. & Popa, D. O. Concurrent surface electromyography and force myography classification during times of
prosthetic socket shift and user fatigue. J. Rehabil. Assist. Technol. Eng. https://​doi.​org/​10.​1177/​20556​68317​708731 (2017).
49. Roland, T., Amsuess, S., Russold, M. F. & Baumgartner, W. Ultra-low-power digital filtering for insulated EMG sensing. Sensors
19, E959. https://​doi.​org/​10.​3390/​s1904​0959 (2019).
50. Bi, L., Feleke, A. G. & Guan, C. A review on EMG-based motor intention prediction of continuous human upper limb motion for
human–robot collaboration. Biomed. Signal Process. Control 51, 113–127. https://​doi.​org/​10.​1016/j.​bspc.​2019.​02.​011 (2019).
51. Calado, A., Soares, F. & Matos, D. A review on commercially available anthropomorphic myoelectric prosthetic hands, pattern-
recognition-based microcontrollers and sEMG sensors used for prosthetic control. In Proc. 2019 IEEE International Conference
on Autonomous Robot Systems and Competitions (ICARSC) 1–6 (2019).

Author contributions
Conceptualization, D.E., E.A., P.B.; methodology, D.E., E.A., J.C., P.B.; formal analysis, E.A.; investigation, D.E.,
E.A., J.C., P.B.; visualization D.E., E.A., J.C.; software, D.E; data curation, D.E., E.A., J.C.; writing—original draft
preparation, review and editing, D.E., E.A., J.C., P.B. All authors have read and agreed to the published version
of the manuscript.

Competing interests
The authors declare no competing interests.

Additional information
Supplementary Information The online version contains supplementary material available at https://​doi.​org/​
10.​1038/​s41598-​023-​33319-4.
Correspondence and requests for materials should be addressed to J.C.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.

© The Author(s) 2023

Scientific Reports | (2023) 13:7768 | https://doi.org/10.1038/s41598-023-33319-4 13

Vol.:(0123456789)