c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

journal homepage: www.intl.elsevierhealth.com/journals/cmpb

A remote quantitative Fugl-Meyer assessment

framework for stroke patients based on wearable
sensor networks
Lei Yu a,b, , Daxi Xiong a , Liquan Guo a , Jiping Wang a
a

Jiangsu Key Laboratory of Medical Optics, Suzhou Institute of Biomedical Engineering and Technology, Chinese
Academy of Sciences, China
b University of Chinese Academy of Sciences, China

a r t i c l e

i n f o

a b s t r a c t

Article history:

To extend the use of wearable sensor networks for stroke patients training and assessment

Received 16 September 2015

in non-clinical settings, this paper proposes a novel remote quantitative Fugl-Meyer assess-

Received in revised form

ment (FMA) framework, in which two accelerometer and seven ex sensors were used to

17 February 2016

monitoring the movement function of upper limb, wrist and ngers. The extreme learning

Accepted 19 February 2016

machine based ensemble regression model was established to map the sensor data to clinical FMA scores while the RRelief algorithm was applied to nd the optimal features subset.

Keywords:

Considering the FMA scale is time-consuming and complicated, seven training exercises

Wearable sensor networks

were designed to replace the upper limb related 33 items in FMA scale. 24 stroke inpatients

Quantitative assessment

participated in the experiments in clinical settings and 5 of them were involved in the exper-

Stroke

iments in home settings after they left the hospital. Both the experimental results in clinical

Upper limb motor function

and home settings showed that the proposed quantitative FMA model can precisely predict

Fugl-Meyer

the FMA scores based on wearable sensor data, the coefcient of determination can reach

Non-clinical settings

as high as 0.917. It also indicated that the proposed framework can provide a potential
approach to the remote quantitative rehabilitation training and evaluation.
2016 Elsevier Ireland Ltd. All rights reserved.

1.

Introduction

A stroke is one of top three causes of mortality and the leading cause of adult disability worldwide [1]. Between 70% and
85% of rst-ever strokes are accompanied by hemiplegia [2].
According to epidemiological statistics [3], the annual stroke
mortality rate is around 1.6 million, approximately 157 per
100,000 in China. Strokes have put enormous mental pressure
and economic burden on our society and families.

Due to the limited resources in rehabilitation centers,

home-based rehabilitation is getting more and more attention.
According to researchers [46], in comparison with inpatient
care, home-based rehabilitation shows no difference in the
effect on any of the outcomes. Moreover, there are fewer constraints on time and space in the home settings, so patients
can practice more frequently, for longer periods of time
and according to their own schedule. Unfortunately, patients
undergoing rehabilitation at home are not able to assess
their own functional state without a physician around. Thus,

Corresponding author at: No. 88, Keling Road, Suzhou, Jiangsu Province 215163, China. Tel.: +86 0512 69588302; fax: +86 0512 69588302.
E-mail address: yul@sibet.ac.cn (L. Yu).
http://dx.doi.org/10.1016/j.cmpb.2016.02.012
0169-2607/ 2016 Elsevier Ireland Ltd. All rights reserved.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

personalized interventions that will maximize the improvements in subjects motor recovery cannot be reached,
which has become the bottleneck of home-based rehabilitation. Additionally, the commonly used assessment scales
have the following two drawbacks: (1) they are susceptible
to subjective factors, the assessment results may different between physicians; (2) they only have several rating
levels and are inuenced by a ceiling effect [7], making
it impossible to exactly detect the improvement in the
movement.
Nowadays, wearable sensor network systems (WSNs)
technology particularly inertial sensors that contain
accelerometers, gyroscopes, and magnetometers can assess
the type, intensity, duration, frequency, and quality of various mobility-related functional activities [815]. Additionally,
some researches combined the inertial measure sensors with
physiological sensors, like ECG [16], sEMG [17], etc. These
sensing systems can introduce new possibilities for continuous, unsupervised, objective monitoring of mobility and
functional activities in clinical and non-clinical settings. From
the aspects of application scenarios, it can be divided into
the following four categories: falling detection [18,19], physical activity monitoring [2022], movement recognition [2325]
and quantitative assessment [2631]. Particularly in the area
of quantitative assessment for stroke patients, there are
many valuable research results have been published. Uswatte
et al. [32] have shown that accelerometer data can provide
clinically-relevant information about upper extremity motor
status via research on 169 stroke survivors. Patel et al. [28,33]
proposed a Random Forests based algorithm to monitor rehabilitation outcomes in stroke patients using accelerometers
attached to the hand, arm and trunk. The authors selected
eight tasks from the Wolf Motor Function Test (WMFT) to estimate the total Functional Ability Scale (FAS) score via analysis
of accelerometer data. Our pilot work has shown that automatic Brunnstrom stage classication can be achieved with
an accuracy of 92.1% by analyzing the accelerometer data
[34]. Zhang et al. [3537] proposed a novel single-index based
assessment approach for quantitative upper limb mobility
evaluation, the experiments collected 145 motion samples
from 21 stroke patients and 8 healthy participants. The results
suggested that the proposed assessment index can not only
differentiate the levels of limb function impairment clearly
but also strongly correlate with the Brunnstrom stages of
recovery.
However, the above researches are all implemented in clinical settings. In other words, to the best of our knowledge, there
is no existing system can remotely assess the motor function
of stroke patients in home settings. Hence in order to help the
stroke patients can do rehabilitation training after they leave
the hospital, this paper proposes a novel remote quantitative Fugl-Meyer assessment system based on wearable sensor
networks.
The rest of the paper is organized as follows. In Section
2, the full view of the proposed framework and modeling
method will be presented. The experimental congurations will be introduced in Section 3. The experimental
results and discussions will be described in Section 4.
Finally, the work and contributes will be concluded in
Section 5.

2.

Methods

2.1.

Overview of system framework

101

The overview of proposed remote quantitative Fugl-Meyer

assessment framework is shown in Fig. 1. The whole system
consists of three parts: the patient client, physician client, and
web server. At the patient client side, patients will wear the
accelerometer and ex sensors rst and then nish the individualized rehabilitation prescriptions which were made by
the physicians. During their training process, the wearable
sensors will record the movement information in real time
and wirelessly transmit to the computer through ZigBee protocol and nally upload to the web server database through
Internet. In our studies, the microcontroller Unit of wearable
sensor is CC2530, which contains an embedded ZigBee module
uses 2.4 G band. The whole wearable sensor is powered with
a Lithium-ion battery, whose capacity is 300 mAh. When the
battery is full charged, the wearable sensor can continue work
about 10 h. Besides the storage function, another important
function of web server is to analyze the movement data and
give an accurate prediction of Fugl-Meyer score (the detailed
modeling method will be introduced in Section 2.3). At the
physician client side, physicians can view the patients training record and Fugl-Meyer score via website and APP (Android
or iOS). Meanwhile, physicians can adjust the individualized
rehabilitation prescriptions according to the patients condition.

2.2.

Simplied upper limb training exercises

As seen in Fig. 1, the core of proposed system is the quantitative Fugl-Meyer assessment (FMA) model, which mapping the

Fig. 1 Overview of system framework.

102

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

sensor data to the clinical FMA scale. Here we mainly consider

the upper limb motor function of stroke patients, especially
focus on the shoulder, elbow, wrist and ngers. Considering
the clinical evaluation with FMA scale is a time-consuming
work, in the previous studies we have designed 7 upper limb
training exercises based on the short FMA (S-FMA) developed
by Hsieh et al. [38] and the guidance of clinical physicians.
From the clinical point of view, the 7 training exercises can
represent the upper limb related 33 items [39] in FMA scale. A
brief description of each training exercise is provided below:
1. Shoulder antexion. Sit down on a chair, horizontal raise the
upper limb of the hemiplegic side as high as possible, then
hold for 5 s and nally move back to the initial position.
2. Shoulder extension. Sit down on a chair, lateral raise the
upper limb of the hemiplegic side as high as possible, then
hold for 5 s and nally move back to the initial position.
3. Forearm pronation and supination. Sit down on a chair,
raise the upper limb of the hemiplegic side to the horizontal
position, then nish the forearm pronation and supination
test.
4. Lumbar touch. Keep the standing posture, move the upper
limb of the hemiplegic side back to touch the lumbar, and
then move back to the initial position.
5. Wrist exion and extension. Put the upper limb of the
hemiplegic side on a at surface, and then nish the wrist
exion and extension exercise.
6. Lateral pinch. Lateral move the thumb of the hemiplegic
side, try to touch the little nger and then move back to
the initial position.
7. Finger touch. Horizontal move the thumb of the hemiplegic
side, try to touch the index nger and then move back to
the initial position.
Among the above seven training exercises, the rst four
training exercises (called exercises 14 below) were applied
to monitoring the movement function of upper limb while
the remaining three training exercises (called exercises 57
below) were used to monitoring the movement function of
wrist and ngers. Hence, for exercises 14, only the accelerometer sensors were needed; for exercises 57, only the ex
sensors were needed. The corresponding movies of these
training exercises are listed in Additional les 17.

2.3.

Quantitative Fugl-Meyer assessment model

Fig. 2 illustrates the whole procedure of establishing the quantitative Fugl-Meyer assessment model.
Due to the fact that the raw sensor data were often noised
for many reasons, such as random noise, packet loss during
wireless transmission, etc. hence, it is necessary to preprocess
the raw sensor data rstly. Considering the movement signals
of stroke patients are low frequency, in this paper we chose
the 5 points smooth method to eliminate the noise.
After the preprocessing step we extract features which capture characteristics such as intensity, orientation, and signal
complexity from the raw sensor data. There are many methods like time-domain, frequency-domain, and so on [40]. Patel
extracted a set of 216 features from 9 body worn sensors, the
detailed information can be seen in Ref. [41]. Based on the

Fig. 2 Flowchart of quantitative Fugl-Meyer assessment

model.

published researches, in this paper we extracted the following

ve features: AMP (amplitude of sensor data); MEAN (mean
value of sensor data); RMS (root mean square value of sensor data); JERK (root mean square value of the derivative of
sensor data); ApEn (approximate entropy of sensor data). The
denition of these features can refer to [42,43]. The AMP feature can describe the movement range while the MEAN feature
mainly describe the orientation of stroke patients upper limb
during the training process. The RMS feature represents the
dynamic energy while the JERK and ApEn features describe
the smoothness and randomness of movement, respectively.
After the meaningful features are all extracted, the next
step is to mapping these features to the clinical Fugl-Meyer
assessment scale. Considering the fact that each of the
designed 7 training exercises can only represents part of
the upper limb movement functions described in FMA scale.
Hence, in this study, by drawing lessons from ensemble
machine learning [44], we established 7 weak regression models for each exercise rst and then ensemble them to build
a comprehensive quantitative Fugl-Meyer assessment model.
There are many linear and nonlinear mapping methods,
such as linear regression, neural networks, and support vector machines (SVM), etc. However, the above methods often
suffer from some drawbacks, like too many parameters, timeconsuming, prone to fall into local minima and so on [45,46].
To avoid these, in this paper, extreme learning machine (ELM)
algorithm was adopted to establish the mapping model.
ELM was rst proposed by Huang et al. [47], which was
applied to the nonlinear mapping of single layer feedforward
network (SLFN). The structure of SLFN is shown in Fig. 3. In
the following section, we will briey introduce the principle
of ELM.
Assume that there are N samples {xi , yi } (i = 1, 2, . . .,
N), where xi = [xi1 , xi2 , . . ., xin ]T Rn1 , yi = [yi1 , yi2 , . . ., yim ]T
Rm1 , m and n are the dimensions of the input and output vector, respectively. Suppose there are L neurons in the

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

hidden layer and let W be L n input weight; B be L 1 bias of

hidden layer neurons and be m L output weight.
The output (Y) of the ELM with L hidden neurons has the
following form

yik =

L


kj gj (W, B, X),

k = 1, 2, . . ., m

H = Y

Step 1: Generate arbitrary input weight wi and bias bi , i = 1, . . ., N

Step 2: Compute the output of neurons in hidden layer according
to Eq. (3)
Step 3: Compute the output weight according to Eq. (5)

= H+ Y

(2)

where Y = [y1 , . . ., yN ]T RNm , = [1 , . . ., L ]T RLm and

H(w1 , . . ., wL , b1 , . . ., bL , x1 , . . ., xN )

ELM Algorithm:

According to optimization theory, the smallest norm the

least-squares solution of the linear system is:

where gj () is the activation function.

Eq. (1) could be written in matrix form as

Table 1 Flow of ELM algorithm.

(1)

j=1

103

g(w1 x1 + b1 )

g(wL x1 + b1 )

..
.

..

..
.

g(w1 xN + b1 )

(3)

g(wL xN + bL )

where wi = [wi1 , wi2 , . . ., win ]T is the weight vector connecting

the ith hidden neuron and the input neurons, i = [i1 , i2 , . . .,
im ]T is the weight vector connecting the ith hidden neuron
and the output neurons, and bi is the threshold of the ith
hidden neuron.
To train an SLFN is simply the equivalent to nding a least of Eq. (2):
squares solution
Y||
||H(w1 , . . ., wL , b1 , . . ., bL )
= min||H(w1 , . . ., wL , b1 , . . ., bL ) Y||

Fig. 3 Structure of single layer feedforward network.

(4)

(5)

where H+ is the pseudo inverse of H.

In summary, the algorithm of ELM consists of three steps,
listed in Table 1.
Compared with traditional backpropagation neural
networks, ELM does not have to iterative adjusting the connecting weights and bias, it maps the training process to a
problem of solving a group of linear equations. Besides, in Ref.
[45], Huang et al. have proved that given any small positive
value > 0 and activation function g:R R which is innitely
N such that
differentiable in any interval, there exists N
for N arbitrary distinct samples, for any wi and bi randomly
chosen, then with probability one, ||H Y|| < .

3.

Experimental congurations

3.1.

Experimental setup

As mentioned above, in this paper we employed two types

of wearable sensors: accelerometer and ex sensors. The
accelerometers were used to monitoring the movement function of upper limb while the ex sensors were used to
monitoring the movement function of ngers and wrist.
The accelerometer sensor chip is ADXL345, whose resolution increases with g range, up to 13-bit resolution at 16 g
(maintaining 3.9 mg/LSB scale factor in all g ranges). These
characteristics enable measurement of inclination changes
less than 1.0 , which is sensitive enough to capture the movement features. The resistance of ex sensor can change from
10 K to 110 K while it is exed from at to 180 state. The
sampling frequency of all sensors was set to 20 Hz mainly
based on the following two considerations: (1) each task needs
1015 s to nish every time, in other words, the frequency of
these tasks is less than 1 Hz. Hence, 20 Hz is enough to capture
the characteristics of these movements; (2) the lower sampling
rate can reduce the power consumption of sensors so that the
wearable devices can continue work longer (about 10 days, 1 h
per day), which is easy to use for those stroke patients because
it does not require frequent charging.
The placements of accelerometer and ex sensors were
illustrated in Fig. 4. From Fig. 4(a), it is clearly to see that the
accelerometer sensors were placed at the geometric center of
upper limbs. 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
epicondyus lateralis humeri was 8 cm. As shown in Fig. 4(b),
there are in total 7 ex sensors were employed to monitoring
the movement function of ve ngers and both the exion

104

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

Fig. 4 Sensor placements on the human body.

Table 2 General information of stroke patients involved

in experiments in clinical settings.
Total (N = 24)
Sex (male/female)
Age (mean standard deviation)
Diagnosis (cerebral hemorrhage/cerebral infarction)
Hemiplegic side (left/right)
Duration after stroke (mean standard deviation)
(months)
Upper limb related FMA score (mean standard
deviation)

16/8
69.4 12.8
7/17
7/17
8.9 4.2
18.3 9.6

and extension movement function of wrist. All the ex sensors were wrapped into a glove so that they were easily for the
patients to wear.

3.2.

Experimental procedure

All the following experiments were approved by the Ethics

Committee of Jiaxing 2nd Hospital, Jiaxing, China. The whole
experiments consist of two parts: clinical and home settings,
respectively. The purpose of experiments in clinical settings was to establish the quantitative Fugl-Meyer assessment
model while the aim of experiments in home settings was to
validate the potential and validity of proposed remote quantitative Fugl-Meyer assessment framework.
24 stroke patients were recruited to participate in the
experiments in clinical settings. The general information of
patients is listed in Table 2, from which we can nd that there
are 16 males and 8 females, the average age was 69.4 12.8
years. All the patients have no severe cognitive and communication problem.
All the patients were required to nish the above mentioned 7 exercises without extra assistance so that their truly
movement function can be reected. To avoid the practice
effect, before the data collection, participants were asked
to practice as few as possible times with the guidance of
physicians to familiar with the whole experiment process
(for example, where the initial position is, and how long for
one posture to hold, etc.). During the experiments, they were
requested to repeat each exercise for 10 times. After the data
collection, each patient was evaluated by two experienced

physicians by using FMA scale, and then the average score

was calculated and set as the nal score for each participant.
As shown in Fig. 1, for the experiments in home settings,
it is required that the participants have a desktop or laptop
computer in their families and the connection to the Internet
is right congured. Among the 24 participants, only 7 patients
satisfy the above requirements. Additionally, due to the lack
of awareness of the importance of rehabilitation, another 2
patients did not agree to the continued rehabilitation in home
settings. Hence, only 5 of the 24 participants were involved
in the experiments in the home settings. They were told to
nish the 7 exercises at home once a week. Considering the
inuence of sensor placement, in this study, besides the wearable sensors, we also provide one camera for each patient and
install remote video conferencing software on the desktop or
laptop computer in their families while they left the hospital.
Every week, the participants wear the sensors under the guidance of physician and technique support engineers through
remote video conferencing software. All the sensor data were
uploaded to the web server database via Internet, and the
predictive Fugl-Meyer score was computed by web server programs. Meanwhile, they were required to go to the hospital
every other month to receive the clinical FMA by physicians.
To avoid the subjective factors between different clinicians, all
the assessment was performed by two assigned experienced
clinicians, and then the average score was calculated and set
as the nal score for each participant. The whole experiments
in home settings lasted for 3 months.
The data sampling and management were implemented
by using the Remote Rehabilitation Training and Assessment
Software (RRTAS), which was developed by our group. The run
environment of RRTAS is Windows 32 bit platform and .Net
Framework 3.5 above.

4.

Results and discussion

4.1.
Generalized performance of quantitative
Fugl-Meyer assessment model
Following the steps described in Fig. 2, we established a quantitative Fugl-Meyer assessment model which maps the wearable

105

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

X1

X2
2
Raw
Denoised

g(m/s )

g(m/s2)

0
-1

20

40

60

80

100

0
-1

120

Raw
Denoised

20

40

60

Y1

g(m/s )

g(m/s )

Raw
Denoised

-1

20

40

60

80

100

Raw
Denoised

0
-0.5
-1

120

20

40

60

Z1

80

100

120

Z2

0.4

0.5

g(m/s )

Raw
Denoised

0.2
g(m/s2)

120

0.5

-0.5

0
-0.2
-0.4

100

Y2

-1.5

80

20

40
60
80
Sampling Points

100

120

Raw
Denoised

0
-0.5
-1

20

40
60
80
Sampling Points

100

120

Flex sensor output

Thumb finger
Raw
Denoised

90

85

80

10

20

30

40
50
60
Sampling Points
Index finger

70

80

90

100

Flex sensor output

70
Raw
Denoised

65
60
55
0

10

20

30

40
50
60
Sampling Points

70

80

90

100

Fig. 5 Comparison between raw and denoised accelerometer sensor data of shoulder antexion exercise.

sensor data to FMA scores. First of all, the raw sensor data
was denoised with 5 point smooth method. Due to the space
limitations, here we only present the comparison of raw and
denoised sensor data of shoulder antexion and nger touch
exercises in Fig. 5. It can be seen that after smooth preprocessing, the denoised signals were smoother than raw signals, and
some outliers caused by packet loss during wireless transmission were eliminated.

As mentioned in Section 2.2, ve types of feature were

extracted (AMP, MEAN, RMS, JERK, and ApEn) for each exercise. Considering each accelerometer sensor has 3 axis (x, y,
and z) signals, hence there are in total 30 features can be
extracted from two accelerometer sensors. In the same way,
there are in total 35 features can be extracted from 7 ex
sensors. Therefore, the structure of ELM model for exercises
14 and exercises 57 is 30-N-1 and 35-N-1, respectively. N

106

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

Exercise 1 (R2 = 0.744)

Exercise 2 (R2 = 0.78)

50

20
40
Exercise 3 (R2 = 0.653)

50
0

20
40
Exercise 5 (R2 = 0.796)

60

50
0

20
40
Exercise 7 (R2 = 0.701)

60

50
0

60
Predictive FMA Scores

Predictive FMA Scores

50

20
40
Exercise 4 (R2 = 0.823)

60

20
40
Exercise 6 (R2 = 0.724)

60

50
0

50
0

0
20
40
60
Comprehensive model (R2 = 0.839)

50

20

40

60

Clinical FMA Scores

20

40

60

Clinical FMA Scores

Fig. 6 Generalized performance of 7 weak models and comprehensive model.

is the number of neurons in hidden layer of SLFN. In this

paper, N is set to 50 for all the 7 ELM weak regression models. Additionally, the structure of nal comprehensive ELM
model is 7-M-1, and here M is set to 10. As described above,
there are 240 data samples for each exercise, among which
200 were selected as training set and the remaining 40 samples were taken as testing set. The generalized performance
of 7 weak models and ensemble comprehensive model are
shown in Fig. 6. It is obviously to see that the all the coefcients
of determination of 7 weak regression models and ensemble
comprehensive model were smaller than 0.8 except exercise
4 (Lumbar touch), which means that the errors between the
predictive FMA scores and clinical FMA scores given by physicians were too high to ignore. Actually, the reason behind this
is that all the 7 weak regression models were built using all
the features, which is a redundant and blind way because not
all the features are equally important for each exercise, only
some of them keep the main information and characteristics.

the selected features kept the main information and characteristic for each exercise. For example, (1) the features from X
axis of two accelerometer sensors were often selected in exercises 1, 2 and 4, which was consistent with the kinematics of
these exercises; similarly, the features from Y and Z axis of
two accelerometer sensors were often selected in exercises 3
and 4; (2) for exercise 5, only the features from two ex sensors
placed on the wrist were selected because during this exercise,
the signals of other ve ex sensors placed on the ngers were
not changed; as well, the features from ex sensor placed on
the thumb were selected for both exercises 6 and 7.
Based on the selected features, we again established the
weak regression model for each exercise and then built a comprehensive quantitative Fugl-Meyer assessment model. The
generalized performance results were shown in Fig. 7. Compared with Fig. 6, it can be seen that the errors between clinical
and predictive FMA scores have reduced a lot; the coefcient
of determination of comprehensive model can reach 0.918.

4.2.
Effects of feature selection on generalized
performance

4.3.
Remote quantitative Fugl-Meyer assessment in
home settings

The above results have shown that, in order to improve the

generalized performance of weak regression models, feature
selection is an indispensible part. In this study, RRelief algorithm [48,49] was applied to nd the optimal features for each
exercise. The feature selection results are listed in Table 3,
in which X1, Y1 and Z1 represent the accelerometer sensor signals from upper arm; X2, Y2 and Z2 represent the
accelerometer sensor signals from forearm; S1S5 represent
the ex sensor signals from thumb to little nger; S6S7 represent the ex sensor signals from wrist exion and extension.
It is obviously to see that after feature selection, the number of selected features have reduced a lot. More importantly,

As mentioned above, the nal goal of this study is to implement the proposed framework in home settings. According
to the experiment described in Section 3.2, the involved ve
stroke patients nished the seven exercises at home once a
week and go to the hospital once a month. The FMA scores
predicted by our framework and physicians were illustrated in
Fig. 8, in which P1P5 denotes the ve stroke patients, respectively. The solid points at the 4th, 8th and 12th week represent
the FMA scores assessed by physicians. From Fig. 8 it is clearly
to see that the errors between predictive and physicians FMA
scores were small enough so that can be ignored. Actually, the
one-way analysis of variance (ANOVA) results of 5 participants

107

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

Table 3 Feature selection results with RRelief algorithm.

Exercise Number
2
3
4

AMP_X1
AMP_Y1
AMP_Z1
AMP_X2
AMP_Y2
AMP_Z2
RMS_X1
RMS_Y1
RMS_Z1
RMS_X2
RMS_Y2
RMS_Z2
MEAN_X1
MEAN_Y1
MEAN_Z1
MEAN_X2
MEAN_Y2
MEAN_Z2
JERK_X1
JERK_Y1
JERK_Z1
JERK_X2
JERK_Y2
JERK_Z2
ApEn_X1
ApEn_Y1
ApEn_Z1
ApEn_X2
ApEn_Y2
ApEn_Z2

Exercise Number
5
6
7
AMP_S1
AMP_S2
AMP_S3
AMP_S4
AMP_S5
AMP_S6
AMP_S7
RMS_S1
RMS_S2
RMS_S3
RMS_S4
RMS_S5
RMS_S6
RMS_S7
MEAN _S1
MEAN _S2
MEAN _S3
MEAN _S4
MEAN _S5
MEAN _S6
MEAN _S7
JERK_S1
JERK_S2
JERK_S3
JERK_S4
JERK_S5
JERK_S6
JERK_S7
ApEn_S1
ApEn_S2
ApEn_S3
ApEn_S4
ApEn_S5
ApEn_S6
ApEn_S7

Different colors show the differences between exercises.

(p1 = 0.613; p2 = 0.127; p3 = 0.482; p4 = 0.084; p5 = 0.581, both large

than 0.05) showed no statistical difference between the predictive results of quantitative assessment model and clinical
FMA scores.
More importantly, it also indicated that instead of the traditional point evaluation, by using the proposed framework, the
detailed training and improvements can be recorded which
can help the physicians to make an individualized training
prescription.

4.4.

Discussion

To validate the generalized performance of ELM algorithm, in

this study SVM algorithm was also applied to establish the
quantitative assessment models. The Gaussian function was
chosen as kernel function, and the parameters were optimized
through grid search. Each model was established for 50 times

Table 4 Generalized performance comparisons

between ELM and SVM.
Models
RMSE
R2
Training time (s)
a

ELM
12.192
0.839
0.226

ELM FSa
6.964
0.918
0.138

SVM
11.875
0.846
4.562

SVM FSa
6.627
0.922
1.085

ELM FS, SVM FS means models are established after the feature
selection process.

and the average comparison results were listed in Table 4. It

is clearly to see that after feature selection, the generalized
performances are improved a lot. The student t-test results
showed that there is no statistically difference between ELM
and SVM models (p = 0.724 > 0.05). Due to the fact that during
the SVM training process, it is needed to solve quadratic programming problems, hence the training time of SVM is much

108

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

Exercise 1 (R2 = 0.872)

Exercise 2 (R2 = 0.862)

50

20
40
Exercise 3 (R2 = 0.879)

50
0

20
40
Exercise 5 (R2 = 0.909)

60

50
0

20
40
Exercise 7 (R2 = 0.871)

60

50
0

60
Predictive FMA Scores

Predictive FMA Scores

50

20
40
Exercise 4 (R2 = 0.921)

60

20
40
Exercise 6 (R2 = 0.884)

60

50
0

50
0

0
20
40
60
Comprehensive model (R2 = 0.918)

50

20

40

60

Clinical FMA Scores

20

40

60

Clinical FMA Scores

Fig. 7 Generalized performance of 7 weak models and comprehensive model with RRelief feature selection.

longer than ELM. However, once the training process has nished, there is no obvious difference between the SVM and
ELM for the prediction process.
The proposed framework has some limitations need to be
concerned, among which ceiling effect is an important one.
From Fig. 8, ceiling effect can be seen in both the clinical physicians assessment and our proposed framework. For example,
the FMA scores of P2 and P4 at the 8th and 12th week were
almost same, which is difcult for physicians to precisely
evaluate the movement function of patients and make individual prescriptions. Hence, it is necessary to develop some
new features and evaluation indexes to describe the detail

Fig. 8 Quantitative FMA scores in home settings.

information of sensor data. To avoid the ceiling effect, many

researchers have proposed some alternative methods. Gladstone et al. [39] stated that expanding the grading system of
each item may have actually been benecial in maximizing
the ability of the scale to detect change and could make the
scale more sensitive to change for those patients at the top
end. In clinimetrics, the optimal number of categories recommended is usually 57. Recently, our study group has designed
some new evaluation indexes based on nonlinear analysis
methods like multi-scale entropy; maximum Lyapunov index,
etc. and the results indicate a potential way to describe
the changes which cannot be shown by clinical assessment
scale.
Another thing must be considered is that the placement
of sensors will inuence the generalized performance of proposed quantitative FMA model. Therefore in the future, we will
go deep into this problem to design a method which can automatic check whether the placement of sensors are correct. For
example, before the experiment starts, we can rst collect a
short time period (for example, 35 s) accelerometer signals
and secondly determine whether the sensors are placed at
the right place and stay at the right position by using pattern recognition methods like template match, articial neural
networks, support vector machine and so on. There are many
published researches in this area, the detailed information can
refer to Refs. [50,51].
Last but not the least, besides the accelerometer and ex
sensor, other types of sensors like gyroscope, magnetometer,
and pressure, etc. should be included to monitoring the movement function, lifestyle of stroke patients. Although many
researchers have been put efforts into these areas, there are
still many problems need to be solved especially when these
technologies are applied in home settings.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

5.

Conclusions
[7]

To extend the use of wearable sensor networks for stroke

rehabilitation in home settings, we proposed a novel remote
quantitative Fugl-Meyer assessment framework, in which two
accelerometer and seven ex sensors were used to capture the
movements of upper limb, wrist and ngers. Extreme learning
machine based ensemble machine learning method was used
to map the sensor data to clinical FMA scores. Meanwhile,
RRelief algorithm was applied to select the optimal features
subset. The experimental results in both clinical and home
settings showed that the proposed quantitative FMA model
can precisely predict the FMA scores according to the wearable
sensor data. It also indicated that the proposed framework can
be implemented in home settings, which provide a potential
approach to the remote rehabilitation training and evaluation.
In the future, we will look deep into the application of multisensor fusion and human computer interaction in the area of
rehabilitation.

[8]

[9]

[10]

[11]

[12]

Conict of interest
The authors declare that they have no proprietary, nancial,
professional, or other personal competing interests of any
nature or kind.

Acknowledgements
The authors thank Doctor Xudong Gu and Jianming Fu who
come from the Rehabilitation Medical Center of Jiaxing 2nd
Hospital, for their valuable suggestion and guidance during
the clinical experiment design and implementation processes.

[13]

[14]

[15]

Appendix A. Supplementary data

[16]

Supplementary data associated with this article can

be found, in the online version, at http://dx.doi.org/
10.1016/j.cmpb.2016.02.012.

references

[1] S.C. Johnston, S. Mendis, C.D. Mathers, Global variation in

stroke burden and mortality: estimates from monitoring,
surveillance, and modelling, Lancet Neurol. 8 (2009) 345354.
[2] B.H. Dobkin, Strategies for stroke rehabilitation, Lancet
Neurol. 3 (2004) 528536.
[3] L. Liu, D. Wang, K.L. Wong, Y. Wang, Stroke and stroke care
in China huge burden, signicant workload, and a national
priority, Stroke 42 (2011) 36513654.
[4] J. Langan, K. DeLave, L. Phillips, P. Pangilinan, S.H. Brown,
Home-based telerehabilitation shows improved upper limb
function in adults with chronic stroke: a pilot study, J.
Rehabil. Med. 45 (2013) 217.
[5] G. Placidi, D. Avola, M. Ferrari, D. Iacoviello, A. Petracca, V.
Quaresima, M. Spezialetti, A low-cost real time virtual
system for postural stability assessment at home, Comput.
Methods Progr. Biomed. 117 (2014) 322333.
[6] C.L. Rau, Y.P. Chen, J.S. Lai, S.C. Chen, T.S. Kuo, F.S. Jaw, J.J.
Luh, Low-cost tele-assessment system for home-based

[17]

[18]

[19]

[20]

[21]

[22]

109

evaluation of reaching ability following stroke, Telemed.

E-Health 19 (2013) 973978.
I.-P. Hsueh, M.-J. Hsu, C.-F. Sheu, S. Lee, C.-L. Hsieh, J.-H. Lin,
Psychometric comparisons of 2 versions of the Fugl-Meyer
Motor Scale and 2 versions of the Stroke Rehabilitation
Assessment of Movement, Neurorehabil. Neural Repair
(2008).
L.K. Simone, N. Sundarrajan, X. Luo, Y.C. Jia, D.G. Kamper, A
low cost instrumented glove for extended monitoring and
functional hand assessment, J. Neurosci. Meth. 160 (2007)
335348.
Motor function
A. Parnandi, E. Wade, M. Mataric,
assessment using wearable inertial sensors, in: 2010 32nd
Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, EMBC10, Buenos Aires, 2010,
pp. 8689.
S.M. Michaelsen, R.P. Gomes, A.P. Marques, L.C. Rodrigues,
N.G. Borges Jr., R. Claudino, M.J. Dos Santos, Using an
accelerometer for analyzing a reach-to-grasp movement
after stroke, Mot. Rev. Educ. Fis. 19 (2013) 746752.
E.V. Olesh, S. Yakovenko, V. Gritsenko, Automated
assessment of upper extremity movement impairment due
to stroke, PLOS ONE 9 (2014).
K. Taraldsen, T. Askim, O. Sletvold, E.K. Einarsen, K.G.
Bjstad, B. Indredavik, J.L. Helbostad, Evaluation of a
body-worn sensor system to measure physical activity in
older people with impaired function, Phys. Ther. 91 (2011)
277285.
S.C. Van Der Pas, J.A. Verbunt, D.E. Breukelaar, R. Van
Woerden, H.A. Seelen, Assessment of arm activity using
triaxial accelerometry in patients with a stroke, Arch. Phys.
Med. Rehabil. 92 (2011) 14371442.
S.J.M. Bamberg, A.Y. Benbasat, D.M. Scarborough, D.E. Krebs,
J. Paradiso, Gait analysis using a shoe-integrated wireless
sensor system, IEEE Trans. Inf. Technol. Biomed. 12 (2008)
413423.
D.O. Olgun, A.S. Pentland, Human activity recognition:
accuracy across common locations for wearable sensors, in:
Proceedings of 2006 10th IEEE International Symposium on
Wearable Computers, Citeseer, Montreux, Switzerland, 2006,
pp. 1114.
E. Skobel, A. Martinez-Romero, B. Scheibe, P. Schauerte, N.
Marx, J. Luprano, C. Knackstedt, 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. 21 (2014)
13321340.
S.H. Roy, M.S. Cheng, S.-S. Chang, J. Moore, G. De Luca, S.H.
Nawab, C.J. De Luca, A combined sEMG and accelerometer
system for monitoring functional activity in stroke, IEEE
Trans. Neural Syst. Rehabil. Eng. 17 (2009) 585594.
L. Schwickert, C. Becker, U. Lindemann, C. Marchal, A.
Bourke, L. Chiari, J. Helbostad, W. Zijlstra, K. Aminian, C.
Todd, Fall detection with body-worn sensors, Z. Gerontol.
Geriatr. 46 (2013) 706719.
M.A. Habib, M.S. Mohktar, S.B. Kamaruzzaman, K.S. Lim,
T.M. Pin, F. Ibrahim, Smartphone-based solutions for fall
detection and prevention: challenges and open issues,
Sensors 14 (2014) 71817208.
C.-C. Yang, Y.-L. Hsu, A review of accelerometry-based
wearable motion detectors for physical activity monitoring,
Sensors 10 (2010) 77727788.
N. Gebruers, C. Vanroy, S. Truijen, S. Engelborghs, P.P. De
Deyn, Monitoring of physical activity after stroke: a
systematic review of accelerometry-based measures, Arch.
Phys. Med. Rehabil. 91 (2010) 288297.
S. Zhang, P. McCullagh, J. Zhang, T. Yu, A smartphone based
real-time daily activity monitoring system, Clust. Comput.
17 (2014) 711721.

110

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 2 8 ( 2 0 1 6 ) 100110

[23] S. Chernbumroong, S. Cang, H. Yu, A practical multi-sensor

activity recognition system for home-based care, Decis.
Support Syst. 66 (2014) 6170.
[24] A. Bulling, U. Blanke, B. Schiele, A tutorial on human activity
recognition using body-worn inertial sensors, ACM Comput.
Surv. (CSUR) 46 (2014) 33.
[25] L. Gao, A. Bourke, J. Nelson, Evaluation of accelerometer
based multi-sensor versus single-sensor activity recognition
systems, Med. Eng. Phys. 36 (2014)
779785.
[26] W. Maetzler, J. Domingos, K. Srulijes, J.J. Ferreira, B.R. Bloem,
Quantitative wearable sensors for objective assessment of
Parkinsons disease, Mov. Disord. 28 (2013) 16281637.
[27] I. Carpinella, D. Cattaneo, M. Ferrarin, Quantitative
assessment of upper limb motor function in Multiple
Sclerosis using an instrumented Action Research Arm Test,
J. Neuroeng. Rehabil. 11 (2014) 67.
[28] S.D. Din, S. Patel, C. Cobelli, P. Bonato, Estimating Fugl-Meyer
clinical scores in stroke survivors using wearable sensors,
in: 2011 Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, EMBC, IEEE,
2011, pp. 58395842.
[29] H.G. Kortier, V.I. Sluiter, D. Roetenberg, P.H. Veltink,
Assessment of hand kinematics using inertial and magnetic
sensors, J. Neuroeng. Rehabil. 11 (2014) 70.
[30] F. Cordella, F.D. Corato, L. Zollo, B. Siciliano, P. van der Smagt,
Patient performance evaluation using Kinect and Monte
Carlo-based nger tracking, in: 2012 4th IEEE RAS & EMBS
International Conference on Biomedical Robotics and
Biomechatronics (BioRob), IEEE, 2012,
pp. 19671972.
[31] M. Khademi, H. Mousavi Hondori, A. McKenzie, L. Dodakian,
C.V. Lopes, S.C. Cramer, Free-hand interaction with leap
motion controller for stroke rehabilitation, in: CHI14
Extended Abstracts on Human Factors in Computing
Systems, ACM, 2014, pp. 16631668.
[32] G. Uswatte, C. Giuliani, C. Winstein, A. Zeringue, L. Hobbs,
S.L. Wolf, Validity of accelerometry for monitoring real-world
arm activity in patients with subacute stroke: evidence from
the extremity constraint-induced therapy evaluation trial,
Arch. Phys. Med. Rehabil. 87 (2006) 13401345.
[33] S. Patel, R. Hughes, T. Hester, J. Stein, M. Akay, J.G. Dy, P.
Bonato, A novel approach to monitor rehabilitation
outcomes in stroke survivors using wearable technology, in:
Proceedings of the IEEE, 98, 2010, pp. 450461.
[34] Y. Lei, W. Ji-ping, F. Qiang, W. Yue, Brunnstrom stage
automatic evaluation for stroke patients using extreme
learning machine, in: Biomedical Circuits and Systems
Conference (BioCAS), 2012 IEEE, IEEE, 2012,
pp. 380383.
[35] Z. Zhang, Q. Fang, X. Gu, Objective assessment of upper limb
mobility for post-stroke rehabilitation, in: IEEE Transactions
on Biomedical Engineering, 2015, p. 1.
[36] Z. Zhang, Q. Fang, X. Gu, Fuzzy inference system based
automatic Brunnstrom stage classication for
upper-extremity rehabilitation, Expert Syst. Appl. 41 (2014)
19731980.

[37] Z. Zhang, L. Liparulo, M. Panella, X. Gu, Q. Fang, A fuzzy

kernel motion classier for autonomous stroke
rehabilitation, IEEE J. Biomed. Health Inform. (2015).
[38] Y.-W. Hsieh, I.-P. Hsueh, Y.-T. Chou, C.-F. Sheu, C.-L. Hsieh, G.
Kwakkel, Development and validation of a short form of the
Fugl-Meyer motor scale in patients with stroke, Stroke 38
(2007) 30523054.
[39] D.J. Gladstone, C.J. Danells, S.E. Black, The Fugl-Meyer
assessment of motor recovery after stroke: a critical review
of its measurement properties, Neurorehabil. Neural Repair
16 (2002) 232240.
[40] M. Zhang, A.A. Sawchuk, A feature selection-based
framework for human activity recognition using wearable
multimodal sensors, in: Proceedings of the 6th International
Conference on Body Area Networks, ICST (Institute for
Computer Sciences, Social-Informatics and
Tele-communications Engineering), Beijing, China, 2011,
pp. 9298.
[41] S. Patel, Quantitative Motor Assessment in Patients with
Mobility Limiting Conditions using Wearable Sensors, The
Department of Electrical and Computer Engineering,
Northeastern University, Boston, 2012, pp. 250.
[42] T. Hester, R. Hughes, D.M. Sherrill, B. Knorr, M. Akay, J. Stein,
P. Bonato, Using wearable sensors to measure motor abilities
following stroke, in: Proceedings of the International
Workshop on Wearable and Implantable Body Sensor
Networks, IEEE Computer Society, 2006, pp. 58.
[43] J. Shin, Y. Park, Y. Lim, M. Park, The application of
approximate entropy for the analysis of irregularity in
ambulatory blood pressure values: PP.25.31, J. Hypertens. 28
(2010) e456.
[44] C. Zhang, Y. Ma, Ensemble Machine Learning: Methods and
Applications, Springer Publishing Company, 2012.
[45] G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning
machine: theory and applications, Neurocomputing 70
(2006) 489501.
[46] M. Kubat, Neural networks: a comprehensive foundation by
Simon Haykin, Macmillan, 1994, ISBN 0-02-352781-7,
Cambridge Univ. Press, 1999.
[47] G.-B. Huang, H. Zhou, X. Ding, R. Zhang, Extreme learning
machine for regression and multiclass classication, IEEE
Trans. Syst. Man Cybern. B: Cybern. 42 (2012) 513529.
[48] I. Kononenko, Estimating attributes: analysis and extensions
of RELIEF, in: Proceedings of the European Conference on
Machine Learning, Springer-Verlag New York, Inc., Catania,
Italy, 1994, pp. 171182.
[49] M. Robnik-Sikonja, I. Kononenko, An adaptation of relief for
attribute estimation in regression, in: Proceedings of the
Fourteenth International Conference on Machine Learning,
Morgan Kaufmann Publishers Inc., 1997, pp. 296304.
[50] J. Liu, L. Zhong, J. Wickramasuriya, V. Vasudevan, uWave:
accelerometer-based personalized gesture recognition and
its applications, Pervasive Mobile Comput. 5 (2009) 657675.
[51] X. Zhang, X. Chen, Y. Li, V. Lantz, K. Wang, J. Yang, A
framework for hand gesture recognition based on
accelerometer and EMG sensors, IEEE Trans. Syst. Man
Cybern. A: Syst. Hum. 41 (2011) 10641076.