An Implementation of Motion Artifacts Elimination for PPG Signal

Processing Based on Recursive Least Squares Adaptive Filter

Chih-Chin Wu, I-Wei Chen, and Wai-Chi Fang, Fellow, IEEE
Department of Electronics Engineering, National Chiao Tung University, 1001 University Road, Hsinchu, Taiwan (R.O.C.)
[email protected]

PPG signal, signal processing methods have to be introduced



Abstract—In Photoplethysmographic (PPG) signals analysis,

the accuracy and stability are highly affected by Motion to circumvent the MAs. [1][2]
Artifacts (MAs) disturbances. In this paper, we adopt an
adaptive and efficient approach based on the developed DC There are some methods have been applied on the PPG
Remover method and Recursive Least Squares (RLS) adaptive signal processing for MAs reduction, such as Fast Fourier
filter for reducing MAs from PPG signals in real time. The Transform (FFT) Filter, Discrete Wavelet Transform (DWT),
experimental results of this work show a high correlation and Least Mean Square (LMS) method, which is a kind of
coefficient between Electrocardiography (ECG)-derived heart adaptive filter algorithm.
rate and PPG-derived heart rate, which is higher than 0.8504 of
FFT Filter retains that main cardiac portion frequency (0.5-
the R value, a high agreement by Bland-Altman analysis in the
4 Hz) and respiratory activity (0.2-0.35 Hz) of PPG and sets
limits of agreement represent the 95% confidence interval and
the standard deviation is 3.81 BPM (Beats Per Minutes). An
other component frequency to zero [3]. However, the
overall PPG signal with higher signal quality is obtained. frequency ranges of MAs are not always fixed, FFT Filter will
Further, the precision of heart rate calculated by PPG is not be applicable if the MAs are in the component frequency
improved. of PPG.
DWT is used to decompose the PPG signals into low and
Keywords—Photoplethysmography ; DC Remover ; Recursive high frequency components of same lengths. Then it
Least Squares filter ; Motion Artifact ; adaptive filter reconstructs to the noise free PPG signal based on both low
and high frequency components. But the mother wavelets are
I. INTRODUCTION determined in advance, it will lead to loss of certain important
physiological information because of the non-stationary and
Photoplethysmography (PPG) has been widely used for non-linear characteristic [4].
clinical setting and home care due to low-cost, non-invasive,
long time monitoring, and easy operation, and can get the two The adaptive filter algorithm LMS assumes that the
important information of vital signs: respiration and heart rates. reference noise received from the accelerometer is statistically
PPG is a simple and inexpensive optical technique that can be correlated with the MA component in the corrupted PPG
used to detect blood volume changes in the microvascular bed signal, while the MA is uncorrelated with the noise-free PPG
of tissue [1]. The arterial pulse in the finger changes the blood signal. A cost function (also termed error signal) is used to
perfusion, which causes a change in optical properties adjust the filter coefficients (or named filter’s weights)
especially in the optical intensity detected after optical continuously in order to improve the signal quality of the
stimulation. PPG is a non-invasive method and can be noise-corrected PPG signal [5].
measured from the surface of the human body, it will not cause This study adopts an adaptive and efficient approach based
any physical harm to subjects. According to the absorption and on the developed DC Remover method and Recursive Least
transmission characteristics, the intensity of red light and Squares (RLS) adaptive filter to reduce MAs in distorted PPG
infrared light passing through the finger will relate to the signals.
changes in the blood volume. PPG signal contains two major
components, which are pulsatile (AC) component and non- DC Remover is a kind of IIR filter design that is often used
pulsatile (DC) component. The AC component is caused by to remove the DC bias interference contained in the signal. The
the change of pulsatile blood volume and has its fundamental purpose of the DC Remover digital filter is to extract the low-
frequency, depending on heart rate. The DC component frequency voltage offset (DC) of PPG signals. RLS is also an
represents the background absorption which is the component adaptive digital filter that can automatically adjust the filter
without a pulsatile signal [2]. Using PPG to monitor cardiac parameters according to the cost function, and compared to the
activity is a low-cost and a convenient method because the LMS algorithm, the RLS algorithm has a faster convergence
light source and phototransistors are easily embedded in a speed and lower error.
wearable system. Therefore, PPG has been widely developed II. SYSTEM DESCRIPTION
in biomedical research and applications.
To improve the overall accuracy, we combine two
However, PPG signal is easily contaminated by Motion algorithms, a DC Remover and a Recursive Least Squares
Artifacts (MAs), interferences caused by subject's movements. adaptive filter. Our work implements the integrated system
In order to track a subject's heart rate by analyzing the sampled

978-1-5090-5803-7/17/$31.00 ©2017 IEEE

with reflection PPG front end Max30100, the analog front-end
circuit, ADXL355, the 3-axis accelerometer, and the Bluno
Microcontroller Unit. The Max30100 is composed of
optimized optics, it combines two LEDs (Red/IR 660/880), a
photodetector, and a 16-bit sigma delta ADC with 125 Hz
sample rate. We extract the raw PPG signals through the
analog front-end circuit and the 3-axis motion signals as the
reference noises through the ADXL335 simultaneously, and
sent them to the Bluno Microcontroller. Fig. 1 describes the
detail data flow of the proposed system.

Fig. 2 DC Remover: The amplitude response under

different parameters of a.

B. Recursive Least Squares Adaptive Filter (RLS)

1).RLS Algorithm
The adaptive filter can use feedback to adjust the filter
coefficients, and the frequency response, as well as the process
involve using a cost function (or error signal) to determine how
the filter coefficients are changed so as to reduce the cost of
the next iterative process of the operation [7].

Fig. 1 A detail data flow of the proposed motion artifacts

elimination system.
Fig. 3 RLS Algorithm Block Diagram
A. DC Remover
In this work, the DC bias in PPG signals is troublesome. The RLS algorithm block diagram is shown in Fig. 3, the
Because the MAs corrupted on PPG signals are mainly formula is as follows:
concentrated in the AC component of the reflection PPG [6],
we apply DC Remover method to separate the DC and AC ( )=∑ = (1.4)
components of the PPG signals, and receive the AC
In the RLS algorithm, the two input signals are primary
component of the PPG signals. The operation formulas are as
input, s(t) and reference input, xt . Those two inputs
follows:
respectively represent the PPG signal with motion noise (AC
= (t) + × ( − 1) (1.1) component), and the accelerometer signal is measured by the
3-axis accelerator parallel to the direction of the finger's blood
( )= ( )− ( − 1) (1.2) flow. This directional acceleration is selected as the reference
signal in that it has a high correlation with the motion noise on
( ) the finger PPG signal [6]. The s1(t) represents the noise-free
( )
= (1.3) PPG signal, and n(t) represents the action noise.
In the equation (1.4), ωk represents the coefficients of the
The PPG sampling point input for each time is represented RLS filter; θ is the n*1 matrix composed of these coefficients
by x(t) in the above equation (1.1). The w(t) is the calculated θ = [ω1, ω2 ... ωn ]T ; φt is the n*1 matrix of acceleration signals
value of the operation process, which will record the DC drift in the x-axis direction φt=[x(t-n),x(t-n+1)…xt]T ; t is the
of the PPG signal in the equation (1.2). The value a is an sampling time. Through the filter coefficient matrix θ and the
operation parameter for controlling the filter cutoff band range X-axis acceleration signal matrix φt, the algorithm performs
and the response speed, and the range is 0 to 1. It can be seen matrix operation to obtain the estimated motion noise . And
from the transfer function (1.3) that the value of a must be less then subtracting the estimated motion noise from the
than 1, and if it is 1, the filtering effect is lost. When a is closer original PPG signal s(t) to obtain the clean PPG signal s'(t)
to 1, the DC cutoff frequency band becomes narrow, and the with respect to the original PPG signal s(t). After several
response speed is slow. Conversely, the DC cutoff frequency experiments, we find that the effect of removing the MAs will
band is wide, and the response speed is fast. As shown in Fig. be saturated when the selected matrix θ is 10 orders or even
2. higher ones, so we set θ to 10 for our PPG signal processing.
2). Updating the Filter Coefficients estimated HR values derived from PPG are very close to the
one derived from ECG when MAs are faint.
After the RLS algorithm recursively calculates the
estimated motion noise each sampling time, the Then the degree of agreement between two datasets is
coefficients in the filter matrix are updated automatically. estimated using Bland-Altman analysis (Fig. 6). In the Bland-
After this operation, the estimated motion noise obtained Altman plot, the corresponding limits of 95% confidence
next time is closer to the actual motion noise. The operations interval of PPI and RRI is [-0.024, 0.024] second.
are shown in equations (1.5) and (1.6).
= + ( − ) (1.5)

= (1.6)

= − (1.7)

In equation (1.5), the new filter coefficient matrix θt will be

the product of the gain vector à and the original coefficient
matrix θ(t-1) plus the cost function . Pt is the correlation
inverse matrix of the input signal φt produced by the process
of calculating the gain vector, as shown in (1.7).
III. EXPERIMENTAL RESULTS
Fig. 5 The Correlation Coefficient analysis between beat-
In order to verify the efficiency of the proposed algorithm, to-beat changes of PPG and ECG from 5 groups MIMIC
two databases are used in this experiment. First one is MIMIC onlineġdatabase. The fitted line is Y=0.99X+0.0046 where X
(Multi-parameter Intelligent Monitoring for Intensive Care) signifies the ECG RRI values, and Y indicates PPG PPI
online database [8] of 5 groups of PPG and ECG sampled valuesļġ the R value is 0.99132 which expresses that the
simultaneously at 125Hz (without MAs), Second is the PPG
estimated HR derived from PPG is very close to the one
signal measured from the 10 subjects by the experimental
system of the 125 Hz sampling rate with MAs and X-axis derived from ECG when MAs are faint.
acceleration signal (parallel to the vector of the finger's blood
flow) and their heart rate collected by the CSI/CRITICARE
8100e nGenuity Bedside Monitor at the stationary chest to be
the ground-truth. The MAs of this experiment set the actions
type as scratches and slight shakings.
The effect of the DC Remover applied on the PPG signal
is showed on the Fig. 4.

Fig. 6 The Bland-Altman agreement analysis between

beat-to-beat changes of PPG and ECG from 5 groups MIMIC
onlineġdatabase. The corresponding limits of 95% confidence
interval is [-0.024, 0.024] second and the standard deviation
Fig. 4 The effect of the DC Remover applied on the PPG is 0.0102s.
signal. The blue line is the PPG rawdata (AC and DC
components), The red line is the processed PPG signal (AC Fig. 7 shows the mean PPG signals of 10 subjects with MAs
component) which is separated from non-constant DC (scratches and slight shakings) in the frequency domain
component. analysis by the FFT. In the Fig. 8b, the frequency component
of MAs appears on the intrinsic frequency of PPG. After the
Fig. 5 shows that the Correlation Coefficient analysis of RLS adaptive filter, the frequency components of MAs are
the 5 MIMIC datasets of PPG-derived PP intervals (PPI) and removed and the PPG characteristic frequency components
ECG-derived RR intervals (RRI), and the result is highly remain.
correlated (R=0.99132) which indirectly means that the
 Fig. 9 The Bland-Altman agreement analysis of the second
database from 10 subjects. The corresponding limits of 95%
confidence interval is [-4.29, 4.26] BPM and the standard
deviation is 3.81 BPM.

IV. CONCLUSION
This paper has presented an efficient algorithm based on
the combination of DC Remover and RLS adaptive filter for
eliminating motion artifacts corrupted on raw PPG signals in
real time. Two kinds of the database were used to test the
algorithm. The results show high correlation, the correlation
coefficient R is higher than 0.8504, and high consistency, the
agreement of the corresponding limits of 95% confidence
interval is [-4.29, 4.26] BPM and the standard deviation is 3.81
Fig. 7 (a) The frequency spectrum of the mean PPG signal BPM. Moreover, all the methods of signal processing are
after the RLS adaptive filter; (b) The frequency spectrum of based on real time design, and have been implemented in
the mean MAs-corrupted PPG signal. hardware design for wearable and home-care applications. It
makes the at-home care possible and lowers the rate of
The Fig. 8 shows the Correlation Coefficient analysis of cardiovascular diseases and medical expenses through a long-
heart rates from the second database with PPG’s MAs after the term monitoring.
RLS processing, the result presents high correlation
(R=0.8504). Fig. 9 is the Bland-Altman plot of heart rates from ACKNOWLEDGMENT
the second database, and the corresponding 95% limits of heart This work was supported in part by the National Science
rate and ground-truth heart rate is [-4.29, 4.26] BPM and the Council of Taiwan, R.O.C., under grant NSC 103-2221-E-
standard deviation is 3.81 BPM. 009 -214. The authors would also like to express their sincere
appreciation to the National Chip Implementation Center for
their chip fabrication and testing service.

REFERENCES
[1] Pan, Huijie, Dogancan Temel, and Ghassan AlRegib. "HeartBEAT:
Heart beat estimation through adaptive tracking." Biomedical and
Health Informatics (BHI), 2016 IEEE-EMBS International Conference
on. IEEE, 2016.
[2] Z. Zhang et al., “Troika: A general framework for heart rate monitoring
using wrist-type photoplethysmographic signals during intensive
physical exercise,” IEEE Trans. Biomed. Eng., vol. 62, no. 2, pp. 522-
531, 2015.
[3] K. A. Reddy, B. George, and V. J. Kumar, “Use of Fourier series
analysis for motion artifact reduction and data compression of
photoplethysmographic signals,” IEEE Trans. Instrum. Meas., vol. 58,
no. 5, pp. 1706–1711, May 2009.
[4] J. J. Liao, S. Y. Chuang, C. C. Chou, C. C. Chang and W. C. Fang, "An
effective photoplethysmography signal processing system based on
Fig. 8 The Correlation Coefficient analysis of the second EEMD method," VLSI Design, Automation and Test(VLSI-DAT),
database from 10 subjects.ġ The R value is 0.8504 which Hsinchu, 2015, pp. 1-4.
presents the high correlation of the heart rate (calculated by [5] G. Comtois, Y. Mendelson and P. Ramuka, "A Comparative Evaluation
of Adaptive Noise Cancellation Algorithms for Minimizing Motion
PPI) and ground-truth. Artifacts in a Forehead-Mounted Wearable Pulse Oximeter," 2007 29th
Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, Lyon, 2007, pp. 1528-1531.
[6] Jiang, Honghui. Motion-artifact resistant design of
photoplethysmograph ring sensor for driver monitoring. Diss.
Massachusetts Institute of Technology, 2003
[7] Y. Ye, Y. Cheng, W. He, M. Hou and Z. Zhang, "Combining Nonlinear
Adaptive Filtering and Signal Decomposition for Motion Artifact
Removal in Wearable Photoplethysmography," in IEEE Sensors
Journal, vol. 16, no. 19, pp. 7133-7141, Oct.1, 2016.
[8] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh,
Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE. PhysioBank,
PhysioToolkit, and PhysioNet: Components of a New Research
Resource for Complex Physiologic Signals. Circulation 101(23):e215-
e220 [Circulation Electronic Pages;
http://circ.ahajournals.org/content/101/23/e215.full]; 2000 (June 13).
PMID: 10851218; doi: 10.1161/01.CIR.101.23.e215