sensors

Article

A Compressed Sensing-Based Wearable Sensor

Network for Quantitative Assessment of
Stroke Patients
Lei Yu 1,2, *, Daxi Xiong 1 , Liquan Guo 1 and Jiping Wang 1
1

Jiangsu Key Laboratory of Medical Optics, Suzhou Institute of Biomedical Engineering and Technology,
Chinese Academy of Sciences, No. 88, Keling Road, Suzhou, Jiangsu 215163, China;
xiongdx@sibet.ac.cn (D.X.); guolq@sibet.ac.cn (L.G.); wangjp@sibet.ac.cn (J.W.)
University of Chinese Academy of Sciences, Beijing 100049, China
Correspondence: yuthreestone@gmail.com; Tel./Fax: +86-512-6958-8302

Academic Editor: Leonhard M. Reindl

Received: 12 December 2015; Accepted: 3 February 2016; Published: 5 February 2016

Abstract: Clinical rehabilitation assessment is an important part of the therapy process because
it is the premise for prescribing suitable rehabilitation interventions. However, the commonly
used assessment scales have the following two drawbacks: (1) they are susceptible to subjective
factors; (2) they only have several rating levels and are influenced by a ceiling effect, making
it impossible to exactly detect any further improvement in the movement. Meanwhile, energy
constraints are a primary design consideration in wearable sensor network systems since they
are often battery-operated. Traditionally, for wearable sensor network systems that follow the
Shannon/Nyquist sampling theorem, there are many data that need to be sampled and transmitted.
This paper proposes a novel wearable sensor network system to monitor and quantitatively assess the
upper limb motion function, based on compressed sensing technology. With the sparse representation
model, less data is transmitted to the computer than with traditional systems. The experimental
results show that the accelerometer signals of Bobath handshake and shoulder touch exercises can be
compressed, and the length of the compressed signal is less than 1/3 of the raw signal length. More
importantly, the reconstruction errors have no influence on the predictive accuracy of the Brunnstrom
stage classification model. It also indicated that the proposed system can not only reduce the amount
of data during the sampling and transmission processes, but also, the reconstructed accelerometer
signals can be used for quantitative assessment without any loss of useful information.
Keywords: compressed sensing; wearable sensor network; quantitative assessment; stroke;
Brunnstrom stage classification

1. Introduction
Stroke, also known as cerebrovascular insult (CVI), cerebrovascular accident (CVA), or brain
attack, is when poor blood flow to the brain results in cell death. Between 1990 and 2010, the number
of strokes which occurred each year decreased by approximately 10% in the developed world and
increased by 10% in the developing world [1]. In 2013, stroke was the second most frequent cause of
death after coronary artery disease, accounting for 6.4 million deaths (12% of the total) [2]. In China,
stroke, with an annual mortality rate of approximately 157 per 100,000, has surpassed heart disease to
become the leading cause of death and adult disability. In addition, China has 2.5 million new stroke
cases each year and 7.5 million stroke survivors [3].
Considering the large population of stroke patients in China and the limited rehabilitation resources
(rehabilitation centers, physicians, therapists), it is an inevitable trend for stroke patients to do rehabilitation

Sensors 2016, 16, 202; doi:10.3390/s16020202

www.mdpi.com/journal/sensors

Sensors 2016, 16, 202

2 of 16

training in home settings. Many previous research results have proven that, in comparison with inpatient
care, home-based rehabilitation shows no difference in the effect on any of the outcomes [4,5]. Moreover,
as there are fewer limitations on the time and space, patients can do rehabilitation training according
to their own schedule in the home settings. However, due to the fact that there are no physicians or
physiotherapists in these home settings to evaluate the motor function of stroke patients and adjust the
prescribed training, how to precisely and automatically assess the motor function without the participation
of physicians has become an important problem that needs to be resolved. Fortunately, the availability of
wearable devices provides a potential approach to solve this problem.
With the development of the Internet of Things (IoT), wearable devices have been widely applied
in the healthcare area. Considering sensor types, inertial measurement sensors such as accelerometers,
gyroscopes, and magnetometers are used individually or together to monitor and analyze the motor
functions of stroke patients [6,7]. In addition, some studies combine inertial measurement sensors with
physiological sensors, like ECG, sEMG, etc. [8,9]. Based on the application scenarios, the following four
categories can be distinguished: fall detection [10,11], physical activity monitoring [1214], movement
recognition [1517] and quantitative assessment [1820]. Particularly in the area of quantitative
assessment for stroke patients, many valuable research results have been published. Patel et al. [21]
proposed a Random Forests-based algorithm to estimate Functional Ability Scale (FAS) scores by using
the signals of six accelerometers placed on the affected arm and the trunk. Based on the same dataset,
Din et al. [20] established a Random Forests model to predict the Fugl-Meyer Assessment (FMA)
scores. Uswatte et al. [22] have shown that just two accelerometers are adequate for assessing whether
rehabilitation has an effect on arm function outside the laboratory. Previous work by our research
group found that accelerometer data can be used to automatically classify the clinical Brunnstrom
stages and some quantitative assessment indexes were designed to evaluate the motor function of
stroke patients [2326].
However, in our previous works, the accelerometer data were wirelessly transmitted from
wearable devices to a receiver using the ZigBee protocol. Because the wearable devices are
battery-operated, the battery life is inversely proportional to the amount of data transmitted. Hence, in
order to extend the battery life and reduce the amount of data during the sampling and transmission
processes, this paper proposes a novel wearable sensor network system based on compressed sensing
technology. The main aim of this paper is to investigate whether the accelerometer signals collected
during the training process of stroke patients could be compressed and whether the reconstruction
errors have any influence on the quantitative assessment models. This paper is organized as follows: in
Section 2, the wearable sensor network, compressed sensing technology and experiment protocols will
be described. The experimental results and corresponding discussion will be described in Section 3.
Finally, the contributions of this work and future work will be presented in Section 4.
2. Materials and Methods
2.1. Compressed Sensing-Based Wearable Sensor Network
As shown in Figure 1, the compressed sensing-based wearable sensor network consists of three
parts: an accelerometer sensor node, ZigBee wireless receiver, and computer. Compared with
traditional accelerometer sensor nodes based on the Nyquist sampling theorem, the sensor node
proposed in this paper has the advantage of compressed sampling, through which the sampling rate
was reduced and the power consumption during the sampling and transmission processes were much
lower than with the traditional method. The ZigBee wireless receiver was connected to the computer
through a USB port and sends the compressed data to the computer. On the computer side, firstly, the
accelerometer signal was reconstructed from the compressed data; secondly, those features that can
represent the motor function of stroke patients were extracted using the time and frequency domain
method; finally, through mapping the features were related to the clinical assessment outcomes given
by physicians, and a quantitative assessment model was built.

Sensors 2016, 16, 202

Sensors 2016, 16, 202

3 of 16
3 of 16

Figure 1.
1. System
of aa compressed
compressed sensing-based
sensing-based wearable
wearable sensor
sensor network.
network.
Figure
System structure
structure of

2.2. Signal Recovery Algorithms

novel signal
signal compression
compression method that depends on the
the
Compressed sensing (CS) [2729] is aa novel
sparsity of signals for compression and reconstruction. The
The following
following Equation
Equation (1) describes the
relationship of
of aa fundamental
fundamental noisy
noisy model:
model:
relationship
yy= x
x`
+v

(1)
(1)

1
In
R N 1 . isisaapart
a raw
accelerometer
signal,
y P RM
data
In this
thispaper,
paper,x P
partofof
a raw
accelerometer
signal,
is the
is compressed
the compressed
that
wirelessly
transmitted
to a remote
receiver
via the
wireless
protocol,
and vand
can be
datawill
thatbe
will
be wirelessly
transmitted
to a remote
receiver
viaZigBee
the ZigBee
wireless
protocol,
v
M N pM ! Nq is a designed sensing matrix that linearly compresses x. Therefore, the
omitted.

P
R
can be omitted.

( ) is a designed sensing matrix that linearly compresses x.

model
usedthe
in model
this paper
a noiseless
expressed
as: expressed as:
Therefore,
usedisin
this papermodel,
is a noiseless
model,

yy= x
x

(2)
(2)
Assume the signal x was sparse under a certain orthogonal space P R N N ; that is, x can be
Assume the signal x was sparse under a certain orthogonal space

; that is, x can be

represented by the following form:
represented by the following form:
x
(3)
(3)
x
=

where = [ 1 , 2 , . . . N ]T was a K-sparse vector pK ! Nq, which means only has K

where elements.
= [1, 2,N]T was a K-sparse vector ( ), which means only has K non-zero
non-zero
elements.
In this paper, for ease of hardware implementation, the sparse Gaussian random matrix was
In this
paper,
ease
hardware
the sparse into
Gaussian
random matrix
was
adopted,
where
the for
value
of of
every
elementimplementation,
was previously embedded
the microcontroller
unit
in
adopted,
where
the value
the
form of
a lookup
table. of every element was previously embedded into the microcontroller unit
in theBased
formon
of athe
lookup
table. proposed by Candes and Donoho [30], the principle of signal recovery
framework
Based
the
framework
proposed
by Candes
and Donoho [30], the principle of signal recovery
was to
solveon
the
following
l1 norm
optimization
problem:
$
was to solve the following l1 norm optimization
problem:
& min||||1

(4)
min
1
%
s.t.y

(4)

s.t. y focused
= on this problem and proposed a series of
have
In the past decades, many researchers
recovery
algorithms,
suchmany
as matching
pursuit
(MP)
[31], on
orthogonal
matching
pursuit (OMP)
[32],
In the
past decades,
researchers
have
focused
this problem
and proposed
a series
of
basis
pursuit
(BP)
[33],
sparse
Bayesian
learning
(SBL)
[34]
and
so
on.
Considering
the
signal
generally
recovery algorithms, such as matching pursuit (MP) [31], orthogonal matching pursuit (OMP) [32],
has
structure
and there
existslearning
intra-block
correlation
elements within
each
basisblock/group
pursuit (BP)
[33], sparse
Bayesian
(SBL)
[34] andamong
so on.the
Considering
the signal
block,
Zhang
et
al.
[35,36]
proposed
a
new
signal
recovery
framework
called
block
SBL
(BSBL)
and
generally has block/group structure and there exists intra-block correlation among the elements
the
results
showed
theetperformance
was obviously
better than
that of
other traditional
methods.
within
each
block, that
Zhang
al. [35,36] proposed
a new signal
recovery
framework
called block
SBL

(BSBL) and the results showed that the performance was obviously better than that of other
traditional methods. Hence, this paper chose the BSBL algorithm to reconstruct the accelerometer

Sensors 2016, 16, 202

4 of 16

Sensors 2016, 16, 202

4 of 16

Hence,
this paper chose the BSBL algorithm to reconstruct the accelerometer signal from compressed
signal from compressed data. The basic principle of the BSBL framework is illustrated in the
data.
The
basic principle of the BSBL framework is illustrated in the Appendix.
Appendix.
2.3. Extreme Learning Machine
2.3. Extreme Learning Machine
Extreme
learning
machine
Huangetetal.
al.[37],
[37],who
whoapplied
applied
it to
Extreme
learning
machine(ELM)
(ELM)was
wasfirstly
firstlyproposed
proposed by
by Huang
it to
thethe
nonlinear
mapping
of
a
single
layer
feedforward
network
(SLFN).
The
SLFN
structure
is
shown
nonlinear mapping of a single layer feedforward network (SLFN). The SLFN structure is shown in in
Figure
2. In
thethe
following
the principle
principleof
ofELM.
ELM.
Figure
2. In
followingsection,
section,we
wewill
willbriefly
briefly introduce
introduce the

Figure2.2.Structure
Structureof
of single
single layer
layer feedforward
Figure
feedforwardnetwork.
network.
T

Assumethere
thereare
areN Nsamples
samplestxi ,y
where xixi rx
xii11 ,, xxii22 ,,
xii,uypii i 1,
1,2,2,
N , where
,,Nq,
Assume
, x, inxin
sTPRn, 1 ,
yi ryi1 , yi2 , , yim sT PTRm1 , m and n are the size of the input and output vector, respectively.
y i yi1 , yi 2 ,, yim , m and n are the size of the input and output vector, respectively.
Equation
(5) is the mathematically expression of a standard single layer feedforward networks (SLFNs),
Equation
is the
mathematically
expression
of a gstandard
single layer feedforward networks
which has L (5)
hidden
neurons
with activation
function
pxq:
(SLFNs), which has L hidden neurons with activation function g x :
H Y
(5)
H = Y
(5)
where Y ry1 , , y N sT P R N m , r1 , , L sT P R Lm and:
where

Y y1 , , y N

1 ,, L

H pw1 , , w L , b1 , , bL , x1 , , x N q

and:
g pw1 x1 ` b1 q . . . g pw L x1 ` b1 q

.
.
.
g w 1 x..1 b1 . . g w L x..1 b1
g pw1 xN ` b1 q
g pw L x N ` b1 q

(6)

H w1 , , w L , b1 , , bL , x1 , , x N
(6)
g w x b g w x b
where wi rwi1 , wi2 , , win sT is the weight vector connecting
the
input neurons
1 N
1
L N and
1 the
ith hidden

neuron, i r i1 , i2 , , im sT Tis the weight vector connecting the output neurons and the ith hidden
w i bi is wthe
is the
weight neuron.
vector connecting the input neurons and the ith
whereand
the
neuron,
threshold
ith hidden
i1 , w
i 2 , , win of
of
T
The training process of an SLFN is simply
the equivalent to finding a least-squares solution
hidden neuron, i i1 , i 2 , , im is the weight vector connecting the output neurons and the
Equation (5):
ith hidden neuron, and bi is the threshold of the ith hidden neuron.
Y|| min||H pw1 , , w L , b1 , , bL q Y||
(7)
||H pw1 , , w L , b1 , , bL q

The training process of an SLFN is simply theequivalent to finding a least-squares solution

to the least-squares regression theory, the best solution of the linear system is:
of According
Equation (5):

H` YH w1 , , w L , b1 , , bL Y
(8)
H w1 ,, w L , b1 ,, bL Y
min
(7)

where H+ is the pseudo inverse of H.

According to the least-squares regression theory, the best solution of the linear system is:

Sensors 2016, 16, 202

5 of 16

Compared with traditional backpropagation neural networks, ELM does not have to iteratively
adjust the connecting weights and bias, as it maps the training process to a problem of solving a group
of linear equations. Besides, in reference [37], Huang et al. have proved that given any small positive
value 0 and activation function g : R R which is infinitely differentiable in any interval, there
r N such that for N arbitrary distinct samples, for any wi and bi randomly chosen, then with
exists N
probability one, ||H Y|| .
In summary, the whole procedure of establishing an ELM model includes three steps, which are
listed in Table 1.
Table 1. Flow of the ELM algorithm.
r
Step 1: Generate random input weight wi and bias bi , i 1, , N.
Step 2: Compute the output of neurons in hidden layer according to Equation (6).
Step 3: Compute the output weight according to Equation (8).

2.4. Experiment Protocols

All the following experiments were approved by the Ethics Committee of Jiaxing 2nd Hospital,
(Jiaxing, China). Twenty-three stroke patients and four physicians were selected to participate in these
experiments. According to the Brunnstrom stage classification criteria, the 23 patients Brunnstrom
stages ranged from II to V. However, considering that patients in stage I cannot move their upper limbs
without extra assistance, patients in stage I were not selected for our experiments. Similarly, due to the
fact that in China, the majority part of stage VI patients leave the hospital due to the high medical costs
and limited rehabilitation resources, we used four physicians instead of stage VI patients to participate
our experiments.
Table 2 lists the general information of the 23 stroke patients, from which it can be clearly seen
that there were 13 males and 10 females, with an age distribution range from 47 to 79. None of
the patients had severe communication and cognitive problems. Before the data collection, all the
experimental processes were demonstrated and the matters needing attention during the experiments
were highlighted in advance by the physicians.
Table 2. The general information of the 23 stroke patients.

Brunnstrom Stage Level

Patients

Sex (M/F)

Hemiplegic Side
(Left/Right)

Limb Dominance
(Left/Right)

II
III
IV
V

2
10
4
7

0/2
5/5
3/1
5/2

2/0
6/4
3/1
2/5

0/2
2/8
0/4
1/6

The accelerometer sensors were placed on the geometric center of the arms, as shown in Figure 3.
The distance between the accelerometer sensor on the forearm and dorsal stripes was 10 cm, while the
distance between the accelerometer sensor on the upper arm and the epicondyus lateralis humeri was
8 cm.

Sensors 2016, 16, 202

6 of 16

Sensors 2016, 16, 202

6 of 16

(a)

(b)

Figure
sensor location.
location.
Figure 3.
3. General
General view
view of
of (a)
(a) the
the accelerometer
accelerometer sensors
sensors and
and (b)
(b) the
the accelerometer
accelerometer sensor

Generally,
motioncapturing
capturing
and
monitoring
applications,
reconstruction
performance
is
Generally, ininmotion
and
monitoring
applications,
reconstruction
performance
is often
often
evaluated
by comparing
the reconstructed
signals
withones
raw through
ones through
the mean
evaluated
by comparing
the reconstructed
signals
with raw
the mean
squaresquare
error
error (MSE).
However,
in thisthe
paper,
signal reconstruction
thesofinal
goal, so the
(MSE).
However,
in this paper,
signalthe
reconstruction
was not thewas
finalnot
goal,
the reconstructed
reconstructed
accelerometer
signals
were further
processed
to establish
an automatic
Brunnstrom
accelerometer signals
were further
processed
to establish
an automatic
Brunnstrom
stage classification
stage
classification
model.
Due
to
the
infidelity
of
the
MSE
for
structured
signals,
it
is
difficult
to
model. Due to the infidelity of the MSE for structured signals, it is difficult to analyze
how the
analyze how
the predictive
accuracy of
the quantitative
is affected by
the
predictive
accuracy
of the quantitative
assessment
model isassessment
affected bymodel
the reconstruction
errors
reconstruction
errorsHence,
measured
MSE. Hence,
it is necessary
to compare
the predictive
accuracy
of
measured by MSE.
it is by
necessary
to compare
the predictive
accuracy
of Brunnstrom
stage
Brunnstrom
stage
classification
models,
which
are
built
based
on
the
reconstructed
and
raw
classification models, which are built based on the reconstructed and raw accelerometer signals,
accelerometer
signals, respectively.
Consequently,
toof
achieve
our goal, all
of the
experimentsinwere
respectively. Consequently,
to achieve
our goal, all
the experiments
were
implemented
two
implemented
in
two
actions:
(1)
Bobath
handshake;
(2)
shoulder
touch.
actions: (1) Bobath handshake; (2) shoulder touch.
2.4.1. Bobath Handshake
experiment
is toisvalidate
the reconstruction
performance
of the BSBL
The purpose
purposeofofthis
this
experiment
to validate
the reconstruction
performance
of algorithm
the BSBL
during the during
Bobath the
handshake
Theexercise.
whole process
of theprocess
Bobath handshake
exercise
can be
algorithm
Bobath exercise.
handshake
The whole
of the Bobath
handshake
divided into
thedivided
following
steps:
exercise
can be
intothree
the following
three steps:
(1)
(2)
(3)

(1)
Sit down
a chair,
the hand
and keep
the thumb
the hemiplegic
side
on top.
Sit down
on a on
chair,
crosscross
the hand
and keep
the thumb
of theofhemiplegic
side on
top.
(2)
Straighten
the
upper
extremities,
lift
above
the
head
and
hold
for
3
s.
Straighten the upper extremities, lift above the head and hold for 3 s.
(3)
Move
the hands
toinitial
the initial
position.
Move
the hands
backback
to the
position.

An additional movie file shows the Bobath handshake exercise in more detail (see
An additional movie file shows the Bobath handshake exercise in more detail (see
Supplementary File 1).
Supplementary File 1).
2.4.2.
2.4.2. Shoulder
Shoulder Touch
Touch
Our
previous studies
studies have
found that
that the
the accelerometer
accelerometer signals
signals of
of the
the shoulder
shoulder touch
touch exercise
exercise
Our previous
have found
can
be applied
applied to
to automatically
automatically classify
classify the
the Brunnstrom
Brunnstrom stage
stage for
for stroke
stroke patients.
patients. To
investigate
can be
To investigate
whether
the
reconstruction
errors
affect
the
predictive
accuracy
of
the
quantitative
assessment
whether the reconstruction errors affect the predictive accuracy of the quantitative assessment model,
model,
the shoulder
touch was
exercise
was completed
by every participant
order
collect the
the shoulder
touch exercise
completed
by every participant
in order to in
collect
thetocompressed
compressed
accelerometer
whole
of the touch
shoulder
touchcan
exercise
can be into
divided
accelerometer
signals. Thesignals.
whole The
process
of process
the shoulder
exercise
be divided
the
into
the
following
four
steps:
following four steps:
(1) Sit down on a chair, and naturally droop the upper limb of the hemiplegic side.
Sit down on a chair, and naturally droop the upper limb of the hemiplegic side.
(2) Raise the upper limb of the hemiplegic side to the horizontal position.
Raise the upper limb of the hemiplegic side to the horizontal position.
(3) Move horizontally to the healthy side shoulder and hold for 5 s.
Move horizontally to the healthy side shoulder and hold for 5 s.
(4) Move back to the initial droop position and take a short break.
Move back to the initial droop position and take a short break.
An additional movie file shows the shoulder touch exercise in more detail [see Supplementary
An additional movie file shows the shoulder touch exercise in more detail [see Supplementary File 2].
File 2].
All patients were required to finish these exercises without extra assistance to reflect their true
movement function. Before the data collection, they were asked to practice several times with the
(1)
(2)
(3)
(4)

Sensors 2016, 16, 202

7 of 16

SensorsAll
2016,
16, 202
patients

of 16
were required to finish these exercises without extra assistance to reflect their7 true
movement function. Before the data collection, they were asked to practice several times with the
guidance
that they
they were
were familiar
familiar with
with the
the whole
whole process.
process. During
During the
the experimental
experimental
guidance of
of physicians
physicians so
so that
processes,
they
were
requested
to
repeat
the
shoulder
touching
exercise
eight
times.
processes, they were requested to repeat the shoulder touching exercise eight times.
The
data sampling
sampling and
and management
management were
were implemented
implemented by
by using
using the
the Remote
Remote Rehabilitation
Rehabilitation
The data
Training
and
Assessment
Software
(RRTAS),
which
was
developed
by
our
group.
The run
Training and Assessment Software (RRTAS), which was developed by our group. The run environment
environment
of
RRTAS
is
Windows
32
bit
platform
and
.Net
Framework
3.5
or
above.
An
additional
of RRTAS is Windows 32 bit platform and .Net Framework 3.5 or above. An additional document file
document
introduces
in more
detail (see Supplementary
File 3).
introduces file
RRTAS
in moreRRTAS
detail (see
Supplementary
File 3).

3. Results
Resultsand
andDiscussion
Discussion
3.1. Accelerometer Signals Compress and Recovery
The raw accelerometer signal of the Bobath handshake exercise is shown in Figure 4, in which
the axes X1,
X1, Y1
Y1and
andZ1
Z1are
arethe
thesignals
signalsofofsensors
sensorsplaced
placed
forearm,
while
axes
onon
thethe
forearm,
while
thethe
axes
X2,X2,
Y2 Y2
andand
Z2
Z2 are
signals
of sensors
placed
onupper
the upper
arm. Figure
From Figure
4, we
that during
the
are
the the
signals
of sensors
placed
on the
arm. From
4, we can
seecan
thatsee
during
the Bobath
Bobath handshake
exercise,
the axes
from
accelerometer
sensorsshow
showperiod
period changes.
changes.
handshake
exercise,
all six all
of six
theofaxes
from
twotwo
accelerometer
sensors
However, 5400 raw sampling data in each axis need to be transmitted to the computer through the
ZigBee wireless protocol during all eight Bobath
Bobath handshake
handshake exercises.
exercises.
Raw Accelerometer Signal
X1

1
0

Y1

-1
1
0

Z1

-1
1
0

X2

-1
1
0

Y2

-1
1
0

Z2

-1
1
0
-1

900

1800

2700
Time Point

3600

4500

5400

Figure
Figure 4.
4. Raw
Raw accelerometer
accelerometer signals.
signals.

As mentioned above, by adjusting the dimension of sensing matrix , the compression ratio
As mentioned above, by adjusting the dimension of sensing matrix , the compression ratio (CR)
(CR) can be changed as follows:
can be changed as follows:
NN M
M
MM
CR
(9)
CR N 1 N
(9)
N
N
Here, we first set the CR to 0.7222 to validate the reconstruction performance of the BSBL
Here, we first set the CR to 0.7222 to validate the reconstruction performance of the BSBL
algorithm. This means that the raw accelerometer signal was compressed to 1500 sampling data.
algorithm. This means that the raw accelerometer signal was compressed to 1500 sampling data. The
The corresponding compressed signal is shown in Figure 5, where it can be seen that the compressed
corresponding compressed signal is shown in Figure 5, where it can be seen that the compressed
signal looks very different from the raw accelerometer signal due to the randomness of the sensing
signal looks very different from the raw accelerometer signal due to the randomness of the sensing
matrix . This characteristic indicates that compressed sensing can not only compress the raw signal
matrix . This characteristic indicates that compressed sensing can not only compress the raw
into a lower dimension compressed signal, but also can encrypt the raw signal so that the privacy of
signal into a lower dimension compressed signal, but also can encrypt the raw signal so that the
the patients is protected.
privacy of the patients is protected.

Sensors 2016, 16, 202

8 of 16

Sensors 2016, 16, 202

8 of 16

Sensors 2016, 16, 202

8 of 16

Z2

Z2

Y2 Y2

X2 X2

Z1

Z1

Y1 Y1

X1 X1

Compressed accelerometer signal

50
0
50
-50
00
0
-50
-200
0
-40
-20 0
40
-40
200
40
0
20 0
0
0
-200
-40
0
-20 0
-40
0
-200
-40
0
-20 0
-40
40
200
40
0
20 0
0
0

Compressed accelerometer signal

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

500

1000

1500

1000

1500

1000

1500

500
500

Time Point

Time
Point
Figure5.5.Compressed
Compressed
accelerometer
signals.
Figure
accelerometer
signals.
Figure 5. Compressed accelerometer signals.

From Figure 4, it is obvious that the raw accelerometer signal is far from sparse and periodic.
From
Figure
4, it difficulty
is obvious that
the raw
signal to
is reconstruct
far from sparse
and
This
creates
a huge
CS accelerometer
recoverysignal
algorithms
the raw
From
Figure
4, it is obvious for
thattraditional
the raw accelerometer
is far from
sparse and periodic.
periodic.
This
creates
a
huge
difficulty
for
traditional
CS
recovery
algorithms
to
reconstruct
the
accelerometer
Fortunately,
the BSBL recovery
framework
proposed
by Zhangthe
[35,36]
This
creates a signals.
huge difficulty
for traditional
CS recovery
algorithms
to reconstruct
raw
raw
accelerometer
signals. Fortunately,
the BSBL recovery
proposed
by Zhang
[35,36]
significantly
outperforms
tradition CS
thanksframework
toframework
its ability
to
explore
exploit
an
accelerometer
signals. Fortunately,
thealgorithms
BSBL recovery
proposed
by and
Zhang
[35,36]
significantly
outperforms
tradition
CS
algorithms
thanks
to
its
ability
to
explore
and
exploit
intra-block correlation
signals. CS
When
the prior
block
partition
significantly
outperformsintradition
algorithms
thanks
to its
ability is
to given,
explorethe
andwhole
exploitraw
anan
intra-block
correlation
in
signals.
When
the
prior
block
partition
is
given,
the
whole
raw
accelerometer
accelerometer
signal is divided
intoWhen
30 blocks
equal
sizepartition
(each block
containsthe5400/30
180
intra-block
correlation
in signals.
the of
prior
block
is given,
whole= raw
signal
is
divided
into
30
blocks
of
equal
size
(each
block
contains
5400/30
=
180
sampling
data).
sampling
data).
The
reconstructed
accelerometer
signals
are
shown
in
Figure
6.
To
validate
the
accelerometer signal is divided into 30 blocks of equal size (each block contains 5400/30 = 180
Thesampling
reconstructed
accelerometer
signals
are
shown
in
Figure
6.
To
validate
the
recovery
performance
recovery
performance
of
BSBL
algorithm,
we
also
reconstructed
the
compressed
signals
with
BP
data). The reconstructed accelerometer signals are shown in Figure 6. To validate the
and
OMP
algorithms,
and
the
correlation
coefficients
of
the
reconstructed
and
raw
signals
of
each
of BSBL
algorithm,
we
also
reconstructed
the
compressed
signals
with
BP
and
OMP
algorithms,
and
recovery performance of BSBL algorithm, we also reconstructed the compressed signals with BP
algorithm
all listedand
inofTable
3. It is clear
that
the of
BSBL
theand
correlation
coefficients
thecorrelation
reconstructed
andby
raw
signals
eachalgorithm,
algorithm
arecorrelation
alloflisted
OMP are
algorithms,
the
coefficients
ofusing
the
reconstructed
and raw the
signals
each in
coefficients
of all
the
reconstructed
signals
higherthe
than
with
the other
two
methods,
Table
3. It is clear
that
by using
theand
the using
correlation
coefficients
of the
reconstructed
algorithm
are
listed
in Table
3.BSBL
It raw
isalgorithm,
clear
thatwere
by
BSBL
algorithm,
the
correlation
which
suggests
thatreconstructed
the BSBL
algorithm
isother
very
effective
for
accelerometer
recovery.
coefficients
ofwere
the
andthe
raw
signals
were
higher
thansuggests
withsignal
thethat
other
and
raw signals
higher
than
with
two
methods,
which
the two
BSBLmethods,
algorithm
which
suggestsfor
that
the BSBL algorithm
is very effective for accelerometer signal recovery.
is very
effective
accelerometer
signal recovery.
X1 X1
Y1 Y1
Z1 Z1
X2 X2
Y2 Y2
Z2 Z2

0.5

Reconstructed Accelerometer Signal

0
1
-1
0
1
-1
0
1
-1
0
1
-1
0
1

0
0.5
-0.5
0
0.5
-0.5
0
0.5
-0.5
0
0.5
-0.5
0
0.5

-1
0
1
-1
0
1

-0.5
0
0.5
-0.5
0
0.5

-1
0
1
-1
0
1
-1
0
1
-1
0
1
-1
0

-0.5
0
0.5
-0.5
0
0.5
-0.5
0
0.5
-0.5
0
0.5
-0.5
0
5400
-0.5
5400

-1

900

1800

900

1800

2700
Time Point
2700
Time Point

3600

4500

3600

4500

Figure 6. Reconstructed accelerometer signals and absolute errors (AE).

Figure6.6.Reconstructed
Reconstructedaccelerometer
accelerometer signals
Figure
signals and
andabsolute
absoluteerrors
errors(AE).
(AE).

AEZ2AE AEY2AE AEX2AE AEZ1AE AEY1

AEY1 AEX1AEX1
Z2
Y2
X2
Z1

Reconstructed Accelerometer Signal

Sensors 2016, 16, 202

Sensors 2016, 16, 202

9 of 16
9 of 16

Table 3. Signal
BPBP
and
OMP
algorithms.
Signalrecovery
recoveryresults
resultsofofBSBL,
BSBL,
and
OMP
algorithms.
Correlation Coefficients between Reconstructed and Raw Signals
Correlation Coefficients between Reconstructed and Raw Signals
Methods
Methods X1
Y1
Z1
X2
Y2
Z2
X1
Y1
Z1
X2
Y2
Z2
BSBL
0.9991
0.9957
0.9965
0.9991
0.9979
0.9969
BSBL 0.87620.9991
0.9957
0.9965
0.9991
0.9979
BP
0.8814
0.8729
0.8651
0.8964 0.99690.9015
BP
0.8762
0.8814
0.8729
0.8651
0.8964
0.9015
OMP
0.9356
0.9521
0.9188
0.9672
0.9248
0.9366
OMP
0.9356
0.9521
0.9188
0.9672
0.9248
0.9366

3.2. Effects of Block Size and CR on Recovery Performance

3.2. Effects of Block Size and CR on Recovery Performance
In this section, we will discuss the effects of block size and compression ratio on the quality of
In this
section, we
will
discuss
the effects
of block
and compression
ratio ratio
on the
quality
signals
reconstructed
from
raw
accelerometer
signals.
Here,size
we chose
the signal to noise
(SNR)
ofassignals
reconstructed
raw accelerometer
signals.
Here,
chose the
signal
noiseinratio
the evaluation
index from
to analyze
the reconstruction
error.
The we
definition
of SNR
is to
shown
Equation
(10):
(SNR)
as the
evaluation index to analyze the reconstruction error. The definition of SNR is shown in
Equation (10):
2

x2
k x k222
SNR 20lg
x x^ 2
k x x2 k2
SNR 20 lg

(10) (10)

where x and x are

the raw and reconstructed signals, respectively. A higher SNR means a smaller
x
x
and
where
reconstruction error.are the raw and reconstructed signals, respectively. A higher SNR means a
smaller
error.
Thereconstruction
effects of block
size on SNR are shown in Figure 7. It is clear that the trends of all six axes
The
effects
of
block
size onsize
SNRincreases,
are shown
in SNR
Figure
7. Itincreases.
is clear that
trends ofwhen
all sixthe
axes
are same; that is, as the block
the
also
Inthe
particular
block
are
same;
that
is,
as
the
block
size
increases,
the
SNR
also
increases.
In
particular
when
the
block
size
size changes from 5 to 20, the corresponding SNR increases quickly. This is due to the fact that the
changes from 5 to 20, the corresponding SNR increases quickly. This is due to the fact that the
sampling rate of the raw accelerometer sensor is 50 Hz, and for stroke patients the duration of the
sampling rate of the raw accelerometer sensor is 50 Hz, and for stroke patients the duration of the
Bobath handshake exercise is about 12 s. Hence, if the block size is too small (for example, a block size
Bobath handshake exercise is about 12 s. Hence, if the block size is too small (for example, a block
less than 10), the signals in each block are almost unchanged because the duration of each block is very
size less than 10), the signals in each block are almost unchanged because the duration of each block
short. Consequently, all of the intra-block correlations of each block are close to 1 and the variances are
is very short. Consequently, all of the intra-block correlations of each block are close to 1 and the
close
to 0. are close to 0.
variances
Although
performanceisisaffected
affectedbyby
the
block
size
in the
block
partition,
Although the
the reconstruction
reconstruction performance
the
block
size
in the
block
partition,
this this
does
not
limit
its
application
to
the
wearable
sensor
network.
Figure
7
also
shows
that
a
wide
range
does not limit its application to the wearable sensor network. Figure 7 also shows that a wide range of
the
block
sizesize
(for(for
example,
from
2020
toto
40)
can
of the
block
example,
from
40)
canlead
leadtotosatisfying
satisfyingresults.
results.
Effects of the block size on SNR
160

Signal to Noise Ratio (SNR)

140
120
100
80
60

Axis
Axis
Axis
Axis
Axis
Axis

40
20
0

10

15

20
25
Block Size

30

35

X1
Y1
Z1
X2
Y2
Z2
40

Figure 7. Effects of the block size on SNR.

Figure 7. Effects of the block size on SNR.

Figure88illustrates
illustrates the
which
wewe
cancan
seesee
thatthat
when
the the
CR increases,
Figure
the effects
effectsof
ofCR
CRon
onSNR,
SNR,from
from
which
when
CR increases,
the SNR decreases, especially when the CR is higher than 0.75. That is to say, if the CR does not
the SNR decreases, especially when the CR is higher than 0.75. That is to say, if the CR does not exceed

Sensors 2016, 16, 202

10 of 16

Sensors 2016, 16, 202

10 of 16

Sensors
16, 202
10 of 16
0.75
(the2016,
length
of compressed
signal cannot
be less
thanbe
1350),
satisfactory
reconstructed
exceed
0.75 (the
length of compressed
signal
cannot
less athan
1350), a quality
satisfactory
quality
signal
can be achieved.
reconstructed
signal can be achieved.
exceed 0.75 (the length of compressed signal cannot be less than 1350), a satisfactory quality
reconstructed signal can be achieved. Effects of the compression ratio on SNR
180
Axis X1
Axis Y1
Axis
Axis X1
Z1
Axis
Axis Y1
X2
Axis
Axis Z1
Y2
Axis
Axis X2
Z2
Axis Y2
Axis Z2

Effects of the compression ratio on SNR

180
160

SignalSignal
to Noise
Ratio (SNR)
to Noise
Ratio (SNR)

160
140
140
120
120
100
100
80
80
60
60
40
40
20
200
0

0.5

0.6

0.5

0.6

0.7
0.8
Compression Ratio
0.7
0.8
Compression
Ratio ratio
the
compression

0.9

0.9

Figure8.8.Effects
Effectsof
of the compression ratioon
Figure
onSNR.
SNR.
Figure 8. Effects of the compression ratio on SNR.

According
compressedsensing
sensingtheory,
theory, the
the relationship
relationship between
satisfy
thethe
According
toto
compressed
betweenM
Mand
andNNshould
should
satisfy
following
inequality:
following
inequality:
According
to compressed sensing theory, the relationship between M and N should satisfy the
N
following inequality:
N

M
K
log
(11)(11)
M Klog K
K
N
M K log
(11)
K the
where
K the
is the
sparsityofofthe
theraw
rawsignal.
signal. Now
Now we
we validate
validate
where
K is
sparsity
the theoretical
theoreticalCR
CRofofthe
theaccelerometer
accelerometer
signals. Figure9 9shows
shows the sparsity
sparsity discrete
discrete cosine transform
(DCT)
coefficients
ofofthetheraw
signals.
transform
(DCT)CR
coefficients
raw
where Figure
K is the sparsity ofthe
the raw signal.
Now we validate
the theoretical
of the accelerometer
accelerometer
signal.
Clearly,
only
a
few
DCT
coefficients
(K

600)
have
large
amplitudes,
while
the
accelerometer
signal.
Clearly,
a few DCT
coefficients
(K 600) (DCT)
have large
amplitudes,
while
signals. Figure
9 shows
theonly
sparsity
discrete
cosine transform
coefficients
of the
rawthe
majority
of
coefficients
havesmall
smallamplitudes.
amplitudes. Hence,
Hence, the
the theoretical
maximum
CR
should
be:
majority
of
coefficients
have
theoretical
maximum
CR
should
be:
accelerometer signal. Clearly, only a few DCT coefficients (K 600) have large amplitudes, while the
majority of coefficients have small amplitudes.
Hence,
theoretical
maximum CR should be:
600
logthe
/ 600
5400
M
CR 1M
min 1 600 log p5400{600q
0.7559
(12)
CRmax max1 min
1 600 log5400
0.7559
(12)
N
5400
/ 600

M
N
5400
min
CRmax 1
1
0.7559
(12)
N
5400
Sparsity of Raw
Signal
20
Sparsity of Raw Signal
20
10
10
0
DCT Coefficients
DCT Coefficients

20
0
-10

200

-10
-20

0
-20

-20
-30

-20
-40

-30
-40

-40
-60

-40
-50
-50

-60
0

900

1800

900

1800

200

400

600

200
400
600
2700
3600
Time Points
2700

3600

800

1000

800
1000
4500
5400
4500

5400

Figure 9. Sparsity of raw Time

accelerometer
signal (axis X1).
Points
Figure 9. Sparsity of raw accelerometer signal (axis X1).

Figurethe
9. Sparsity
raw accelerometer
signal
This result is essentially
same as of
Figure
8, which means
that(axis
the X1).
more sparser the raw signal
is, the higher CR can be achieved.
This result is essentially the same as Figure 8, which means that the more sparser the raw signal
This result is essentially the same as Figure 8, which means that the more sparser the raw signal
is, the higher CR can be achieved.
is, the higher CR can be achieved.

Sensors 2016, 16, 202

11 of 16

Sensors 2016, 16, 202

3.3. Effects
of Compressed Sensing on Quantitative Assessment Model

11 of 16

Effects
of Compressed
Sensing
Quantitative
Model
In3.3.
the
previous
experiment,
weonshowed
thatAssessment
the accelerometer
signals of rehabilitation exercises
can be compressed
and
precisely
reconstructed.
Another
question
is
whether
the reconstructed
signals
In the previous experiment, we showed that the accelerometer signals of rehabilitation
exercises
affect can
the predictive
accuracy
of the quantitative
assessment
Toisanswer
this,
carried out the
be compressed
and precisely
reconstructed.
Anothermodel.
question
whether
thewe
reconstructed
signals
affect the predictive
accuracy
of the
quantitative
model.
we to
following
experiment:
as mentioned
above,
there
are manyassessment
assessment
scalesToinanswer
clinicalthis,
settings
carried
out the following
experiment:
as mentioned
above,
there are many
assessment
scales
in
evaluate
the movement
function
of stroke patients,
such as
the Fugl-Meyer
assessment
scale,
Functional
clinical
settings
to
evaluate
the
movement
function
of
stroke
patients,
such
as
the
Fugl-Meyer
ability scale, Brunnstrom stage classification, Action Research Arm Test, and so on. Compared with
Functional ability scale, Brunnstrom stage classification, Action Research Arm
other assessment
assessmentscale,
scales,
the Brunnstrom stage classification tool only has six levels, hence it has the
Test, and so on. Compared with other assessment scales, the Brunnstrom stage classification tool
advantage of being easy to use and time saving. In our previous work [26], we have proven that the
only has six levels, hence it has the advantage of being easy to use and time saving. In our previous
Brunnstrom
stage of stroke patients can be automatically classified through the shoulder touch exercise
work [26], we have proven that the Brunnstrom stage of stroke patients can be automatically
by using
the
ELM
method.
classified
through
the shoulder touch exercise by using the ELM method.
Brunnstrom Stage III

Accelerometer(Axis Y1)

Accelerometer(Axis Y1)

Brunnstrom Stage II

-0.5

-1

50

100
Time Point

150

-0.4
-0.6
-0.8
-1
0

200

50

100
Time Point

(a)

(b)

Accelerometer(Axis Y1)

Accelerometer(Axis Y1)

0.5

-0.5

-1
50

100
Time Point

200

Brunnstrom Stage V

Brunnstrom Stage IV

150

150

200

(c)

-0.5

-1
0

50

100
Time Point

150

200

(d)

Accelerometer(Axis Y1)

Brunnstrom Stage VI
0

-0.5

Raw Signal
Reconstructed Signal

-1
0

50

100
Time Point

150

200

(e)
Figure 10. (a)(e) Reconstructed and raw accelerometer signals from Brunnstrom stages II to VI.

Figure 10. (a)(e) Reconstructed and raw accelerometer signals from Brunnstrom stages II to VI.

Due to space limitations, here we present only the reconstructed and raw axis Y1 accelerometer

Due to space limitations, here we present only the reconstructed and raw axis Y1 accelerometer
signals ranges from Brunnstrom stage II to VI in Figure 10. It is clear that as the Brunnstrom stage
signals ranges from Brunnstrom stage II to VI in Figure 10. It is clear that as the Brunnstrom stage
increases, the reconstruction error decreases. This is because the lower the Brunnstrom stage, the
increases,
thethe
reconstruction
This
thefinish
lower
Brunnstrom
stage, the
weaker
stroke patientserror
motordecreases.
function, and
heis
orbecause
she cannot
thethe
shoulder
touch exercise
weaker the stroke patients motor function, and he or she cannot finish the shoulder touch exercise

Sensors 2016,
2016, 16,
16, 202
202
Sensors

12of
of16
16
12

freely. Additionally, with the Brunnstrom stage decreases, the smoothness of accelerometer signals
freely.
Additionally,
with
the Brunnstrom
stage
the smoothness of accelerometer signals
becomes
worse, which
brings
about difficulty
in decreases,
signal recovery.
becomes
worse, which
difficulty in signal
recovery.
To investigate
the brings
effectsabout
of reconstruction
error on
the predictive accuracy of the Brunnstrom
investigate model,
the effects
of reconstruction
error on the
predictivemodel
accuracy
thereconstructed
Brunnstrom
stageTo
classification
we established
the quantitative
assessment
withofthe
stage
classification
model,
we established
the quantitative
modelofwith
reconstructed
and raw
accelerometer
signals,
respectively.
To ensure assessment
the consistency
our the
comparison,
the
and
raw
accelerometer
signals,
respectively.
To
ensure
the
consistency
of
our
comparison,
the
feature
feature extraction and ELM parameters were all set to be the same. More specifically, 12 features
extraction
wereamplitude
all set to be(AMP)
the same.
More
specifically,
12 features
including
root
including and
root ELM
meanparameters
square (RMS),
from
three
axes of two
accelerometer
sensors
mean
square (RMS),
amplitude
(AMP)
from
three the
axesoutput
of twoofaccelerometer
sensors
were
extracted
were extracted
and set
as the input
of ELM
model,
the ELM model
was the
Brunnstrom
and
as the
input
ELM model,
the output
of was
the ELM
was
the Brunnstrom
stage
and
stagesetlevel,
and
the of
number
of hidden
neurons
set tomodel
30. The
training
set contains
152level,
samples
the
of hidden
neurons
was set to 30. The training set contains 152 samples and the testing set
andnumber
the testing
set contains
38 samples.
contains
samples.
The 38
results
are shown in Figure 11 and Table 4, from which it can be seen that of the total 38
The in
results
are shown
Figure
11 and
Table 4, whose
from which
it can be
seen were
that ofaccurately
the total
samples
the testing
set, inthere
were
34 samples
Brunnstrom
stages
38
samples
the testing
set, there were
samples
whose Brunnstrom
stages
were accurately
predicted
byinusing
the reconstructed
signal,34while
35 samples
were accurately
predicted
using the
predicted
usingindicated
the reconstructed
signal, while 35
samples
accurately
using the
raw signal,bywhich
that the reconstruction
error
alonewere
has no
influencepredicted
on the accuracy
of
raw
signal, whichassessment
indicated that
the reconstruction
aloneofhas
no influence
on theresult,
accuracy
of the
the quantitative
model.
Actually, the error
analysis
variance
(ANOVA)
listed
in
quantitative
assessment
model.
the analysis
of difference
variance (ANOVA)
result,
listed inaccuracy
Table 5,
Table 5, indicates
that there
is noActually,
statistically
significant
between the
predictive
indicates
is no statistically
significant
between
the
predictive accuracy
the raw
of the rawthat
andthere
reconstructed
Brunnstrom
stage difference
classification
model.
Consequently,
we canofconclude
and
Brunnstrom
stage
Consequently,
can conclude
that the
that reconstructed
the reconstruction
error has
no classification
influence on model.
the establishment
of awe
quantitative
assessment
reconstruction
error has no influence on the establishment of a quantitative assessment model.
model.

Quantitative Assessment Model Accuracy (%)

Comparison of compressed and raw signals on quantitative assessment model accuracy

100
Raw Signal
Compressed Signal
95

90

85

80

75

70

II

III

IV
V
Brunnstrom Stage

VI

Total

Figure
accuracy.
Figure 11.
11. Comparison
Comparison of
of compressed
compressed and
and raw
raw signals
signals on
on quantitative
quantitative assessment
assessment model
model accuracy.
Table
raw signals
signals on
on quantitative
quantitative assessment
assessment model
model accuracy.
accuracy.
Table 4.
4. Comparison
Comparison of
of compressed
compressed and
and raw

Brunnstrom Stage

Samples in Testing Set

II

Brunnstrom Stage

III
IV
V
VI

II
III
IV
V
VI
Total

Total

Predictive Accuracy

Samples in Testing Set

Accuracy
Raw Model Predictive
Compressed
Sensing Model
Raw Model
Compressed Sensing Model

100 (2/2)

100 (2/2)

2
14 14
9 9
9
9 4
4 38

100 (2/2)
85.7
85.7 (12/14)
(12/14)
88.8 (8/9)
(8/9)
88.8
100 (9/9)
100
100 (9/9)
(4/4)
92.1
100(35/38)
(4/4)

100 (2/2)
78.6 (11/14)
78.6 (11/14)
100 (9/9)
100 (9/9)
88.8 (8/9)
88.8 (8/9)
100 (4/4)
89.5 (34/38)
100 (4/4)

38

92.1 (35/38)

89.5 (34/38)

Table 5. Analysis of variance results.

Source

Table 5. Analysis of variance results.

Sum of Squares
df
Mean Square

Source
Sum of Squares
Between Groups
0.0132
Between Within
GroupsGroups
0.013296.9737
Total
96.9868
Within Groups
96.9737
Total
96.9868

df 1
1 74
74 75
75

Prob > F

Mean
Square 0.01
F 0.9205 Prob > F
0.01316
1.31046
0.01316
0.01
0.9205
1.31046

Sensors 2016, 16, 202

13 of 16

4. Conclusions
To reduce the amount of data in a wearable sensor network during the sampling and transmission
processes, this paper proposes a novel wearable sensor network based on compressed sensing
technology, and applied it to the monitoring and quantitative assessment of stroke patients upper
limb motor function. The experimental results showed that the proposed method can not only
compress and precisely reconstruct the raw accelerometer signals, but can also apply the reconstructed
signals to establish the Brunnstrom stage automatic classification model. It also indicated that the
proposed system can provide a theoretical basis for individualized and remote rehabilitation. However,
considering the fact that the clinical experiment only recruited 23 stroke patients, which is too small
a number to verify the statistical reliability and validity of the proposed quantitative assessment
model, in the future, we will gather more clinical data and improve the generalization performance
of the model. Additionally, in this paper, the compressed signals need to be reconstructed on the
computer side before the analysis process. In the future, we will try to investigate whether the analysis
process can be directly implemented in the compressed domain. If yes, this could reduce the off-line
computation burden and make the on-node analysis more possible.
Supplementary Materials: Additional File 1: Bobath handshake exercise (Movies, MP4 format), Additional File 2:
Shoulder touch exercise (Movies, MP4 format), Additional File 3: Introduction of Remote Rehabilitation Training
and Assessment Software.
Acknowledgments: This research is partly supported by the Jiangsu Province Science and Technology Support
Program (Funds Number: BE2012654). The authors thank Drs. Xudong Gu and Jianming Fu who come from the
Rehabilitation Medical Center of Jiaxing 2nd Hospital, for their valuable suggestions and guidance during the
clinical experiment design and implementation processes. The authors also thank Zhilin Zhang for guidance and
suggestions about the compressed sensing and BSBL algorithm.
Author Contributions: LY carried out the study design, signal processing and manuscript drafting. DxX
participated in the design of the study. LqG carried out the design of the wearable sensor network and
accelerometer sensor node. JpW participated in the software and algorithm design. All authors read and
approved the final manuscript.
Conflicts of Interest: The authors declare that they have no competing interest.

Abbreviations
The following abbreviations are used in this manuscript:
CS
WSN
UE
WMFT
FAS
RIP
MP
OMP
BP
SBL
BSBL
CR
SNR
ELM
DCT
RRTAS

Compressed Sensing
Wearable Sensor Network
Upper Extremity
Wolf Motor Function Test
Functional Ability Scale
Restricted Isometry Property
Matching Pursuit
Orthogonal Matching Pursuit
Basis Pursuit
Sparse Bayesian Learning
Block Sparse Bayesian Learning
Compression Ratio
Signal to Noise Ratio
Extreme Learning Machine
Discrete Cosine Transform
Remote Rehabilitation Training and Assessment Software

Sensors 2016, 16, 202

14 of 16

Appendix
Without loss of generality, assume x can be divided into g blocks, i.e.:
T
x rx
1 , , xd1 , , xd g1 `1 , , xd g s
looooomooooon
looooooooomooooooooon
x1T

xTg

where di p@iq are not necessarily identical. Among the g blocks, only k pk ! gq blocks are nonzero but
their locations are unknown.
Each block xi PRdi 1 is assumed to satisfy a parameterized multivariate Gaussian distribution:
p pxi ; i , Bi q N p0, i Bi q , i 1, , g
where i is a non-negative parameter controlling the block-sparsity of x. When i =0, the ith block
becomes zero. Bi PRdi di is a positive definite matrix, capturing the correlation structure of the ith
block. Under the assumption that blocks are mutually uncorrelated, the prior of x is:
p px; ti , Bi ui q N p0, 0 q
(
where 0 =diag 1 B1 , , g Bg . Assume the noise vector satisfies p pv; qN p0, Iq, where is a
positive scalar. Therefore the posterior of x is given by:

p xy; , ti , Bi ui1 N p x , x q
with:

1
x 0 T I ` 0 T
y

0 `

1
1 T

Once the parameters , ti , Bi ui1 are estimated, the Maximum-a-Posteriori (MAP) estimate of x,
^

denoted by x, can be directly obtained from the mean of the posterior, i.e.:

1
^
x 0 T I ` 0 T
y
The parameters can be estimated by a Type II maximum likelihood procedure. This is equivalent
to minimizing the following cost function:

2log p py|x; qp px; ti , Bi ui q dx

1
L pq fi

log I ` 0 T ` yT I ` 0 T
y
!
)
g
where denotes all the parameters, i.e., fi , ti , Bi ui1 .
References
1.

2.

3.

Feigin, V.L.; Forouzanfar, M.H.; Krishnamurthi, R.; Mensah, G.A.; Connor, M.; Bennett, D.A.; Moran, A.E.;
Sacco, R.L.; Anderson, L.; Truelsen, T.; et al. Global and regional burden of stroke during 19902010: Findings
from the Global Burden of Disease Study 2010. The Lancet 2014, 383, 245255. [CrossRef]
Naghavi, M.; Wang, H.; Lozano, R.; Davis, A.; Liang, X.; Zhou, M.; Vollset, S.E.; Ozgoren, A.A.; Abdalla, S.;
Abd-Allah, F.; et al. Global, regional, and national age-sex specific all-cause and cause-specific mortality
for 240 causes of death, 19902013: A systematic analysis for the Global Burden of Disease Study 2013.
The Lancet 2015, 385, 117171.
Liu, L.; Wang, D.; Wong, K.L.; Wang, Y. Stroke and stroke care in China huge burden, significant workload,
and a national priority. Stroke 2011, 42, 36513654. [CrossRef] [PubMed]

Sensors 2016, 16, 202

4.

5.

6.
7.
8.

9.

10.

11.

12.
13.

14.
15.
16.
17.
18.
19.

20.

21.

22.

23.
24.

15 of 16

Langan, J.; DeLave, K.; Phillips, L.; Pangilinan, P.; Brown, S.H. Home-based telerehabilitation shows
improved upper limb function in adults with chronic stroke: A pilot study. J. Rehabil. Med. 2013, 45, 217220.
[CrossRef] [PubMed]
Chen, J.; Jin, W.; Zhang, X.X.; Xu, W.; Liu, X.N.; Ren, C.C. Telerehabilitation Approaches for Stroke Patients:
Systematic Review and Meta-analysis of Randomized Controlled Trials. J. Stroke Cerebrovasc.Dis. 2015, 24,
26602668. [CrossRef] [PubMed]
Hadjidj, A.; Souil, M.; Bouabdallah, A.; Challal, Y.; Owen, H. Wireless sensor networks for rehabilitation
applications: Challenges and opportunities. J. Netw. Comput. Appl. 2013, 36, 115. [CrossRef]
Patel, S.; Park, H.; Bonato, P.; Chan, L.; Rodgers, M. A review of wearable sensors and systems with
application in rehabilitation. J. NeuroEng. Rehabil. 2012, 9, 117. [CrossRef] [PubMed]
Skobel, E.; Martinez-Romero, A.; Scheibe, B.; Schauerte, P.; Marx, N.; Luprano, J.; Knackstedt, C. Evaluation
of a newly designed shirt-based ECG and breathing sensor for home-based training as part of cardiac
rehabilitation for coronary artery disease. Eur. J. Prev. Cardiol. 2014, 21, 13321340. [CrossRef] [PubMed]
Roy, S.H.; Cheng, M.S.; Chang, S.S.; Moore, J.; De Luca, G.; Nawab, S.H.; De Luca, C.J. A combined sEMG
and accelerometer system for monitoring functional activity in stroke. IEEE Trans. Neural Syst. Rehabil. Eng.
2009, 17, 585594. [CrossRef] [PubMed]
Schwickert, L.; Becker, C.; Lindemann, U.; Marchal, C.; Bourke, A.; Chiari, L.; Helbostad, J.L.; Zijlstra, W.;
Aminian, K.; Todd, C. Fall detection with body-worn sensors. Zeitschrift fr Gerontologie und Geriatrie 2013,
46, 706719. [CrossRef] [PubMed]
Habib, M.A.; Mohktar, M.S.; Kamaruzzaman, S.B.; Lim, K.S.; Pin, T.M.; Ibrahim, F. Smartphone-based
solutions for fall detection and prevention: challenges and open issues. Sensors 2014, 14, 71817208.
[CrossRef] [PubMed]
Yang, C.C.; Hsu, Y.L. A review of accelerometry-based wearable motion detectors for physical activity
monitoring. Sensors 2010, 10, 77727788. [CrossRef] [PubMed]
Gebruers, N.; Vanroy, C.; Truijen, S.; Engelborghs, S.; De Deyn, P.P. Monitoring of physical activity after
stroke: A systematic review of accelerometry-based measures. Arch. Phys. Med. Rehabil. 2010, 91, 288297.
[CrossRef] [PubMed]
Zhang, S.; McCullagh, P.; Zhang, J.; Yu, T. A smartphone based real-time daily activity monitoring system.
Clust. Comput. 2014, 17, 711721. [CrossRef]
Chernbumroong, S.; Cang, S.; Yu, H. A practical multi-sensor activity recognition system for home-based
care. Decis. Support Syst. 2014, 66, 6170. [CrossRef]
Bulling, A.; Blanke, U.; Schiele, B. A tutorial on human activity recognition using body-worn inertial sensors.
ACM Comput. Surv. (CSUR) 2014, 46. [CrossRef]
Gao, L.; Bourke, A.K.; Nelson, J. Evaluation of accelerometer based multi-sensor versus single-sensor activity
recognition systems. Med. Eng. Phys. 2014, 36, 779785. [CrossRef] [PubMed]
Maetzler, W.; Domingos, J.; Srulijes, K.; Ferreira, J.J.; Bloem, B.R. Quantitative wearable sensors for objective
assessment of Parkinsons disease. Mov. Disord. 2013, 28, 16281637. [CrossRef] [PubMed]
Carpinella, I.; Cattaneo, D.; Ferrarin, M. Quantitative assessment of upper limb motor function in Multiple
Sclerosis using an instrumented Action Research Arm Test. J. NeuroEng. Rehabil. 2014, 11. [CrossRef]
[PubMed]
Din, S.D.; Patel, S.; Cobelli, C.; Bonato, P. Estimating Fugl-Meyer clinical scores in stroke survivors using
wearable sensors. In Proceedings of 2011 Annual International Conference of the Engineering in Medicine
and Biology Society (EMBC), Boston, MA, USA, 30 August3 September 2011; pp. 58395842.
Patel, S.; Hughes, R.; Hester, T.; Stein, J.; Akay, M.; Dy, J.G.; Bonato, P. A novel approach to monitor
rehabilitation outcomes in stroke survivors using wearable technology. Proc. IEEE. 2010, 98, 450461.
[CrossRef]
Uswatte, G.; Foo, W.L.; Olmstead, H.; Lopez, K.; Holand, A.; Simms, L.B. Ambulatory monitoring of arm
movement using accelerometry: An objective measure of upper-extremity rehabilitation in persons with
chronic stroke. Arch. Phys. Med. Rehabil. 2005, 86, 14981501. [CrossRef] [PubMed]
Zhang, Z.; Fang, Q.; Gu, X. Fuzzy inference system based automatic Brunnstrom stage classification for
upper-extremity rehabilitation. Exp. Syst. Appl. 2014, 41, 19731980. [CrossRef]
Zhang, Z.; Fang, Q.; Gu, X. Objective Assessment of Upper Limb Mobility for Post-stroke Rehabilitation.
IEEE Trans. Biomed. Eng. 2015. [CrossRef] [PubMed]

Sensors 2016, 16, 202

25.
26.

27.
28.
29.
30.
31.
32.
33.
34.
35.
36.

37.

16 of 16

Zhang, Z.; Liparulo, L.; Panella, M.; Gu, X.; Fang, Q. A Fuzzy Kernel Motion Classifier for Autonomous
Stroke Rehabilitation. IEEE J. Biomed. Heal. Inform. 2015. [CrossRef] [PubMed]
Yu, L.; Wang, J.; Fang, Q.; Wang, Y. Brunnstrom stage automatic evaluation for stroke patients using extreme
learning machine. In Proceedings of the 2012 IEEE on Biomedical Circuits and Systems Conference (BioCAS),
Hsinchu, Taiwan, 2830 November 2012.
Donoho, D.L. Compressed sensing. IEEE Trans. Inf. Theory 2006, 52, 12891306. [CrossRef]
Baraniuk, R.G. Compressive sensing. IEEE Sign. Process. Mag. 2007, 24, 118121. [CrossRef]
Candes, E.J.; Romberg, J.K.; Tao, T. Stable signal recovery from incomplete and inaccurate measurements.
Commun. Pure Appl. Math. 2006, 59, 12071223. [CrossRef]
Tsaig, Y.; Donoho, D.L. Extensions of compressed sensing. Sign. Process. 2006, 86, 549571. [CrossRef]
Mallat, S.G.; Zhang, Z. Matching pursuits with time-frequency dictionaries. IEEE Trans. Sign. Process. 1993,
41, 33973415. [CrossRef]
Tropp, J.; Gilbert, A.C. Signal recovery from random measurements via orthogonal matching pursuit.
IEEE Trans. Inf. Theory 2007, 53, 46554666. [CrossRef]
Chen, S.S.; Donoho, D.L.; Saunders, M.A. Atomic decomposition by basis pursuit. SIAM J. Sci. Comput. 1998,
20, 3361. [CrossRef]
Wipf, D.P.; Rao, B.D. Sparse Bayesian learning for basis selection. IEEE Trans. Sign. Process. 2004, 52,
21532164. [CrossRef]
Zhang, Z.; Rao, B. Extension of SBL algorithms for the recovery of block sparse signals with intra-block
correlation. IEEE Trans. Sign. Process. 2013, 61, 20092015. [CrossRef]
Zhang, Z.; Jung, T.-P.; Makeig, S.; Rao, B. Compressed sensing for energy-efficient wireless telemonitoring
of noninvasive fetal ECG via block sparse Bayesian learning. IEEE Trans. Biomed. Eng. 2013, 60, 300309.
[CrossRef] [PubMed]
Huang, G.-B.; Zhou, H.; Ding, X.; Zhang, R. Extreme learning machine for regression and multiclass
classification. IEEE Trans. Syst. Man Cybern. Part B: Cybern. 2012, 42, 513529. [CrossRef] [PubMed]
2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons by Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).