The Combined Effects of Adaptive Control and Virtual Reality

on Robot-Assisted Fine Hand Motion Rehabilitation in Chronic

Stroke Patients: A Case Study

Xianwei Huang, PhD, Fazel Naghdy, PhD, Golshah Naghdy, PhD,

Haiping Du, PhD, and Catherine Todd, PhD

Robot-assisted therapy is regarded as an effective and reliable method for the de-
livery of highly repetitive training that is needed to trigger neuroplasticity following
a stroke. However, the lack of fully adaptive assist-as-needed control of the robotic
devices and an inadequate immersive virtual environment that can promote active
participation during training are obstacles hindering the achievement of better
training results with fewer training sessions required. This study thus focuses on
these research gaps by combining these 2 key components into a rehabilitation
system, with special attention on the rehabilitation of fine hand motion skills. The
effectiveness of the proposed system is tested by conducting clinical trials on a
chronic stroke patient and verified through clinical evaluation methods by mea-
suring the key kinematic features such as active range of motion (ROM), finger
strength, and velocity. By comparing the pretraining and post-training results, the
study demonstrates that the proposed method can further enhance the effective-
ness of fine hand motion rehabilitation training by improving finger ROM, strength,
and coordination. Key Words: Stroke rehabilitation—robot assisted therapy—assist-
as-needed control—virtual reality—rehabilitation gaming system—clinical assessment.
© 2018 National Stroke Association. Published by Elsevier Inc. All rights reserved.

Introduction task is challenging enough.3 This approach holds promise

for the improvement or even recovery of the affected
Stroke can result in paralysis, speech impairment, loss
motion skills. However, conventional therapies (CTs) such
of memory and reasoning ability, coma, or even death.1
as physical and occupational therapy can be time-
Complete or partial loss of movability in an upper ex-
consuming and labor intensive, requiring long training
tremity (UE) is the most commonly reported impairment
sessions and incurring high costs. Moreover, it is diffi-
after suffering a stroke, which can hinder the perfor-
cult to effectively evaluate the effectiveness of CTs
mance of activities of daily living (ADLs) and significantly
quantitatively and objectively. Robot-assisted therapy (RT)4
undermine the quality of life (QOL) of the stroke patient.2
can provide a highly repetitive movement training and
Recent neuroplasticity studies indicate that a highly re-
overcome the shortcomings of CTs through the reduc-
petitive and task-specific training may induce changes
tion of associated labor and the offer of an accurate,
in the brain, and this response can be optimized if the
automated movement control and the quantification of
rehabilitation performance.5
From the Universality of Wollongong, Northfields Ave, Wollongong,
NSW 2522, Australia.
Multisensory feedback has been proven to be critical
Received July 28, 2016; accepted August 16, 2017. in re-establishing the neural pathways damaged by stroke
Address correspondence to Xianwei Huang, PhD, University of and closing the sensor motor loop.6 Virtual reality (VR)
Wollongong, Wollongong 2522, Australia. E-mail: xh962@ technology has recently been widely used in the field of
uowmail.edu.au.
augmented stroke rehabilitation. VR allows users to in-
1052-3057/$ - see front matter
© 2018 National Stroke Association. Published by Elsevier Inc. All
teract with a multisensory simulated environment and
rights reserved. receive real-time feedback on performance, thus allow-
https://doi.org/10.1016/j.jstrokecerebrovasdis.2017.08.027 ing patients to correct their motion. This, in turn, can

Journal of Stroke and Cerebrovascular Diseases, Vol. 27, No. 1 (January), 2018: pp 221–228 221
222
facilitate repetition, intensity, and task-oriented training,
all of which promote voluntary active motion. There-
fore, VR-based RT offers the potential to specifically
promote and/or enhance functional movement recovery.7
VR systems can provide safe, ecological, and individu-
alized 3-dimensional environments where patients can
perform specific actions to achieve a goal. Another ad-
vantage of a VR intervention is that patients can perceive
such interventions as enjoyable exercise games rather than
treatment methods and, thereby, increase motivation and
treatment compliance.8 VR can also be used as an as-
sessment method by recording and objectively measuring
the performance of patients and their behavioral re-
sponses within the virtual world. VR-based RT thus has
the potential to utilize motor learning principles in re-
lation to task-oriented training. However, the intensity
and dose–response aspects of a VR game–based inter-
vention training with evidence-based efficacy and clear
objectives and outcomes need to be further determined.9
The aim of this study was to further evaluate the ef-
fectiveness of the developed VR-based RT treatments with
an adaptive control method deployed in a robotic train-
ing device. In addition, the study explores more effective
methods to transform the achieved improvements to the
performance of ADL tasks in real life. Special attention
is paid to fine hand motion rehabilitation, which in-
volves small, precise, and coordinated movements of the
fingers and requires the integration of muscular, skele-
tal, and neurological functions. This is because fine hand
motion is vital for more delicate ADL tasks such as eating,
drinking, and personal hygiene and is necessary for the
improvement of the QOL of stroke patients.
Materials and Methods
RT Platform
In this study, a 5 degrees of freedom hand rehabilita-
tion robotic device named Amadeo (Tyromotion GmbH,
Graz, Austria)10,11 was used as the experimental plat-
form. Amadeo (Tyromotion GmbH) can provide position-
based passive and active assistive training modes that
emphasize the flexion and extension of each finger. The
moving finger slides are attached to the fingers using a
small magnetic disc and an adhesive tape for connec-
tion to the robot. The slides then transfer, bend, or stretch
movements to the fingers. The VR-based RGS is imple-
mented with Amadeo (Tyromotion GmbH), which is
composed of a finger rehabilitation robot and a PC with
LCD displaying the virtual scenarios in front of the user
as shown in Figure 1.
The major limitation of Amadeo’s (Tyromotion GmbH)
standard training protocol is that it provides position control
with constant assistive force intensity regardless of the
actual need of the patient. In clinical practice, patients
with different levels of active mobility skills require various
levels of assistive support. Also, even for the same patient,
X. HUANG ET AL.
Figure 1. Training conducted on a stroke patient using Amadeo.
this constant assistive force intensity can be excessive to
that required for the movement of the finger. As a result,
the patient may end up following the finger slides pas-
sively, which makes the training task less challenging and
compromises the effectiveness of the training. In order
to achieve optimal training results and to reduce the re-
quired training sessions and associated costs, a
reinforcement learning neural network–based AAN adap-
tive control was designed and validated through computer
simulation and experimental work. The notion of the AAN
control method is to provide exact assistive force inten-
sity along the intended motion trajectory only when it
is needed, thus keeping the training challenging and en-
gaging. This is achieved by observing the force and position
feedback state at the current time-step and making pre-
dictions of the next time-step state based on the feedback
data. A detailed design and online learning process of
the proposed AAN control method can be found in Huang
et al.12
Participant
Considering the final evaluation, comparisons and con-
clusions will be based on the performance of the patient
during clinical tests; the condition of the subject before
the study may affect the outcome of the rehabilitation
process. Thus, it is important to set unified require-
ments for recruited patients to ensure that they start with
similar conditions. The Fugl-Meyer Assessment (FMA) and
the Motor Assessment Scale (MAS) were used to assess
patients’ abilities during the evaluation sessions. In order
to maintain consistency, we defined the following inclu-
sion requirements for the recruited subjects:
1) Age range: 50-75 (patients around the same age were
preferred).
EFFECTS OF ADAPTIVE CONTROL AND VIRTUAL REALITY ON ROBOT-ASSISTED REHABILITATION 223
2) Severity of impairment: motor abilities suggested
by MAS score:
In section 6 (upper arm): 3-5;
Section 7 (hand movements): 2-4.
3) Language spoken: English
4) Good cognition: suggested by the Rowland Uni-
versal Dementia Assessment Scale or Mini Mental
State Examination score of 26 (out of 30) or more.
The general inclusion principle is to allow space where
the patients can experience the expected changes during
the training sessions. Therefore, patients with a “medium
level” of hand motor skills were preferred. Also, at the
beginning of all sessions, we needed to confirm that pa-
tients who met these suggested inclusion criteria were
capable of operating the Amadeo (Tyromotion GmbH)
robotic device under guidance, otherwise they were not
included. More specifically, subjects with the following
conditions would not be included for this study (exclu-
sion criteria):
(1) non–stroke-caused functional deficits of an arm
and hand motor function or a history of more than
one stroke clinically;
(2) other neurological disorders such as Parkinson’s
disease;
(3) severe cognitive dysfunction;
(4) visual and hearing impairment;
Figure 2. Flow diagram of intervention process and protocol. Abbrevia-
(5) contracture or deformities in the UE; tions: FMA, Fugl-Meyer Assessment; MAS, Motor Assessment Scale; ROM,
(6) previous history of epilepsy. range of motion.
The exclusion criteria were kept to a minimum in order
to evaluate the feasibility of using RT among a variety
of patients. As a result, 1 54-year-old male stroke patient of “active mode therapy” in which the hand motion is
was found to satisfy the criteria; he suffered from hemiple- assisted by the robot; and 10 minutes of games that in-
gia caused by a stroke 52 months prior to the start of cluded active training in a simulated environment.
this study. The patient had limited movability in his right Intervention Phase B: 6 weeks of the proposed train-
arm, and his right-hand function was severely compro- ing protocol. After intervention Phase A, the patient was
mised by the stroke. The patient signed a written informed subjected to a novel dynamic adaptive assistive force based
consent to participate and a written informed consent for on a reinforcement learning neural network control al-
the publication of his clinical results. gorithm and the newly developed VR training games.
Each session included 10 minutes of “passive mode” train-
Intervention ing, 10 minutes of “adaptive AAN mode” training, and
10 minutes of a “new VR-based RGS” training.
The patient undertook a total of 18 weeks of training In the “new task-oriented VR-based RGS” for inter-
using the Amadeo (Tyromotion GmbH) hand rehabilita- vention Phase B, task-oriented interactive VR games such
tion robot, with 1 session per day and 3 days per week as object manipulation and simulated cooking scenarios
(a total of 54 training sessions) with each training session were used to enhance active participation and to promote
lasting for 30 minutes (1620 minutes in total). The re- the transformation of the acquired motor skills into abil-
search aim of this study was to investigate whether the ities to perform ADLs in real life. The process and training
patient is able to continue improving even after the po- protocols of the clinical trials is shown in Figure 2.
tential progress (if any) of Amadeo’s (Tyromotion GmbH) At the end of each session, the active range of motion
standard training protocol. Thus, the intervention was de- (ROM) and force intensity of fingers (both extending and
signed to have 2 separate phases. The training protocols grasping) were examined using Amadeo (Tyromotion
were as follows: GmbH) embedded sensors. Three clinical tests were con-
Intervention Phase A: 12 weeks of Amadeo (Tyromotion ducted at the beginning of all sessions (week 0), at the
GmbH) standard training protocol consisting of 10 minutes end of the 12 weeks of training, and at the end of all
of “passive mode training” in which the hand is stimu- the 18 weeks of RT training. During these tests, fine hand
lated in continuous passive motion therapy; 10 minutes motor skills of each subject were assessed using standard
224
clinical evaluation procedures: the FMA hand score and
the MAS.
VR-Based RGS and Transferring
Environment
To make the training less dull and monotonous as well
as to motivate the patient’s voluntary and active partic-
ipation in the training, 2 interactive VR-based RGS and
1 transferring virtual environment (VE) were designed.
The customized training program was developed with
close collaboration between clinicians and developers. When
implemented on Amadeo (Tyromotion GmbH), the highly
repetitive but functional tasks serve as an enriched mo-
tivating environment with challenging and customizable
difficulty levels. The RGS strengthens the extending and
grasping of each finger movement, and the VE simu-
lates object manipulation in real life. In these VR systems,
real-time force and position signals were incorporated into
the feedback loop to provide immediate feedback on the
patient’s performance.
RGS—Flying Bird
Flying bird (Fig 3, A) is an interactive game coded using
JavaScript and Html5. In the game, the user controls the
height of a bird flying through a series of vertical columns,
and immediate feedback is sent through an Ethernet cable
using UDP protocol. The score, indicating the user’s per-
formance, is calculated as the total number of columns
the bird can pass through without collision. This game
is featured with position mode and force mode, where
Figure 3. VR-based RGS and transferring VE. (A) Flying bird VR game. (B) Spaceship VR game. (C) Transferring VE-simulated supermarket.
(D) Transferring VE kitchen and cooking scenario. Abbreviations: VE, virtual environment; VR, virtual reality.
X. HUANG ET AL.
the object in the game is controlled by feedback posi-
tion signal or force signal.
RGS—Spaceship
Spaceship (Fig 3, B) is coded in JavaScript and Html5.
In the game, the user controls the position of a space-
ship to avoid incoming meteorites and bombs and to collect
virtual awards. The score is calculated as the total length
of time that the user manages to continue playing and
represents the user’s performance. The game is featured
with both position and force modes, and the level of dif-
ficulty can be adjusted according to the abilities of the
subject. The VR-based RGS and VE screenshots are shown
in Figure 3.
Transferring VE
Unlike the RGS mentioned above, which aims at im-
proving the overall strength and ROM of the hand and
fingers, the transferring VE aims to transfer the achieved
improvements in terms of ROM and force intensity to
motor skills to carry out ADL tasks in real life and thus
improve the QOL of the patients. The VE is thus fea-
tured with lifelike virtual scenarios that aim to strengthen
skills such as movement planning, target reaching, and
object manipulation and classification, especially the move-
ment output force intensity and movement accuracy of
each finger (suggested by finger position).
The VE consisted of a 2-dimensional simulated super-
market and kitchen setting to increase the engagement
of patients. The transferring VE focuses on improving
EFFECTS OF ADAPTIVE CONTROL AND VIRTUAL REALITY ON ROBOT-ASSISTED REHABILITATION
strength and active ROM in the context of a functional
reaching movement. The subject is required to reach and
grasp a simulated object and perform certain ADL tasks
(i.e., buy cooking ingredients, open the oven, set the alarm
clock) with the paretic hand. Prior to each training session,
the game settings are calibrated so as to be customized
to the participant’s active ROM of the affected fingers.
Participants are instructed to pick up virtual cooking in-
gredients from a supermarket shelf (see Fig 3, C) with a
limited amount of virtual money and then to use the pur-
chased ingredients to cook a turkey in a simulated kitchen
setting (see Fig 3, D).
To achieve these tasks, the screen is divided into 5 ver-
tical sections, denoting the realm of 5 independent fingers.
When the subject’s finger reaches the position of an object,
corresponding actions are triggered if the detected force
is larger than the preset adjustable threshold value, which
can be adjusted according to the actual abilities of a certain
subject. The real-time position and orientation and the
flexion and abduction of each of the fingers are trans-
lated into 2-dimensional movement. Moreover, the objects
are placed at different heights to accommodate the active
ROM of each finger.
Results and Analysis
A number of parameters associated with fine hand
motion rehabilitation including spasticity, reflexes, level
of voluntary control, and function movement were evalu-
ated. Kinematic measurements, including hand movement
speed and movement duration, were calculated using data
collected by the robot.
Clinical Assessment Methods
Considering the focus on fine hand motion rehabili-
tation in this study, and the conclusion would be drawn
based on the pretraining and post-training performance
of the subject, choosing suitable quantitative evaluation
methods with a focus on fine hand movements was thus
critical in this research. Four evaluation methods satis-
fying these requirements were chosen to assess the outcome
of the rehabilitation, namely active range of movement
(ROM), active force output, FMA hand subsection, and
MAS Hand Movement Score. Active ROM and force were
Table 1. Pretraining and post-training results
Force (N)
Assessment Flexion Extension ROM (%)
First (0 week) 12.2 ± .3 5.8 ± .4 38.6 ± .6
Second (12 week) 20.4 ± .2 11.5 ± .2 73.5 ± .4
Third (18 week) 22.3 ± .1 13.7 ± .2 79.3 ± .3
Abbreviations: FMA, Fugl-Meyer Assessment; MAS, Motor Assessment Scale; ROM, range of motion.
225
measured using the embedded position and force sensor
of Amadeo (Tyromotion GmbH). The FMA and the MAS
were chosen because they are widely used and easy to
conduct and administer.13
The FMA is a performance-based index designed to
assess motor function and balance in patients following
a stroke. FMA has been used clinically as well as in re-
search to determine disease severity, describe motor
recovery, and plan and assess treatment interventions. The
FMA assesses isolated voluntary and synergistic move-
ment patterns, grasp, and reflex activity. Scoring in the
FMA is made on the basis of observing the subject’s ability
to complete certain predefined tasks, and a 3-point scale
is used to measure performance (0 = unable; 1 = partial;
2 = performs fully) on 7 test items (total possible score = 14
points) in the hand subsection. A higher FMA score in-
dicates higher motor ability.14
As the secondary outcome measure, the MAS is a
performance-based scale developed to assess everyday
motor function. In contrast to the FMA, the MAS employs
a task-oriented approach to evaluate the performance of
specific functional tasks rather than isolated patterns of
movement. The MAS hand movement is scored on a
7-point scale from 0 to 6, where 6 indicates the optimum
motor behavior. During testing, subjects are required to
perform each task 3 times, and the best result is recorded.
Results
The effectiveness of the proposed method is verified
through the performance of the patient by comparing the
pretraining and post-training results. Four widely used
clinical measurement methods were also used to compare
the results. In order to ensure the validity and consis-
tency of the data, each test was repeated 3 times, and
the mean value of the readings was chosen as the mea-
sured score. All measurement data are presented as
mean ± standard deviation along with the statistical sig-
nificance (P value). The results are shown in Table 1.
As it can be seen from Table 1, the test results show
improvements in all of the 4 measurement methods. Sig-
nificant improvements are observed between the first and
second assessments, and mild improvements are achieved
between the second and third assessments. As suggested
FMA-hand (0-14) MAS-hand movements (0-6) P value
4.3 ± .6 2.3 ± .3 .05
9.3 ± .3 4.6 ± .3 .02
11.0 ± 0 5.3 ± .3 .01
226 X. HUANG ET AL.
by the FMA, the hand score has improved from 4.3 to
9.3 in Phase A, and the improvement continues to 11.0
at the third assessment in Phase B. The MAS hand move-
ment score measures changed from 2.3 to 4.6 in 12 weeks
and slightly increased to 5.3 in 18 weeks, in comparison
with the score measured at Phase A.
Between the results obtained in Phases A and B (sug-
gested by the second and third assessments), apart from
the improvement in force, ROM, FMA, and MAS scores,
the statistical changes are also worth noticing. It can also
be observed that the standard deviation and statistical
significance (P value) are dropped, which means smaller
variations and more consistency. This can be inter-
preted as that the patient can perform the training tasks
more smoothly and with fewer variables within each test.
Figure 5. Output force evolvement.

Statistical Analysis
As mentioned in section 2, 3 clinical assessments were
conducted using the FMA and the MAS at weeks 0, 12,
and 18. According to the recorded data, the FMA and
the MAS are shown in Figure 4.
As can be seen from Figure 4, the performance of the
subject improved in both Phase A (weeks 0-12) and Phase
B (weeks 12-18), and this trend can be seen using both
FMA and MAS evaluation methods. However, the mag-
nitude of the improvement between Phase A and Phase
B is not the same. Actually, the slope of recovery trend
in Phase A is roughly twice that of Phase B, indicating
that the progress in Phase A is more obvious than that
in Phase B. However, due to the limitation that there are
only 3 data sets available for the FMA and MAS assess-
ments, we cannot identify during which sessions these
improvements occurred. Active ROM and force output
data analyses are thus needed to conduct a more de-
tailed conclusion.
At the end of each training session, finger flexion and
extension force and active ROM tests were performed,
Figure 4. FMA and MAS assessment results. Abbreviations: FMA, Fugl-
Meyer Assessment; MAS, Motor Assessment Scale.
and the results were recorded. Therefore, for each of these
evaluation methods, there are 54 data sets available for
a more detailed analysis. More importantly, the data not
only provide a comparison between pretraining and post-
training but also give information on when during the
training the change actually occurred. Therefore, this assists
in identifying the exact optimal length of the training
session required in future studies.
As can be seen in the force output in Figure 5, both
the flexion (grasping) and extension (extending) output
force graphs indicate growth in both Phase A and Phase
B. The fitted lines suggest that after a fast and obvious
rise at the beginning of Phase A, the growing momen-
tum seems to halt into a stage where there is almost no
improvement in the latter half of Phase A after 24 tests
(week 8). And this trend continues until the beginning
of Phase B. However, the curve follows a similar but less
obvious pattern, and the growth almost stabilizes in the
latter half of Phase B as well. This trend can be seen in
both flexion and extension lines. The efficiency of the pro-
posed rehabilitation methods was further examined by
analyzing the active ROM testing results, as shown in
Figure 6.
The ROM analysis shows a significant improvement
in the first half of Phase A; however, after 6 weeks (18
tests) since the beginning of the training sessions, the
growth slows down gradually to be almost flat. In Phase
B (after 36 tests), the ROM initially improves slowly but
stabilizes around 79 after 4 weeks (48 tests shown in the
graph) since the introduction of the proposed hybrid re-
habilitation training protocol.
From the above results, we can interpret that a more
intuitive and engaging training protocol, as adopted in
Phase B, can fulfill the task of providing AAN control
as well as the task-oriented VR-based immersive train-
ing for the subject. It can promote further achievement
(shown in Phase B) after the halted growth using the
Amadeo (Tyromotion GmbH) standard training proto-
col (as shown in Phase A). However, we have to admit
EFFECTS OF ADAPTIVE CONTROL AND VIRTUAL REALITY ON ROBOT-ASSISTED REHABILITATION
Figure 6. Active ROM assessment results. Abbreviation: ROM, range
of motion.
that the improvement in Phase B is also limited because
there are physical limitations in the stroke patient that
hindered him to achieve further improvements in both
Phase A and Phase B.
Discussions and Conclusions
In this study the advantages of force-rendered AAN
adaptive control,15 VR8-based RGS,16 and transferring VE17
were combined with the aim of optimizing the training
effectiveness and enhancing the rehabilitation efficiency.
The chronic stroke subject who underwent the pro-
posed rehabilitation approach showed improvement in
clinical evaluation methods using the FMA and the MAS,
as well as kinematic characteristics suggested by active
ROM and output force intensity. The result of this study
was consistent with the previous work using the same
Amadeo (Tyromotion GmbH) robotic device and its stan-
dard training protocol.11,18 Improvements were observed
in the FMA, output force, and active ROM after the train-
ing sessions. In addition to these achievements, what it
is worth noticing in this study is that the patient can further
improve after the halted trend using the Amadeo
(Tyromotion GmbH) standard training protocol. This sug-
gests that the proposed research protocol can help stroke
patients to achieve better training results.
As revealed by many studies, the proximal improve-
ments in the arm do not necessarily transfer to the distal
arm or vice versa and that arm improvements do not
manifest as improved ADL performance.11 Therefore, rather
than focusing on the proximal UE, namely shoulder,
forearm, and elbow movement rehabilitation in previ-
ous studies, this study targeted the distal hand and finger
as the targeted rehabilitation parts. As suggested by active
ROM, the force and coordination produced by the hand
and finger are critical in performing delicate ADLs, which
may lead to improved QOL.
227
Compared with previous studies, where only acute
(within 1 month after stroke)19,20 or subacute (1-6 months)21
patients were included, this case study was based on a
chronic patient (>6 months). A similar rehabilitation
treatment22 conducted on chronic stroke patients was found
to potentially contribute toward increased upper limb motor
recovery. Thus, the findings of this research would be of
interest to chronic stroke patients as well as the re-
search groups exploring this category of patients. However,
in any benchmarking between methods, the assump-
tion and criteria used in the studies, such as intervention
time, number of training sessions per week, duration of
training sessions, type of stroke, affected brain side, time
poststroke, VR tasks, feedback, and severity of lesion, may
affect the final outcome and should be taken into account.
Finally, we also need to note that the test results are
based on 1 subject only, which makes it difficult to totally
rule out invalid test data. Therefore, larger-scale clinical
trials, following the same inclusion and exclusion crite-
rion, should be conducted on more stroke patients to enable
the collection of more clinical test data and the estab-
lishment of a confirmed conclusion. Further studies should
also explore whether the outcome achieved through the
proposed method can be maintained over a long period.
References
1. Riddoch MJ, Humphreys GW, Bateman A. Stroke.
Physiotherapy 1995;81:689-694.
2. Scrivener K, Sherrington C, Schurr K. A systematic review
of the responsiveness of lower limb physical performance
measures in inpatient care after stroke. BMC Neurol
2013;13:4.
3. Stewart KC, Cauraugh JH, Summers JJ. Bilateral
movement training and stroke rehabilitation: a systematic
review and meta-analysis. J Neurol Sci 2006;244:89-95.
4. Worthen-Chaudhari LC, McDowell SK, Brewer BR.
Poststroke upper extremity rehabilitation: a review of
robotic systems and clinical results. Top Stroke Rehabil
2007;14:22-44.
5. Hallett M. Recent advances in stroke rehabilitation.
Neurorehabil Neural Repair 2002;16:211-217.
6. Poling GL, Weisenberger JM, Kerwin T. The Role of
Multisensory Feedback in Haptic Surface Perception. In
11th Symposium on Haptic Interfaces for Virtual
Environment and Teleoperator Systems (HAPTICS’03).
2003.
7. Llorens R, Noé E, Colomer C, et al. Effectiveness, usability,
and cost-benefit of a virtual reality-based telerehabilitation
program for balance recovery after stroke: a randomized
controlled trial. Arch Phys Med Rehabil 2015;96:418-425,
e2.
8. Saposnik G, Levin M, Outcome Research Canada
(SORCan) Working Group. Virtual reality in stroke
rehabilitation: a meta-analysis and implications for
clinicians. Stroke 2011;42:1380-1386.
9. Darekar A, McFadyen BJ, Lamontagne A, et al. Efficacy
of virtual reality-based intervention on balance and
mobility disorders post-stroke: a scoping review.
J Neuroeng Rehabil 2015;12:46.
10. Sale P, Lombardi V, Franceschini M. Hand robotics
rehabilitation: feasibility and preliminary results of a
228
robotic treatment in patients with hemiparesis. Stroke Res
Treat 2012;2012:1-5.
11. Hwang CH, Seong JW, Son DS. Individual finger
synchronized robot-assisted hand rehabilitation in subacute
to chronic stroke: a prospective randomized clinical trial
of efficacy. Clin Rehabil 2012;26:696-704.
12. Huang X, Naghdy F, Du H, et al. Reinforcement learning
neural network (RLNN) based adaptive control of fine
hand motion rehabilitation robot. In 2015 IEEE Conference
on Control Applications (CCA). 2015.
13. Sivan M, O’Connor RJ, Makower S, et al. Systematic
review of outcome measures used in the evaluation of
robot-assisted upper limb exercise in stroke. J Rehabil
Med 2011;43:181-189.
14. Bosecker C, Dipietro L, Volpe B, et al. Kinematic robot-
based evaluation scales and clinical counterparts to
measure upper limb motor performance in patients with
chronic stroke. Neurorehabil Neural Repair 2010;24:62-69.
15. Marchal-Crespo L, Reinkensmeyer DJ. Review of control
strategies for robotic movement training after neurologic
injury. J Neuroengineering Rehabil 2009;6:20.
16. da Silva Cameirao M, Bermúdez I Badia S, Duarte E,
et al. Virtual reality based rehabilitation speeds up
functional recovery of the upper extremities after stroke:
a randomized controlled pilot study in the acute phase
X. HUANG ET AL.
of stroke using the rehabilitation gaming system. Restor
Neurol Neurosci 2011;29:287-298.
17. Lohse KR, Hilderman CGE, Cheung KL, et al. Virtual
reality therapy for adults post-stroke: a systematic review
and meta-analysis exploring virtual environments
and commercial games in therapy. PLoS ONE 2014;9:
e93318.
18. Sale P, Mazzoleni S, Lombardi V, et al. Recovery of hand
function with robot-assisted therapy in acute stroke
patients. Int J Rehabil Res 2014;37:236-242.
19. Masiero S, Celia A, Rosati G, et al. Robotic-assisted
rehabilitation of the upper limb after acute stroke. Arch
Phys Med Rehabil 2007;88:142-149.
20. Masiero S, Armani M, Rosati G. Upper-limb robot-assisted
therapy in rehabilitation of acute stroke patients: focused
review and results of new randomized controlled trial.
J Rehabil Res Dev 2011;48:355.
21. Colombo R, Sterpi I, Mazzone A, et al. Robot-aided
neurorehabilitation in sub-acute and chronic stroke: does
spontaneous recovery have a limited impact on outcome?
Neurorehabilitation 2013;33:621-629.
22. Sale P, Bovolenta F, Agosti M, et al. Short-term and
long-term outcomes of serial robotic training for
improving upper limb function in chronic stroke. Int J
Rehabil Res 2014;37:67-73.