A Human-Computer Interface based on Electromyography

Command-Proportional Control

Sergey Lobov

, Nadia Krilova

, Innokentiy Kastalskiy

, Victor Kazantsev

and Valeri A. Makarov

1,2

Lobachevsky State University of Nizhny Novgorod, Gagarin Ave. 23, 603950, Nizhny Novgorod, Russia

Instituto de Matemática Interdisciplinar, Applied Mathematics Dept., Universidad Complutense de Madrid,

Avda Complutense s/n, 28040, Madrid, Spain

Keywords: Electromyography, Human-Computer Interface, Pattern Classification, Artificial Neural Networks.

Abstract: Surface electromyographic (sEMG) signals represent a superposition of the motor unit action potentials that

can be recorded by electrodes placed on the skin. Here we explore the use of an easy wearable sEMG

bracelet for a remote interaction with a computer by means of hand gestures. We propose a human-

computer interface that allows simulating “mouse” clicks by separate gestures and provides proportional

control with two degrees of freedom for flexible movement of a cursor on a computer screen. We use an

artificial neural network (ANN) for processing sEMG signals and gesture recognition both for mouse clicks

and gradual cursor movements. At the beginning the ANN goes through an optimized supervised learning

using either rigid or fuzzy class separation. In both cases the learning is fast enough and requires neither

special measurement devices nor specific knowledge from the end-user. Thus, the approach enables

building of low-budget user-friendly sEMG solutions. The interface was tested on twelve healthy subjects.

All of them were able to control the cursor and simulate mouse clicks. The collected data show that at the

beginning users may have difficulties that are reduced with the experience and the cursor movement by

hand gestures becomes smoother, similar to manipulations by a computer mouse.

1 INTRODUCTION

Recent years witness a rapidly growing interest to

the development of devices controlled by

electromyographic (EMG) signals through a human-

machine interface. There have been proposed

interfaces controlling personal computers (PC)

(Chowdhury et al., 2013; “Myo™ Gesture Control

Armband”, 2013), mobile and humanoid robots

(Wang et al., 2012; Lobov et al., 2015a; Lobov et

al., 2015b), powered prostheses (Roche et al., 2014;

Hahne et al., 2014), and exoskeletons (Kiguchi and

Hayashi, 2012; Singh and Chatterji, 2013; Mironov

et al., 2015), among others. Despite technical differ-

rences in the implementation, such devices in gene-

ral exploit quite similar controlling strategies (see,

e.g., Peerdeman et al., 2011; Roche et al., 2014).

The simplest approach uses a single-channel

recording of the bioelectrical activity of a muscle

and applies either proportional (gradual) (Bottomley

and Cowell, 1964) or trigger-like (Kobrinskiy, 1960)

transformation to generate the controlling output.

Multi-channel setups allow for simultaneous

treatment of the activity of several muscles and, in

general, are more promising due to higher number of

degrees of freedom involved in the analysis. Then,

commands sent to an external device can be

evaluated either by a regression over EMG signals

or by a classification EMG in terms of the classical

pattern recognition problem (Kiguchi and Hayashi,

2012; Roche et al., 2014).

Some of the proposed techniques have been

implemented in commercially available devices. For

example, the wearable bracelet MYO™ (Thalmic

Labs Inc.) employs classification of five hand

gestures for managing a personal computer (PC)

(“Myo™ Gesture Control Armband”, 2013). This

bracelet, however, does not implement sEMG-based

proportional control of a PC cursor. Instead, it uses

measurements of spatial coordinates of a hand. This

imposes restrictions for the use of the device by

disabled people, e.g. by amputees.

The powered prostheses available on the market

support either single channel or multichannel

regression strategies generating controlling output

(Roche et al., 2014). At the time being, pattern

Lobov, S., Krilova, N., Kastalskiy, I., Kazantsev, V. and Makarov, V.

A Human-Computer Interface based on Electromyography Command-Proportional Control.

DOI: 10.5220/0006033300570064

In Proceedings of the 4th International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX 2016), pages 57-64

ISBN: 978-989-758-204-2

recognition methods have not been implemented yet

due to a limited number of possible types of

movements. Although the regression based methods

showed important advances, there also have been

revealed weak points. For example, the setup tuning

procedure usually takes relatively long time and

requires specific instruments for measuring the exact

position of different parts of the arm (Fougner et al.,

2012; Jiang et al., 2012; Hahne et al., 2014).

Artificial neural networks (ANNs) have also

widely been used to solve both sEMG classification

and regression problems. Their accuracy is rather

high in comparison to other approaches (see, e.g.,

Peerdeman et al., 2011; Baspinar et al., 2013).

However, the adaptation of ANNs to commercially-

ready human-computer interfaces is still an open

problem that also requires investigation of the user

experience and potential restrictions.

In this work we describe a low-cost human-

computer interface that uses multichannel sEMG

signals processed by ANN. An output of the ANN is

then used as a controlling signal for a PC. Thus, the

gesture classification and regression are strongly

overlapping processes made by the ANN. Moreover,

the ANN learning can be accomplished relatively

fast and requires no special measurement techniques.

Using this interface a user can move a mouse cursor

on a computer screen by hand movements (muscle

contraction) and simulate mouse clicks. Then

disabled people and amputees can use such an

interface in their daily life.

2 METHODS

2.1 Subjects and Testing Task

For experimental purpose we recruited 12 healthy

volunteers of either sex from 21 to 41 years old. The

study complied with the Helsinki declaration

adopted in June 1964 (Helsinki, Finland) and revised

in October 2000 (Edinburg, Scotland). The Ethics

Committee of the Lobachevsky State University of

Nizhny Novgorod approved the experimental

procedure. All participants gave their written

consent.

After machine learning of the interface ANN

(individual for each subject) all participants were

asked to move remotely (using hand gestures) a PC

pointer in Windows OS and to perform the

following testing task: open calculator application,

type "2 + 2 =", find the result and close the

application. Each participant performed the testing

task twice to examine two types of the learning

procedures (see below). Then each participant

described verbally his/her user experience.

All subjects had no previous experience with

sEMG interfaces, except one who used MYO

bracelet for two weeks. For this "experienced" user

an additional test was developed. The user was

asked to connect four points on the computer screen

forming a diamond: a) by using a computer mouse

and b) by the sEMG interface. Then the performance

of "diagonal” movements was evaluated as

=1–





−







, (1)

were 



and 



are the lengths of the curves drawn

by means of the sEMG interface and computer

mouse, respectively.

Figure 1: The hardware-software system MyoCursor. User

with a MYO bracelet placed on forearm controls a PC.

The bracelet transmits eight sEMG signals through a

Bluetooth interface to the PC equipped with MyoCursor

software.

2.2 Myocursor Hardware-Software

Setup

Figures 1 and 2 show the developed hardware-

software system, called MyoCursor. The system

consists of a MYO bracelet worn on a forearm of the

user and a PC with specially designed software (Fig.

1). The bracelet has eight equispaced sensors

acquiring myographic signals at F = 1 kHz rate.

However, our tests have shown that it cannot deliver

data with the frequency above 300 Hz. Missing data

are filled in by returning previously sampled values.

The sEMG signals are sent through a Bluetooth

interface to a PC. We use the MYO SDK to access

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics

raw eight-channel data, while the built-in software

of the bracelet is disabled. Acquired signals are then

processed by MyoCursor software in real-time (Fig.

2). The software performs the recognition of hand

gestures and estimates the muscle efforts that finally

control the cursor in Windows OS (Microsoft Inc.)

in a way similar to that one can achieve with

ordinary computer mouse.

Figure 2: MyoCursor interface. The software processes in

real time the sEMG signals and generates commands

controlling the mouse cursor. Left-top: controlling hand

gestures, red button marks the recognized gesture (wrist

flexion). Right: Example of eight sEMG signals

corresponding to the performed gesture (vertical and

horizontal axes are in mV and s, respectively). Left-

bottom: Controlling toolbars.

2.3 Basic Hand Gestures Imitating

Mouse Manipulations

Natural hand gestures can be extremely rich. For the

human-computer interface we have selected the

following seven static hand gestures as basic motor

patterns: 1) hand at rest is used for eliminating the

cursor trend (see below); 2) hand clenched in a fist

simulates mouse-left click; 3,4) wrist flexion and

extension imitates the cursor movement to the left

and to the right, respectively; 5,6) radial and ulnar

deviations simulate up and down cursor movements,

respectively; and 7) extended palm (fingers together

or separately) is used for imitation of the mouse-

right click. An artificial neural network (see below)

should learn sEMG patterns associated with these

basic motor patterns. For machine learning we

adopted two procedures:

i) A user performs two series each consisting

seven basic gestures.

ii) A user performs two series as in (i) and four

additional gestures that are combination of pair

gestures 3-6 (e.g. simultaneous wrist flexion

(3) and radial deviation (5), which serves for

diagonal left-up movement).

In either case the users performed each gesture

during 2-3 seconds with a 2-3 seconds relaxing

pause between gestures.

2.4 Signal Analysis and Neural

Network

We divide in real-time the sEMG data flow, x(t),

into 100 ms time windows (

(



)

∈ℝ



). Windowing

is performed every 50 ms. At this rate an artificial

neural network performs calculations and provides

the cursor controlling signal (Fig. 3).

At the first step the root mean square (RMS) of

the EMG activity over 100 ms time window is

evaluated:



(



)







(

−

)







(2)

where N = 0.1F is the number of samples in time

window. The RMS data, as a composite feature of

the current hand gesture, are fed into an ANN with

one hidden layer containing eight neurons (Fig. 3,

but see also Sect. 3.2). Each network neuron applies

weighted sum over its inputs and uses sigmoidal

activation function to generate the output, y:

=

(

∙

)

,()=







(3)

where∈ℝ



is the vector of synaptic weights

related to the given neuron and dot stands for inner

product. The learning, i.e., adjustment of the

neuronal weights {w}, is achieved by the standard

back-propagation algorithm (Rumelhart et al., 1985).

For training and testing purposes we use sets

containing 40-60 samples corresponding to each

class. The classification error is calculated both for

the training and for testing sets. It serves as a

criterion for stopping the learning procedure as soon

as the error starts increasing on test samples. In

average the learning process on an optimized ANN

requires about 5000 training epochs and takes less

than 1 min on a standard Intel Core i5 PC. In case of

examining different parameters of the ANN (e.g. the

number of neurons in the hidden layer) the learning

time changes accordingly (could be higher or lower).

For the first learning procedure, (i), each gesture

corresponds to a single target class (Fig. 2). This

facilitates the learning procedure since each output

neuron should produce binary output: 1 for its own

A Human-Computer Interface based on Electromyography Command-Proportional Control

Figure 3: Data flux in the MyoCursor system. Raw sEMG

activity is mapped into cursor movements and mouse

clicks in Windows OS. First RMS and MAV activity is

evaluated in a 100 ms time window. The RMS pattern is

fed into the input layer of an artificial neural network with

one hidden layer. Every 50 ms the network output from

seven neurons provides two commands for mouse-like

clicking and four commands for cursor movements. The

latter are multiplied by the MAV to gain the cursor speed.

class and 0 for the others. To accommodate

compound gestures added in the second learning

procedure, (ii), we used the target value 1/

√

for

the two output neurons participating in the

corresponding compound gesture. Such choice of the

target value ensures the generation of a compound

vector output with unitary length when both neurons

are activated (i.e., adding two orthogonal vectors of

this length results in a unitary length vector).

2.5 Proportional Control of Cursor

Once the learning is deemed finished, online

controlling of the Windows interface can be enabled.

The cursor movement along the X-axis (Y-axis) is

proportional to the difference of the output neurons

responsible for the gestures "left" and "right" ("up"

and "down", Fig. 3). This difference is a step-like

function, which is not optimal for the cursor

manipulation. To introduce a proportional control

we employ an approach similar to that described by

Lobov et al. (2015b).

We estimate the muscle effort by evaluating the

mean absolute value (MAV) averaged over all EMG

sensors:



(



)









(

−









(4)

where K is the number of sEMG channels (in our

case K = 8). Then the cursor speed can be set

proportional to the MAV (Fig. 3). However, due to

some intrinsic jitter in the muscle tone we usually

have observed a slow involuntary cursor drift. To

eliminate this artifact, the trend defined by relaxed

hand state is subtracted from the cursor controlling

signals. Thus, we define the cursor velocity by:

()=(



()−





)(



()−





)

(5)

where H is the Heaviside step function and



ℎ

is the

drift threshold, corresponding to A(t) evaluated over

time intervals with hand at rest.

Finally, the cursor displacement, , along the X-

and Y-axes is given by:







(





−



)

,





(





−



)

, (6)

where



, 



, 



, and 



are the network output (Fig.

3) corresponding to the gestures “move right”,

“move left”, “move up”, and “move down”,

respectively.

3 RESULTS

To perform the testing task described in Methods a

user should be able to move the PC cursor on the

screen and to simulate clicks of mouse buttons. We

then implemented the hardware-software setup that

replaced a physical computer mouse by a virtual

pointer controlled by the human-computer interface

based on sEMG signals (Fig. 1).

3.1 Mouse Clicks

We associated the left and right clicks of “mouse

buttons” to two single hand gestures (see Methods).

The gesture selection was not trivial. Indeed, the

gestures assigned to the clicks must differ

significantly from gestures for cursor movements.

Otherwise, the ANN may confuse them, which can

significantly diminish the user experience (usually

clicks should be done at precise cursor positions).

At the beginning we performed experiments

using nine gestures, including besides those

described in Methods the supination and pronation.

Then the latter two were refused since supination

was badly recognized due to the electrode

localization (around forearm), while pronation was

confused sometimes with the forearm flexion.

To optimize the ANN performance we ran the

process of gesture recognition on the same set of

data varying the number of neurons in the hidden

layer of the ANN and the learning rate. Figure 4

shows the results. The ANN error drops significantly

between 4 and 8 neurons and then stays unchanged,

while the learning time increases (Fig. 4A). Thus,

we selected a network with eight hidden neurons for

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics

further experiments. We also observe that the

learning rate 0.01 optimizes both the learning error

and the learning time. Thus, this value has been used

in all experimental tests of the interface.

Figure 4: Performance of the artificial neural network

(mean squared error and learning time) at classifying hand

gestures with different number of neurons in the hidden

layer (A) and different values of the learning rate (B). In

case (A) the learning rate was set to 0.01. Error bars show

the standard error.

3.2 Cursor Movement: Naïve

Approach

A naïve approach to control the cursor movements

can be implemented by the event coding similar to

that used for the mouse clicks described above. We

can set the speed of the cursor movement to a

constant. Then the user will use gestures to start and

stop the movement. The drawback of such a strategy

resides in the inevitable trade off between the speed

(responsiveness of the interface) and the accuracy of

the cursor movement. Our experiments have shown

that this strategy significantly downgrades the user

experience. Nevertheless, we used these data to

Table 1: Gesture classification error and time of execution

of the user task (see Methods). User’s body types: asthenic

(a), normosthenic (n), and hypersthenic (h).

Subject number;

sex; body type;

age (years)

Classification

error, %

Task execution

time, s

rigid

classes

fuzzy

classes

rigid

classes

fuzzy

classes

1 male n 28 0.4 6.6 46 56

2 female n 28 2.1 7.3 83 110

3 male n 41 3.8 4.8 44 71

4 female n 40 0.8 6.7 64 300

5 male n 28 1.1 8.5 55 235

6 female n 21 1.2 8 76 207

7 male n 35 0.7 4.8 47 60

8 female h 21 7.8 12 169 257

9 female a 26 2.2 5.8 103 109

10 female n 28 1.4 3.1 113 179

11 male n 21 5.6 9.7 113 113

12 female h 23 4.6 5.8 120 100

mean ± s.e. 2.6±0.7 6.9 ± 0.7 86 ± 11 150 ± 23

evaluate the classification accuracy that can be

achieved in real tasks (Table 1, column “rigid

classes”). Our results confirmed that the mean ANN

error (2.6± 0.7%) is low enough for implementing

the sEMG interface.

3.3 Proportional Control of Cursor

As abovementioned, to achieve a flexible cursor

movement we aim at a combined command-

proportional control with two degrees of freedom. In

this case the cursor movement direction is defined

by gestures, while its speed is controlled by the

degree of muscle contraction (MAV), which is

almost equivalent to the palm angle. This may

significantly improve the user experience.

Experiments conducted with twelve subjects

showed that all users were able to move the cursor

and successfully simulate left and right mouse

clicks. Then we studied the performance in cases of

using rigid and fuzzy classes (see also Sect. 2.3).

3.3.1 Rigid Classes

After the network training with rigid correspondence

between hand gestures and cursor movements, all

users managed to control the cursor. Performing the

testing task (Sect. 2.1) took from 44 s to 169 s

depending on the user with the mean 86 ± 11 s

(Table 1, “rigid classes”). Nevertheless, after the test

users reported a number of repetitive difficulties: i)

Performing the task using the sEMG-interface was

much harder than using a physical mouse. For

comparison, the same test performed by using a

hardware mouse was 10 times faster in average; ii)

Diagonal cursor movements, requiring simultaneous

displacement along the X and Y axes, were

significantly more difficult than movements

involving one axis only.

3.3.2 Fuzzy Classes

An important limiting factor of the “rigid classes”

scheme resides in the intrinsic feature of the

standard neural network approach, i.e., sharp

boundaries among classes. It leads to the “winner

takes all” phenomenon and difficulties in smooth

controlling the cursor movement. The cursor usually

follows a steps-like trajectory advancing in X or Y

directions separately instead of a smooth diagonal

movement (Fig. 5, green curve). To overcome this

problem we introduced fuzzy class overlapping (see

Methods).

The implementation of fuzzy classes indeed

facilitated the diagonal movements of the cursor.

A Human-Computer Interface based on Electromyography Command-Proportional Control

However, our tests showed that only 4 out of 12

users found this way better than using the rigid

classes approach. In average the error of gesture

identification increased to 6.9 ± 0.7% (Table 1).

Moreover, the testing task execution time

significantly increased to 150 ± 23 s. Subjectively

this performance downgrade the users explained by

the need of making unnatural gestures. For example,

simultaneous wrist extension and ulnar deviation

(required for the diagonal right-down movement)

were reported as a pattern complex to perform.

Then, an increase in wrong classification of

compound gestures leaded to the cursor movement

in wrong direction. This, in turn, increased the test

time. Table 2 summarizes the subjective user

experience and comparison of both schemes.

Figure 5: Representative example of line drawing by an

experienced user. The task consists in connecting blue

circles by a cursor by following directions shown by blue

arrows. Grey, green, and red curves mark cursor traces

corresponding to the use of a physical mouse, sEMG with

ridged classes, and sEMG with fuzzy classes interfaces,

respectively.

3.3.3 Performance of Experienced User

Since the results of experimental tests were quite

unexpected, we hypothesized that the inconvenience

of working with the sEMG-interface with fuzzy

classes might be explained by the absence of the

experience of dealing with such an interface. Indeed,

all subjects were used to common mouse interface,

while working with sEMG may require some

preliminary practice. Thus, we selected one of the

users and asked him to work with the sEMG-

interface regularly during two weeks. Then we

repeated the testing task.

Figure 5 shows the drawing made by this user

employing three different interfaces: 1) Common

mouse (grey curve); 2) sEMG with rigid classes

(green curve); and 3) sEMG with fuzzy classes (red

curve). As expected, the training significantly

improved the sEMG performance. Taking the

performance of the mouse interface as 100%, we

obtained 75.1% for the "diagonal” performance (see

Methods) by using MyoCursor with rigid classes and

92.5% for MyoCursor with fuzzy classes. Thus,

training may improve significantly the user

experience and the user may approach the

performance close to the mouse interface.

Table 2: Subjective user experience with different types of

interfaces.

User comments

Preferred

method

Critical remarks

1 rigid

- Cursor drift

- Direction of movement coincides badly

with the desired direction

- Mouse clicks are difficult because of the

high threshold

2 fuzzy Clicks provoke cursor jumps

3 fuzzy

Fuzzy method is easier to move the cursor

diagonally

4 rigid –

5 rigid –

6 rigid

"Right-down" movement is confused with

plain "down"

7 rigid

- "Right-down" is badly detected

- 2nd test is done with a tired hand

8 fuzzy

- "Left" movement is confused with plain

"down" and "rest"

- If the hand is not relaxed before click, then

the cursor goes down

9 rigid "Right-up" is confused with plain "right"

10 fuzzy

- "Right-down" is confused with plain

"down"

- Clicks are complicated

11 rigid

- "Right-down" is confused with plain

"down"

- Clicks are complicated

12 rigid

- "Right-down" is confused with plain

"down"

4 DISCUSSION

In this work we have proposed a human-computer

interface based on a real-time recording and

processing of the surface electromyographic signals.

The interface allows controlling a PC with Windows

OS by natural hand gestures. The signal acquisition

has been implemented through an easy wearable

commercially available sEMG bracelet. This,

together with simplified software learning

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics

procedure, enables building low-budget and user-

friendly sEMG solutions that may also be useful for

disabled people and amputees.

The main difference in the algorithmic part of

our approach with existing methods based on the

regression techniques (Roche et al., 2014; Hahne et

al., 2014; Fougner et al., 2014) is the use of an

artificial neural network performing the gesture

classification. The ANN is trained at the beginning

by a relatively small set of simple hand gestures:

seven or eleven gestures depending on the method

type. This allows avoiding long lasting tuning

process common for the regression approaches,

which stems from gradual sampling of changes of

muscle tension in different movements and their

combinations. Once the ANN has been trained, it

can detect commands for simulating the right and

left mouse clicks, and for moving cursor on the PC

screen. Using an estimate of the mean muscle effort

we have implemented a proportional control of the

cursor movement. Thus, the user can easily change

the cursor velocity and hence the movement

precision by “applying” more or less effort to the

gesture.

We have tested the method on twelve healthy

subjects of either sex. To do it we implemented two

types of the cursor controlling strategies: “rigid”

classes with four individual gestures for moving

right, left, up, and down; and "fuzzy" classes with

additional compound gestures for diagonal

movements. In both cases all subjects were able to

control cursor. Our experience suggests that the

fuzzy approach is potential preferable (see Fig. 5).

However, the experimental results have shown that

in average the controlling performance decreases for

this approach, despite a theoretically attractive

possibility to move the cursor diagonally.

The subjective evaluation of the user experience

has suggested that, on the one hand, the performance

reduction can be linked with the requirement to

perform unnatural gestures (for example

simultaneous wrist extension and ulnar deviation).

On the other hand, we can anticipate that in the

fuzzy case there may exist a competition in the

output layer of the ANN, which may have negative

influence on the cursor controlling function. Thus,

we can alert the reader on the necessity of future

research involving optimization of the set of gestures

and the ANN architecture.

In the present study specific features of the users

(e.g. the degree of fitness) have been left out due to

small size of the data set. However, the collected

data allow us foreseeing that the type of constitution

may play an important role in the success of the

human-computer interface. For example, Table 1

suggests that hypersthenics may show worst results,

though statistically significant conclusions require

additional experiments.

Another point for discussion is the user readiness

to a specific control of a PC by gestures. It is worth

noting that all subjects had no previous experience

in the use of such type of interfaces, while all of

them used the common mouse interface in their

daily life. Therefore, fair comparison between the

mouse and gesture types of interfaces requires either

special sampling over subjects (for example the use

of elderly, with no experience with PC) or training

subjects to use the MyoCursor system before testing.

An experiment with one user has shown that the user

training may improve significantly the ability to use

the fuzzy sEMG-interface in such a way that its

performance may approach the performance of the

mouse interface (92% vs 100% performance reached

in the test).

ACKNOWLEDGEMENTS

This work was supported by the Russian Ministry of

Education and Science under the Federal Program

(unique identification number RFMEFI58114

X0011).

REFERENCES

Baspinar, U., Varol, H. S., & Senyurek, V. Y. (2013).

Performance comparison of artificial neural network

and Gaussian mixture model in classifying hand

motions by using sEMG signals. Biocybernetics and

Biomedical Engineering, 33(1), 33-45.

Bottomley, A. & Cowell, T. (1964). An artificial hand

controlled by the nerves. New Scientist, 21, 668-71.

Chowdhury, A., Ramadas, R., & Karmakar, S. (2013).

Muscle computer interface: A review. In A.

Chakrabarti & R. V. Prakash (Eds.), ICoRD’13,

Lecture Notes in Mechanical Engineering (pp. 411-

421). Springer India.

Fougner, A., Stavdahl, O., Kyberd, P. J., Losier, Y. G., &

Parker, P. A. (2012). Control of upper limb prostheses:

terminology and proportional myoelectric control – a

review. IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 20(5), 663-667.

Hahne, J. M., Biessmann, F., Jiang, N., Rehbaum, H.,

Farina, D., Meinecke, F. C., Muller, K. R., & Parra, L.

C. (2014). Linear and nonlinear regression techniques

for simultaneous and proportional myoelectric control.

IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 22(2), 269-279.

A Human-Computer Interface based on Electromyography Command-Proportional Control

Jiang, N., Vest-Nielsen, J. L., Muceli, S., & Farina, D.

(2012). EMG-based simultaneous and proportional

estimation of wrist/hand kinematics in uni-lateral

trans-radial amputees. Journal of Neuroengineering

and Rehabilitation (9).

Kiguchi, K. & Hayashi, Y. (2012). An EMG-based control

for an upper-limb power-assist exoskeleton robot.

IEEE Transactions on Systems; Man, and Cybernetics,

Part B: Cybernetics, 42(4), 1064-1071.

Kobrinskiy, A. E. (1960). Bioelectrical control of

prosthetic devices. Herald of the Academy of Sciences

USSR, 30, 58-61.

Lobov, S. A., Mironov, V. I., Kastalskiy, I. A., &

Kazantsev, V. B. (2015a). A spiking neural network in

sEMG feature extraction. Sensors, 15(11), 27894-

27904.

Lobov, S. A., Mironov, V. I., Kastalskiy, I. A., &

Kazantsev, V. B. (2015b). Combined use of

command-proportional control of external robotic

devices based on electromyography signals.

Sovremennye tehnologii v medicine [Modern

Technologies in Medicine], 7(4), 30-38.

Mironov, V. I., Lobov, S. A., Kastalskiy, I. A., &

Kazantsev, V. B. (2015). Myoelectric control system

of lower limb exoskeleton for re-training motion

deficiencies. Lecture Notes in Computer Science,

9492, 428-435.

Myo

Gesture Control Armband - Wearable Technology

by Thalmic Labs (2013). Thalmic Labs Inc. Retrieved

May 26, 2016, from www.myo.com

Peerdeman, B., Boere, D., Witteveen, H., Huis in ’t Veld,

R. H., Hermens, H., Stramigioli, S., Rietman, H.,

Veltink, P., & Misra, S. (2011). Myoelectric forearm

prostheses: State of the art from a user-centered

perspective. Journal of Rehabilitation Research and

Development, 48(6), 719-738.

Roche, A. D., Rehbaum, H., Farina, D., & Aszmann, O. C.

(2014). Prosthetic myoelectric control strategies: A

clinical perspective. Current Surgery Reports, 2(3), 1-

11.

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1985).

Learning internal representations by error propagation.

Parallel Distributed Processing (pp. 318-362). San

Diego: La Jolla Institute for Cognitive Science,

California University.

Singh, R. & Chatterji, S. (2012). Trends and challenges in

EMG based control scheme of exoskeleton robots - a

review. International Journal of Scientific &

Engineering Research, 3(8), 1-8.

Wang, B., Li, Z., Ye, W., & Xie, Q. (2012). Development

of human-machine interface for teleoperation of a

mobile manipulator. International Journal of Control,

Automation and Systems, 10(6), 1225-1231.

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics