Soft Exoskeleton Glove for Hand Assistance Based on

Human-machine Interaction and Machine Learning

Xiaoshi Chen*, Li Gong*, Lirong Zheng, Zhuo Zou
State Key Laboratory of ASIC and System, Fudan University, Shanghai, China
E-mail: { lrzheng, zhuo}@fudan.edu.cn

Abstract—This paper proposes a human machine interaction Some work has been done in the field of exoskeleton
system in the field of stroke rehabilitation, based on the concept gloves [8-11]. M. Cortese et al. [12] develop a novel
of mirror therapy (MT). It aims to improve the hand motor mechatronics master-slave setup for hand rehabilitation. It
function of stroke patients, enabling a true synchronization consists of a sensor glove acting as a remote master and a
between the affected hand and non-affected hand (healthy hand) powered hand exoskeleton as a slave. The master unit can
for the stroke patient. It consists of a soft exoskeleton glove, a control the slave unit to assist the affected hand to complete
surface electromyography (sEMG) signal collecting armband mirror gestures, including “Rest”, “Pinch”, and “Lateral
and machine learning (ML) algorithms. The glove is developed grasp”. J. H. Bae et al. [13] design a wearable hand
by integrating low-power motors to provide force strength for
rehabilitation robot, DULEX-II. The DULEX-II is with three
the hand movement. Unlike the rigid exoskeleton devices, the
degrees of freedom, which can assist both wrists and all
glove is comfortable to wear and lightweight, so it is more
suitable for rehabilitation training of stroke patients in daily life. fingers movements. Serpelloni et al. [14] design a a soft
The armband collects the sEMG signals for pattern recognition robotic glove that is driven by EMG for hand mirror therapy.
by the ML algorithms. In the experiment, four subjects perform The system uses a brushless motor to push and pull the cable,
10 hand gestures to collect data for model training. A which drives the finger movement.
comparison of data preprocessing is conducted to find the In this paper, a rehabilitation system for a power-assisted
optimal data segmentation method and feature vector sets. A
glove control employing a commercial wearable MYO
series of pattern recognition algorithms are developed and
armband is proposed. The system consists of a soft
assessed in different aspects, including prediction accuracy,
training time and predicting time. All 10 gestures can be
exoskeleton glove, a sEMG signal collecting armband and ML
recognized in offline mode with an accuracy up to 99.4%. The algorithms. The armband is worn on a non-affected hand to
control of soft exoskeleton glove in real-time manner is also collect the sEMG signals to sense the hand motion of the
carried out, and the accuracy is 82.2%. The experiment result patient. Then the ML algorithms recognize what motions the
demonstrates the feasibility of the proposed system. The hand made and send classification results to the exoskeleton
innovations and limitations of the work are discussed at the end glove. The exoskeleton glove is controlled to assist the
of the paper. affected hand to accomplish the mirror movements. The
contribution of this work includes: provide the exoskeleton
Keywords—soft exoskeleton glove, surface electromyography glove which the patients can wear on the affected hand with
(sEMG), mirror therapy, machine learning higher comfort and safety than rigid exoskeletons. The patient
can take rehabilitation training at home without being
I. INTRODUCTION confined to the medical environment. Besides, the system has
Stroke is a life-threatening disease, caused by blockage of customized 10 kinds of fine-grained training gestures for
a blood vessel to the brain or bleeding in or around the brain mirror therapy, aiming to train each finger individually and
[1]. About 55% to 75% of survivors have serious disabilities train multiple fingers in coordination. Moreover, the accuracy
such as paralysis and aphasia [2-4]. Among them, the most and training time of different ML algorithms are analyzed and
common is the loss of part or all of the hand’s motor function compared.
[5], resulting in patients unable to take care of themselves.
The rest of this paper is organized as follows. The
Hand dysfunction can seriously affect the quality of patients’
framework of the proposed system is described in Section II.
life, which means more demand is needed on the hand
The methodology of gesture recognition and control of the soft
rehabilitation.
exoskeleton glove are introduced in Section III. Then the
Mirror therapy (MT) is one of the traditional methods for experiment results are evaluated and discussed in Section IV.
improving hand motor function, which is simple, effective, Finally, Section V concludes the whole paper.
and inexpensive [6]. The traditional mirror therapy refers to
placing a mirror in front of the patient. The patient observes II. FRAMEWORK OF THE PROPOSED SYSTEM
the image of the non-affected hand in the mirror, and imagines A. Overview of the System Architechture
the affected hand making the same action, so as to achieve the
therapeutic purpose through optical illusion and visual The architecture of the system implemented in this study
feedback. Mirror therapy is now extended to another method, is shown in Fig. 1. The proposed system consists of the
bilateral movement [1, 7]. In bilateral movement, the following parts: a human-machine interaction module
rehabilitation equipment is used to assist the affected hand to collecting sEMG signals from the user, a pattern recognition
make mirror gestures as the non-affected hand, generating module classifying different hand gestures, and an
direct visual feedback and promoting the rehabilitation of the exoskeleton glove providing power assistance.
patient’s hands. Studies have shown that bilateral The conventional mirror therapy blocks the visual input to
simultaneous exercise, for mirror therapy, is more effective “cheat” the user, and the affected hand does not make any
than unilateral exercise. movement in fact. In this paper, we propose a rehabilitation
therapy scenario that enables a true synchronization between
the affected hand and non-affected hand for the stroke patient.
* X. Chen and L. Gong contributed equally to this work The stroke patient wears MYO armband on the non-affected

978-1-7281-5871-6/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: San Jose State University. Downloaded on September 01,2021 at 21:28:19 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. Architecture of the stroke rehabilitation system. Fig. 3. System architecture of the glove.

Fig. 4. The illustration of 10 hand gestures.

Fig. 2. The soft exoskeleton glove.

forearm and wears the exoskeleton glove on the affected hand.

The user’s non-affected hand performs a gesture such as
bending fingers or fist. The MYO armband recognizes the
gesture of the non-affected hand through the proper ML
algorithms. Then the corresponding commands are sent to the
exoskeleton control. With the help of exoskeleton drive
modules, the affected hand can perform the same gesture. The Fig. 5. The sEMG signal segmentation.
hand perception module measures the motion state of the
affected hand in real-time, providing feedback signals to the
control module and correcting the fine motion of the affected III. METHODOLOGY
hand. The advantage of the proposed rehabilitation therapy
scenario is that the patient can not only get visual feedback A. The Human Machine Interaction
from watching the affected hand, but also perform the same During the experiment, the subjects are asked to sit on a
actions on the affected hand assisted by exoskeleton glove. chair in a relaxed position, and wear the MYO armband on the
non-affected forearm. The sampling rate of the MYO armband
B. Glove Design is 200 Hz. The subjects are required to make the 10 specific
In this paper, we implement a soft exoskeleton glove that hand gestures, as shown in Fig 4. Each of the gestures should
enables power assistance for the affected hand. The be held for more than 10 seconds for recording. Then the same
exoskeleton glove mainly consists of the following parts: a gesture is repeated. To avoid muscle tension, enough rest is
soft glove, pressure and flex sensors, an electronic control arranged for each time of recording. Each subject completes
module, a power-assisted module, and a wireless module. Fig. 20 sEMG data record. Therefore total 80 of sEMG data
2 shows the placement of sensors and wires inside the glove. recordings are obtained. The collected data is for both model
Fig. 3 is the architecture of the glove control system. We training and testing.
select STM32F103 as the core of the control logic. A sensor B. Pattern Recognition
interface is implemented to process the signals from pressure
sensors and flex sensors, including signal filtering, 1) Data segmentation
amplification, and digitization. To provide enough power In order to fully utilize the sEMG signal, we chose the
strength to the fingers, the micro motors are used. We overlapped segmentation method [17], as shown in Fig. 5.
implement five H-Bridge circuits to drive the motors. The The length of the data window and time increment are two
details of the actuation design can be found in [15, 16]. important factors affecting the accuracy of the model training.
Considering the balance between power consumption and A longer data window contains more effective information
wearable requirement, we select two 18650 Li-ion batteries while decreases the real-time performance. A shorter time
that have a total capacity of 23.040 Wh. Therefore, the increment window results in more segmented data set, but it
exoskeleton glove can fulfill the requirement of rehabilitation also increases the complexity of model training. In this sense,
in daily life. we conduct an experiment on how to choose the parameters

Authorized licensed use limited to: San Jose State University. Downloaded on September 01,2021 at 21:28:19 UTC from IEEE Xplore. Restrictions apply.
 Fig. 7. Classification result of a signal feature.
Fig. 6. Accuracy using different overlapped and data windows.

classification accuracy. The optimal feature set contains MAV,

for data segmentation. We compare the impact of different WL, RMS, and STD.
lengths of increment and data windows. SVM is selected as
the classifier in this test. The results are shown in Fig. 6. 3) Pattern recognition
In this session, we evaluate a variety of ML algorithms,
From the Fig. 6, we can find that with the increase length including supervised learning and ensemble learning. The
of data window, the accuracy increases synchronously. purpose of the evaluation is to find the most suitable
However, the growth rate of accuracy slows down algorithms in offline and real-time classifications, respectively.
significantly after the data window length reaches 200 ms.
The decision tree is a very common ML algorithm. One of
What’s more, there is no big difference between the two
the advantages of the decision tree is its interpretability. It is
curves. It indicates that the time increment length has little also fast for fitting and low on memory usage. How to set the
influence on the result as long as there are enough samples in number of leaves and branches in the tree is an issue worth
training set. In addition, it is recommend that the control considering, which is a balance between the model accuracy
delay should be less than 300 ms in real-time [18], so the and robustness.
segmentation length is set no longer than 300 ms.
LDA is another useful ML algorithm in solving
2) Feature extraction classification problems. It has shown the features of
The amplitude, energy, and frequency of the sEMG signal computing efficiency [19] and robust [20]. The goal of LDA
vary with the degree of muscle contraction. It is not practical is to project the data set onto the line so that the similar data is
if the raw data is fed to the classifier directly for pattern as close as possible while different types of data are as far
recognition. In this session, we evaluated the performance of away as possible. The maximized target of LDA is listed in
different features. A comparison is presented to show the most the equation below:
effective features for pattern recognition.
ɘ் ܵ௕ ɘ
There are seven time domain features involved in the ‫ܬ‬ൌ (1)
ɘ் ܵఠ ɘ
comparison: mean absolute value (MAV), zero crossings
(ZC), waveform length (WL), slope sign changes (SSC), root Where ܵఠ is within-class scatter matrix, and ܵ௕ is between-
mean square (RMS), variance (VAR) and standard deviation class scatter matrix.
(STD). SVM has shown a strong capability in classifying
With the features above, we made a comparison to choose biomedical signals in previous researches [14, 21]. The key
the optimal feature set in pattern recognition. The rule of the point of SVM is to find a hyperplane that has the maximum
comparison is as follows. At first, the ML algorithm runs with margin for the support vectors. The equation is
every single feature. To avoid the impact of a specific ݂ሺ‫ݔ‬ሻ ൌ ɘ் ߶ሺ‫ݔ‬ሻ ൅ ܾ (2)
algorithm, three ML algorithms are involved, including SVM,
kNN and Subspace kNN. Once the performance of each Where ߱is a weight vector, ߶ሺ‫ݔ‬ሻis the kernel map, andܾis
an offset.
feature is achieved, the features are sorted according to the
classification accuracy. Fig. 7 shows the result of running kNN is another popular ML algorithm for pattern
with a single feature. recognition. It is easy to implement with high accuracy. In
To evaluate the feature performance in different classifiers, this study, a variety of parameter combinations are tested,
the average value is calculated. It is shown that most of the including the number of neighbors, distance weight, etc. The
features achieve a relative high accuracy except ZC (42.1%, ensemble learning is also evaluated in this study. The
35.6%, and 41.4%) and SSC (18.20%, 14.00%, and 15.80%). serialization or parallelization method can be generated
From Fig. 7 we can find that there is no significant difference depending on whether there is a strong dependency between
for MAV, WL, RMS, and STD. The accuracy is quite close to the base learners. The typical algorithm for each method is
each other, around 88% respectively. Then we made a few boosting and bagging. By combining many learners it is
feature sets by combing features together to find the highest possible to obtain a more accurate result than a single learner.

Authorized licensed use limited to: San Jose State University. Downloaded on September 01,2021 at 21:28:19 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. RESULTS OF HAND GESTURE RECOGNITION

Discrim
Neural
Tree inant SVM KNN Ensemble learning
network
analysis
Machine learning
algorithm Linear Fine Subspace hidden
Fine Linear Quadrat Cubic Fine Cosine Cubic Weighte Boosted Bagged Subspac
Discrimi Gaussia Discrimin layer:
Tree SVM ic SVM SVM KNN KNN KNN d KNN Trees Trees e KNN
nant n SVM ant 27, 21

Accuracy 83.9% 68.0% 92.2% 98.6% 99.1% 99.3% 99.4% 98.4% 98.5% 99.2% 64.9% 98.6% 65.7% 99.4% 99.3%

Input Training
16.20 3.99 122.11 106.88 106.52 364.63 36.87 36.72 1090.00 35.68 302.30 87.26 32.48 578.86 31.00
pattern 1 Time (s)
Predict
6.516 7.386 52.635 60.439 58.212 89.681 15.356 30.213 41.555 13.055 25.248 34.637 42.508 152.022 4.947
Time (ms)
Accuracy 81.0% 66.4% 90.7% 98.1% 98.7% 98.8% 98.9% 97.7% 97.9% 98.6% 64.0% 98.0% 64.3% 99.0% 98.3%
Input Training
14.84 2.65 112.64 101.79 103.52 404.70 32.38 32.09 1007.60 32.35 272.23 80.92 28.43 524.30 15.00
pattern 2 Time (s)
Predict
7.142 7.617 52.938 60.638 60.020 89.027 13.880 30.087 40.577 13.693 22.369 36.009 42.543 148.797 5.160
Time (ms)
Accuracy 78.4% 64.7% 88.6% 97.0% 98.0% 97.8% 98.0% 96.5% 96.7% 97.6% 62.1% 97.0% 62.6% 98.2% 97.3%
Input Training
15.85 3.00 122.66 109.62 112.05 456.55 32.83 32.28 1037.80 33.05 275.87 86.57 29.77 536.22 22.00
pattern 3 Time (s)
Predict
8.406 6.925 53.187 57.515 56.942 84.308 13.601 32.071 39.567 12.544 22.867 36.082 40.919 144.385 5.561
Time (ms)

Input pattern 1: Data length = 300ms, time increment = 50ms; Input pattern 2: Data length = 250ms, time increment = 50ms; Input pattern 3: Data length = 200ms, time increment = 50ms

It is also worth noting that the ensemble classifiers may lower than others, from 88.6% to 92.2% among the three
consume much more time compared with other classifiers. different input patterns. As to the kNN, different parameters
Therefore how to implement a proper ensemble classifier is result in similar results, which are fluctuating from 96.4% to
an essential issue in ensemble learning. 99.4% among different input patterns and distance metrics.
The performance of kNN and SVM are quite similar in terms
At last, a feedforward neural network is implemented. of accuracy. The ensemble learning also shows an interesting
The number of nodes for the input layer is equal to the result. The subspace kNN has the highest accuracy (99.4%)
number of feature extracted vectors. The hidden layer and among all the classifiers, which is slightly higher than kNN.
output layer transfer function is tan-sigmoid function. There Due to the combination of weak learners, the Bagged Trees
are two layers of hidden layer, with 27 nodes and 21 nodes, can reach 98.6% when input is at 300ms, which is far more
respectively. Scaled conjugate gradient backpropagation is higher than the fine tree (83.9%). In contrast, the Boosted
used for network training. The weight for each node is Trees’ accuracy is not only lower than that of the Bagged
adjusted by using a gradient descent strategy. Trees, but also much lower than that of the Fine Tree.
Therefore, we believe that ensemble learning can achieve
C. Hand Gesture Classification Result higher classification accuracy than a single classifier only
As discussed in the previous section, a variety of ML through a proper combination of multiple learners. In hand
classifiers are applied, so we can find the most suitable gesture recognition, a parallelization method can achieve a
classifiers for gesture recognition and glove control. The result better performance than the serialization method. Last but not
of hand gesture recognition is presented in Table I. least, the neural network also shows very high classification
accuracy (99.3%), which is comparable with SVM, kNN and
In this study, 3 types of input patterns are evaluated, the Bagged Trees.
difference is the length of the data window. Obviously, a
longer piece of data contains more sEMG signals, making it The training time and predict time is another important
easier for gesture recognition. Considering the constraint ofaspect of evaluating classifier performance. The Fine Tree and
LDA has quite short training time and predict time. Especially
real-time performance, the input pattern length is set from 200
ms to 300 ms. As mentioned in the previous section the time the LDA has the shortest training time around 3 seconds,
which shows its computing efficiency. However, due to their
increment is related to the total number of training sets, so it
is set to 50 ms. lower accuracy, they are not the most suitable classifiers.
SVM and kNN have quite high accuracies as discussed above.
Accuracy is the most important factor in evaluating ML The training time of kNN is obviously shorter than SVM’s
algorithms. In general, the accuracy increases with the except Cubic kNN. The real-time predict time of SVM with
increase of segmented data length. From Table I we can find fine Gaussian kernel function reaches 84.308 ms while input
that the decision tree’s accuracy and LDA’s accuracy are data length is 200ms, which hardly meets the real-time
lower than that of other classifiers, around 80%, and 65% performance requirement. As to ensemble learning, the
respectively. As to SVM, we conduct the experiment with a training time is much longer than that of the corresponding
variety of kernel functions. Compared with the decision tree single classifier due to the combination of multiple weak
and LDA, SVM reaches a relatively high accuracy. The learners. The Bagged Trees get a balance between accuracy
accuracy of SVM with Gaussian function kernel is up to 99.3% and predict time with a predicted time around 36 ms. Although
while the data length is 300ms. Linear kernel’s accuracy is the accuracy of Subspace kNN is the highest its real-time

Authorized licensed use limited to: San Jose State University. Downloaded on September 01,2021 at 21:28:19 UTC from IEEE Xplore. Restrictions apply.
Fig. 8. The illustration of 10 hand gestures, the gestures are controlled by
the hand wearing MYO armband.

predict time is also extremely long, varying from 144.385 ms

to 152.002 ms. Together with the length of the input data, the Fig. 9. Results of real-time classification accuracy.
total latency exceeds the constraint of 300 ms. Therefore it is
not suitable for real-time controlling in this study. At last, the
neural network has shown very good real-time performance. the gesture for the same reason. Another reason is that the
Its real-time predict time is the shortest among all classifiers, forearm muscle of the human body has a coupling effect, and
and at the same time, the high classification accuracy is any finger movement will generate some additional force on
guaranteed. other fingers. These gestures are similar to each other, which
leads to misclassifications. Another important issue is the
With the discussion and analyzation above, we can find calibration of the sEMG signal collecting. Although wearing
that there is a slightly accuracy loss with the decrease of data the position of the armband is recorded for each subject, there
window length. Considering the wireless transmission latency is still a certain deviation of the electrode position, which is
between the armband and PC, as well as PC and the likely to introduce a few input errors during testing.
exoskeleton glove, we selected the classifiers based on the
data length of 200 ms, so that the real-time performance can IV. EVALUATION AND DISCUSSION
be guaranteed. Four classifiers, fine kNN, Bagged Trees, Compared with other work, the innovation of the work is
Cubic SVM and neural network, are used for real-time control the application of soft exoskeleton glove in the proposed
of the soft exoskeleton glove. rehabilitation system. In the conventional mirror therapy, the
D. Control of Soft Exoskeleton Glove patient looks at the non-affected hand movement in the mirror
and images the movement of the affected hand. While in this
The real-time control of soft exoskeleton glove is
work, there is an authentic movement on the affected hand.
accomplished using the classifiers derived in the previous
Compared with the rigid exoskeletons, the soft exoskeleton
section. The purpose of the experiment is to assess the
glove is lightweight and comfortable for users, the affected
performance of the classifiers in real-time manner and the
hand can have a more intuitive experience. Thus a better
practical applications of mirror therapy. The soft exoskeleton
rehabilitation result can be achieved in our proposed system.
glove is mounted to a wooden prosthesis which cannot
provide any force by itself. Fig. 8 displays a series of There are some limitations of the system. One potential
snapshots of the wooden prosthesis assisted by the soft limitation of the work is that the classifier of ML is not
exoskeleton glove. universal for all users. The data set obtained for the ML
algorithm is only from four subjects, although each gesture is
In the beginning, all fingers are in a relaxed state. The non-
repeated a few times for enlarging the data set. A possible
affected forearm wearing the armband is also in a relaxed state.
solution is to create a larger data set by collecting sEMG
Once the muscle onset, the classified gesture result is
signals from more subjects. Also, the classifiers can be further
transmitted to the glove, thus the corresponding fingers are
evaluated, considering their generalization ability.
bent by the integrated power-assisted cable. Each gesture is
repeated 10 times. The performed gestures and classified Furthermore, sensor information fusion is also an issue
gestures are both recorded to calculate the recognition rate. worth considering. Due to the limitation of the sEMG signal,
the recognition accuracy of some fine movement is not high.
The result of real-time classification accuracy is illustrated
One approach to identifying more gestures with high accuracy
in Fig. 9. As we can see some gestures remain the recognition
is to fuse sensor information from other dimensions. It can be
accuracy as high as that of the offline classification. However,
a future work of putting flexible sensors on the surface of the
the recognition rate of some gestures is significantly lower
hand, detecting biomedical signals during the hand movement.
than the results of offline classification, such as EI, EMI, and
Therefore a hybrid sensor signals can be used to determine the
FTI. The recognition rate of these gestures is between 60%
precise motion of a human hand.
and 80% in all four classifiers. In general, the average
accuracy for the four classifiers is 79.14%, 77.49%, 81.03%, Additionally, although the glove is capable of applying
and 82.19%, respectively. There are some reasons for the force to each finger of the affected hand, the direction of the
lower accuracy in real-time classification. First, the transient force is fixed. Currently, the glove is not able to help extend
state of sEMG signals is also involved in the classification fingers, it can only help bend fingers. Experiments are
after the muscle onset, although a threshold is set to reduce its conducted to implement a bidirectional power assistance
impact. The misclassification is very likely to occur at the glove. However, this may lead to a more complex cable
beginning of the gesture. And it is also observed at the end of routing inside the glove, as well as more motors integrated to

Authorized licensed use limited to: San Jose State University. Downloaded on September 01,2021 at 21:28:19 UTC from IEEE Xplore. Restrictions apply.
provide power strength in the opposite direction, which [8] N. F. Roslan, A. M. M. Ali, and R. Suhairi, "Design of an artificial Arm
significantly increases the weight of the entire system. For the with hand gripper for telemedicine system," i Int. Conf. Eng. Technol.
Technopreneuship, ICET, pp. 284-288, 2014.
future work, it is worth keeping a balance between the system
[9] A. M. M. Ali and M. M. A. Jamil, "Biomechatronics design of a novel
functionality and complexity, fulfilling the requirement of artificial arm," BMEiCON - Biomed. Eng. Int. Conf., 2013.
portable and wearable in our proposed system. [10] H. S. Cao and D. G. Zhang, "Soft robotic glove with integrated sEMG
sensing for disabled Ppeople with hand paralysis," IEEE Int. Conf.
V. CONCLUSION Robot. Biomim., ROBIO, 2016, pp. 714-718.
This study presents a hand rehabilitation system consisting [11] L. Gerez, J. Chen and M. Liarokapis, "On the Development of Adaptive,
of a portable soft exoskeleton glove and a human-machine Tendon-Driven, Wearable Exo-Gloves for Grasping Capabilities
interface based on ML. The glove can provide external power Enhancement," IEEE Robot. Autom., vol. 4, no. 2, pp. 422–429, 2019.
strength to each finger to help the rehabilitation of the affected [12] M. Cortese, M. Cempini, P. R. D. Ribeiro, S. R. Soekadar, M. C.
Carrozza, and N. Vitiello, "A mechatronic system for robot-mediated
hand. The hand gesture recognition interface is implemented hand telerehabilitation," IEEE-ASME Trans. Mechatron, vol. 20, no. 4,
with accuracy in offline mode (99.4%) and real-time mode pp. 1753-1764, Aug 2015.
(82.2%), respectively. The result shows the feasibility of the [13] J. H. Bae, Y. M. Kim, and I. Moon, "Wearable hand rehabilitation robot
proposed system, and the limitations and future works are also capable of hand function assistance in stroke survivors," in Proc. IEEE
discussed. RAS EMBS Int. Conf. Biomed. Rob. Biomechatronics, BioRob, 2012,
pp. 1482–1487.
REFERENCES [14] M. Serpelloni, M. Tiboni, M. Lancini, S. Pasinetti, A. Vertuan, and M.
Gobbo, "Preliminary study of a robotic rehabilitation system driven by
[1] Z. Yue, X. Zhang, and J. Wang, "Hand rehabilitation robotics on
EMG for hand mirroring," in IEEE Int. Symp. Med. Meas. Appl.,
poststroke motor recovery," Behav Neurol, vol. 2017, p. 20, 2017
MeMeA - Proc., 2016, pp. 1–6.
[2] A. Basteris, S. M. Nijenhuis, A. H. A. Stienen, J. H. Buurke, G. B.
[15] X. Chen et al., "A wearable hand rehabilitation system with soft
Prange, and F. Amirabdollahian, "Training modalities in robot-
gloves," in IEEE Trans. Ind. Inform., in press.
mediated upper limb rehabilitation in stroke: a framework for
classification based on a systematic review," J Neuroeng Rehabil, vol. [16] M. Xiloyannis, L. Cappello, D. B. Khanh, S.-C. Yen, and L. Masia,
11, Jul 10 2014. “Modelling and design of a synergy-based actuator for a tendon-driven
soft robotic glove,” in Proc. IEEE RAS EMBS Int. Conf. Biomed. Rob.
[3] A. Chowdhury et al., "Active physical practice followed by mental
Biomechatronics, BioRob, 2016, pp. 1213–1219.
practice using BCI-Driven hand exoskeleton: A pilot trial for clinical
effectiveness and usability," IEEE J Biomed Health Inform, vol. 22, no. [17] N. Nilsson, B. Hakansson, and M. Ortiz-Catalan, "Classification
6, pp. 1786-1795, Nov 2018. complexity in myoelectric pattern recognition," J Neuroeng Rehabil,
vol. 14, Jul 10 2017.
[4] D. S. Nichols-Larsen, P. C. Clark, A. Zeringue, A. Greenspan, and S.
Blanton, "Factors influencing stroke survivors' quality of life during [18] B. Hudgins, P. Parker, and R. N. Scott, "A new strategy for
subacute recovery," Stroke, vol. 36, no. 7, pp. 1480-4, Jul 2005. multifunction myoelectric control," IEEE Trans. Biomed. Eng, vol. 40,
no. 1, pp. 82-94, Jan 1993.
[5] S. L. Wolf et al., "Effect of constraint-induced movement therapy on
upper extremity function 3 to 9 months after stroke - The EXCITE [19] K. Englehart and B. Hudgins, "A robust, real-time control scheme for
randomized clinical trial,", Jama-J Am Med Assoc, vol. 296, no. 17, pp. multifunction myoelectric control," IEEE Trans. Biomed. Eng, vol. 50,
2095-2104, Nov 1 2006. no. 7, pp. 848-854, Jul 2003.
[6] C. Y. Wu, P. C. Huang, Y. T. Chen, K. C. Lin, and H. W. Yang, [20] P. Kaufmann, K. Englehart, and M. Platzner, "Fluctuating EMG signals:
"Effects of mirror therapy on motor and sensory recovery in chronic investigating long-term effects of pattern matching algorithms," Annu.
stroke: a randomized controlled trial," Arch Phys Med Rehabil, vol. 94, Int. Conf. IEEE Eng. Med. Biol. Soc., EMBC, 2010, pp. 6357-6360
no. 6, pp. 1023-30, Jun 2013. [21] M. A. Oskoei and H. Hu, "Support vector machine-based classification
[7] A. Boos, Q. Y. Qiu, G. G. Fluet, and S. V. Adamovich, "Haptically scheme for myoelectric control applied to upper limb," IEEE Trans.
facilitated bimanual training combined with augmented visual Biomed. Eng, vol. 55, no. 8, pp. 1956-65, Aug 2008.
feedback in moderate to severe hemiplegia," in Proc. Annu. Int. Conf.
IEEE Eng. Med. Biol. Soc. EMBS, 2011, pp. 3111-3114.

Authorized licensed use limited to: San Jose State University. Downloaded on September 01,2021 at 21:28:19 UTC from IEEE Xplore. Restrictions apply.