This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

Portable EMG Data Acquisition Module for Upper Limb

Prosthesis Application
Sidharth Pancholi and Amit M. Joshi, Member, IEEE

Abstract— Electromyography (EMG) signals are gaining four channel EMG acquisition system was developed based on
popularity to develop the prosthetics. In the paper, an efficient instrumentation amplifier (INA326) and operational amplifiers
multi-channel EMG signal acquisition system has been proposed for initial signal smoothening. However each channel requires
for upper limb prosthetic application. Various arm exercises
have been performed to obtain EMG signals from 5 different arm an individual INA326 and operational amplifiers (AD8603)
muscles for the validation of developed hardware. The muscle’s for conditioning which makes the system cumbersome and
position has been selected by palpation method. Furthermore, the complex. The system based on instrumentation amplifiers
classification algorithms have been examined for seven different require multiplexer to combine the data from multiple
activities. Total 29 subjects have been chosen (25 intact and 4 channels into single channel for wireless transmission, which
Amputees) to acquire the EMG data by these activities. To makes system expensive. Active electrode based system also
classify the recorded EMG dataset, 9 TDM (time domain) and 7
FDM (frequency domain) features have been extracted. A proposed but each channel requires separate cable and
comparative analysis of different classifiers is presented for interfaces this make system expensive. The power
different muscle position of electrodes. The signal processing and consumption of this kind of acquisition system is also very
classification algorithms have been processed in MATLAB 2016a. high [11]. System based on ADS1294 IC has utilized up to
The accuracy of classification ranges for different classification four channels EMG signal acquisition in the paper [12]. Since
algorithms from 57.69 % to 99.92 % for all subjects. it has only 4 differential inputs, therefore it can acquire less
muscle activities at a time. A small EEG/EMG recorder was
Index Terms—Acquisition; Classification; LDA; k-NN;
Prosthetic; sEMG; SVM; QDA. developed in the paper [13]. It required separate ADC which
has only 12 bit resolution, and fixed sampling rate. A FPGA
I. INTRODUCTION based hybrid EMG signal acquisition system was also
proposed for effective EMG measurement and separation
Electromyography (EMG) signals are becoming one of the
processing. The main drawback of the system was delay time
vital biological parameters which have wide range of
[14, 15]. In this paper, a multichannel system is being
applications in biomedical engineering, prosthetic device
proposed for prosthetic device, which is wireless, cost
development, human machine interaction (HMI) and
effective and compact in design. The prototyping cost of this
rehabilitation devices [1]. The electromyogram (EMG) is a
system is approx. INR 5000 (78 USD) which includes; analog
bio-potential signal which is acquired by electrodes through a
front end, DSP processor, passive components, Electrodes or
muscle fiber skin to observe the muscle activity. It is also
leads, PCB designing and wireless module.
associated with neural signal, sent from spinal cord to muscles
The pattern recognition for EMG consists of two parts,
[2, 3].The voltage range of EMG signal is 50 uV to 100 mV
features (attributes) extraction and classification (machine
and the frequency is varied from 10 Hz to 500 Hz for surface
learning). In the feature extraction process, the dimensionality
Electromyography(sEMG) [4].The sEMG signal is acquired
of the sEMG dataset is reduced to form a feature vector. This
by applying noninvasive electrodes.
is helpful to retrieve the useful information and also useful to
Many people lost their limbs due to accidents; some
remove an unwanted data. The accuracy and classification
people don’t have their limbs by birth. These people cannot
time in pattern recognition depend upon these features [15].
perform their daily routine work such as holding or moving an
The features are extracted through the segmentation of sEMG
object, eating etc. One solution for these amputees is EMG
signal rather than extracting through individual sample in
based artificial limb or Prosthetic device [5-8]. EMG based
order to keep structural information.
prosthetic device uses EMG signal to control the action of
There are evident gaps between devices and control
artificial limb with the help of motor [9]. This device includes
approaches of prosthetic limb. The large data set is required
data acquisition module, machine learning algorithm for EMG
for effective pattern classification and if the number of classes
pattern recognition and mechanical structure for controlling.
increases then huge amount of training data is needed to
Although many signal processing software and high level
improve classification results. In this study, a visual feedback
EMG signal acquisition systems are available in the market
mechanism is used for the training of amputees to improve
they are expensive (e.g. NeXus-10, BIONOMADX). There are
cognitive abilities of brain. The perception of the missing limb
systems developed, based on instrumentation amplifiers
is still attached to the body, is called phantom limb. Almost
(INA2141 or INA128), are difficult in design and provide time
60% to 80% amputees feel this sensation of phantom limb.
delay in signal processing for multichannel acquisition [10]. A

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

Sometimes, this sensation is painful. Muscles activation channel can be analyzed. In the proposed system, the
signals usually present in the residual part of the arm for the important parts of acquisition module (differential amplifier,
amputated limb. Brain always looking for the signals from the filtering, and analog to digital converter) are in a single chip
amputated part, However visual feedback system is introduced (ADS1298). For the overvoltage and surge protection each
for reducing the many chronic neurological disorders related channel of the system has been featured with the protection
to the phantom limb pain [16-19]. This visual feedback system circuit, which is shown in Fig.2. In this protection circuit
can be utilized in improvement of pattern classification of the schottky diodes and a second order passive low pass filter
missing limb for prosthetic application [20-22]. have been used. Schottky diodes have high switching speed
Proposed device is validated for arm activity recognition, and the low pass filter has cut-off frequency is 1 KHz.
where 25 healthy subjects (15 Male and 10 Female) and 4
amputees have been considered for EMG data acquirement. In
the subsequent stage, machine learning strategies have been
used for recognition of different arm exercises. Five muscles
have been chosen for acquiring EMG data and 16 features
(time and frequency domain) have been calculated. Further,
EMG data have been classified for different arm activities
including all feature sets as well as an individual feature set of
each channel.
In the proposed work, a bio-potential acquisition module
and a java based GUI is developed to display the real time
data. The hardware module is wireless, portable, power
efficient and has eight channels. Section II presents the system
architecture for the proposed system. Section III emphasis on
proposed data acquisition model. Section IV describes Fig. 2. over voltage protection circuit
overlapping windowing technique Section V covers feature There are several benefits of such types of bio-potential
extraction process, while section VI explains the EMG pattern acquisition systems [18]:
classification. Results are discussed in Section VII and a) The system has extremely low power consumption in
conclusion is derived in Section VIII. order of 0.75mW/channel.
b) The selection of sampling frequency may be done as
II. SYSTEM ARCHITECTURE per necessity.
The acquisition module of EMG signal comprises of c) A safety feature is also provided for Lead off
different parts such as electrodes, differential Amplification, Detection.
antialiasing filter, conversion of analog signal into digital d) Programmable gain provides the flexibility to select
signal and display system. Electrodes or leads act as a channel appropriate gain.
between muscle tissue skin and inputs of amplifier [17]. e) System is smaller in size that can be used as a
The hardware module is based on ADS1298 IC (Texas wearable device in future.
Instruments) and ARM cortex M4 series processor, shown in For the wireless communication Bluetooth module has been
Fig. 1. used of 2.4 GHz frequency ISM Band. This module provides
optimal balance between power demand and transmission in
the short range communication and use Gaussian frequency
shift keying modulation. The operating temperature of this
module is -20o C to +74o C. Further, a java based GUI is
developed which shows real time data. The signals are
acquired at 4000 samples/sec on Intel core i3 computer with
2.54 GHz and 3.80 RAM.

III. PROPOSED DATA ACQUISITION PROTOCOL

The impedance of skin should be less in order to acquire the
meaningful information from the EMG signal. Therefore, the
skin of subject should be cleaned before acquisition of EMG
signal. Dead skin cells and hairs must be removed from the
Fig. 1. Block diagram representation of proposed system
skin surface. For this, abrasive gel should be used to remove
The ARM cortex M4 processor is widely used in DSP the dead skin [23, 24]. Also moisture content can weaken the
applications; it is cost effective and has low power signal, so there should not be any sweat drops on the skin.
consumption. GUI displays real time EMG signal and also Main procedure that has been carried out for acquisition of
provides the data in requisite format so that the signal of every signal from surface is described below:

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

TABLE I positions of electrode placement are shown in Table

MUSCLE POSITIONS FOR ACQUISITION OF EMG SIGNAL II.
The Intact subjects were asked to perform each experiment,
S. Age Gender Height Weight Reason for with their right hand in comfortable position. While amputees
No. Amputation asked to execute movements as naturally as possible with both
the limbs missing and intact, by looking in the mirror and
1. 32 Male 172 cm 79 Kg Accident
normally without any mirror. Mirror was used to create
2. 44 Male 168 cm 88 Kg Cancer reflection of the undamaged limb, it makes a perception in the
3. 52 Male 177 cm 77 Kg Accident subject mind that both arm is doing same exercise. This create
4. 38 Male 171 cm 90 Kg Accident visual feedback to the human brain, this is shown in Fig.5.
Five muscles have been chosen for the study using palpation
method [25].
TABLE II
MUSCLE POSITIONS FOR ACQUISITION OF EMG SIGNAL
S.No. Channels Muscles Name
1. Channel 1 Extensor Carpi Ulnaris Muscle
2. Channel 2 Extensor Digitorum Communis muscle
3. Channel 3 Extensor Carpi Radialis Longus muscle
4. Channel 4 Flexor Farpi Radialis Muscle
5. Channel 5 Biceps Brachii Muscle
6. Ref. Electrode Elbow

Fig. 4. Performed Activities (i) Hand Open (ii) Hand Close (iii) Wrist Flexion
Total 25 healthy subjects (10 females and 15 males) and 4 (iv) Wrist Extension (v) Soft Gripping (vi) Medium Gripping (vii) Hard
Gripping
Amputees have been enlisted for acquisition Remaining arm
percentage is calculated as the ratio between the length of
amputated arm and the length of contralateral forearm from

Fig. 3. Amputee 1: Wrist disarticulation Fig. 5. Data acquisition setup for the amputees

the elbow. All the amputees are wrist disarticulated; Fig. 3 IV. EMG DATA SEGMENTATION
shows the picture of amputated hand of subject 1.
a) Subsequently, subjects were asked to sit at a desk on There is a need for data segmentation when data is
laboratory chair. Seven upper-limb activities processed for the feature extraction; in these segmentations
(exercises) have been performed five to six times for data is considered as quasi- stationary. Two main parts should
normal subjects and four time for amputees: i. Hand be considered properly in order to design appropriate data
open (HO), ii. Hand closed (HC), iii. Wrist extension segmentation strategy; namely the window length
(WE), iv. Wrist flexion (WF), v. Soft gripping (SG), (segmentation length) and the amount of overlapping between
vi. Medium gripping (MG), and vii. Hard gripping two consecutives segments. The segmentation window length
(HG), shown in Fig 4. has trade-off between classifier accuracy and classification
b) The training has been provided to overcome response time. As the segmentation length increases the
cognitive effort prior to acquisition. classification performance improves, on the other hand
c) A band pass filter of frequency range of 20 Hz to 500 classification length must be small enough to satisfy the hard
Hz and an amplification gain 12x have been applied. real time performance of the myoelectric control. Many
Sampling frequency is fixed at 4000 Hz. experiments have been carried out by researchers with
d) The EMG dataset of the right arm exercise has been different segment length. In the literature, a delay of 150~250
prepared through five muscle positions, using bipolar ms interval is considered for EMG activity segmentation [26].
Ag-AgCl (Silver – silver chloride) electrodes 2 mm In the presented work, window size has been taken of 250 ms
diameter spaced 12 to 36 mm apart. Muscles (1000 samples) with window shift of 25 ms, shown in Fig. 6.

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

quadratic discriminant analysis (QDA), Support Vector

Machine (SVM), Random Tree (RT) and Random Forest
(RF). In the k-NN classification predication of the test data is
evaluated by measuring distance from the training samples. In
this work, to determine the class of the testing data Euclidean
distance is calculate and k=10 has been considered.
Classification through LDA shows more robustness while not
being trained iteratively compared to SVM, KNN, RT and RF.
In the LDA classification technique discrimination function is
used to determine the class and each class is considered an
identical.
Fig. 6 Feature extraction process by applying overlapping window technique TABLE III
EXTRACTED FEATURES FOR THIS STUDY
V. FEATURE EXTRACTION
Various noise and artifacts are likely to be merged along S.No. Time Domain S.No. Frequency Domain
with EMG signals at the time of acquisition, thus the required Features (TDM) Features (FDM)
information remains mixed inside the raw EMG signal. The 1. Integrated EMG 1. Mean Frequency
use of these types of signals directly into the classification (IEMG) (MNF)
process decreases the classifier’s efficiency, so there is a need 2. Mean Absolute Value 2. Median Frequency
to extract useful information or features from raw EMG (MAV) (MDF)
signal. The set of features has to be applied by naming or 3. Modified mean absolute 3. Peak frequency Power (PKF)
value type 1 (MMAV 1)
labeling for supervised machine learning. This is called feature 4. Modified mean absolute 4. Mean frequency Power
vector preparation, which is applied into the classification value type 2 (MMAV 2) (MFP)
process. A useful feature vector must contain the valuable 5. Wilson Amplitude 5. Frequency Median
information and remove the irrelevant information (noise) [27, (WAMP) (FMD)
6. Root Mean Square 6. Frequency Mean
28]. Several EMG features have been proposed by the (RMS) (FMN)
scientists to increase the accuracy of classification. In this 7. Waveform Length 7. Modified Frequency Median
paper two feature group, time and frequency domain, have (WL) (MFMD)
been examined, shown in Table III [29, 30]. 8. Zero Crossing
There are some unwanted elements in feature matrix, (ZC)
9. Simple Square Integral
which are not required in the processing; these parameters or (SSI)
attributes can be removed by feature selection techniques.
Dimensionality reduction of feature matrix is a method to The classification algorithm of QDA can be realized as a
reduce the dimension of an original feature vector, while more common version of LDA, which splits the classes by a
maintaining the information of discrimination and eliminating quadratic plane instead of linear plane, as executed in LDA
the remaining useless data. This reduces the time of [24-27]. In the QDA covariance matrix is calculated for each
computation for data classification [31-34]. In the proposed class separately. Both LDA and QDA are interchangeable and
work, the feature selection is carried out by manual channel perform well when the feature size is large. Random forest is a
selection method. In this work, total 5 channels data has been Decision tree based classifier which is composed of a number
considered for the classification and subsequently the of decision tree classifier. Random forest classifier shows
classification accuracy of each channel has been calculated. better result of classification, however it requires larger time
than LDA and QDA [37]. SVM classification consists of a
VI. EMG PATTERN CLASSIFICATION hyper surface in the space that is used for classification of the
In the feature extraction process EMG signal represents as a data set. The performance of classifier would be higher when
Feature vector, which is fetched as an input to the classifier. the margin of hyper surface from nearest training points of all
As EMG signals is not directly applied to the classifier, due to classes is greater. It is initially design to classify binary classes
the randomness of the EMG signal and also larger dimension. but many approaches introduced to determine multiclass
The Electromyogram classification process mainly depends on problems also [38]. In this work, RBF (Radial Basis Function)
the quality of the extracted features. Feature extraction step in has been used as SVM kernel. Random tree classifier is sort of
the classification, enhances the concentration of information decision tree classification technique, which takes less time
of the signal [35, 36]. In the presented work, different than random forest. Classifier data set has been divided into
classification techniques have been used to understand which training data set (50% of all data), testing data set (30% of all
classifier has better performance for the activity classification. data) and validation data set (20% of all data). For the
Six different types of machine learning algorithms have been classification of amputee’s data, training Data used which has
chosen for comparative analysis of classification: k-nearest obtained by visual feedback and testing data has taken
neighbors (k-NN), linear discriminant analysis (LDA), normally (without visual feedback).

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

Fig. 8. All 5 channels wave form representation for Wrist Extension (WE)
Fig. 7. All 5 channels wave form representation for Hand Open (HO) activity: A
activity: A (channel 1), B (channel 2), C (channel 3), D (channel 4), E
(channel 1), B (channel 2), C (channel 3), D (channel 4), E (channel 5
(channel 5)

Fig. 9. All 5 channels wave form representation for Hard Gripping (HG) activity: Fig. 10. Channel 5 (Biceps Brachii muscle) EMG signal for HO, WE and
A (channel 1), B (channel 2), C (channel 3), D (channel 4), E (channel 5) HG activities

VII. RESULT AND DISCUSSION

There are different external and internal causes of low to 500 Hz range. Furthermore, three different activities
frequency noise that can reduce the quality of EMG signal waveform: activity HO, activity WE, activity HG, have been
during acquisition. The two external sources of noise, EMG presented for all 5 channels for analysis purpose, Fig.7 to Fig.
leads motion artifact and power line interference. These two 9 respectively. EMG signal is continuously generated in
noises can be almost removed with suitable circuit design. channel 1 (Extensor Carpi Ulnar muscle) because this muscle
The two external noise sources generate in the electronics is always stretched in rest condition for activity WE and
circuitry of the amplification system, the skin-electrode activity HG. By Performing activity HG, EMG signal of larger
interface (electrochemical noise) and thermal noise artifact. amplitude has been produced in the channel 2 (Extensor
Combination of these two noises produce the baseline noise Digitorum Communis muscle). The rest condition of activity
which is sensed whenever a sensor is appended to the skin. WE and WF is same and produce continuous EMG signal in
The movement artifact noise is generated at Electrode-skin channel 4 (Flexor Farpi Radialis Muscle). However, the
interface. This is due to: (a) the muscle movement occurs channel 1 produces larger amplitude for these two activities.
below the skin [39], and (b) when a force impulse moves via EMG signal is generated on the channel 5 (Biceps Brachii
muscle and skin underlying the sensor causing a movement at muscle) by performing gripping activities, shown in Fig. 10.
the electrode skin interface [40, 41]. The consequential time- The Amplitude of EMG signal varies with respect to intensity
varying voltage is formed across the two electrodes, which is of gripping. Harder the gripping, greater would be the
the main concern of noise sources and needs attention. In our amplitude of EMG signal. In the next step for classification,
preliminary work, it has been analyzed that Butterworth band total 80 features (5 channels), are considered for all 29 subject
pass filter of 6th order has fast settling time and it exhibits less data. Where, the Random Forest, k-NN and Random tree
overshoot. So this filter is applied for the removing high classifiers have shown better results.
frequency noise, base line error and motion artifacts of 20 Hz

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

Fig. 11. Classification accuracy Comparison for different classification algorithm Fig. 12. Classification accuracy Comparison for different classification algorithm
including all performed activities: All (All channels), ch 1 (channel 1), ch 2 including HO, HC, WE and WF activities: All (All channels), ch 1 (channel 1),
(channel 2), ch 3 (channel 3), ch 4 (channel 4), ch 5 (channel 5) ch 2 (channel 2), ch 3 (channel 3), ch 4 (channel 4), ch 5 (channel 5)

Fig. 13. Classification accuracy Comparison for different classification algorithm Fig. 14. Classification accuracy Comparison for different classification algorithm
including SG, MG and HG activities: All (All channels), ch 1 (channel 1), ch 2 including all activities for prosthetic: All (All channels), ch 1 (channel 1), ch 2
(channel 2), ch 3 (channel 3), ch 4 (channel 4), ch 5 (channel 5) (channel 2), ch 3 (channel 3), ch 4 (channel 4), ch 5 (channel 5)

The classification results through LDA and QDA are also training process [36-38]. LDA classifiers increase its
satisfactory. Further classification accuracy has been performance when the huge number of features is used [39].
calculated by each channel (16 features each channel). The LDA with 5 channels has shown better classification accuracy
channel 5 has better performance as compare to other when activities are quite different. For all Activities
channels, shown in Fig 11. The total arm activities have been classification, the accuracy ranges corresponds all channels
divided into two categories. Activities HO, HC, WF, WE are 99.54% maximum for random forest and minimum 75.38%
included in first category, rest of three gripping activities (SG, for LDA classifier. In case of individual channels, channel 2
MG, HG) are included in second category. Same procedure is and channel 5 shows better classification performance.
followed on these two categories separately as explained Channel 2 and channel 5 both shows maximum performance
above. The classification performance for both the category is in case of random forest classifier which is 92.07% and
shown in Fig. 12 and Fig. 13 respectively. The classification 91.56% respectively. However channel 5 has minimum of
accuracy is increased for each category. The contribution of 57.68% for LDA classifier.
channel 5 in classification accuracy is maximum for first A. Category 1 EMG pattern classification
category and contribution of channels 1 and 4 are maximum in In this category the classification accuracy has been improved
second category. The classification accuracy of first category for all the classification algorithms. Some of HO activities
is better than second category. The activities performed in first misclassified and predicted as HC activity and vice-versa. In
category are quite different from each other’s, but the case of all channels all the algorithms shows better
activities of second category resemble each other to some classification performance maximum (99.92%) achieved by
extent. By dividing the total activities into two categories, the random forest classifier and minimum (89.76%) by the LDA.
classification accuracy of discriminant analysis (LDA, QDA) classifiers. So for the myoelectric control these kinds of
is improved. These classification algorithms (LDA, QDA) are activity can be easily classified in the short time.
less complex, simpler to implement and require less time in B. Category 2 EMG pattern classification

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

In this category, different gripping activity has been training model prepared with visual feedback setup and tested
performed. Maximum classification accuracy is achieved by it with normal data, the classification accuracy significantly
Random forest (99.70%) and minimum by LDA (82.63%). improved by 87% with LDA and by 85% with Random forest
Although, less amplitude signal is generated on the channel 5 as shown in Fig. 14. In the both the cases, healthy subjects and
(Biceps Brachii muscle) in comparison of other channels, by the amputees the random forest classifier has performed
performing all activities but it has shown better classification satisfactory and consistent. Confusion matrix of the Random
results. forest classifier for both the cases healthy and amputees’
C. EMG pattern classification for amputees pattern classification have been shown in Table IV and Table
The classification performance is evaluated separately for V respectively. Even the feature matrix size is less but it
Amputee’s data, highest accuracy has been achieved by LDA consume more time than other classification algorithms.
classifier with all five muscles signals, and random forest
I. CONCLUSION
classifier has better performance. Channels 5 (Biceps Brachii
The paper presents a wireless and compact EMG data
Muscle) has performed better for different EMG patterns.
acquisition module for prosthesis application. This system has
Lowest classification patterns detected by the SVM classifier.
advantage of online visualization of signal during EMG
When training models are prepared with normal data then
activity and also passes the data in required format (.xls, .txt
classification accuracy has been achieved at most 78% with
and .csv etc.) for the offline processing.
random forest classifier and 73% by LDA. In the case of

TABLE IV TABLE V
CONFUSION MATRIX FOR HEALTHY SUBJECTS WITH RANDOM FOREST CLASSIFIER CONFUSION MATRIX FOR AMPUTEES WITH RANDOM FOREST CLASSIFIER

True HO HC WF WE SG MG HG True HO HC WF WE SG MG HG

Pred. Pred.

HO 1584 0 0 0 10 8 0 HO 86 4 0 0 10 8 0

HC 4 3378 0 0 4 0 2 HC 0 71 0 0 1 2 4

WF 0 0 2316 0 0 0 0 WF 0 0 73 0 0 1 0

WE 0 2 0 1352 0 4 0 WE 0 0 1 57 0 1 0

SG 8 0 0 0 1234 0 4 SG 2 0 0 0 47 18 2

MG 2 2 0 0 8 1538 10 MG 0 2 0 0 18 50 4

HG 12 2 0 0 10 10 1698 HG 1 3 2 0 3 4 60

The proposed system has capability to record several EMG for the feature matrix preparation. The k-NN and Random
based activities for analysis purpose. In this paper, five forest classifier have better performance in case of smaller
channels have been considered for EMG data acquisition. feature vector size. However, LDA has excellent performance
for larger feature vector size. The results show that the
TABLE VI
proposed system has potential to recognize the real time
COMPARISON OF COST WITH OTHER AVAILABLE DEVICES
activity for the upper limb arm. Cost comparison has been
CHANNE PRICE IN COMMERCIAL shown in the Table VI.
S.NO. PRODUCT
LS MARKET
PHYSIOLAB [40] $90 REFERENCES
1. 1
($90X8) FOR 8 CHANNELS
MYOWARE MUSCLE [1] Reaz, M. BI, M. S. Hussain, and Faisal Mohd-Yasin. "Techniques of
$37.95 EMG signal analysis: detection, processing, classification and
2. SENSOR [41] 1
($37.95X8) FOR 8 CHANNELS applications." Biological procedures online 8.1 (2006): 11-35.
NEXUS-10-MKII [42] [2] Merletti, Roberto, and Loredana R. Lo Conte. "Surface EMG signal
3 (INCLUDE SOFTWARE 4 $2900 processing during isometric contractions." Journal of Electromyography
COST) and Kinesiology 7.4 (1997): 241-250.
[3] Stashuk, Dan. "EMG signal decomposition: how can it be accomplished
4. PROPOSED DEVICE 8 PROTOTYPING COST ( $78)
and used?." Journal of Electromyography and Kinesiology 11.3 (2001):
151-173.
In the classification process, the accuracy has been calculated [4] Farina, Dario, and Roberto Merletti. "A novel approach for precise
for all five channels as well as each available channel. During simulation of the EMG signal detected by surface electrodes." IEEE
Transactions on Biomedical Engineering 48.6 (2001): 637-646.
feature extraction process, total 16 features have been selected [5] Doerschuk, Peter C., Donald E. Gustafon, and Alan S. Willsky. "Upper
extremity limb function discrimination using EMG signal analysis."
IEEE Transactions on Biomedical Engineering 1 (1983): 18-29.

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JSEN.2018.2809458, IEEE Sensors
Journal

[6] Erik Scheme MSc, PEng, and PEng Kevin Englehart PhD. in Proc. of the ACM Conf. on Human Factors in Computing Systems,
"Electromyogram pattern recognition for control of powered upper-limb ser. CHI ’15, 2015, pp. 3653–3656.
prostheses: State of the art and challenges for clinical use." Journal of [28] S. Benatti, F. Casamassima, B. Milosevic, E. Farella, P. Schonle, S.
rehabilitation research and development 48.6 (2011): 643. Fateh, T. Burger, Q. Huang, and L. Benini, “A versatile embedded
[7] Avizzano, Carlo Alberto, Emanuele Ruffaldi, and Massimo Bergamasco. platform for emg acquisition and gesture recognition,” Biomedical
"A novel wearable biometric capture system." Control and Automation Circuits and Systems, IEEE Transactions on, vol. 9, no. 5, pp. 620–630,
(MED), 2014 22nd Mediterranean Conference of. IEEE, 2014. 2015.
[8] Liu, Jing Z., et al. "Accessory hardware for neuromuscular [29] De Luca, Carlo J., et al. "Filtering the surface EMG signal: Movement
measurements during functional MRI experiments." Magnetic artifact and baseline noise contamination." Journal of biomechanics 43.8
Resonance Materials in Physics, Biology and Medicine 13.3 (2002): (2010): 1573-1579.
164-171. [30] Hsueh, Ya-Hsin, Chieh Yin, and Yan-Hong Chen. "Hardware System
[9] C. Castellini and P. van der Smagt, “Surface emg in advanced hand for Real-Time EMG Signal Acquisition and Separation Processing
prosthetics,” Biological cybernetics, vol. 100, no. 1, pp. 35–47, 2009. during Electrical Stimulation." Journal of medical systems 39.9 (2015):
[10] A.Pradhan et al., "Acquisition and classification of EMG using a dual- 88.
channel EMG biopotential amplifier for controlling assistive [31] Gokgoz, E., and Subasi, A., Effect of multiscale PCA de-noising on
devices," 2016 IEEE Annual India Conference (INDICON), Bangalore, EMG signal classification for diagnosis of neuromuscular disorders. J.
2016, pp. 1-5 Med. Syst. 38(4):1–10, 2014.
[11] Guo, Weichao, et al. "Development of a multi-channel compact-size [32] Winslow, J., Jacobs, P.L., Tepavac, D., Fatigue compensation during
wireless hybrid SEMG/NIRS sensor system for prosthetic FES using surface EMG. J. Electromyogr. Kinesiol. 13:555–568, 2003.
manipulation." IEEE Sensors Journal 16.2 (2016): 447-456. [33] . Keller, T., Curt, A., Popovic, M.R., Signer, A., Dietz, V., Graspingin
[12] de Oliveira, Igor Harres, and Alexandre Balbinot. "Portable high lesioned tetraplegic subjects using the EMG controlled
electrocardiograph based on the integrated circuit ADS1294 using an neuroprosthesis. Neurorehabilitation 10:251–255, 1998.
android application as interface." Health and Technology 5.2 (2015): [34] Langzama, E., Isakovb, E., Mizrahi, J., Evaluation of methods for
147-154. extraction of the volitional EMG in dynamic hybrid muscle activation. J.
[13] Di Pascoli, Stefano, et al. "Design and implementation of a wireless in- Neuroeng. Rehabil. 3:27, 2006
ovo EEG/EMG recorder." IEEE transactions on biomedical circuits and [35] Langzama, E., Nemirovskyc, Y., Isakovb, E., Mizrahi, J., Muscle
systems 7.6 (2013): 832-840 enhancement using closed-loop electrical stimulation: volitional versus
[14] Hsueh, Ya-Hsin, Chieh Yin, and Yan-Hong Chen. "Hardware System induced torque. J. Electromyogr. Kinesiol. 17:275–284, 2007.
for Real-Time EMG Signal Acquisition and Separation Processing [36] K. Englehart and B. Hudgins, "A robust, real-time control scheme for
during Electrical Stimulation." Journal of medical systems 39.9 (2015): multifunction myoelectric control," in IEEE Transactions on Biomedical
88. Engineering, vol. 50, no. 7, pp. 848-854, July 2003. doi:
[15] H. Huang et al., “An analysis of emg electrode configuration for targeted 10.1109/TBME.2003.813539.
muscle reinnervation based neural machine interface,” Neural Systems [37] Zardoshti-Kermani, Mahyar, et al. "EMG feature evaluation for
and Rehabilitation Engineering, IEEE Transactions on, vol. 16, no. 1, movement control of upper extremity prostheses." IEEE Transactions on
pp. 37–45, 2008 Rehabilitation Engineering 3.4 (1995): 324-333.
[16] Hsu, Chih-Wei, and Chih-Jen Lin. "A comparison of methods for [38] Kim, Kang Soo, et al. "Comparison of k-nearest neighbor, quadratic
multiclass support vector machines." IEEE transactions on Neural discriminant and linear discriminant analysis in classification of
Networks 13.2 (2002): 415-425. electromyogram signals based on the wrist-motion directions." Current
[17] Khoddami, Seyyedeh Maryam, Noureddin Nakhostin Ansari, and Applied Physics 11.3 (2011): 740-745.
Shohreh Jalaie. "Review on laryngeal palpation methods in muscle [39] Ramachandran, Vilayanur S., and Eric L. Altschuler. "The use of visual
tension dysphonia: validity and reliability issues." Journal of Voice 29.4 feedback, in particular mirror visual feedback, in restoring brain
(2015): 459-468. function." Brain 132.7 (2009): 1693-1710.
[18] Instruments, Texas. "ADS129x low-power, 8-Channel, 24-Bit analog [40] https://www.amazon.com/PhysioLab-Biosignal-Module-Mini-size-
front-end for biopotential measurements." datasheet) SBAS459K isolation/dp/B013OUBLI2?th=1
January (2010). [41] https://www.sparkfun.com/products/13723
[19] Cipriani, Christian, et al. "Online myoelectric control of a dexterous [42] https://www.mindmedia.com/products/nexus-10-mkii
hand prosthesis by transradial amputees." IEEE Transactions on Neural
Systems and Rehabilitation Engineering 19.3 (2011): 260-270 Sidharth Pancholi has completed M -Tech
[20] Junior, Jose Jair Alves Mendes, et al. "Robotic Arm Activation using from Thapar University, Patiala, India in
Surface Electromyography with LABVIEW." IEEE Latin America May 2016. Currently, he is pursuing Ph -D
Transactions 14.8 (2016): 3597-3605. from Malaviya National Institute of
[21] L. H. Smith, L. J. Hargrove, B. A. Lock and T. A. Kuiken, "Determining Technology, Jaipur, India. He is working in
the Optimal Window Length for Pattern Recognition-Based Myoelectric biomedical engineering domain. He has two
Control: Balancing the Competing Effects of Classification Error and years of industry experience. His current
Controller Delay," in IEEE Transactions on Neural Systems and research interests include biomedical signal
Rehabilitation Engineering, vol. 19, no.2, pp.186-192, April2011. doi: processing, Prosthetic device development and embedded systems.
10.1109/TNSRE.2010.2100828
[22] A. Pantelopoulos and N. G. Bourbakis, “A survey on wearable sensor- Amit M Joshi has Completed his Ph.D from
based systems for health monitoring and prognosis,” IEEE Trans. on National Institute of Technology, Surat,
Systems, Man, and Cybernetics, vol. 40, no. 1, pp. 1–12, 2010. India in August,2015. Currently, He is
[23] Phinyomark, Angkoon, Pornchai Phukpattaranont, and Chusak working as Assistant Professor in
Limsakul. "Feature reduction and selection for EMG signal Electronics & Communication dept. at
classification." Expert Systems with Applications 39.8 (2012): 7420- National Institute of Technology, Jaipur
7431.. since July 2013. He has eight years of
[24] Houjie, Wei, Guan Jin'an, and Fang Hao. "Portable Biopotential Signal teaching experience and two years of
Acquisition System of ADS1298 Anolog Front-end [J]." industry experience. His current research interests include biomedical
Microcontrollers & Embedded Systems 2 (2012): 012. signal processing, VLSI DSP Systems and Embedded system design.
[25] M. A. Oskoei and H. Hu, “Myoelectric control systems: A survey,”
Biomedical Signal Proc. and Control, vol. 2, no. 4, pp. 275–294, 2007.
[26] D. Farina et al., “The extraction of neural information from the surface
emg for the control of upper-limb prostheses: Emerging avenues and
challenges,” Neural Systems and Rehabilitation Engineering, IEEE
Transactions on, vol. 22, no. 4, pp. 797– 809, 2014.
[27] F. Haque, M. Nancel, and D. Vogel, “Myopoint: Pointing and clicking
using forearm mounted electromyography and inertial motion sensors,”

1558-1748 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.