Mobile Networks and Applications (2022) 27:728–738

https://doi.org/10.1007/s11036-019-01323-6

Detection and Removal of Motion Artifacts in PPG Signals

David Pollreisz1 · Nima TaheriNejad1

Published online: 8 August 2019

© The Author(s) 2019

Abstract
With the rise of wearable devices, which integrate myriad of health-care and fitness procedures into daily life, a reliable
method for measuring various bio-signals in a daily setup is more desired than ever. Many of these physiological
parameters, such as Heart rate (HR) and Respiratory Rate (RR), are extracted indirectly and using other signals such as
Photoplethysmograph (PPG). Part of the reason is that in some cases, such as RR measurements, the devices which directly
measure them are cumbersome to wear and thus, rather impractical. On the other hand, signals, such as PPG from which the
RR can be extracted, are not very clean. This poses a challenge on reliable extraction of these metrics. The most important
problem is that they are corrupted by motion artifacts. In this paper, we review the state of the art algorithms which are used
to detect and filter motion artifacts in PPG signals and compare them in terms of their performance. The insight provided
by this paper can help the scientists and engineers to obtain a better understanding of the field and be able to use the most
suitable technique for their work, or come up with innovative solutions based on existing ones.

Keywords Motion artifact · PPG · Wearable devices · Health-care systems

1 Introduction propagate through the system and lead to false diagnoses.

One of these fundamental and widely-used sensors is the
The face of health-care systems across the globe is Photoplethysmograph (PPG) because it is easy to use, cheap
changing thanks to Wearable Health-care Systems (WHS) and can help in extracting many health-care parameters of
and Internet of Things (IoT), and their benefits such as cost interest. PPG measures the blood pulse wave from which the
effectiveness and the extended information they provide heart rate, its variations and even the respiratory rate can be
[1–6]. Their applications ranges from daily well-being extracted. Through physiological mechanisms respiration
purposes to emotion recognition [7, 8], Early Warning Score modulates the blood volume pulse in three different ways:
(EWS) [4, 9, 10], and detection of epileptic seizures [11]. Amplitude Modulation (AM), Baseline Wanderer (BW)
Moreover, typical medical devices in the health-care domain and Frequency Modulation (FM) [14]. From these features
that are present in a hospital are expensive and need respiratory rate can be extracted.
trained practitioners to operate them. In contrast, wearable The largest problem with the proper extraction of
devices can be operated by general public, cost little, and these health parameters is that the PPG signals are often
can be deployed to perform their job inside and outside measured during various kinds of movement and therefore
hospitals [4, 6]. This can reduce hospitalization and play an are corrupted with motion noise. This noise can appear in
important role for the aging population. [4, 6, 10, 12]. the form of unruly signals of large amplitudes in the PPG
However, one of the main challenges that WHSs face is signals. It is also reflected in the frequency domain and
that of accuracy and noise [4, 10, 13]. This is an important overlaps with the frequency range of breath or heart rate
factor since many of these devices have limited number of [15]. Luckily, another fundamental sensor which is often
sensors and many of the extracted information are indirectly integrated in wearable devices is an accelerometer which
obtained through those sensors. Thus, their inaccuracy can helps in detection and removal of this noise. With the
accelerometer (or similar reference signals) it is possible to
measure movement and link it to the part of the respiratory
 Nima TaheriNejad signal which is corrupted with motion artifacts [16, 17].
nima.taherinejad@tuwien.ac.at With a reference signal (such as the acceleration) it is easier
to remove the motion noise. This is particularly important
1 Technische Universität Wien, Vienna, Austria when the noise lays in the same frequency band as the signal
Mobile Netw Appl (2022) 27:728–738 729

of interest [16]. Therefore, traditional filtering methods

may not work well since they cannot distinguish between
the sought after signal and the movement noise.
Nevertheless, there are some possibilities to alleviate this
issue (contamination of the signal due to motion artifacts)
even if no acceleration signal is available. Mainly, thanks to
the fact that if there is no noise, some statistical values are
almost the same over time. Thus, it is possible to calculate
values such as skewness and kurtosis [18], set them as
thresholds and compare them to the upcoming periods and
consequently mark parts of the PPG signal as corrupt. The
disadvantage of these methods is that it only marks the
faulty parts and cuts them out entirely. To overcome this
problem it is possible to generate a syntethic reference
signal out of the corrupted PPG signal [19] with the use of
Empirical Mode Decomposition (EMD).
In this paper, we review the state of the art regarding
all aforementioned methods as well as their advantages and
disadvantages. This should provide designers with a good
insight into design challenges and possible solutions in their
specific applications. The rest of the paper is structured as
follows: The methods for detection or reduction of motion
artifacts which do not use acceleration data are presented
in Section 2. Methods using synthetic reference signals
are reviewed in Section 3 and those which measure the
acceleration signal using a dedicated sensor are presented Fig. 1 Flowchart of MA removal algorithm with no acceleration data
in Section 4. Afterwards the result of all different methods [18]
are summarized and compared in the Section 5. Finally,
Section 6 concludes the paper.
as kurtosis, skewness, and standard deviation for each cycle
are almost equal. These values are then calculated and set as
2 No acceleration data thresholds for the comparison algorithm.

 n
If there is no accelerometer or other reference signals 1 
std =  (xi − x̄)2 (1)
available, the motion artifacts can mainly be marked and cut n
i=1
out - but not filtered out. In this section, we review three
1 n
different methods for this purpose. (xi − x̄)3
skewness = n i=1 3 (2)
std
2.1 A statistical approach 1 n
(xi − x̄)4
kurtosis = n i=1 4 (3)
std
The simplest approach does not use any extra sensors for
acceleration or try to reproduce it. As proposed in [18], If there are movement, the amplitude of the PPG signal
they use only statistical parameters of the PPG signal to changes greatly and consequently the statistical parameters
detect and cut out signal parts contaminated with Movement rise above the formerly set thresholds and the signal is
Artifacts (MAs). As seen in Fig. 1, the first step of the marked as corrupt. The corrupted signal is then cut out of
algorithm is to filter the corrupted signal using a band-pass the original signal and only the clean signal is left.
filter with the passing band of 0.5 to 6Hz, which is the main The method was tested with 10 healthy subjects who had
frequency band of interest for heart rate. to perform four different tasks: (i) no movement, (ii) finger
Afterwards it is segmented based on the signal period movement, (iii) wrist movement, and (iv) elbow movement.
and from the segmented signal the standard deviation, They recorded the PPG signal with a TI’s AFE4400-
skewness, and kurtosis is calculated using Eqs. 1, 2, and 3, SPO2EVM sensor, which is a chest strap. According to a set
respectively, where x̄ is the mean value. If there is no threshold, a part of the signal is marked as corrupted with
movement during the recording, the statistical values such movement and cut out as seen in Fig. 2. They compare their
730 Mobile Netw Appl (2022) 27:728–738

frequency complex demodulation. For the next step, the

complex demodulation is calculated by using the center
frequencies of the previously obtained time-frequency
spectrum. In a clean signal, the dominant frequency for the
heart rate is nearly continuously present and between 0.5 -
3.5 Hz but in a corrupted signal the dominant frequency can
shift due to the noise and thus, lead to a wrong Heart rate
(HR) detection. Out of these differences a threshold system
was developed to mark signals as moderately corrupted and
corrupted. After the signal is marked as corrupted it can be
deleted as in the method before.
For this work 200 subjects took part of the experiments
to test the algorithm and 521 recordings were made, each
one minute long. The used signal was recorded using an
iPhone camera and the PPG signal was extracted based
on the average of 50x50 pixels of the green band. Out of
Fig. 2 a Original PPG signal, b PPG after pre-processing, c detected
these signals three features were extracted: root mean square
movement, d cut-out algorithm applied [18]
of successive differences, Shannon entropy and sample
entropy. These features are then used in a linear Support
results with an Independent Component Analysis (ICA)- Vector Machine (SVM) classifier to detect irregular heart
Least Mean Square (LMS) algorithm. The corrupted PPG rhythms. The SVM was trained with 82 clean signals and
signal had a mean Standard Deviation (STD) error of 7.16 tested with 898 segments (30 seconds each). Without the
± 0.36 Beat Per Minute (BPM) for elbow movement, 6.12 noise detection algorithm out of 449 recordings 156 were
± 0.65 BPM for wrist movement and 11.02 ± 0.56 BPM misclassified and with the noise detection only 29 were false
for finger movement. The ICA-LMS method had an error positives.
of 6.39± 1.46 BPM for elbow movement, 6.41 ± 1.38 for
wrist movements and 10.6 ± 1.67 for finger movement. 2.3 Discrete wavelet transform
Therefore, the statistical method had a mean error of 6.58 ±
1.32 BPM for elbow movement, 6.22 ± 0.8 for wrist In the Discrete Wavelet Transform (DWT) method, it is tried
movement and 10.77 ± 1.09 for finger movement. Overall, to identify the heart rate out of the PPG signal [21] without
for all three motions, the proposed algorithm had a mean any reference signal, even if it is corrupted with motion
error of 7.85 BPM, compared to the original 8.1 BPM of the artifacts. This is a method that uses a few measurements of
corrupted signal. noise to generate an estimate of unknown states. First, as
seen in Fig. 3, the data is pre-processed. In this step, the DC
2.2 Variable frequency complex demodulation components are separated from the AC component and with
a wavelet transformation the PPG signal is first decomposed
Here we describe another method which does not use and afterwards the components in the bandwidth of 0.39 -
a reference signal but an algorithm based on Variable 12.5 Hz are reconstructed.
Frequency Complex Demodulation (VFCDM) [20]. This In the next step, the following features are extracted:
detection method looks at dominant peak amplitudes and standard deviation of peak-to-peak amplitudes as in Eq. 1,
dominant frequency components. Each 30 seconds of the standard deviation of peak-to-peak intervals, the kurtosis as
PPG signal was first band-pass filtered to 0.5Hz - 10Hz calculated in Eq. 3, and Mean Absolute Deviation (MAD)
using a 6th order zero-phase Butterworth filter, detrended, of peak-to-peak amplitudes as in
and then normalized by the maximum. On this signal a
peak detection method is used to mark all the peaks of  |An,i , −A¯n |
N
the PPG signal. If there is a gap between the finger and MADn = (4)
N
the camera (their PPG sensor) this induces motion artifacts. i=1
Such motion artifacts are marked in the 30 second windows
of the PPG signal with the pre-processing filtering and where An is the peak-to-peak amplitude and N is the number
peak detection. Afterwards, to detect artifacts the VFCDM of peak-to-peak intervals in the signal.
method is used to differentiate spectral characteristics of An SVM is then used to classify the data sets after
the noise from the clean PPG signal. The first step involves training and a 10-fold cross-validation. To remove motion
obtaining the initial time-frequency spectrum using fixed artifacts Kalman filter is used. The filter observes a series
Mobile Netw Appl (2022) 27:728–738 731

Fig. 3 The Flowchart of the

DWT algorithm with one path
for training and the other to test
it [21]

of corrupted signals to generate an estimate of the unknown This loop has to be repeated until r(t) has only one
state. First the controlled process x, is calculated by extrema. Afterwards, the spectrum of each IMF based
on the predefined frequency range has to be computed.
xk = Axk−1 + Buk−1 + wk−1 , (5)
The last step to generate the reference noise signal is
and a measurement, z, is described by identifying the desired signal portion range and eliminating
IMF corresponding to the desired frequency components of
zk = H xk + vk , (6) the PPG.
where wk and vk are random variables that represent For the “adaptive step-size LMS algorithm” block, seen
the process and measurement noise. The matrix A is the in Fig. 4, an estimate of the gradient to search the minimum
state transition matrix and the matrix B is the optional error on the surface is used. The error of each step is
input model for the control uk . Matrix H relates to the reduced by updating the step size parameter. To implement
measurement zk , as shown in Eq. 6. the adaptive filter the following steps as seen in Eq. 8 to
To test their algorithm, eleven subjects took part in Eq. 12 are used.
the measurements and they had to perform two different y(n) = wT (n)u(n) (8)
motions; (i) keep their hand still for one minute, and (ii)
e(n) = d(n) − y(n) (9)
wave their hand. This was done to identify and remove
motion artifacts from two PPG sensors that were attached w(n + 1) = w(n) + μe(n)u(n) (10)
to the left and right index finger. During their analysis, they μ(n + 1) = μ(n) + ρe(n)γ (n)u(n) H
(11)
found out, that for their tests a window length of seven d
seconds had the best results. With their method, they could γ (n) = (w(n)) (12)
dμ(n)
reduce the absolute bias from 13.97 BPM to 6.87 BPM.
where u(n) is the generated reference signal, y(n) the filter
output, w(n) the filter coefficients, e(n) the error, d(n) the
PPG signal (which would be the desired output if there is no
3 Synthetic reference data
noise), μ the step size, γ H the gradient vector and ρ controls
the step size parameter.
3.1 EMD technique
Figure 5 is an example of the results of the algorithm
designed in [19]. The test subjects had to perform three
Another possible solution without an accelerometer signal is
different movements with their finger: (i) horizontal motion,
to generate a reference signal from the corrupted PPG signal
(ii) vertical motion, and (iii) bending. For the measurement
using Complex Empirical Mode Decomposition (CEMD)
technique [19]. To generate the reference noise signal the
following steps have to be done. First, all the local minima
and maxima of the originals signal (x(t) = d(t) = S(n) +
N(n)) need to be found. The next step is to envelope all
the maxima (umax) and all the minima (umin). After all
the envelope signals are generated, the mean value, m(t), is
calculated:
umax + umin
m(t) = . (7)
2
The value of the mean is then subtracted from the original
signal: h(t) = x(t) − m(t). The new signal h(t) is
decomposed into Intrinsic Mode Function (IMF) by sifting
process until h(t) meets the IMF conditions. The next step Fig. 4 Adaptive filter using CEMD technique to generate the reference
is to identify the quasi-residue function r(t) = x(t) − c. signal out of the PPG signal [19]
732 Mobile Netw Appl (2022) 27:728–738

Fig. 5 a Original signal, b

generated movement signal, c
original signal minus the
movement noise [19]

of the PPG signal a clip-on type PPG sensor with an forgetting factor, λ, are optimized. The output is sliced into
Ni-DAQ Pad-6015 data acquisition system was used. To windows and fast Fourier transformed.
compare the original corrupted signal with the signal In [22], only 10 data sets were used to test the algorithm.
without noise, the peak to peak values of the PPG signal For their experiments the subjects had to walk, jump, and
are compared. For corrupted PPG signal with horizontal run. Two PPG sensors, as seen in Fig. 6, that were used
motion the mean error is 0.426 ± 0.087 BPM, with vertical were Hamamatsu S9706. They put them on the hip to
movement 0.514 ± 0.107, and for bending motion 0.459 ± measure the PPG signal and the movement as well. Since
0.067. In comparison, the proposed algorithm had an error they had only 10 data sets a cross-validation was performed
of 0.379 ± 0.036 BPM for horizontal movement, 0.435 ± to alleviate this problem. They first selected the n-th data set
0.059 for vertical movement, and 0.363 ± 0.131 for bending for validation and use the rest as a training set. Afterwards,
motion. Overall, the mean error of the proposed algorithm the parameters were chosen for the training set and on
is 0.392 BPM compared to 0.466 BPM of the (originally) the chosen validation set the error, En , was calculated.
corrupted signals. These two steps were repeated for all data sets and at the
end the average error was calculated. To find the optimal
3.2 Dual PPG sensor parameter values, the authors swept the K values from 5 to
50, and β from 0.01 to 0.9. With the optimal parameters,
A similar approach that uses synthesized reference signal the algorithm can reduce root mean square error by 62% for
was shown in [22]. A second PPG sensor is used to generate the walking, 83% for the running, and 79% for the jumping.
the movement signal. The second PPG sensor is positioned That is, a Root Mean Square Error (RMSE) of 6.5 BPM
a few millimeters1 away from the skin, as seen in Fig. 6a, after reduction and 28.26 BPM RMSE before reduction.
so it only measures when the subject is in motion. In Fig. 6b
one can see typical outputs of both sensors, first few seconds
without motion and afterwards with motion. 4 Measured acceleration data
This algorithm starts by band-pass filtering both of the
recorded signals between 0.8 to 5 Hz with a Hanning 4.1 LMS filtering
window of 128 points and a sampling rate of 10 samples per
second. Afterwards, an adaptive filter, as seen in Fig. 7, is One of the methods based on acceleration sensors to remove
applied. The u(n) input on Fig. 7 measures the movement motion artifacts is presented by [16]. The flow diagram
and the d(n) is the PPG signal from the sensor which is of the proposed algorithm is shown in Fig. 8. First, the
fully in touch with skin. The adaptive filter tries to bring raw PPG signal is band-pass filtered with a 4th order
the difference between u(n) and d(n) down to zero. To Butterworth Infinite Impulse Response (IIR) filter in the
do that a Normalized Least Mean Square (NLMS) and range of 0.3–5 Hz. For the motion data filtering block
Recursive Least Square (RLS) algorithm are applied and a Singular Value Decomposition (SVD) is used which
two parameters, namely the number of taps, K, and the generates a motion artifact reference for the adaptive filter.
In the third block of the algorithm, the in-band removal
takes place. There, they use a modified LMS adaptive filter,
1 We could not find what is the exact number for this gap. where the coefficients, h(n), are updated based on the least
Mobile Netw Appl (2022) 27:728–738 733

Fig. 6 a Application of the two

PPG sensors on the skin, where
the second sensor is a few
millimeter away. b The typical
output of a two PPG setup [22]

mean error, e(n). An identical filter is placed in the reference up to 12-15km/h. A dual channel PPG alongside a three-
signal path to adjust the weights. This adjustment is called axis accelerometer sensor was used to record the data from
X-LMS. In Eqs. 13–16 the X-LMS calculations are shown, the wrist with a sampling rate of 125Hz. With the X-LMS
method they had an error of 1.37 BPM.
yc (n) = w T (n) ∗ u(n) : Output generation (13)
e(n) = d(n) − yC (n) : Error calculation (14) 4.2 RLS filtering
I −1
uC ∗ (n) = ci∗ ∗ u(n − i − M + 1) (15)
i=0 A similar approach is proposed by [17]. Instead of X-LMS
w(n + 1) = w(n) + μ ∗ uC ∗ (n)e(n) : Weight updates (16) they use a DC Remover and an RLS adaptive filter. The
where u(n), d(n) and e(n) are input, desired output and first step of the algorithm, as shown in Fig. 10, is the DC
error, respectively. w(n) are the weights of the estimated removal. This IIR filter removes the offset of the PPG signal
filter and μ the step size. ci∗ represents the coefficients and thus the RLS adaptive filter has a faster convergence
of estimated filter for compensation, and uC ∗ (n) is the speed compared to an LMS algorithm. The output of the
reference signal as one can see in Fig. 9. In the last part, DC remover is the AC component of the PPG signal. The
the peak tracking takes place with some adaptive threshold following, Eqs. 17–19, describe the DC remover.
levels. This is done with the Slope Sum Method (SSM),
w(t) = x(t) + a ∗ w(t − 1) (17)
where they use a window size of 2-3 seconds to detect onset
and offset of each peak. y(t) = w(t) − w(t − 1) (18)
For the experiments in [16], 12 different subjects took Y (Z) 1 − Z −1
= (19)
part and they had to do different exercises such as running X(Z) 1 − a ∗ Z −1

Fig. 7 Adaptive filtering using a

second PPG sensor, located a
few millimeters afar from the
skin of the subject [22]
734 Mobile Netw Appl (2022) 27:728–738

Fig. 8 Flow diagram of the

algorithm that uses SVD and
LMS adaptive filter to remove
motion artifacts [16]

Fig. 9 The flowchart of the

LMS adaptive filter [16]

Fig. 10 Flow chart of a motion

artifact removal algorithm that
uses a 3-axis accelerometer to
remove the movement
artifacts [17]
Mobile Netw Appl (2022) 27:728–738 735

Fig. 11 Flow chart of the

algorithm which uses an X-LMS
adaptive filter [17]

In these equations, w(t) is the value of the operation To test their algorithm, in [17], the authors used
process which records the DC drift, and a controls the filter two different databases. The first one is Multi-parameter
cut-off frequency. When a becomes closer to 1, the slope of Intelligent Monitoring for Intensive Care (MIMIC) and a
the filter response becomes sharper and only the frequencies second set of signals that were measured from 10 subjects,
that disturb the signal are damped, however, when it is 1, where subjects did small movements such as scratching
the filter effect will be lost. After the signal is removed of themselves or shaking slightly. In Fig. 12, we can see the
its DC values the RLS adaptive filter comes into play. spectrum of a PPG signal after the algorithm removed the
In Fig. 11, one can see that s(t) is the primary input, motion noise. From the database of the 10 subjects the
xt is the reference input of the accelerometer sensor, s1 (t) results of the Bland Altman plot provides a range limit of
is the noise-free PPG signal and n(t) is the motion noise. -4.29 to 4.26 for the difference of the heart rate and the
Equation 20 shows the formula for calculating the estimated ground truth, with a STD of 3.91 BPM.
motion noise, where ωK T represents the coefficients of the

filter, Θ is the matrix of these coefficients, φ is the matrix of 4.3 Hankel matrix filtering
the accelerometer data, n̂(t) is the estimated motion noise,
and t is the sampling time. The authors of [23] try to overpass previous techniques with
Motion Artifact Removal (MAR) and Adaptive Tracking

t (AT). In [23] they use datasets from subjects who were
n̂(t) = T
ωK ∗ x K = Θ T ∗ φt (20) running. The flow chart of the proposed algorithm can be
k=0 seen in Fig. 13.
Fig. 12 a Cleaned PPG signal,
b raw PPG signal [18]
736 Mobile Netw Appl (2022) 27:728–738

In the first step, the pre-processing, the signals are

windowed, filtered and a Hankel data matrix is constructed.
The model for the acceleration data is shown in Eq. 21,
where X, Y , Z are the three axes of the accelerometer signal
and N, and L represent the observations of the three axes
accelerometer. H can be decomposed by the SVD and its
eigenvalues are calculated by

H = [XY Z] ∈ RNx3L (21)

The same is done with the PPG signal (g) whose model is
shown in Eq. 22, in which e is the heart rate and m is the
movement artifact.

g = m + e, g ∈ RN (22)
Fig. 14 The proposed algorithm with different starting points (shown
The corresponding Hankel data matrix of Eq. 22 can be in red lines) and ground truth (the black line) [23]
decomposed using SVD. The result of the decomposed
process are two orthonormal subsets. When compared with
To validate the algorithm, the authors of [23] tested it
the artifact component, it shows that the artifact and the
on 25 test subjects where the subjects wore a wrist type
heart rate signal belong to two orthonormal sub-spaces.
PPG sensors and performed different kind of exercises.2 In
The next step is to find the spectral peaks of the vectors.
Figure 14, one can see that even for different starting points
This is easily done using the calculation in Eq. 22 since the
(red lines) all lines come to the ground truth (black line,
cardiac frequency is the dominant frequency in the matrix.
an Electrocardiography (ECG) derived signal). The green
To minimize the error, when motion comes into play, a
line is the estimation of the proposed algorithm. They tested
rough estimation of the joint probability density function
their algorithm on subjects that were running and overall
of the heart rate versus the motion artifact frequency is
had an average absolute error of 2.26 BPM.
performed. This probability function is used to separate the
fundamental harmonics of the heart rate from the motion
artifact.
5 Results and discussion

In this section, we review and discuss the results of the

previously presented algorithms. The results are presented
following the same section division as before. In Table 1
we have inserted an overview of the reviewed algorithms
and their results. This table provides information on which
algorithm is only for detection and which also removes the
motion artifact through filtering. The column Acc. shows
which algorithm uses an accelerometer signal and which
one does not. A pointer on the used method, number
of subjects, and a summary of the results are the other
information seen in the table. This information can help
design engineers to choose the most suitable method for
their application, given its constraints and requirements.
For example, if no accelerometer sensor is available
and the mere detection of the motion artifacts suffices,
a statistical threshold algorithm [18] can be used. This
algorithm only uses statistical calculations like kurtosis or
standard deviation. If these values exceed their respective

Fig. 13 Flow chart of a system that uses the Hankel matrix to remove
motion artefacts [23] 2 We could not find exactly what kind of exercises were used.
Mobile Netw Appl (2022) 27:728–738 737

Missing information are shown by “NA”. In the App. (Approach) column “D” stands for detection and “R” for removal of the noise from the signal. In the Acc. (Acceleration data), “Syn.” stands
thresholds, this means there were motion artifacts in the

Finger movement, wrist move-

Scratches and slight shaking

Walking, running and jump-
motion and bending of the
Horizontal motion, vertical
signal and it can be marked as such. The marked signal

ment and elbow movement

can be then cut out or if they are used for any health
measurement, the result could be marked as potentially

Different exercises
Waving the hand
inaccurate. The Peak to Peak Error (PPE) for this method
is 7.85 BPM compared to the corrupted PPG signal with
Movement

8.1 BPM PPE. Another way to detect motion artifacts

Running
finger.
without any accelerometer sensor is to look at the frequency

ing
components with VFCDM as seen in Section 2.2. This only
needs some values without movement to set some ground
TI’s AFE4400SPO2EVM

truth and compare these values to every value to come. If

Wearable PPG sensor

the signal is corrupted it looks different in the frequency

Hamamatsu S9706
Ni-DAQ Pad-6015

domain, just as in the time domain. With this approach

iPhone camera

out of 449 recordings only 29 were false positive, in other

from DigiO2

words, 6.45% miclassification in total.

Sensor

However, we bear in mind that it is still possible to go

NA
NA
NA

beyond mere detection without an accelerometer too. For

example, despite signal corruption, using Kalman filter and
RMSE 28.26 BPM

DWT, it is possible to decompose the signal and reconstruct

PPE 13.97 BPM

PPE 0.466 BPM

the good parts [21]. Thus, the motion artifacts can be

PPE 8.1 BPM

removed and one can extract the heart rate or Respiratory

Rate (RR). The PPE of this method is 6.87 BPM comapred
Before

to 13.97 BPM without any filtering. It is also possible to

NA
NA
NA
NA

generate a reference signal out of a corrupted PPG signal

with EMD [19]. After the reference signal is generated, an
Mean absolute error
6.45% false positive
Peak to Peak Error

adaptive filter can be used to remove motion artifacts. With

(PPE) 7.85 BPM

RMSE 6.5 BPM

PPE 0.392 BPM

PPE 1.37 BPM

PPE 6.87 BPM

this method the PPE can be reduced to 0.392BPM from

of 2.26 BPM

0.466 BPM. In case a second PPG sensor is available, it can

STD 3.81
Outcome

be used to generate a movement signal [22]. The second

PPG is applied a few millimeters away from the skin, such
that if the subject moves, the distance from the sensor to skin
is reduced or increased and therefore the signal changes.
Subjects

Using the difference of the two PPG signals a movement

200

NA
NA

signal and the weights of an adaptive filter can be extracted.

12
10
12
10

11

The RMSE can be reduced to 6.5 BPM compared to 28.26

Table 1 Overview of reviewed algorithms, their method and results

Acc.

BPM before filtering.

Syn.
Syn.

Yes
Yes
Yes
No

No
No

With an accelerometer data it is easier to differentiate

between contaminated and motion-artifact-free signal and
App.

remove it. In one approach, out of the accelerometer data

D

R
R
R

a reference signal is generated with an SVD to adjust the

weights of an adaptive filter [16]. This approach has the
DC remover and RLS

lowest error of all proposed algorithms with only 1.37 BPM.

SVD and X-LMS
NLMS and RLS

Compared to the LMS algorithm of [16] a faster algorithm

MAR and AT
CEMD LMS

is the RLS [17]. It starts by removing the DC component

Evaluation
Statistical

VFCDM

for synthetic reference generation

Method

of the signal and applies an RLS adaptive filter to the

DWT

signal. Although it gains in speed, it loses in precision.

This approach predicts the heart rate with a confidence of
95% in the interval of [-4.29–4.26] BPM of the ground-
Raghuram et al. [19]

Tanweer et al. [16]

truth value and has a standard deviation of 3.81 BPM. The

Hanyu and Xiao-

Bashar et al. [20]

Lin and Ma [21]

Bacá et al. [23]

Hara et al. [22]

most complex algorithm uses MAR and AT [23]. Out of

Wu et al. [17]

the filtered signals (PPG and accelerometer) the Hankel

hui [18]

matrix has to be calculated. Out of this matrix, the noise

Paper

part has to be calculated and subtracted from the original

738 Mobile Netw Appl (2022) 27:728–738

signal. It does not have the lowest error value (their error 8. Pollreisz D, TaheriNejad N (2017) A simple algorithm for emotion
rate is 2.26 BPM), however, they claim that they have the recognition, using physiological signals of a smart watch. In: 2017
39th annual international conference of the ieee engineering in
most robust approach of all (which we could not evaluate
medicine and biology society (EMBC), pp 2353–2356
independently since they do not provide metrics such as 9. Götzinger M et al (2016) Enhancing the early warning score
STD). system using data confidence. In: International conference on
wireless mobile communication and healthcare. Springer, pp 91–
99
10. Götzinger M, Azanpour A, Azimi I, Taherinejad N, Rahmani
6 Conclusion AM (2017) Enhancing the self-aware early warning score system
through fuzzified data reliability assessment. In: International
In this paper, we reviewed eight different algorithms conference on wireless mobile communication and healthcare.
Springer
that either detect motion artifacts or filter them based
11. Forooghifar F, Aminifar A, Alonso DA (2018) Self-aware
on a synthesized or measured reference signal. We wearable systems in epileptic seizure detection, pp 7
presented their core idea and features, and summarized 12. Nia AM et al (2015) Energy-efficient long-term continuous
their characteristics and results to provide a large picture personal health monitoring. IEEE Transactions on Multi-Scale
Computing Systems 1(2):85–98
of the literature and existing methods. We hope that this
13. TaheriNejad N (2019) Wearable medical devices: challenges and
study can be used by engineers to make a better choice in self-aware solutions. In: IEEE life sciences newsletter, pp 5–6
design regarding what algorithm to use in which situations. 14. Charlton P et al (2018) Breathing rate estimation from the
Although we did our best to provide a comparison between electrocardiogram and photoplethysmogram: a review. IEEE
Reviews in Biomedical Engineering PP(99):1–1
different methods, we keep in mind that to have a fair
15. Liu J, Chen J, Jiang H, Jia W, Lin Q, Wang Z (2018) Activity
and proper comparison, one needs to implement all the recognition in wearable ECG monitoring aided by accelerometer
algorithms and test them on the same dataset under the data. In: 2018 IEEE international symposium on circuits and
same conditions and restrictions. We consider this outside systems (ISCAS), pp 1–4
the scope of this paper and as an interesting potential future 16. Tanweer KT, Hasan SR, Kamboh AM (2017) Motion artifact
reduction from PPG signals during intense exercise using filtered
work. x-LMS. In: 2017 IEEE international symposium on circuits and
systems (ISCAS), pp 1–4
Funding Information Open access funding provided by TU Wien 17. Wu CC, Chen IW, Fang WC (2017) An implementation of motion
(TUW). artifacts elimination for PPG signal processing based on recursive
least squares adaptive filter. In: 2017 IEEE biomedical circuits and
Open Access This article is distributed under the terms of the
systems conference (BioCAS), pp 1–4
Creative Commons Attribution 4.0 International License (http://
18. Hanyu S, Xiaohui C (2017) Motion artifact detection and
creativecommons.org/licenses/by/4.0/), which permits unrestricted
reduction in PPG signals based on statistics analysis. In: 2017
use, distribution, and reproduction in any medium, provided you give
29th Chinese control and decision conference (CCDC), pp 3114–
appropriate credit to the original author(s) and the source, provide a
3119
link to the Creative Commons license, and indicate if changes were
19. Raghuram M, Sivani K, Reddy KA (2016) Use of complex
made.
EMD generated noise reference for adaptive reduction of motion
artifacts from PPG signals. In: 2016 international conference on
electrical, electronics, and optimization techniques (ICEEOT),
References pp 1816–1820
20. Bashar SK, Han D, Soni A, McManus DD, Chon KH (2018)
Developing a novel noise artifact detection algorithm for
1. Dohr A et al (2010) The internet of things for ambient assisted smartphone PPG signals: preliminary results. In: 2018 IEEE
living. In: Proceedings of the international conference information EMBS international conference on biomedical health informatics
technology: new generations (BHI). IEEE, pp 79–82
2. Atzori L et al (2010) The internet of things: a survey. Computer 21. Lin WJ, Ma HP (2016) A physiological information extraction
networks 54(15):2787–2805 method based on wearable PPG sensors with motion artifact
3. Miorandi MD (2012) Internet of things: vision, applications and removal. In: 2016 IEEE international conference on communica-
research challenges. Ad Hoc Networks 10(7):1497–1516 tions (ICC), pp 1–6
4. Anzanpour A et al (2017) Self-awareness in remote health 22. Hara S et al (2017) Parameter optimization of motion artifact
monitoring systems using wearable electronics. In: Proceedings of canceling PPG-based heart rate sensor by means of cross
design and test europe conference (DATE), Lausanne, Switzerland validation. In: 2017 11th international symposium on medical
5. Parego P, Rahmani AM, Taherinejad N (eds.) (2017) Wireless information and communication technology (ISMICT), pp 73–
communication and mobile healthcare. Lecture Notes of the Insti- 76
tute for Computer Sciences, Social Informatics and Telecommu- 23. Bacá A et al (2015) CARMA: a robust motion artifact reduction
nications Engineering. Springer International Publishing algorithm for heart rate monitoring from PPG signals. In:
6. Yin H, Akmandor AO, Mosenia A, Jha NK (2018) Smart 2015 23rd European signal processing conference (EUSIPCO),
healthcare. Foundations and Trends. In: Electronic design pp 2646–2650
automation, vol 2(14), pp 401–466
7. TaheriNejad N, Pollreisz D (2016) Assessment of physiological
signals during happiness, sadness, pain or anger. In: 6th Publisher’s Note Springer Nature remains neutral with regard to
MobiHealth jurisdictional claims in published maps and institutional affiliations.