EMG based Control of Transtibial Prosthesis

Anoosha Anis

, M. Abbas Irshad

, Syed M. Hamza

, Noman Naseer

, Hammad Nazeer

, Andrian

Department of Mechatronics Engineering, Air University, Islamabad, Pakistan

Faculty of Technology and Computer Science, Universitas Prima Indonesia, Medan, Indonesia

Keywords: Transtibial Amputation, Active Control, Gait, EMG driven, Indirect BCI

Abstract: Amputation is defined as the loss of a limb. Transtibial amputation is the amputation below the knee. The

purpose of this research is to develop an Electromyography (EMG) based control to mimic the three positions

of an ankle. The EMG signals are extracted using eight channel Myo Armband on the tibialis muscles on

eight subjects. These signals correspond to the two extreme positions of an ankle and a rest position. The

features are extracted and K-Nearest Neighbour is used as a classifier to differentiate between the extreme

positions with 98.75 % training classification accuracy. The classified signals are then used to control the

prosthesis which mimics the ankle movement. This research can be applied to rehabilitate the ankle and help

the people with lower limb amputations.

1 INTRODUCTION

A prosthesis is an artificial device that is developed

to replace the function of a lost limb. There are two

major types of limb prostheses: Upper-extremity

prostheses and Lower-extremity prostheses. Upper-

extremity prostheses include prostheses for trans-

radial amputation, trans-humeral amputation, wrist

dis-articulation, elbow dis-articulation and shoulder

dis-articulation. Whereas, lower-extremity prosthesis

includes prostheses for hip disarticulation,

transfemoral amputation, knee disarticulation,

transtibial amputation, ankle disarticulation and

partial foot amputation.

Moreover, they can be divided into Active and

Passive depending on the use of external power

(Windrich et al., 2016). The passive prosthesis does

not contain any electronic or mechanical moving part.

These prostheses are mostly used for cosmetic

purposes and provide the basic functions like

pushing, pulling and supporting (Maat et al., 2018).

An active prosthesis includes externally powered

devices. They consist of sensors in contact with the

skin, which pick up the signals from the skin and in

turn control the devices/ actuators, which in turn

controls the movement (Windrich et al., 2016).

The intuitive control can be developed using

different techniques like Surface Electromyography

(sEMG) (Anil and Sreeletha, 2019), Ultrasound

imaging(González and Castellini, 2013),

electroencephalography (EEG) (Bright et al., 2016),

Force myography (FMG) (Cho et al., 2016),

Implantable Myoelectric Sensors (Pasquina et al.,

2015) and Targeted Muscle Reinnervation

(TMR)(Cheesborough et al., 2015). Out of these

techniques sEMG, ultrasound imaging, EEG, FMG

are non-invasive techniques whereas, Implantable

Myoelectric sensor and TMR are invasive

techniques(Turnip, Soetraprawata and Kusumandari,

2013).

The Myoelectric signals are produced due to

variations in the state of muscle fibre. The variation

in electric potential in the motor neurons is detected

by the electrodes as an EMG signal. So greater the

variation/contraction will be, greater the amplitude of

the recorded voltage will be. The EMG signals can be

detected either by intrusive/intramuscular or non-

intrusive technique. The intrusive technique involves

the use of needle EMG electrodes by inserting them

into the muscle under examination. The advantage of

this technique is that it reduces the muscle noise and

thus produces more accurate results (Waris and

Kamavuako, 2018). Whereas, the non- intrusive

technique uses surface EMG electrodes which uses

the surface-based detection technique for the EMG

signal. This technique does involve more muscle

noise but is preferred over the previous method as it

does not involve any special formalities and

procedure. Moreover, the latest research and

technology has resulted in more sensitive sensors

which can capture the signals from the skin much

Anis, A., Irshad, M., Hamza, S., Naseer, N., Nazeer, H. and Andrian, .

EMG based Control of Transtibial Prosthesis.

DOI: 10.5220/0009464200740081

In Proceedings of the International Conference on Health Informatics and Medical Application Technology (ICHIMAT 2019), pages 74-81

ISBN: 978-989-758-460-2

very accurately without the need of inserting them

(Del Vecchio et al., 2017).

The aim of this research is to encounter all the

issues faced due to passive prosthesis by developing

a physical prototype of an ankle foot prosthetic active

in nature to exhibit its benefits to a transtibial

amputee. This includes development and

implementation of methodology for EMG signal

acquisition as well as classification and identification

of intuitive signal for the lower limb prosthesis

control. The prosthesis mimics the position of the

ankle on the basis of the EMG signals that are

classified using the KNN classifier.

The conducted research has various applications

in the field of Robotics, Bio-medical Industry and

Health sector. Such as, to help people with transtibial

amputations to become independent and in defence

sector to help the soldiers with amputations to return

to normal life.

2 METHODOLOGY

The methodology of this project is divided into four

major stages. The first step includes acquisition of

data from the subjects. In the second step, the

acquired data is then processed and different

statistical features are extracted and fed to the

classifier. The third step includes the training of

classifier on the extracted features and the last step

involves the testing of the classifier on the real-time

data. The methodology of this process can be seen in

the figure 1.

2.1 Data Acquisition

Myo Armband was utilized for the purpose of signal

acquisition. It consists of eight EMG electrodes and

and an in-built bluetooth for data transmission. The

use of multiple channels improves the accuracy. It

also comprises of accelerometer and a gyroscope. The

transmitted signals can be captured via Myo SDK on

MATLAB. The figure 2 shows Myo Armband sensor.

Figure 2: Myo Armband sensor.

It was worn by the subject just below the knee that

continuously read the muscle data and sent it via the

in-built Bluetooth to the laptop present right next to

the subject in form of a vector. The laptop’s Bluetooth

received the incoming data and passed it to the

MATLAB. Figure 3 and 4 shows position of sensor.

Figure 1: Experimental setup.

EMG based Control of Transtibial Prosthesis

Figure 3: Location of electrodes on muscles (frontal view.)

Figure 4: Location of electrodes on muscles (dorsal view).

The paradigm decided for this project is as follows:

 Number of classes(actions) = 2

 Number of subjects = 8

 Number of male subjects = 6

 Number of female subjects = 2

 Age group = 20-25 years

 Number of activities per subject per

trial= 10

 Total time for each trial = 30s

 All subjects are healthy

2.2 Feature Extraction and Accuracy

The Myo Armband provides filtered signals so after

receiving the data from Myo Armband, Savitzky-

Golay filter was applied to smoothen the signal

(Christov, Raikova and Angelova, 2018). Figure 5

shows un-filtered EMG signal whereas, Figure 6

shows filtered EMG signal.

Figure 5: Un-filtered EMG signal.

Figure 6: Filtered EMG signal.

After smoothing the signal, the features were

extracted. The features help reducing the input data

into less but useful data (Phinyomark, Khushaba and

Scheme, 2018). They provide useful information

about the signal; therefore, features need to be

selected carefully. Different statistical features were

extracted such as standard deviation, root mean

square, mean absolute value, zero crossings, maxima

and minima (Hong, Khan and Hong, 2018). The

classifiers Linear discriminant analysis (LDA),

Support Vector Machine (SVM) and k-nearest

neighbours (KNN) were trained separately on each of

the above feature. Maximum accuracy was achieved

using Root Mean Square as a feature on KNN.

2.3 Real-time Classification and

Application Interface

In this phase, the data is classified real-time on the

trained classifier. It is also called Online

classification. When the amplitude being calculated

exceeds the threshold of 0.4, the wave is sent for

ICHIMAT 2019 - International Conference on Health Informatics and Medical Application Technology

feature extraction and then the features are fed to the

classifier. The classifier then classifies the gesture

into respective class and then, based on the output

class, the prosthetic foot moves up or down. The

process is shown in figure 7.

Figure 7: Application interface.

3 MODELING AND SIMULATION

3.1 Gesture

There are two gestures i.e. up and down. The rest

position was distinguished on the basis of amplitude.

If the calculated amplitude was greater than 0.4 then

it meant gesture and its features were extracted and

passed on to the classifier. Otherwise, it meant rest

(Reaz, Hussain and Mohd-Yasin, 2006). The total

number of classes is two as seen in Figure 8-10.

3.2 Data Collection

The Data was collected via Myo Armband which

continuously sent the data wirelessly to the laptop at

the sampling frequency of 200 Hz. The data was

received in the form of a nx8 vector where ‘n’

depends on the duration of the activity and eight

represents number of electrodes. It was then plotted

and processed using MATLAB. The Figure 11- 14

shows the raw data obtained for rest, up and down

gesture on each electrode as well as a combination of

all the electrodes.

Figure 8: Rest position.

Figure 9: Up gesture.

Figure 10: Down gesture.

3.3 Classification

The classifier was used to differentiate between the

two classes on the basis of input features(Qureshi et

al., 2016). A classifier was needed that could satisfy

the following conditions: easy to understand, need

EMG based Control of Transtibial Prosthesis

only acceptably short calculation time and have a

more than decent predictive power(Simbolon et al.,

2016). Three classification algorithms Support

Vector Machine (SVM), Linear Discriminant

Analysis (LDA) and K-Nearest neighbour (KNN)

were trained, tested and validated.

Figure 11: Raw EMG signal for rest.

Figure 12: Raw EMG signal for up gesture.

Figure 13: Raw EMG signal for down gesture.

Figure 14: Raw signals for rest, up and down gesture.

In Support Vector Machine, each data item is

plotted as a point in space where the dimension of the

space depends on the number of features(Turnip et

al., 2016). The classification is then performed by

finding the hyper-plane that separates the classes

well. The new data is then plotted in the same space

and the category is decided on the basis of the side of

the gap where they fall(Alkan and Günay, 2012). An

offline accuracy of 88.7% was achieved using SVM.

The LDA predicts the class of the set of data by

calculating the probability for each class. The

probability is estimated using the Bayes Theorem.

The class with the highest probability is selected as

the output class (Alam and Arefin, 2018).LDA is a

general form of Fisher’s linear discriminant, which is

used in statistics and pattern recognition

problems(Naseer et al., 2016). An offline accuracy of

96.4%was achieved using LDA.

KNN is another method used for classification

which makes use of the fact that the similar things

exist in close proximity. The input data is assigned the

class of majority of its closest neighbours. It is easy

to implement and can be used for both the regression

and classification problems. The advantages of this

algorithm are that it doesn’t need to make various

assumptions and to build a separate model(Altın and

Er, 2016). The only con is that it becomes slower as

the number of examples increase(Khan et al., 2018).

An offline accuracy of 98.7% and an online accuracy

of 90% was achieved. The KNN was chosen since it

provided the highest classification accuracy and fast

calculation time. The results can be seen in the Figure

15.

4 RESULTS

When the muscle is at rest then the calculated

amplitude is less than 0.4 and no action is performed

ICHIMAT 2019 - International Conference on Health Informatics and Medical Application Technology

Figure 15: Percentage accuracy versus features on three different classifiers.

at the output. So, if the foot is up then the prosthetic

also stays up and if the foot is down then the

prosthetic remains down at rest.

When the calculated amplitude is greater than 0.4

and a change in the muscle signals is detected, then

the signal is sent for feature extraction and then to the

classifier. Depending on the classification result, a

command is generated.

The output command generated from the

MATLAB is sent to the Arduino using Arduino

hardware support package on MATLAB. On the basis

of the classified signal the MATLAB generates the

forward, backward or rest command to Arduino

which controls the actuator integrated with it and

moves the ankle up, down or keeps it in rest state.

In the Figure 16-18, one can observe the output of

prosthetic foot against each of the gesture.

Figure 16: Rest position.

Figure 17: Up gesture.

Figure 18: Down gesture.

5 CONCLUSIONS AND FUTURE

RECOMMENDATIONS

This paper presents an improved technique to control

a transtibial prosthesis using EMG signals and

machine learning algorithms.

EMG based Control of Transtibial Prosthesis

The EMG signals were obtained using Myo

Armband. Features such as root mean square were

extracted and fed to the classifier and the classifiers

were then trained, tested and validated on different

features. It was found that using root mean square as

a feature and KNN as a classifier gave maximum

accuracy. An offline accuracy of 98.75% and an

online accuracy of 90% was achieved(Anil and

Sreeletha, 2019).

However, the prosthesis can be made more

natural-like by mimicking the human gait that can be

done by increasing the number of output classes on

the basis of the angle of the ankle joint.

Moreover, the classifier performance can be

improved by increasing the number of training data

set. The combination of different features can also be

implemented to increase the accuracy. The artificial

neural networks or other deep learning techniques

may be applied to improve the accuracy.

Although, the prime objective of controlling the

prosthetic was achieved, closed loop control must be

introduced for precise and robust control of the joint.

More work needs to be done to make the

prosthesis portable by the introduction of single board

computer like Beaglebone, Raspberry Pi and UDOO.

REFERENCES

Alam, M. S. and Arefin, A. S. (2018) ‘Real-Time

Classification of Multi-Channel Forearm EMG to

Recognize Hand Movements using Effective Feature

Combination and LDA Classifier’, Bangladesh Journal

of Medical Physics, 10(1), pp. 25–39. doi:

10.3329/bjmp.v10i1.39148.

Alkan, A. and Günay, M. (2012) ‘Identification of EMG

signals using discriminant analysis and SVM

classifier’, Expert Systems with Applications. Elsevier

Ltd, 39(1), pp. 44–47. doi:

10.1016/j.eswa.2011.06.043.

Altın, C. and Er, O. (2016) ‘Comparison of Different Time

and Frequency Domain Feature Extraction Methods on

Elbow Gesture’s EMG’, European Journal of

Interdisciplinary Studies, 5(1), p. 35. doi:

10.26417/ejis.v5i1.p35-44.

Anil, N. and Sreeletha, S. H. (2019) ‘EMG Based Gesture

Recognition Using Machine Learning’, Proceedings of

the 2nd International Conference on Intelligent

Computing and Control Systems, ICICCS 2018. IEEE,

(Iciccs), pp. 1560–1564. doi:

10.1109/ICCONS.2018.8662987.

Bright, D. et al. (2016) ‘EEG-based brain controlled

prosthetic arm’, Conference on Advances in Signal

Processing, CASP 2016, pp. 479–483. doi:

10.1109/CASP.2016.7746219.

Cheesborough, J. E. et al. (2015) ‘Targeted muscle

reinnervation and advanced prosthetic arms’, Seminars

in Plastic Surgery, 29(1), pp. 62–72. doi: 10.1055/s-

0035-1544166.

Cho, E. et al. (2016) ‘Force myography to control robotic

upper extremity prostheses: A feasibility study’,

Frontiers in Bioengineering and Biotechnology,

4(MAR), pp. 1–12. doi: 10.3389/fbioe.2016.00018.

Christov, I., Raikova, R. and Angelova, S. (2018)

‘Separation of electrocardiographic from

electromyographic signals using dynamic filtration’,

Medical Engineering and Physics. Elsevier Ltd, 57, pp.

1–10. doi: 10.1016/j.medengphy.2018.04.007.

González, D. S. and Castellini, C. (2013) ‘A realistic

implementation of ultrasound imaging as a human-

machine interface for upper-limb amputees’, Frontiers

in Neurorobotics, 7(OCT), pp. 1–11. doi:

10.3389/fnbot.2013.00017.

Hong, K. S., Khan, M. J. and Hong, M. J. (2018) ‘Feature

Extraction and Classification Methods for Hybrid

fNIRS-EEG Brain-Computer Interfaces’, Frontiers in

Human Neuroscience, 12(June), pp. 1–25. doi:

10.3389/fnhum.2018.00246.

Khan, R. A. et al. (2018) ‘FNIRS-based Neurorobotic

Interface for gait rehabilitation’, Journal of

NeuroEngineering and Rehabilitation. Journal of

NeuroEngineering and Rehabilitation, 15(1), pp. 1–17.

doi: 10.1186/s12984-018-0346-2.

Maat, B. et al. (2018) ‘Passive prosthetic hands and tools:

A literature review’, Prosthetics and Orthotics

International, 42(1), pp. 66–74. doi:

10.1177/0309364617691622.

Naseer, N. et al. (2016) ‘Determining optimal feature-

combination for LDA classification of functional near-

infrared spectroscopy signals in brain-computer

interface application’, Frontiers in Human

Neuroscience, 10(MAY2016), pp. 1–10. doi:

10.3389/fnhum.2016.00237.

Pasquina, P. F. et al. (2015) ‘First-in-man demonstration of

a fully implanted myoelectric sensors system to control

an advanced electromechanical prosthetic hand’,

Journal of Neuroscience Methods. Elsevier B.V., 244,

pp. 85–93. doi: 10.1016/j.jneumeth.2014.07.016.

Phinyomark, A., Khushaba, R. N. and Scheme, E. (2018)

‘Feature extraction and selection for myoelectric

control based on wearable EMG sensors’, Sensors

(Switzerland), 18(5), pp. 1–17. doi:

10.3390/s18051615.

Qureshi, N. K. et al. (2016) ‘Comparison of classification

performance for fNIRS-BCI system’, 2016 2nd

International Conference on Robotics and Artificial

Intelligence, ICRAI 2016, pp. 54–57. doi:

10.1109/ICRAI.2016.7791228.

Reaz, M. B. I., Hussain, M. S. and Mohd-Yasin, F. (2006)

‘Techniques of EMG signal analysis: Detection,

processing, classification and applications’, Biological

Procedures Online, 8(1), pp. 11–35. doi:

10.1251/bpo115.

Simbolon, A. I. et al. (2016) ‘An experiment of lie detection

based EEG-P300 classified by SVM algorithm’,

Proceedings of the 2015 International Conference on

Automation, Cognitive Science, Optics, Micro Electro-

ICHIMAT 2019 - International Conference on Health Informatics and Medical Application Technology

Mechanical System, and Information Technology,

ICACOMIT 2015, pp. 68–71. doi:

10.1109/ICACOMIT.2015.7440177.

Turnip, A. et al. (2016) ‘EEG-based brain-controlled

wheelchair with four different stimuli frequencies’,

Internetworking Indonesia Journal, 8(1), pp. 65–69.

Turnip, A., Soetraprawata, D. and Kusumandari, D. E.

(2013) ‘A Comparison of EEG Processing Methods to

Improve the Performance of BCI’, International

Journal of Signal Processing Systems, 1(1), pp. 63–67.

doi: 10.12720/ijsps.1.1.63-67.

Del Vecchio, A. et al. (2017) ‘Associations between motor

unit action potential parameters and surface EMG

features’, Journal of Applied Physiology, 123(4), pp.

835–843. doi: 10.1152/japplphysiol.00482.2017.

Waris, A. and Kamavuako, E. N. (2018) ‘Effect of

threshold values on the combination of EMG time

domain features: Surface versus intramuscular EMG’,

Biomedical Signal Processing and Control. Elsevier

Ltd, 45, pp. 267–273. doi: 10.1016/j.bspc.2018.05.036.

Windrich, M. et al. (2016) ‘Active lower limb prosthetics:

A systematic review of design issues and solutions’,

BioMedical Engineering Online. BioMed Central,

15(3), pp. 5–19. doi: 10.1186/s12938-016-0284-9.

EMG based Control of Transtibial Prosthesis