Advanced EMD method using Variance
Characterization for PPG with Motion Artifact
Bo Pang, Ming Liu, Xu Zhang, Peng Li, Zhaolin Yao,
Xiaohui Hu and Hongda Chen
State Key Laboratory on Integrated Optoelectronics
Institute of Semiconductors, Chinese Academy of Sciences
Beijing, China
Email: liuming@semi.ac.cn
Qi Gong
School of Electronic and Information Engineering
South China University of Technology
Guangdong, China
Abstract—Motion Artifact (MA) reduction is an important
part in Photoplethysmography (PPG) signal processing for
wearing devices. The heavily-corrupted period in PPG can
hardly be rebuilt by frequency domain methods. This paper
proposes an advanced time-frequency analysis method based on
Empirical Mode Decomposition (EMD) using variance
characterization of extrema. In this way, the computing costs are
largely decreased by picking out the corrupted period, while the
wave clusters found in it for estimation can help reduce error
detection rate. The result shows our method is accurate in pulse
rate estimation for heavily-corrupted PPG signals. The average
relative error of our method is1.03% , which is a result of data
from PhysioBank MIMIC II waveform database.
Keywords—PPG; motion artifact; EMD; variance
characterization
I. I
NTRODUCTION
Nowadays, continuous physiological monitoring has been a
significant research orientation in the area of disease diagnosis,
prevention and health assessment. The collected physiological
signals are important reflections of the healthy state of human
body and the processing for noninvasive continuous
monitoring of human physiological parameters has become an
important research area in physiological signal recognition.
Heart Rate (HR) is a key physiological indicator to show
the health condition of heart. There are several methods to
measure the HR, and among them, the PPG [1] has been
widely-applied for wearable devices because of its
noninvasiveness and convenience. PPG is an optical method to
measure the change of peripheral blood volumetric. However,
the most serious challenge of PPG processing is how to deal
with motion artifacts (MA). MA can be involved by the finger
movement, tremble, and probe mismatch etc. Other than power
frequency interference and baseline drift, the frequency band of
MA is generally overlapping with that of PPG signals (The
frequency band of PPG is from 0.5 to 4 Hz in general [2]).
Hence, the simple frequency domain processes, like band-pass
filter, fail in dealing with MA and we need some time-
frequency analysis methods for the heavily-corrupted PPG
signals.
In this paper, we proposed an advanced method based on
EMD and variance characterization to process the heavily-
corrupted PPG waves. The novel points of this work can be
summarized as below:
(1) The computing cost is reduced through recognizing
the corrupted districts automatically by using variance
characterization.
(2) By retrieving the wave clusters that can be used for
estimation from the corrupted period, the error
detection rate is reduced.
Many papers have studied on the methods of reducing MA
effects in physiological signals, not just PPG signal.
Nevertheless, most of them focus on the signals corrupted by
slight motion artifacts. Malte Kirst et al. proposed a MA
reduction method for ECG using discrete wavelet transform
(DWT). The MA was reduced by reconstructing the signal
from the modified coefficients after zeroing or modifying parts
of the coefficients. But according to our work, DWT could do
little for rebuilding the PPG signal as the spectrum of PPG is
not so concentrated like ECG signals [5]. By using EMD, Qian
Wang et al. have proposed an efficient way to remove MA
from PPG, but they still couldn’t process the heavily-corrupted
signals by just adding up the first several intrinsic mode
functions (IMF) [6]. M. Raghu Ram et al. came up with a novel
AS-LMS Adaptive filter basing on the long-tested LMS filter
method in non-stationary signal processing. But as discussed
above, the frequency domain methods are not so ideal in MA
reduction. MA usually overlaps with the clean PPG in
frequency domain, especially in heavily-corrupted situations
[7].
II. A
LGORITHM
A. Empirical Mode Decomposition
The EMD is a part of Hilbert-Huang transform, which is an
adaptive time-frequency method. It was proposed for analyzing
fluid mechanics not quite long before biomedical signal
processing took it into applications [3]. It is confirmed that the
EMD method is highly effective and versatile for noisy non–
linear and non–stationary signals involving the physiological
signals, since it is based on the local characteristic time scale.
The key of this method is that we can decompose any
complicated time-series into several IMFs [4]. The IMFs can
be seen as a complete and nearly orthogonal basis of the
978-1-5090-2959-4/16/$31.00 2016 IEEE