Hindawi

Journal of Healthcare Engineering

Volume 2019, Article ID 6283279, 7 pages
https://doi.org/10.1155/2019/6283279

Research Article
Noise-Robust Heart Rate Estimation Algorithm from
Photoplethysmography Signal with Low
Computational Complexity

JaeWook Shin and Jaegeol Cho

Department of Medical and Mechatronics Engineering, Soonchunhyang University, Asan, Republic of Korea

Correspondence should be addressed to Jaegeol Cho; [email protected]

Received 15 February 2019; Accepted 20 March 2019; Published 21 May 2019

Academic Editor: Patrick Boissy

Copyright © 2019 JaeWook Shin and Jaegeol Cho. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This paper introduces a noise-robust HR estimation algorithm using wrist-type PPG signals that consist of preprocessing block,
motion artifact reduction block, and frequency tracking block. The proposed algorithm has not only robustness for motion noise
but also low computational complexity. The proposed algorithm was tested on a data set of 12 subjects and recorded during
treadmill exercise in order to verify and compare with other existing algorithms.

1. Introduction this paper proposes a PPG sensor with low complexity and
an algorithm based on a 3-axis acceleration sensor to esti-
Recently, as interest in health increases, there is a growing mate the heart rate. To evaluate the performance of the
demand for users to continuously diagnose diseases or to proposed algorithm, we compared the performance of the
manage disease by measuring biosignals. In order to meet proposed algorithm with that of the existing algorithms.
the needs of users, wearable pace measurement devices
based on photoplethysmography (PPG) sensors have been 2. Methods
commercialized in many companies [1].
The PPG is a sensor that measures changes in blood In this paper, the algorithm consists of three stages in order
vessel contraction and expansion using LEDs and photo- to estimate the heart rate during exercise in the wearable
diodes. It can be used to measure the heart rate and oxygen device based on the PPG sensor. The first stage is pre-
saturation in a noninvasive manner and is widely used in processing the input PPG sensor data and 3-axis acceleration
wearable devices. However, in case of PPG sensor signals in data. The second stage is to remove the MA noise from the
wearable devices, the heart noise estimation error may be PPG sensor signal. The last stage is the frequency tracking to
caused by motion artifact (MA) due to body movements. estimate the heart rate in the motion-free signal. The
Various algorithms have been developed to overcome this flowchart of the proposed algorithm is shown in Figure 1.
problem. Conventional algorithms mainly use PPG sensor
signals of different wavelengths to remove motion noise 2.1. Data Set. In this paper, we tested the proposed heart rate
from the PPG sensor signals or effectively remove motion estimation algorithm using 12 data sets in the IEEE Signal
noise using acceleration sensor signals and measure heart- Processing Cup 2015 database [2]. We compared the heart
beat [2–8]. rate with the output of the ECG signals based on the data set.
However, the existing algorithms use various signals and
use complex algorithms. Therefore, the existing algorithms
are difficult to use in wearable devices with constraints of 2.2. Normalized Least Mean Squares Algorithm. The nor-
price, power, and system size. To overcome these drawbacks, malized least mean squares (NLMS) algorithm is widely used
2 Journal of Healthcare Engineering

Previous HR

IIR band-
PPG signal Preprocessing: Motion artifact
pass filter
down-sampling reduction:
Accelerometer signals band-pass filtering adaptive filter

HR estimation:
Final HR in BPM adaptive frequency
tracking

Figure 1: Block diagram of the proposed algorithm.

because it has a simple calculation among various adaptive Preprocessed PPG

filters and ease of implementation [9, 10]. We consider data
d(n) derived from an unknown system: ACC X-axis Wx
T +
d(n)  u (n)w + v(n), (1) –
ACC Y-axis Wy Σ Σ Clean PPG
where w is an unknown system that we expect to estimate,
u denote the input vector, v accounts for the measurement ACC Z-axis Wz
noise, and n is the iteration number. Assume that the
unknown system order is M, w and u are the M-di- Figure 2: Adaptive filter for motion artifact reduction.
mensional column vectors. The coefficient w(n)  of the
adaptive filter is updated using the difference e(n) between sign ux (n)
the adaptive filter output signal y(n) and the desired w x (n − 1) + μ
x (n)  w sign(e(n)),
sign uTx (n)sign ux (n)
signal d(n) for the input signal u(n) so that the square
mean error is minimized. The NLMS algorithm can be (3)
expressed as
signuy (n)
  − 1) + μ T
u(n) y (n)  w
w y (n − 1) + μ sign(e(n)),
signuTy (n)signuy (n)
w(n)  w(n e(n),
u (n)u(n)
(2)
(4)
 − 1),
e(n)  d(n) − y(n)  d(n) − uT (n)w(n
sign uz (n)
where μ is the step size, 0 < μ ≤ 1. z (n)  w
w z (n − 1) + μ sign(e(n)),
sign uTz (n)sign uz (n)
(5)
2.3. Low Computational Complexity MA Reduction
Algorithm. The PPG signal includes noise-free signals and e(n)  d(n) − y(n), (6)
the MA that is generated due to the movement of the body.
Because MA is highly correlated with the acceleration sensor y(n)  signuTx (n)
wx (n − 1) + signuTy (n)
wy (n − 1)
signals, a clean PPG Signal can be obtained to remove a
signal having a high correlation with the acceleration sensor + signuTz (n)
wz (n − 1),
from the PPG signal. Therefore, the corrupted PPG signals (7)
are used as desired signal d(n) and 3-axis accelerometer
signals are used as input signal u(n) to reduce the MA as where sign(·) denotes the sign function and ux , uy , and uz
shown in Figure 2. denote x-axis, y-axis, and z-axis accelerometer signal,
The conventional NLMS algorithm requires 3M + 1 respectively.
multiplication when the order of the adaptive filter is M. Due to use of only the sign of the input signal vector and
Despite the small computational complexity of the NLMS, the error, the proposed algorithm requires only M multi-
an algorithm with a small computational complexity is re- plications because the multiplication required in Equations
quired for wearable systems due to price, power, and system (3)–(5) and (7) can be calculated by adding. Therefore, the
size limitations. In order to overcome this drawback, we algorithm can be implemented with a small amount of
propose an adaptive noise cancellation algorithm which can computation compared to the existing NLMS algorithm. In
have similar performance with low computational com- particular, calculation time can be further shortened for
plexity as follows: using a CPU without a floating point unit.
Journal of Healthcare Engineering 3

2.4. Adaptive Frequency Tracking. We used an oscillator- The output of IIR band-pass filter dhr (n) is used as the
based adaptive notch filter (OSC-ANF) algorithm [11] to desired signal for adaptive filter instead of d(n) in the MA
estimate the heart rate using the PPG signal that passed reduction step.
through the MA reduction stage. The OSC-ANF algorithm is
based on a second-order IIR band-pass filter and traces the
strongest frequency of the signal. The OSC-ANF algorithm 2.6. Performance Measurement. To verify the performance
operates as follows: of the proposed algorithm, 12 data sets of IEEE Signal
Processing Cup 2015 database were used. The data set
􏽢 (n)(1 + β)x(n − 1) − βx(n − 2)
x(n) � α used provides the reference heart rate measured from the
electrocardiogram as well as the PPG sensor signal and
+ 0.5(1 − β)(sign(e(n)) − sign(e(n − 2))), the acceleration sensor signal. To compare the performance
of the algorithm, we used the two methods that average
eα (n) � x(n) − 2􏽢α(n)x(n − 1) + x(n − 2), absolute error and average absolute error percentage as
follows:
Px (n) � 1 − μa 􏼁Px (n − 1) + μa x2 (n − 1), 1 N 􏼌􏼌􏼌 􏼌􏼌
Error1 � 􏽘 􏼌HRest (n) − HRtrue (n)􏼌􏼌,
N n�1
x(n − 1) 􏼌􏼌 􏼌􏼌 (11)
􏽢 (n + 1) � α
α 􏽢 (n) + μa e (n),
2Px (n) α 1 N 􏼌􏼌HRest (n) − HRtrue (n)􏼌􏼌
Error2 � 􏽘 .
N n�1 HRtrue (n)
ω(n + 1) � cos−1 (􏽢α(n + 1)),

fs 3. Results and Discussion

HRest (n + 1) � ω(n + 1) × × 60,
2π 3.1. Parameter Settings. In order to reduce the computa-
(8) tional complexity, we use down-sampled PPG and accel-
where ω(n + 1) is the estimated frequency, HRest (n + 1) erometer signal that are resampled 125 Hz to 25 Hz. Figure 3
is the estimated HR in BPM, fs is the sampling rate, μa is shows the average absolute error of the proposed algorithm
the step size, and β controls the 3 dB bandwidth of the 2nd with various filter tap lengths which used MA reduction step.
order IIR band-pass filter. As can be seen, the proposed algorithm has best perfor-
mance when the adaptive filter order is 21 (M � 21). Pa-
rameter setting of the proposed algorithm is summarized in
2.5. Noise-Robust Adaptive Frequency Tracking. To improve Table 1.
the tracking performance of the OSC-ANF algorithm under
highly noisy environments, we propose the noise-robust 3.2. Performance of the Proposed Algorithm. In this paper, we
OSC-ANF (NR-OSC-ANF) algorithm that is derived by verified the performance of the proposed heart rate esti-
noise-robust adaptive filter concept [12, 13] as follows: mation algorithm using 12 data sets in the IEEE Signal
1 L−1 Processing Cup 2015 database. Error1 and Error2 were
α(n) � 􏽘α 􏽢 (n − l), obtained for each set and compared with other algorithms by
L l�0
comparing the heart rate output through the three-stage
algorithm and the ECG signal-based heart rate provided by
eα (n) � x(n) − 2α(n)x(n − 1) + x(n − 2), the data set. Figure 4 shows that the proposed algorithm can
(9) sufficiently remove motion artifacts even with low com-
2 putational complexity. Figure 5 is the HR tracking results
Px (n) � 1 − μa 􏼁Px (n − 1) + μa x (n − 1),
plot on test data set 08 and set 09 with ECG-based HR. The
x(n − 1) estimated HR form PPG signal matches with ECG-based HR
􏽢 (n + 1) � α(n) + μa
α e (n). satisfactorily.
2Px (n) α
Tables 2 and 3 show that the performances of other
By using the average of the past estimated frequencies, existing algorithms and the proposed algorithm do not differ
the NR-OSC-ANF algorithm makes improved frequency greatly. Although the proposed algorithm does not have best
tracking performance in low signal-to-noise ratio (SNR) performance compared with other algorithms, it is con-
environments. sidered to be worthy of an algorithm for use in a wearable
In addition, to improve MA reduction performance, we device because of its low computational complexity. The
further use IIR band-pass filter, the preprocessed PPG signal proposed algorithm requires only few multiplication for
by estimated α􏽢 (n + 1), as follows: preprocessing and NR-OSC-ANF.
Figure 6 shows Bland–Altman plot for the training data
􏽢 (n) 1 + βhr 􏼁d(n − 1) − βhr d(n − 2)
dhr (n) � α set. In this case, the limits of agreement were [−3.97, 5.04]
(10)
+ 0.5 1 − βhr 􏼁(d(n) − d(n − 2)). BPM. Figure 7 indicates the scatter plot between the ground
4 Journal of Healthcare Engineering

18
16

Average absolute error

14
12
10
8
6
4
2
0
05 10 15 20 25 30 35 40
Filter order
Figure 3: Filter order selection.

Table 1: Parameter setting.

Algorithm Parameters
MA reduction algorithm M  21, μ  0.0001
NR-OSC-ANF L  5, β  0.95, μa  0.025
IIR band-pass filter βhr  0.8

PPG signal ACC signal (Y-axis)

200 200
Heart rate (BPM)

Heart rate (BPM)

150 150

100 100

50 50

0 0
50 100 150 200 250 50 100 150 200 250
Time (sec) Time (sec)
(a) (b)
e(n) Sign(e(n))

200 200
Heart rate (BPM)

Heart rate (BPM)

150 150

100 100

50 50

0 0
50 100 150 200 250 50 100 150 200 250
Time (sec) Time (sec)
(c) (d)
Figure 4: Continued.
Journal of Healthcare Engineering 5

x(n)

200

Heart rate (BPM)

150

100

50

0
50 100 150 200 250
Time (sec)
(e)

Figure 4: Frequency spectrogram of various signals.

160 160
150 150
140 140
Heart rate (BPM)

Heart rate (BPM)

130 130
120 120
110 110
100 100
90 90
80 80
70 70
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time (sec) Time (sec)
ECG HR ECG HR
PPG HR PPG HR
(a) (b)

Figure 5: HR estimation results for (a) set 08 and (b) set 09.

Table 2: Error1 results of the proposed algorithm and the existing algorithms.
Data set TROIKA [2] JOSS [3] NLMS + OSC-ANFc [7] Combination of adaptive filters [8] Proposed algorithm
1 2.29 1.33 1.75 1.34 1.33
2 2.19 1.75 1.94 0.70 1.92
3 2.00 1.47 1.17 0.66 0.83
4 2.15 1.48 1.67 0.70 1.03
5 2.01 0.69 0.95 0.63 0.54
6 2.76 1.32 1.22 0.86 1.44
7 1.67 0.71 0.91 0.66 0.65
8 1.93 0.56 1.17 0.58 0.56
9 1.86 0.49 0.87 0.52 0.43
10 4.70 3.81 2.95 2.46 2.51
11 1.72 0.78 1.15 1.21 0.83
12 2.84 1.04 1.00 0.74 1.79
Av. ± std 2.34 ± 0.79 1.29 ± 0.86 1.40 ± 0.58 0.92 ± 0.52 1.16 ± 0.62
6 Journal of Healthcare Engineering

Table 3: Error2 results of the proposed algorithm and the existing algorithms.
Data set TROIKA [2] NLMS + OSC-ANFc [7] Combination of adaptive filters [8] Proposed algorithm
1 1.90 1.59 1.17 1.06
2 1.87 1.99 0.70 2.18
3 1.66 1.02 0.57 0.72
4 1.82 1.51 0.63 0.97
5 1.49 0.75 0.49 0.41
6 2.25 1.05 0.67 1.23
7 1.26 0.72 0.50 0.50
8 1.62 1.04 0.50 0.50
9 1.59 0.76 0.46 0.38
10 2.93 0.93 1.56 1.59
11 1.15 0.79 0.80 0.57
12 1.99 0.79 0.55 1.21

25 Conflicts of Interest
Estimated HR-ground truth HR

20
15 The authors declare that there are no conflicts of interest
10 regarding the publication of this paper.
5.04 (+1.96SD)
5
0.54
0
–5
–3.97 (–1.96SD) Acknowledgments
–10
–15 This work was supported by the Soonchunhyang University
–20 Research Fund (No. 20170809). This research was also
–25
40 60 80 100 120 140 160 180 200 supported by Basic Science Research Program through the
Mean ground truth HR and estimated HR National Research Foundation of Korea (NRF) funded by
the Ministry of Education (NRF-2018R1D1A1B07050427).
Figure 6: Bland–Altman plot.

180
References
y = 0.9953x + 1.178 [1] D. J. Terbizan, B. A. Dolezal, and C. Albano, “Validity of seven
160
r 2 = 0.9911
Estimated HR (BPM)

commercially available heart rate monitors,” Measurement in

140 Physical Education and Exercise Science, vol. 6, no. 4,
120 pp. 243–247, 2002.
[2] Z. Zhang, Z. Pi, and B. Liu, “TROIKA: a general framework for
100 heart rate monitoring using wrist-type photoplethysmo-
80 graphic signals during intensive physical exercise,” IEEE
Transactions on Biomedical Engineering, vol. 62, no. 2,
60 pp. 522–531, 2015.
60 80 100 120 140 160 180
[3] Z. Zhang, “Photoplethysmography-based heart rate moni-
Ground truth HR (BPM) toring in physical activities via joint sparse spectrum re-
Figure 7: Scatter plot. construction,” IEEE Transactions on Biomedical Engineering,
vol. 62, no. 8, pp. 1902–1910, 2015.
truth HR and estimated HR. The fitted line was [4] D. Dao, S. M. A. Salehizadeh, Y. Noh et al., “A robust motion
y  0.9953x + 1.178, where x is the ground truth HR and y is artifact detection algorithm for accurate detection of heart
the estimated HR. rates from photoplethysmographic signals using time-
frequency spectral features,” IEEE Journal of Biomedical
and Health Informatics, vol. 21, no. 5, p. 1, 2017.
4. Conclusions [5] Y. Fujita, M. Hiromoto, and T. Sato, “PARHELIA: particle
This paper presents a noise-robust HR estimation algorithm filter-based heart rate estimation from photoplethysmo-
graphic signals during physical exercise,” IEEE Transactions
using PPG signals that have not only robustness for motion
on Biomedical Engineering, vol. 65, no. 1, p. 1, 2018.
noise but also low computational complexity. In order to [6] D. Zhao, Y. Sun, S. Wan, and F. Wang, “SFST: a robust
verify the performance of the proposed heart rate estimation framework for heart rate monitoring from photo-
algorithm, we compared with other existing algorithms plethysmography signals during physical activities,” Bio-
using the IEEE Signal Processing Cup 2015 database. medical Signal Processing and Control, vol. 33, pp. 316–324,
2017.
Data Availability [7] S. Fallet and J.-M. Vesin, “Robust heart rate estimation using
wrist-type photoplethysmographic signals during physical
The data used to support the findings of this study are exercise: an approach based on adaptive filtering,” Physio-
available from the corresponding author upon request. logical Measurement, vol. 38, no. 2, pp. 155–170, 2017.
Journal of Healthcare Engineering 7

[8] K. R. Arunkumar and M. Bhaskar, “Heart rate estimation

from photoplethysmography signal for wearable health
monitoring devices,” Biomedical Signal Processing and Con-
trol, vol. 50, pp. 1–9, 2019.
[9] A. Sayed, Fundamentals of Adaptive Filtering, Wiley, New
York, NY, USA, 2003.
[10] S. Haykin, Adaptive Filter Theory, Prentice-Hall, Upper Saddle
River, NJ, USA, 4th edition, 2002.
[11] H. Liao, “Two discrete oscillator based adaptive notch filters
(OSC ANFs) for noisy sinusoids,” IEEE Transactions on Signal
Processing, vol. 53, no. 2, pp. 528–538, 2005.
[12] Y. S. Choi, S. E. Kim, and W. J. Song, “Noise-robust nor-
malised subband adaptive filtering,” Electronics Letters,
vol. 48, no. 8, pp. 11-12, 2012.
[13] H. Zhao, Z. Zheng, Z. Wang, and B. Chen, “Improved affine
projection subband adaptive filter for high background noise
environments,” Signal Processing, vol. 137, pp. 356–362, 2017.
 International Journal of

Rotating Advances in
Machinery Multimedia

The Scientific
Engineering
Journal of
Journal of

Hindawi
World Journal
Hindawi Publishing Corporation Hindawi
Sensors
Hindawi Hindawi
www.hindawi.com Volume 2018 http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi
Engineering
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

VLSI Design
Advances in
OptoElectronics
International Journal of

International Journal of
Modelling &
Simulation
Aerospace
Hindawi Volume 2018
Navigation and
Observation
Hindawi
www.hindawi.com Volume 2018
in Engineering
Hindawi
www.hindawi.com Volume 2018
Engineering
Hindawi
www.hindawi.com Volume 2018
Hindawi
www.hindawi.com www.hindawi.com Volume 2018

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Hindawi Hindawi Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018