Vous êtes sur la page 1sur 5

Heart Rate Measurement from Wearable

Photoplethysmographic Sensor Using Spot and


Tracking Methods
P Madhan Mohan, V Nagarajan, Member, IEEE, Sounak Ranjan Das and Rajesh M
Abstract In this paper we try to explore the concepts of
Heart Rate Measurement using Photoplethysmographic (PPG)
optical sensor. There is a vast array of processing algorithms that
can be applied to the raw PPG signal to get the Heart Rate (HR).
But in most of the cases these signals are corrupted by motion
which makes it very difficult to extract the actual HR. In this
paper we find out how HR can be obtained in Real-Time for Spot
HR measurement and then move on to Motion HR. While doing
this we also try to use as less resources as possible in Spot HR
measurement. For Spot HR measurement we have taken
measurement from a PPG sensor while sitting steady for short
intervals since for Spot Measurement we need the signal only up
to a good quality. For Motion HR estimation we have generated
test data from subjects wearing the sensor and doing mild
exercises. During the experiments we obtained a success rate of
around 98% with our test data.
KeywordsPhotoplethysmography; Heart Rate Measurement;
Filters; Motion artifact removal; Wearable sensors, Signal
Decomposition

I.

INTRODUCTION

Wearable electronics is growing into modern electronics


industry at a very fast rate. Wearable devices like smart
watches, fitness bands, head mounted displays etc. are
flooding the consumer electronics section. One of the major
wearable segments is the Heart Rate Measurement. Heart Rate
is a key factor of the human body from which we can derive
various parameters like calories burnt, sleep quality, stress
level, respiratory functions and much more. Most wearable
Heart Rate Monitors (HRM) use the Photoplethysmography
[1], [2] sensor which is a non-invasive type HR sensor based
on the reflection or refraction of Infrared light from the skin.
Infrared is absorbed by the blood and hence the blood flowing
just below the skin through the arteries or veins absorbs the
light. This absorbed light is then detected by a photo detector.

P. Madhan Mohan , Research Scholar, Sathyabama University/Jasmin


Infotech Pvt. Ltd., Faculty of Electrical and Electronics, Chennai, Tamil
Nadu, India.
Dr V. Nagarajan Member, IEEE, Professor and Head, Adhiparasakthi
Engineering College, Melmaruvathur, Tamilnadu, India.
Sounak Ranjan Das, Student, M.Tech Embedded Systems, NIELIT,
Calicut, Kerala, India.
Rajesh M, Scientist/Engineer C, NIELIT Calicut, Kerala, India

Based on the absorption of the infrared light the photo detector


gives an analog voltage. This voltage is then passed through
the analog backend circuits through some low level analog
filters and an ADC to give digital outputs. In most of the
modern fitness bands they use a wrist type PPG sensor where
the sensor is placed on top or below the wrist and is integrated
on to the band.
In order to accurately determine the HR the signal needs to be
cleaned and enhanced to a great extent. Most of the modern
research in this area is going on in processing the signal to get
a clean waveform. The process can be divided into five basic
steps i.e.

Baseline Removal

Band-pass Filtering

Signal Decomposition

Motion Artifact Removal & Reconstruction

HR estimation and tracking

For the time being we will neglect steps 3 and 4 since we will
deal with only Spot HR detection. Later in our discussion we
will include these steps to see how our algorithm performs in
Motion HR detection.
Normally human HR varies from 50 to 250 Beats Per Minute
(BPM) [17], which is equal to 0.8 to 4.17 Hz. Hence we will
select the bands of our filters in this range. Although the
baseline removal can be ignored as it will be done by the
band-pass filter but we are keeping it for an extra layer of
enhancement. The baseline removal is nothing but a low pass
filter to get rid of the dc and the baseline components of the
signal. After these two steps the signal will be pretty much
ready for HR extraction unless it is corrupted by motion. So
we will limit our resources like window time, FFT points, etc
to see what are the minimum resources required to give
acceptable HR estimation in Real-Time.
In the course of developing the algorithm we faced a lot of
challenges. Since we are developing the algorithm for RealTime implementation we had to use efficient and simple
algorithm that require less computation. We have gone
through some research papers which suggest some complex
algorithms involving complex matrix operations like SVD [6],
wavelet denoising, EMD [8][9] etc. But when we tried to
implement this algorithm it took quite a bit of more time than
the ones we used in our algorithm.

I.

BLOCK DIAGRAM OF ALGORITHM


M

Fig. 1. Shows the block diagram of the com


mplete algorithm
along with motion artifact removal. As m
mentioned in the
introduction we have first set a time winddow on the PPG
signal which will be used for processing. Affter it is done we
will move to the next window. This is folloowed by baseline
removal and filtration which we will discusss in the coming
sections. For motion artifact removal wee have used an
adaptive NLMS filter. After all the filtration H
HR is estimated

Fig. 1. Block Diagram of the Algorithm

using FFT method and it is send for peak traccking. This is how
our algorithm works and the following secttions will explain
these steps one by one.
II.

SPOT HR ESTIMATTION

A. The Problems and Motivations


The PPG signal consists of several componennts (Fig. 2.) due to
skin tissue, bones, capillaries, venous bloodd, arterial blood,
etc. Actual PPG signal may be obtained by subtracting these
components from the original signal.

Fig. 2. Components of a PPG Signaal

The next problem is the motion which m


may be caused by
complex hand or wrist movements which ccauses an air gap
between the sensor and the skin tissues. But running complex
motion algorithms every time will resultt in wastage of
processing power and time. In order to overcoome this problem
we decide to switch between algorithms bassed on the signal
quality. When the signal quality is above a certain limit we
may switch to Spot HR estimation. This mayy be given by the
signal SNR (Signal to Noise Ratio) or thee posture of the

wearer based on the accelerometer data. In this paper we will


consider the SNR of the signal as a measurement
m
of quality.
Several methods are proposed for the baseline removal and
band-pass filtering of the PPG signaal. Some of the methods are
like Median Filter [3], Kalman Filtter [4], frequency filtering,
FIR Filters, and several other techniques [5]. In our
experiment we experimented with all
a kinds of filters and we
chose a combination of these filterrs to obtain a good signal.
Since each method has its own draawback hence one method
alone was not sufficient to clean th
he signals. For example if
we only do moving average theree is a problem regarding
window selection. Since our signal varies
v
from 0.4Hz-4Hz it is
difficult to set a proper averaging window
w
for the filter. Next
is Frequency filtering. In this metho
od the FFT of the signal is
found and the undesired frequenciess are removed. But this can
be a problem at lower frequenciess because there may be a
chance of frequency leakage and the filter may distort the
signal by removing significant dataa. Hence we will use this
filter after we have sparsified the sig
gnal. Next is the FIR Filter
which is the most commonly used filter in these kinds of
applications. But for our application
n we need to design an FIR
Filter with steep cutoffs. Hence we make use of Kaiser
Window on our filter design. In the next section we will
discuss in detail about this method.
B. Baseline Removal and Band-passs Filtering
PPG signals comprise of DC and AC components. The DC
v
blood flow etc. But
components are due to skin tissues, venous
our main interest is the AC com
mponent. So the simplest
method to get rid of the DC com
mponents is to subtract the
signal by its own mean. This can be achieved by a process
called detrend in MATLAB. Fig. 3. Shows how the signal
looks after detrending. A suitable window
w
may be chosen for
averaging the signal. After detrend
ding the signal the entire
signal is now at zero level. Now we can apply various filters to
this signal. For spot HR a simple FIR
F filter works well. But
when the complexity of the signal increases in case of motion
artifacts we need to apply some exttra layers of pre processing
to enhance the signal. For now we will be using an FIR filter
with Kaiser Window.
In case of a normal FIR filter it is veery difficult to estimate the
ideal number of taps to get a good frequency
f
response. In this
case we needed a steep frequency cutoff. But our experiments
with 32 tap FIR filter showed som
me leakage and there was a
slight trend in the signal which resu
ulted in a peak at the lower
frequencies. Whereas the use of Kaaiser Window gave a very
good response and the frequency cutoffs were steep. One
drawback of this method is that therre is a phase lag and some
unwanted oscillations in the beeginning. These can be
overcome by a zero-phase forward and backward digital filter
but that will require the data size to be three times the number
of filter coefficients which is not a viable option for spot
signals. We will use it when we worrk with motion artifacts but
in case of spot signals it will be a waaste of time and processing
power.

window with good SNR. Once we get a good signal we get the
heart rate from (1). After displaying the heart rate we store it
for reference in the next window.
2) Peak Selection: This process starts after the first
window has been detected. The preceding windows use the
HR detected in the previous window and searches for a peak
in the vicinity since HR may change by a maximum of 10
BPM. Once the peak is detected it is displayed and the bin
location from FFT is stored for the next window.

(a)

(b)
Fig. 3. (a) Raw PPG signal (b) PPG signal after detrending.

C. HR Estimation and Tracking


Before Estimating the HR we need to do some more
preprocessing on the signal. As stated in [6] we need to
sparsify the signal in order to get each and every frequency in
a single frequency bin. Therefore
must be such that
/
60 must be less than or equal to 1. In order to do so
and keeping in mind the hardware constraints we decimate the
signal by a factor of 10 which brings sampling frequency
to
100/10 = 10. This allows us to use a 1024 point FFT which
makes each bin to contain 0.58 BPM.
After decimating the filtered signal a 1024 point FFT is
applied to the signal. From the FFT we choose the highest
peak to be the frequency corresponding to the heart rate. But
in some cases there may be some more peaks near the edge of
the spectrum which results in false HR selection. In order to
avoid this we do a frequency filtering to the FFT spectrum.
Any frequency outside the desired range is set to 0 and the
peak is estimated from the remaining spectrum.
Once the peak is obtained HR is given by the formula
60
Where,

Frequency peak position from FFT


Sampling frequency of the signal
Decimation factor
Number of FFT points.
For spot heart rate estimation this is enough to get the HR
but if it is for tracking HR measurement we need to do an
extra step that is tracking the HR. It is often noticed in some
cases that even after so much of filtering the heart rate is not
estimated properly and there is a huge difference in heart rate
from the previous measurement. According to study the
change in HR rarely exceeds 10 BPM within 2 seconds. Hence
in case of tracking we need to follow the trajectory of the heart
rates and select the peaks accordingly.
There are mainly three steps in tracking as mentioned in[]
i.e.
1) Initialization: This is the first window of HR estimation.
It is assumed that this window of PPG signal has a good PPG
waveform. We have modified this step a little bit. We
calculate the SNR of the filtered PPG signal and if the SNR is
good enough then we proceed or we wait until we get a

3) Verification: This process keeps track of the trajectory


of the heart rate i.e. the heart rate is increasing or decreasing
or whether the algorithm is missing any estimation. All in all it
verifies whether the heart rate is deviating from its path and
tries to correct it by increasing the search area in peak
selection.
III.

MOTION ARTIFACT REMOVAL

In this section we will briefly discuss about the motion artifact


removal from the PPG signal due to mild motions as this
section is still under research and we will be publishing our
proceedings in the future papers.
Motion artifacts are distortions caused in the signal due to
complex hand movements which results in a gap between the
sensor and the users skin. A lot of research is going on in this
field and some commonly used techniques are adaptive filters,
selective removal of frequency components by EMD [8][9],
wavelet, ICA, etc.
The adaptive filter technique is a very commonly used
technique which makes use of the accelerometer data. In our
experiment we make use of an NLMS [7] adaptive filter as
shown in Fig. 4. We use the accelerometer signal as the
reference noise signal and the filtered PPG data as the signal
corrupted with noise. It is to be noted that the accelerometer
data passes through all the pre processing steps as the original
PPG signal.

Fig. 4. Adaptive Noise canceller with reference noise.

Although this method removes some of the noise due to


motion we still need to improve on the process. Fig. 4. Shows
how an NLMS adaptive filter cleans up the signal using data
from the accelerometers.
Another method of removing motion artifacts is by signal
decomposition and sparse signal reconstruction. Some of the
common methods used in signal decomposition are Empirical

Mode Decomposition (EMD) [8], [9], Single-channel


Independent Component Analysis (SCICA) [10], [11],

are recorded for a short period of time like around 8 to 10


seconds each based on the signal quality.
The motion signals are taken with the same equipments and
the subject made simple hand movements like moving the
wrist up down, sitting standing up, moving hand sideways. In
our future proceedings we will be experimenting with more
complex motions like running and heavy exercise and try to
develop an algorithm to successfully determine the HR.

(a) Sampling Point

(b) Frequency Hz

(c) Sampling Point

B. Algorithm Settings
The sampling frequency of the sensor has been set to 100
Hz, the window size for the spot signals have been set to 5 sec.
and for motion signals it has been set to 9 sec with 1 sec step
interval. The down sampling factor for both the algorithms has
been set to 10. The search area or for HR tracking in motion
files have been set to 24. This number has been found by trial
and error method. Any value from 24 to 32 seemed to work
fine.
C. Performance Measurement
The performance of the algorithm is measured by
comparing the estimated HR with the reference HR and
deciding whether it is a pass or fail. The algorithm is a pass
if
100
10. After running the algorithm on the
.

entire database the number of pass cases is calculated and a


total success rate is calculated by the following
.
100.
formula
.

(d) Frequency Hz
Fig. 5. (a)Raw PPG Signal, (b) FFT Spectrum of Raw PPG signal, (c) PPG
signal after baseline removal and adaptive filtering, (d) FFT spectrum of
filtered signal with the point denoting the position of HR.

Singular
Spectrum
Analysis
[12],
[13],
wavelet
decomposition, etc. A reconstruction technique used in [6] is
Sparse Signal Reconstruction [14]-[16]. We experimented
with EMD, wavelet decomposition, ICA and SSA. But most
of the techniques have very heavy computational load and
some didnt give as good results as adaptive filtering. We tried
SSA and SSR and also the FOCUSS [17], algorithm as
described in [6] but it seems to have Singular Value
Decomposition (SVD) implemented inside it which is a very
time consuming computation. Even on an x86 architecture it
takes a decent amount of time to calculate SVD of an 8sec
long signal. So we decided to stick with adaptive filtering
using NLMS filter.
IV.

EXPERIMENTAL RESULTS

A. Data Recording
A single-channel PPG signal, a three-axis accelerometer and
signal from a Polar sensor has been used simultaneously to
record the signals. The Polar is used as a reference and been
used for analysis. For Spot HR measurements the subject was
made to sit steady without any significant motion and signals

D. Results
Fig. 6. Shows the result of the Spot HR estimation algorithm.
We were able to achieve a success rate of around 98% with
this algorithm. If we set the performance measurement limits
to less than 5% then we get a success rate of about 96%. There
are a total of 112 test files in this result.

Fig. 6. Estimation result for Spot HR measurement.

Fig. 7. Shows the result of motion HR estimation algorithm. It


is to be noted that the Spot Estimation gives one HR for each
case or file but the motion HR estimation needs a long signal
since it is a tracking type algorithm and so in the figure the x
axis denotes time for motion HR and test case for Spot HR.
We were able to achieve a success rate of about 90% which
we can see from the plot.

[7]

[8]

Fig. 7. Estimation results for Motion HR measurement.

V.

DISCUSSION

From the above experiments and observations it is evident


that Spot HR estimation can be accurately achieved by proper
usage of traditional filters like the FIR Filter. For HR
estimation with motion artifacts in the signal we need some
kind of algorithm that can properly model the noise from the
accelerometer signals. We plan to investigate into this area
further in our future work and try to bring out good and
accurate HR estimation in case of complex motions like
running or heavy exercise.
VI.

CONCLUSION

In this paper we propose an algorithm for Spot HR


estimation using wrist type PPG-sensor and compare the
results to more standard medical equipment like the Polar
finger type PPG-sensor. In our experiments we found our
algorithm to give a success rate of about 98%. Whereas in our
experiments with tracking HR estimation with the subject
performing mild hand motions we were able to achieve near
80% accuracy. We need to further improve our algorithm to
work perfectly for complex motions like running and heavy
exercise so that it can be implemented in Real-Time for smart
watch or fitness band like applications. We are also referring
some current papers and publications in this subject to get a
better understanding of the subject matters.

References
[1]

[2]

[3]

[4]

[5]

[6]

J. Allen, Photoplethysmography and its application in clinical physiological measurement, Physiological measurement, vol. 28, no. 3,
pp.R1R39, 2007
A. Kamal, J. Harness, G. Irving, and A. Mearns, Skin photoplethysmographya review, Computer methods and programs in biomedicine,
vol. 28, no. 4, pp. 257269, 1989.
Awodeyi, A.E., Alty S.R., Ghavami M., Median Filter Approach for
Removal of Baseline Wanderer in Photoplethysmography Signals,
Modelling Symposium(EMS), 2013 European, pp. 261-264, 2013.
Mneimneh, M.A., Yaz, E.E., Johnson, M.T., Povinelli, R.J. , An
adaptive kalman filter for removing baseline wandering inECG signals,
Computers in Cardiology, 2006, pp. 253-256, 2006.
Awodeyi, Adewale Emmanuel, Alty Stephen R., Ghavami Mohammad,
On Filtering of Photoplethysmography Signals, Bioinformatics and
Bioengineering (BIBE), 2014 IEEE International Conference, pp. 175178, 2014.
Zhilin Zhang, Zhouyue Pi, Benyuan Liu R. Nicole, TROIKA, A
General Framework for Heart Rate Monitoring Using Wrist-Type
Photoplethysmographic (PPG) Signals During Intensive Physical

[9]

[10]

[11]
[12]
[13]

[14]

[15]
[16]
[17]

Exercise, Biomedical Engineering, IEEE Transactions, vol. 62, Issue. 2,


pp. 522-531, 2015.
R. Yousefi, M. Nourani, S. Ostadabbas, and I. Panahi, A motiontolerant adaptive algorithm for wearable photoplethysmographic
biosensors, IEEE Journal of Biomedical and Health Informatics, vol.
18, no. 2, pp. 670681, 2014.
Kevin T Sweeney, Damien Kearney, Tom as E Ward, Shirley Coyle,
Dermot Diamond, Employing Ensemble Empirical Mode
Decomposition for Artifact Removal: Extracting Accurate Respiration
Rates from ECG Data during Ambulatory Activity, 35th Annual
International Conference of the IEEE EMBS Osaka, Japan, 3 - 7 July,
2013, pp. 977-980.
Hae-Kyung Jung, Do-Un Jeong, Development of Wearable ECG
Measurement System Using EMD for Motion Artifact Removal, ,
Computing and Convergence Technology (ICCCT), 2012 7th
International Conference, pp. 299-304, 2012.
C. J. James and D. Lowe, Extracting multisource brain activity from a
single electromagnetic channel, Artificial Intelligence in Medicine, vol.
28, no. 1, pp. 89104, 2003.
M. E. Davies and C. J. James, Source separation using single channel
ica, Signal Processing, vol. 87, no. 8, pp. 18191832, 2007.
N. Golyandina, V. Nekrutkin, and A. A. Zhigljavsky, Analysis of time
series structure: SSA and related techniques. CRC Press, 2001.
T. Harris and H. Yuan, Filtering and frequency interpretations of
singular spectrum analysis, Physica D: Nonlinear Phenomena, vol. 239,
no. 20, pp. 19581967, 2010.
I. F. Gorodnitsky and B. D. Rao, Sparse signal reconstruction from
limited data using FOCUSS: a re-weighted minimum norm algorithm,
IEEE Trans. on Signal Processing, vol. 45, no. 3, pp. 600616, 1997.
D. Donoho, Compressed sensing, IEEE Transactions on Information
Theory, vol. 52, no. 4, pp. 12891306, 2006.
M. Elad, Sparse and Redundant Representations: From Theory to
Applications in Signal and Image Processing. Springer, 2010.
Liangyou Chen, Andrew T. Reisner, and Jaques Reifman, Automated
Beat Onset and Peak Detection Algorithm for
Field-Collected
Photoplethysmograms, 31st Annual International Conference of the
IEEE EMBS Minneapolis, Minnesota, USA, September 2-6, 2009, pp.
5689-5692.

Vous aimerez peut-être aussi