SPECIAL SECTION ON DATA-ENABLED INTELLIGENCE FOR DIGITAL HEALTH

Received May 19, 2019, accepted June 3, 2019, date of publication June 7, 2019, date of current version July 16, 2019.
Digital Object Identifier 10.1109/ACCESS.2019.2921568

Simultaneous Human Health Monitoring and

Time-Frequency Sparse Representation
Using EEG and ECG Signals
WANGPENG HE 1 , (Member, IEEE), GENG WANG1 ,
JIE HU1 , CHENG LI1 , (Student Member, IEEE),
BAOLONG GUO1 , (Senior Member, IEEE),
AND FENGPING LI2
1 School of Aerospace Science and Technology, Xidian University, Xi’an 710071, China
2 Institute of Laser and Optoelectronics Intelligent Manufacturing, Wenzhou University, Wenzhou 325035, China
Corresponding author: Fengping Li (lfp@wzu.edu.cn)
This work was supported in part by the Laser Manufacturing and Additive Manufacturing Project of National Key Research and
Development Program of China under Grant 2018YFB1108000, in part by the Wenzhou Municipal Key Science and Research
Program under Grant ZG2017003, in part by the National Natural Science Foundation of China under Grant 51805398,
and in part by the National Natural Science Foundation of Shaanxi Province under Grant 2018JQ5106.

ABSTRACT In the field of human health monitoring, intelligent diagnostic methods have drawn much
attention recently to tackle the health problems and challenges faced by patients. In this paper, an efficient and
flexible diagnostic method is proposed, which enables the simultaneous use of a machine learning method
and sparsity-based representation technique. Specifically, the proposed method is based on a convolutional
neural network (CNN) and generalized minimax-concave (GMC) method. First, measured potential signals,
for instance, electroencephalogram (EEG) and electrocardiogram (ECG) signals are directly inputted into
the designed network based on CNN for health conditions classification. The designed network adopts small
convolution kernels to enhance the performance of feature extraction. In the training process, small batch
samples are applied to improve the generalization of the model. Meanwhile, the ‘‘Dropout’’ strategy is
applied to overcome the overfitting problem in fully connected layers. Then, for a record of the interested
EEG or ECG signal, the sparse representation of useful time-frequency features can be estimated via the
GMC method. Case studies of seizure detection and arrhythmia signal analysis are adopted to verify the
performance of the proposed method. The experimental results demonstrate that the proposed method
can effectively identify different health conditions and maximally enhance the sparsity of time-frequency
features.

INDEX TERMS Feature extraction, convolutional neural network, deep learning, sparse representation,
health monitoring.

I. INTRODUCTION
With the development of society and the improvement of
economy, humans have been paying more and more atten-
tion to their health conditions [1]–[3]. Meanwhile, data-
driven intelligent technologies have achieved great success
in tackling the health problems and challenges faced by
patients [4]–[6], such as introducing EEG signals to mon-
itor and prevent epilepsy, encephalitis and intracranial
The associate editor coordinating the review of this manuscript and
approving it for publication was Weihong Huang.
tumors [7], [8], adopting ECG signals to identify and diag-
nose arrhythmia, heart disease and myocardial infarction
(MI) [9], [10]. These methods use complex medical equip-
ment to obtain potential signals of human organs, so that it
could analyze and recognize various lesion [11]. Therefore,
it is of great significance to extract representative information
from these complicated signals to monitor human health
conditions [12].
In recent years, much work has been devoted to develop-
ing health monitoring techniques [13]–[15], such as spec-
trum analysis, wavelet transform. Among them, artificial

2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 7, 2019 Personal use is also permitted, but republication/redistribution requires IEEE permission. 85985
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals
intelligence (AI) related studies have attracted much attention
from researchers [16]. It could efficiently extract features and
improve the accuracy of health condition recognition. As one
of the representative algorithms in deep learning, CNN is able
to adaptively capture useful information from raw input sig-
nal through multiple non-linear transformations and approxi-
mate complex non-linear functions. Various related methods
based on CNN have been proposed recently. Ullah et al. [17]
proposed an ensemble of pyramidal one-dimensional
CNN model, which was suitable for epilepsy detection.
Acharya et al. [18] proposed a CNN algorithm for the auto-
mated detection of a normal and MI ECG beats. In [19],
a deep learning architecture was proposed to identify epileptic
seizure and achieved higher accuracy.
Currently, another popular research topic in this area is
focused on signal sparse representation [20]. The `1 norm
regularization is a classic method which is convex and
can induce sparsity most effectively [21]. However, the `1
norm tends to underestimate high-amplitude components of
the true signal. Non-convex regularizations are also widely
used, which can induce sparsity without underestimating
high-amplitude components [22]. Whereas, the formulated
cost function is generally non-convex and has suboptimal
local minima. The sparse representation method based on
GMC method adopts a class of non-convex penalty as reg-
ular terms to maintain the convexity of the cost function
and can maximally induce the sparsity of estimated results.
In [23], a vehicle active safety model is presented for vehi-
cle speed control based on driver vigilance detection using
low-cost, comfortable, wearable EEG sensors and sparse
representation. Wang et al. [24] proposed a novel frame-
work to extract compact and discriminative features from
ECG signals for human identification based on sparse rep-
resentation of local segments, which could capture local and
global structural information to characterize the ECG signals
more precisely. She et al. [25] proposed a novel approach
called FDDL-ELM, which combines the discriminative
power of extreme learning machine (ELM) with the recon-
struction capability of sparse representation and achieve
superior performance than the other existing algorithms.
However, quite few studies have been conducted on the
jointly intelligent health conditions classification and use-
ful features representation for human health monitoring.
Therefore, this paper proposes an efficient and flexible diag-
nostic method based on CNN and GMC method, which
simultaneously achieves conditions classification and time-
frequency sparse representation of EEG or ECG signals.
Firstly, a CNN model is designed for human health conditions
classification. The network adopts small convolution kernels
to enhance the performance of feature extraction. Meanwhile,
the ‘‘Dropout’’ strategy is applied to overcome the overfitting
problem in fully connected layers. In the training process,
small batch samples are applied to improve the generalization
of the model. Then, for a record of interested EEG or ECG
signal, the sparse representation of useful time-frequency
features can be obtained by the GMC method.
85986
FIGURE 1. Architecture of CNN.
The main advantages of the proposed method can be gener-
alized as follows: (1) it can jointly solve the problem of health
conditions classification and time-frequency sparse represen-
tation; (2) the designed flexible CNN model has excellent fea-
ture extraction ability and can automatically extract essential
features from EEG or ECG signals, which greatly increases
the classification accuracy; (3) the sparse features hidden in
EEG or ECG signals can be represented in time-frequency
domain for observing the useful features distribution conve-
niently by the GMC method. It utilizes a non-convex penalty
to maintain the convexity of the cost function and induce
a higher level of sparsity without underestimating the high-
amplitude components like the `1 norm.
The rest of this paper is structured as follows.
Section 2 briefly introduces the basic theory of CNN
and time-frequency sparse representation via GMC method
respectively. Section 3 presents the proposed health monitor-
ing method based on CNN and GMC method. In Section 4,
experiments are implemented to validate the effectiveness
and superiority of the proposed method. The conclusions are
given in Section 5.
II. BASIC THEORY OF CNN AND GMC METHOD
A. BASIC THEORY OF CNN
The forward-transmission process of CNN aims to extract
representative features and output the classification result.
Generally, it consists of convolutional layers, pooling layers
and fully connected layers [26]. The original image is divided
into several classifications after the forward-transmission
process, as shown in Figure 1.
The backward-transmission process optimizes the network
by calculating the error between the output and the target
values. Loss function and parameter optimization method are
important in this process since they can effectively update the
parameters of each layer.
1) CONVOLUTIONAL LAYER
A convolutional layer is responsible for extracting represen-
tative features from input signals. It is made up of multi-
ple convolution kernels, and each neuron of the convolution
kernels has a weight and a bias [27]. Each neuron in the
convolutional layer is connected to multiple neurons in the
region close to the previous layer. The size of the region
depends on the size of the convolution kernel. Convolution
operations are performed between the output of the previous
layer and the convolution kernel in the convolutional layer.
The results of the operations are output to the following layer
through an activation function. The progress of feature maps
VOLUME 7, 2019
W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals

in a convolutional layer can be defined as follows:
X ally and the primary way to optimize parameters is stochastic
Xjl = f ( Xil−1 ∗ wlij + blj )
i∈Mj
where Xjl denotes the j th neuron of the l th layer, Mj is the
j th convolution area of the l − 1 th layer, Xjl−1 represents a
neuron of Mi , wlij refers to the weight matrix of the convolu-
tion kernel, bli denotes a trainable bias, ∗ is the computation
of convolution and f (·) represents the activation function.
2) POOLING LAYER
After feature extraction in the convolutional layer, the signals
are input into the pooling layer for feature selection and
information filtering. There are two commonly used pool-
ing operations, maximum pooling and average pooling [30],
which merely reduce dimensions. A pooling layer has neither
parameters nor weights updating. In the pooling layer, each
non-overlapping area of the input signals with n ∗ n size is
pooled. The maximum or average value is selected as the
result. Finally, the original signal is reduced by n times in
both dimensions.
3) FULLY CONNECTED LAYER
A fully connected layer in CNN is equivalent to a hidden
layer in a traditional feed forward neural network. It is usually
located at the last part of the hidden layers of CNN and only
transfers signals to other fully connected layers. The feature
map is converted into a one-dimensional vector in fully con-
nected layer and then sent into the following layer through an
activation function. The output of a fully connected layer is
obtained by:
yl = f (wl ∗ x l−1 + bl )
where yl refers to the output of the l th layer, x l−1 denotes the
one-dimensional vector of the l − 1 th layer, wl is a trainable
weight, bl represents a trainable bias and f (·) refers to the
activation function.
4) LOSS FUNCTION
A loss function measures the error between the output of
the network and the correct result. Commonly adopted loss
functions are mean-square error function, cross-entropy func-
tion, negative log likelihood function. In this paper, the cross-
entropy function is chosen as the loss function, which can be
defined as
n
1X
E =− [yk lntk + (1 − yk )ln(1 − tk )]
n
k=1
where n represents the number of samples, t refers to the
predicted value and y denotes the true value.
5) PARAMETER OPTIMIZATION METHOD
After calculating by the abovementioned loss function,
the error has to be transmitted back. In the neural network,
parameter optimization is to minimize the loss function actu-
(1) gradient descent (SGD). The method updates the parameters
by calculating the gradient of the loss function, which has a
faster convergence speed than other methods. The algorithm
is described as
∂E
θ0 = θ − η (4)
∂θ
where E denotes the loss function in (3), θ is the weight and
bias before updating, θ 0 represents the weight and bias after
updating, η refers to the learning rate.
6) ACTIVATION FUNCTION
In the neural network, extracted features are input into the
following layer through an activation function. If there is no
activation function, the output of each layer is a linear func-
tion of the input signal of this layer. Therefore, the activation
function is adopted to add nonlinear factors into the network.
It allows the neural network to approximate any nonlinear
function and make it available in many nonlinear modelS.
Generally, the commonly adopted activation functions are
sigmoid functions, RELu functions.
The sigmoid function is widely used in the hidden layers of
ANN and the fully connected layers of CNN. Its geometry is
an S-shaped curve. The RELu function is commonly adopted
in the convolutional layers of CNN, which can make the
backward-transmission process more efficient and simplify
computing. The formula of the Sigmoid function and the
RELu function are described, respectively, as follows:
1
f (x) = (5)
1 + e−x
f (x) = max(0, x) (6)
(2)
7) DROPOUT STRATEGY
Overfitting is a widespread problem in training, which results
in poor performance on the testing datasets. Hence, this paper
applies the dropout strategy to overcome that problem as well
as gradient vanishing in deep neural networks. In the learning
process of a neural network, some neurons of hidden layers
are randomly ignored. Since the ignored neurons in each iter-
ation of the training process are different, the importance of
each neuron is balanced. On the contrary, a few high-weight
neurons will completely control the output if the dropout
strategy is not adopted. As a result, this method effectively
reduces the structural risk of the network.
B. TIME-FREQUENCY SPARSE REPRESENTATION VIA
(3) GMC METHOD
This section introduces a GMC method for time-frequency
sparse representation. The GMC penalty is a class of non-
convex penalties for sparse-regularized linear least squares
that maintains the convexity of the cost function to be mini-
mized and promotes more strongly sparsity than `1 norm [28].
In time-frequency domain, features are extracted by estimat-
ing the short-time Fourier transform (STFT) coefficients,

VOLUME 7, 2019 85987

 W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals

which can be obtained by minimizing the least squares

cost function. The cost function can be formulated
as F : RN → R
1
F(x) = ky − Axk22 + λψB (x) λ > 0 (7)
2
where y ∈ RN is the observed signal, the matrix A ∈ RM ×N
represents the inverse STFT operator and ψB (x) : RN → R
denotes the GMC penalty (regularization term). The
GMC penalty ψB represents a multivariate generalization
of the Huber function defined by infimal convolution. Even
FIGURE 2. The generalized minimax-concave penalty for the matrix B.
though the penalty is non-convex, it is easy to prescribe this
penalty so as to maintain the convexity of the cost function.

1) GMC PENALTIES
The scaled Huber function s : R → R is defined as
sb (x) := s(b2 x)/b2 , b 6= 0 (8)
where b is a scaled variable and

 1 x 2,
 |x| ≤ 1 FIGURE 3. The MC penalty for several values of b.
s(x) = 2 1 (9)
|x| = , |x| ≥ 1

2
For b = 0, the function is defined as s0 := 0. For b 6 = 0, The penalty is parameterized by a scalar b or a matrix B,
the scaled Huber function is given by and the convexity of F depends on b or B being suitably
 prescribed. The MC penalty for several values of the scaling
1
 b2 x 2 ,
 |x| ≤ 1/b2 parameter is shown in Figure 3.
sb (x) = 2 1 (10)
|x| −
 , |x| ≥ 1/b 2
2) CONVEXITY CONDITION OF THE OBJECT FUNCTION
2b2
The scaled Huber function can be written as Using GMC penalty in (15), the cost function F can be
n o expressed as
sb (x) = min |v| + 12 b2 (x − v)2 (11)
v∈R F(x)
λ
 
The minimax-concave (MC) penalty function is defined as 1
= ky − Axk22 +λ kxk1 − min λ kvk1 + kB(x − v)k22
2 2

1 v∈RN
|x| − x 2 , |x| ≤ 1

λ
 
φ(x) = |x| − s(x) = 1 2 (12) 1 2 2
= max ky − Axk2 +λ kxk1 −λ kvk1 − kB(x − v)k2
 ,
 |x| = 1 v∈RN 2 2
2 1 T T
Therefore, the scalar scaled penalty can be defined as = x (A A−λB B)x + λ kxk1 + max h(x, v)
T
(17)
2 v∈RN
φb (x) = |x| − sb (x) (13) where the function h is affine in x. The last term of (17) is
Note that φ0 (x) = |x|. For b 6 = 0, guaranteed to be convex as it is the pointwise maximum of a
 set of convex functions. Thus, F is convex if AT A − λBT B is
1
|x| − b2 x 2 , |x| ≤ 1/b2
 positive semidefinite. In other words, the cost function F is
φb (x) = 1
2 (14) convex if

 , |x| ≥ 1/b2
2b2 1 T
BT B  A A (18)
Let B ∈ RM ×N , the generalized Huber function sB (x) can be λ
defined as B can be set
n o
sB (x) = min kvk1 + 21 kB(x − v)k22 (15) γ /λA, 0≤γ ≤1
p
B= (19)
v∈RN
A multivariate generalization of the minimax-concave The parameter γ controls the non-convexity of the GMC
(GMC) penalty can be defined as penalty. If γ = 0, then B = 0 and the penalty reduces to
the `1 norm. If γ = 1, then BT B = (1/λ)AT A, and the
ψB (x) = kxk1 − sB (x) (16)
non-convexity of the penalty reaches maximum. In practice,
The GMC penalty function is illustrated in Figure 2. the range of 0.5 ≤ γ ≤ 0.8 is used.

85988 VOLUME 7, 2019

W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals
TABLE 1. The iterative algorithm for gmc method.
3) OPTIMIZATION ALGORITHM
An effective approach for the global minimizer of the sparse-
regularized cost function is to formulate the optimization
problem:
 
1 2
(x̃, ṽ) = arg min max F(x, v) = ky − Axk2 +λψB (x, v)
x∈RN v∈RN 2
(20)
The solution can be obtained using the forward-backward
(FB) algorithm, and the iterative algorithm is shown
in Table 1.
III. THE PROPOSED METHOD BASED ON CNN AND GMC
METHOD
In this section, an efficient and flexible diagnostic method
based on CNN and the GMC method is presented for human
health monitoring. The proposed method enables the simul-
taneous use of machine learning method and sparsity-based
representation technique. More specifically, potential signals
(e.g., EEG and ECG) of human organs are directly input into
the proposed model for processing. Firstly, a network based
on CNN is designed for human health conditions classifica-
tion. Then, for a record of interested EEG or ECG signal,
the sparse representation of useful time-frequency features
can be obtained and enhanced by the GMC method. The
overall framework of the proposed health monitoring method
is demonstrated in Figure 4. The main steps are summarized
as follows:
Step 1: Collect EEG or ECG signals from human organs.
Segment the signals and construct the training datasets and
testing datasets.
Step 2: Design the network based on CNN for health
conditions classification. Initialize the parameters of weights,
biases and learning rates. In the formulated network, small
one-dimensional convolution kernels are adopted in the con-
volutional layers to improve the performance of feature
extraction. In full connection layers, the dropout strategy is
adopted to invalidate the weights of some neurons, so that it
could overcome the overfitting problem.
VOLUME 7, 2019
Step 3: Feed the training datasets into the formulated net-
work. Calculate the error between the output of the last layer
and the target values. In this process, small batch samples
are applied for training to improve the generalization of the
network.
Step 4: Transmit the error backward and update the weights
along with biases layer by layer. Transmit forward and back-
ward iteratively until the output error meets the predefined
termination criterion.
Step 5: Feed the testing datasets into the well trained
network to classify the health conditions.
Step 6: Construct the time-frequency representation oper-
ator and specify the related parameters in the optimization
problem formulated by the GMC method. Note that the STFT
is adopted and implemented as a normalized tight frame in
this paper.
Step 7: For a record of interested EEG or ECG signal,
the sparse representation of useful time-frequency features
can be extracted via the presented GMC method. Notice that
the GMC method uses a non-convex penalty to maintain the
cost function is strictly convex and induce stronger sparsity.
IV. EXPERIMENTAL DEMONSTRATIONS
In this section, the effectiveness of the proposed intelligent
health monitoring method based on CNN and GMC method
is verified on the experimental datasets. The program of the
formulated model is developed in python 3.6.7 with Tensor-
flow deep learning library and run on Windows 10 with a
RTX 2080 Ti GPU.
A. CLASSIFICATION OF EPILEPTIC CONDITIONS USING
EEG SIGNALS
1) DATA PREPARATION
EEG signals used in this experiment are collected by Andrze-
jak et al. at Bonn University, Germany [29]. These segments
were selected and cut out from continuous multi-channel
EEG recordings after visual inspection for artifacts. The
dataset consists of five sets, and each set containing 100 sin-
gle channel EEG segments with a duration of 23.6s. The
sampling rate is set at 173.61 Hz. The health conditions of the
five sets are listed in Table 2 respectively. Typical samples of
the measured EEG signals corresponding to the five types of
health conditions.
In the five conditions above, 500 samples need to be
divided into training datasets and testing datasets. In order
to ensure the accuracy of the experiment, these samples were
shuffled and 400 of them were randomly selected as training
samples. Then, the remainder of 100 samples were introduced
as testing samples. Table 3 shows the sample size and labels
for different health conditions. The labels are applied to com-
pare with the output of the formulated network to calculate
error and optimize parameters.
2) MODEL CONSTRUCTION
The proposed network based on CNN in this paper is
depicted in Figure 6, which has three convolutional layers,
85989
 W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals

TABLE 2. The health conditions of the five datasets.

FIGURE 4. Framework of the proposed human health monitoring method.

three polling layers followed by two fully connected layers TABLE 3. Health conditions samples distribution.
and a softmax layer in the end. The convolution layers and
pooling layers aim to extract representative features from the
input signals. The fully connected layers and a softmax layer
is to identify the machinery condition.
In the proposed model, the input one-dimensional signal is
4096 ∗ 1. After three convolutional layers with multiple 7 ∗ 1
small convolution kernels and three pooling layers, the ouput
signals are sixty-four signals of 512 ∗ 1. Then, the feature
maps are flattened and connected with a fully connected layer.
After two fully connected layers, 128 ∗ 1 signal is output to
the next layer. Finally, the original data is divided into five
patterns through the softmax classification layer.
RELu function performs activation in the convolutional
layer. The activation function of the fully connected layer 3) EXPERIMENTAL RESULT
is the Sigmoid function. The pooling layers introduces the As mentioned above, there are 100 testing datasets in total.
maximum pooling method for feature selection and informa- 20, 13, 25, 25 and 17 datasets are in the five condition
tion filtering. The ‘‘Dropout’’ strategy is adopted to over- respectively. The testing datasets are used to verify the clas-
come the overfitting problem in the fully connected layer. sification accuracy of the proposed method. After repeated
The loss function taken in the model is the cross- entropy training, the classification accuracy of the testing datasets
function. Back-propagation is employed for the updates of will be output by the last layer. Figure 7 demonstrates the
model parameters where the SGD method is used. learning classification accuracy of the proposed method based on a
rate is set to 0.0001 and batch size is set to 1. confusion matrix. The row and the column of the confusion

85990 VOLUME 7, 2019

W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals
FIGURE 5. Typical samples of the measured EEG signals corresponding to
the five types of health conditions.
matrix represent the prediction label and the true label
respectively.
In the 100 testing datasets, 97 samples are accurately clas-
sified and 3 samples are misjudged. Three samples of B, C
and D are identified incorrectly as A, D and C condi-
tion respectively. Hence, the proposed method shows A
classification accuracy of 97%. The classification accuracy
FIGURE 6. The detailed architecture of the proposed network based on CNN.
VOLUME 7, 2019
of the A and E condition can achieve 100%. The accuracy
for B, C, D can also reach to 92.3%, 96% and 96% respec-
tively. It can be seen that the proposed method obtains a
desirable result evidently.
The classification results of the testing datasets are pre-
sented in Figure 8 for better visualization of the performance.
The testing datasets distribution of A and E condition is
concentrated and almost gathered at the same point, which
indicates that the proposed method has a great classification
effect on these two conditions. However, the distribution
of B, C and D condition is a little scattered, which indicates
that the classification effects of the method on these three
conditions need to be improved.
As previously discussed, the method proposed in this
paper shows a desirable performance on the testing
datasets. It is fairly evident that the method can be effec-
tively used for classification of epileptic conditions using
EEG signals.
B. ARRHYTHMIA ANALYSIS USING ECG SIGNALS
In this section, the experiment adopts the ECG signals from
Massachusetts Institute of Technology and Beth Israel hos-
pital (MIT-BIH) arrhythmia database [30]. The sampling
frequency is set to 360 Hz. A short-time Fourier transform
(STFT) with 50% overlapping segments is adopted. For the
inverse STFT operator A, we use a short window 32 samples.
For the GMC penalty, we set λ = 0.1 so as to minimize
the root-mean-square error (EMSE) and set B as in (19) with
γ = 0.8. A record of measured time domain waveform and
its time-frequency spectrogram are shown in Figure 9 and
Figure10, respectively. The optimized STFT coefficients
(time-frequency sparse representation) illustrated
in Figure 11.
85991
 W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals
FIGURE 7. The multi-classification confusion matrix of the proposed
classification method.
FIGURE 8. Classification result visualization.
FIGURE 9. The time domain waveform of the ECG signal.
FIGURE 10. The time-frequency spectrogram of the interested ECG signal.
It can be observed from Figure 11 that the GMC solu-
tion extracted sparse time-frequency spectrogram via the
GMC method has fewer extraneous noise artifacts. More-
over, the characteristic features are clearly revealed in the
time-frequency domain without the interferences of noisy
components. Notice that in this subsection, the results
of an interested signal are extracted for the purpose of
exhibiting useful sparse features in time-frequency domain.
85992
FIGURE 11. The extracted sparse time-frequency spectrogram via the
GMC method.
Future diagnostic conclusion or post-processing based on the
extracted results can be conducted flexibly.
V. CONCLUSION
In this paper, an efficient and flexible diagnostic method
based on CNN and the GMC method is proposed for human
health monitoring. The proposed method enables the simul-
taneous use of machine learning method and sparsity-based
representation technique. It extracts features directly from
potential signals (e.g., EEG and ECG) of human organs.
Firstly, a flexible CNN model is designed for human health
conditions classification. In the formulated network, small
one-dimensional convolution kernels are adopted in the con-
volutional layers to effectively improve network perfor-
mance. In the training process, small batch samples are
applied for training to improve the generalization of the
model. Besides, the ‘‘Dropout’’ strategy is applied to over-
come the overfitting problem in fully connected layers. Then,
for a record of interested EEG or ECG signal, the sparse rep-
resentation of useful time-frequency features can be achieved
and enhanced by the GMC method.
The analyzed results of two experimental cases with dif-
ferent potential signals, namely seizure detection using EEG
signals and arrhythmia analysis using ECG signals, demon-
strate the superiority of the proposed method. The proposed
method achieves a satisfactory accuracy and can jointly
extracts useful time-frequency sparse features from interested
EEG or ECG signals, which is practical and convenient in
human health monitoring.
Future work will focus on developing automated machine
learning techniques to adaptively meet the requirements of
different diagnostic applications. Acceleration algorithms
will also be adopted to guarantee the intelligent classification
network satisfying real-time applications. Moreover, data-
driven sparse represented technique will be studied which can
avoid the problem of manually specifying parameters.
REFERENCES
[1] Q. Zhang, D. Zhou, and X. Zeng, ‘‘HeartID: A multiresolution convolu-
tional neural network for ECG-based biometric human identification in
smart health applications,’’ IEEE Access, vol. 5, pp. 11805–11816, 2017.
[2] I. Megiddo, A. Colson, D. Chisholm, T. Dua, A. Nandi, and
R. Laxminarayan, ‘‘Health and economic benefits of public financing of
epilepsy treatment in India: An agent-based simulation model,’’ Epilepsia,
vol. 57, no. 3, pp. 464–474, 2016.
VOLUME 7, 2019
W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals
[3] P. Swami, T. K. Gandhi, B. K. Panigrahi, M. Tripathi, and S. Anand,
‘‘A novel robust diagnostic model to detect seizures in electroencephalog-
raphy,’’ Expert Syst. Appl., vol. 56, pp. 116–130, Sep. 2016.
[4] C. Venkatesan, P. Karthigaikumar, A. Paul, S. Satheeskumaran, and
R. Kumar, ‘‘ECG signal preprocessing and SVM classifier-based abnor-
mality detection in remote healthcare applications,’’ IEEE Access, vol. 6,
pp. 9767–9773, 2018.
[5] U. R. Acharya, H. Fujita, M. Adam, O. S. Lih, V. K. Sudarshan,
T. J. Hong, J. E. W. Koh, Y. Hagiwara, C. K. Chua, C. K. Poo, and
T. R. San, ‘‘Automated characterization and classification of coronary
artery disease and myocardial infarction by decomposition of ECG signals:
A comparative study,’’ Inf. Sci., vol. 337, pp. 17–29, Jan. 2017.
[6] B. Chen, Y. Li, and N. Zeng, ‘‘Centralized wavelet multiresolution for exact
translation invariant processing of ECG signals,’’ IEEE Access, vol. 7,
pp. 42322–42330, 2019.
[7] A. Sharmila and P. Geethanjali, ‘‘DWT based detection of epileptic seizure
from EEG signals using naive Bayes and K-NN classifiers,’’ IEEE Access,
vol. 4, pp. 7716–7727, 2016.
[8] J. Liu, K. Narasimhan, V. Elamaran, N. Arunkumar, M. Solarte, and
G. Ramirez-Gonzalez, ‘‘Clinical decision support system for alcoholism
detection using the analysis of EEG signals,’’ IEEE Access, vol. 6,
pp. 61457–61461, 2018.
[9] D. Craven, B. McGinley, L. Kilmartin, M. Glavin, and E. Jones, ‘‘Adap-
tive dictionary reconstruction for compressed sensing of ECG sig-
nals,’’ IEEE J. Biomed. Health Inform., vol. 21, no. 3, pp. 645–654,
May 2017.
[10] S. Raj and K. C. Ray, ‘‘Sparse representation of ECG signals for auto-
mated recognition of cardiac arrhythmias,’’ Expert Syst. Appl., vol. 105,
pp. 49–64, Sep. 2018.
[11] M. Rizwan, B. M. Whitaker, and D. V. Anderson, ‘‘AF detection from ECG
recordings using feature selection, sparse coding, and ensemble learning,’’
Physiol. Meas., vol. 39, no. 12, Dec. 2018, Art. no. 124007.
[12] J. Li, G. Deng, W. Wei, H. Wang, and Z. Ming, ‘‘Design of a real-time
ECG filter for portable mobile medical systems,’’ IEEE Access, vol. 5,
pp. 696–704, 2017.
[13] A. Bhattacharyya, R. B. Pachori, A. Upadhyay, and U. R. Acharya,
‘‘Tunable-Q wavelet transform based multiscale entropy measure for auto-
mated classification of epileptic EEG signals,’’ Appl. Sci., vol. 7, no. 4,
p. 385, 2017.
[14] B. van Ginneken, ‘‘Fifty years of computer analysis in chest imaging: Rule-
based, machine learning, deep learning,’’ Radiological Phys. Technol.,
vol. 10, no. 1, pp. 23–32, 2017.
[15] Y. Yu and Z. Sun, ‘‘Sparse coding extreme learning machine for classifi-
cation,’’ Neurocomputing, vol. 261, pp. 50–56, Oct. 2017.
[16] H. Greenspan, B. van Ginneken, and R. M. Summers, ‘‘Deep learning
in medical imaging: Overview and future promise of an exciting new
technique,’’ IEEE Trans. Med. Imag., vol. 35, no. 5, pp. 1153–1159,
May 2016.
[17] I. Ullah, M. Hussain, E.-U.-H. Qazi, and H. Aboalsamh, ‘‘An auto-
mated system for epilepsy detection using EEG brain signals based
on deep learning approach,’’ Expert Syst. Appl., vol. 107, pp. 61–71,
Oct. 2018.
[18] U. R. Acharya, H. Fujita, S. L. Oh, Y. Hagiwara, J. H. Tan, and
M. Adam, ‘‘Application of deep convolutional neural network for auto-
mated detection of myocardial infarction using ecg signals,’’ Inf. Sci.,
vol. 415, pp. 190–198, Nov. 2017.
[19] R. San-Segundo, M. Gil-Martín, L. F. D’Haro-Enríquez, and
J. M. Pardo, ‘‘Classification of epileptic EEG recordings using signal
transforms and convolutional neural networks,’’ Comput. Biol. Med.,
vol. 109, pp. 148–158, Jun. 2019.
[20] M. Luo and K. Zhang, ‘‘A hybrid approach combining extreme learning
machine and sparse representation for image classification,’’ Eng. Appl.
Artif. Intell., vol. 27, pp. 228–235, Jan. 2014.
[21] M. Yang, L. Zhang, X. Feng, and D. Zhang, ‘‘Sparse representation based
Fisher discrimination dictionary learning for image classification,’’ Int. J.
Comput. Vis., vol. 109, no. 3, pp. 209–232, Sep. 2014.
[22] W. He, B. Chen, N. Zeng, and Y. Zi, ‘‘Sparsity-based signal extraction
using dual Q-factors for gearbox fault detection,’’ ISA Trans., vol. 79,
pp. 147–160, Aug. 2018.
[23] Z. Zhang, D. Luo, Y. Rasim, Y. Li, G. Meng, J. Xu, and C. Wang, ‘‘A vehi-
cle active safety model: Vehicle speed control based on driver vigilance
detection using wearable EEG and sparse representation,’’ Sensors, vol. 16,
no. 2, p. 242, Feb. 2016.
VOLUME 7, 2019
[24] J. Wang, M. She, S. Nahavandi, and A. Kouzani, ‘‘Human iden-
tification from ECG signals via sparse representation of local seg-
ments,’’ IEEE Signal Process. Lett., vol. 20, no. 10, pp. 937–940,
Oct. 2013.
[25] Q. She, K. Chen, Y. Ma, T. Nguyen, and Y. Zhang, ‘‘Sparse representation-
based extreme learning machine for motor imagery EEG classification,’’
Comput. Intell. Neurosci., vol. 2018, Oct. 2018, Art. no. 9593682.
[26] S. U. Amin, M. Alsulaiman, G. Muhammad, M. A. Bencherif, and
M. S. Hossain, ‘‘Multilevel weighted feature fusion using convolutional
neural networks for EEG motor imagery classification,’’ IEEE Access,
vol. 7, pp. 18940–18950, 2019.
[27] R. M. Mehmood, R. Du, and H. J. Lee, ‘‘Optimal feature selection and deep
learning ensembles method for emotion recognition from human brain
EEG sensors,’’ IEEE Access, vol. 5, pp. 14797–14806, 2017.
[28] I. Selesnick, ‘‘Sparse regularization via convex analysis,’’ IEEE Trans.
Signal Process., vol. 65, no. 17, pp. 4481–4494, Sep. 2017.
[29] R. G. Andrzejak, K. Lehnertz, C. Rieke, F. Mormann, P. David, and
C. E. Elger, ‘‘Indications of nonlinear deterministic and finite dimensional
structures in time series of brain electrical activity: Dependence on record-
ing region and brain state,’’ Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat.
Interdiscip. Top., vol. 64, p. 061907, 2001.
[30] MIT-BIH Arrhythmia Database. Accessed: May 17, 2019. [Online]. Avail-
able: https://physionet.org/physiobank/database/mitdb/
WANGPENG HE received the B.S. and Ph.D.
degrees in mechanical engineering from Xi’an
Jiaotong University, Xi’an, China, in 2007 and
2016, respectively. In 2014, he was appointed as
a Visiting Scholar with the New York Univer-
sity, New York, NY, USA. He joined the School
of Aerospace Science and Technology, Xidian
University, where he is currently an Assistant
Professor. His research interests include signal
processing, machine vision, sparsity-based signal
processing, machine learning, and machinery fault diagnosis.
GENG WANG received the B.S. degree in mea-
surement and control technology and instrument
major from the School of Electro-mechanical
Engineering, Xidian University, in 2018, where
he is currently pursuing the M.S. degree with the
School of Aerospace Science and Technology. His
main research interests include deep learning and
neural networks.
JIE HU received the B.S. degree in automa-
tion engineering major from the Xi’an Univer-
sity of Architecture and Technology, Xi’an, China,
in 2018. She is currently pursuing the M.S. degree
with the School of Aerospace Science and Tech-
nology, Xidian University. Her main research
interests include sparsity-based signal processing
and machinery fault diagnosis.
CHENG LI received the B.S. degree from the
School of Electro-mechanical Engineering, Xidian
University, Xi’an, China, in 2013, and the M.S.
degree from the School of Aerospace Science and
Technology, Xidian University, in 2017, where he
is currently pursuing the Ph.D. degree with the
Intelligent Control and Image Engineering Insti-
tute. He focuses on researching in multi-scale
image representation.
85993
 W. He et al.: Simultaneous Human Health Monitoring and Time-Frequency Sparse Representation Using EEG and ECG Signals

BAOLONG GUO received the B.S., M.S., and FENGPING LI is currently a Distinguished
Ph.D. degrees from Xidian University, in 1984, Professor with the Institute of Laser and Opto-
1988, and 1995, respectively, all in communication electronics Intelligent Manufacturing, Wenzhou
and electronic systems. University, Wenzhou, China. His research fields
From 1998 to 1999, he was a Visiting Scholar mainly include adaptive control algorithm and
with Doshisha University. He is currently a Full machine vision, laser processing application, and
Professor with the School of Aerospace Sci- industrial control systems.
ence and Technology, Xidian University, Xi’an,
China. His current research interests include pat-
tern recognition, intelligent information process-
ing, image processing, and video communication. He has published over
100 academic papers and organized several international conferences as the
Co-Chair.

85994 VOLUME 7, 2019