ECG signal enhancement based on improved denoising auto-encoder
Peng Xiong
a
, Hongrui Wang
a,b
, Ming Liu
b
, Suiping Zhou
c
, Zengguang Hou
d
, Xiuling Liu
b,
n
a
College of Electronic and Information Engineering, Yanshan University, Qinhuangdao, China
b
Key Laboratory of Digital Medical Engineering of Hebei Province, College of Electronic and Information Engineering, Hebei University, Baoding, China
c
School of Science and Technology, Middlesex University, UK
d
Institute of Automation, Chinese Academy of Sciences, Beijing, China
article info
Article history:
Received 2 October 2015
Received in revised form
6 February 2016
Accepted 27 February 2016
Keywords:
Denoising auto-encoder (DAE)
ECG signal denoising
Wavelet transform (WT)
Deep neural network (DNN)
abstract
The electrocardiogram (ECG) is a primary diagnostic tool for examining cardiac tissue and structures. ECG
signals are often contaminated by noise, which can manifest with similar morphologies as an ECG
waveform in the frequency domain. In this paper, a novel deep neural network (DNN) is proposed to
solve the above mentioned problem. This DNN is created from an improved denoising auto-encoder
(DAE) reformed by a wavelet transform (WT) method. A WT with scale-adaptive thresholding method is
used to filter most of the noise. A DNN based on improved DAE is then used to remove any residual noise,
which is often complex with an unknown distribution in the frequency domain. The proposed method
was evaluated on ECG signals from the MIT-BIH Arrhythmia database, and added noise signals were
obtained from the MIT-BIH Noise Stress Test database. The results show that the average of output signal-
to-noise ratio (SNR) is from 21.56 dB to 22.96 dB, and the average of root mean square error (RMSE) is
less than 0.037. The proposed method showed significant improvement in SNR and RMSE compared with
the individual processing with either a WT or DAE, thus providing promising approaches for ECG signal
enhancement.
& 2016 Elsevier Ltd. All rights reserved.
1. Introduction
Cardiovascular disease is one of the most common threats to
human health and is characterized by a high disease incidence,
disability rate, and mortality. Development of telemedicine now
allows for remote electrocardiogram (ECG) monitoring and pro-
vides information for the diagnosis and treatment of cardiovas-
cular diseases. Specifically, remote ECG monitoring systems can
improve early diagnosis accuracy of coronary disease states. Tel-
emedicine practices have recently expanded to include qualitative
and quantitative analysis of arrhythmia and evaluation of rest-
enosis in patients who have suffered a myocardial infarction.
The ECG signal from remote ECG monitoring systems provides
an effective method for detecting heart diseases. The ECG signal is
measured by surface electrodes placed on the skin of patients.
However, the signal is often corrupted by large amounts of noise
from muscle artifacts (MA), electrode motion (EM), and baseline
wander (BW). These noise signals may be confused with the
ectopic ECG signals (Rahman et al., 2011). Hence, noisy ECG signals
should be enhanced by removing the noise components for further
processing such as feature extraction and pattern recognition.
Many researchers have reported on different techniques for
denoising ECG signals such as empirical mode decomposition
(EMD) (Karagiannis and Constantinou, 2011; Kabir and Shahnaz,
2012), adaptive filtering (Sajjad et al., 2012; Rahman et al., 2012;
Moradi et al., 2014), and the wavelet method (Awal et al., 2012,
2014; KapilTajane and Pitale, 2014; Reddy et al., 2009; Gokhale,
2012; Smital et al., 2013).
Karagiannis and Constantinou (2011) studied the performance
of EMD on ECG signals. Noisy ECG signals were processed with
EMD in order to extract the intrinsic mode functions (IMFs). The
algorithm implemented by Karagiannis deduced a 95% bound for
the white Gaussian noise in an ECG time series. Kabir and Shahnaz
(2012) proposed a windowing method in conjunction with EMD in
order to reduce the noise from the initial IMFs while preserving
the QRS complex and yielding a relatively clean ECG signal.
However, it is known that the Hilbert transform intrinsic to EMD
cannot separate similar frequency signals. Therefore, this method
of noise reduction may filter out P-waves and T-waves from the
signal, potentially resulting in misdiagnosis.
Rahman et al. (2012) proposed several adaptive recurrent filters
for remote ECG signal denoising, such as a signed regressor algo-
rithm, a normalized least mean square (LMS) and an error non-
linear signed regressor LMS. Simulation results from Rahman
confirmed that the performance of sign-based algorithms is better
than the LMS method. However, adaptive filters often require a
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/engappai
Engineering Applications of Artificial Intelligence
http://dx.doi.org/10.1016/j.engappai.2016.02.015
0952-1976/& 2016 Elsevier Ltd. All rights reserved.
n
Corresponding author.
E-mail address: liuxiuling121@hotmail.com (X. Liu).
Engineering Applications of Artificial Intelligence 52 (2016) 194–202