2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept.

21-24, 2016, Jaipur, India

Development of low cost EMG data acquisition

system for Arm Activities Recognition
Sidharth Pancholi Ravinder Agarwal
Electrical and Instrumentation Engineering Department Electrical and Instrumentation Engineering Department
Thapar University Thapar University
Patiala, Punjab, India Patiala, Punjab, India
sid.2592@gmail.com ravinder_eeed@thapar.edu
Abstract—Electromyography (EMG) signals are becoming expensive (e.g. BIONOMADX, NeXus-10) [6] and also system
continuously more important in many fields, including developed based on ICs INA128 or INA2141 are complex in
biomedical/clinical, prosthesis, human machine interaction and design and provide time delay in signal processing [7-9].
rehabilitation devices. In the present study, to meet the requisites Hence there is a need of a system having lower price and
of EMG data acquisition systems, a high resolution, and highly
competitive eight channel system has been developed, which is
compact in design. In this paper a system of bio-signal
cost efficient and compact as compared to commercially available acquisition is presented, which is portable, power efficient and
systems. To validate the developed system, EMG signals have have eight channel. This system transfers bio-signals by serial
been acquired from various muscles for different arm activities peripheral interface to the microcontroller and there decoding
and also machine learning techniques have been utilized for of SPI signals occur. This system displays real time data
activity recognition. For the current study 8 Male and 4 Female acquisition on graphical user interface and also delivers the
healthy subjects have been selected. For classification purpose, data in required format so that the signals of each channel can
various time and frequency domain features have been extracted be easily analyzed. This work provides a potential solution for
and a comparative study of different classification techniques is global market where cost effectiveness and portability is a key
presented. The classification accuracy ranges from 43.64% to
92.61% for different classification algorithms. For this piece of
issue.
work MATLAB 15a is utilized for signal processing and machine Further soft computing techniques have been utilized for
learning. collecting information to separate the different arm based EMG
activities with existing algorithms for comparative analysis of
Index Terms— Electromyogram, Feature extraction, Random machine learning. To be successful in classification and
Forest, Decision Tree, Support Vector Machines (SVM), Linear recognition of the surface EMG signals, three main cascaded
Discriminant Analysis (LDA). modules should be considered which consist of data processing
(filtering), feature extraction and classification module [8]. In
I. INTRODUCTION
the future, this system can also be used to control a robotic arm
EMG signal is useful for clinical application and basic in the field of real-time processing.
research investigations in many areas. Amplitude of the EMG
signal is varying from 0-10mV and the frequency is around
0~1000 Hz [1]. However, the useful EMG signals which are
included important data has frequency range 20 to 500 Hz
[1,2].
There are two types of electromyography: needle
electromyography, it is an invasive technique and involves the
use of needle electrodes that are introduced into the muscular
tissue by the physician , recording its activities and second one
is surface electromyography or sEMG, it delivers an unharmed,
effortless and non-invasive method for the objective
quantification of the energy of the muscle [3].
It is complex to design a methodology which can process
EMG signal throughout the electrical stimulation in real-time
by the reason of separating volitional EMG signal and stimulus
EMG signal involves many issues, such as ES interference and
the effect of filtering problems. Recently, for the hybrid EMG
signals, different system methodology and processing
techniques have been projected for signal analysis [3-5]. II. MATERIAL AND METHODS
The objective of this work is to produce cost effective
multi-channel acquisition system (shown in Fig. 1). The A. Hardware
Fig. 1.Development
Basic steps involved in classification problem
commercially available system now days in market are

978-1-5090-2029-4/16/$31.00 @2016 IEEE 2465

 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

Common acquisition systems of EMG signals are based on B. Electrode Placement and Protocol For Data Acquisition
three main parts, differential amplification, filtering process a) Five channels have been chosen out of eight channels that
and an analog-digital conversion. Electrodes act as the are available in hardware.
interface between muscle tissue and EMG signal amplifier [9]. b) Data have been collected from 12 subjects of age between
In the implemented system these three important parts of 22 to 26 (8 males and 4 females), total seven activities
acquisition system are in a single IC. Development of this

Fig. 4. Real time representation of EMG signals on GUI

have been performed by each subject.

c) The EMG data has been collected from five positions on
Fig. 2. Developed EMG data Acquisition system
the right arm using bipolar Ag/AgCl electrodes [2].
Muscle positions of electrode placement are: Extensor
EMG signal acquisition system is based on ADS 1298 IC Carpi Ulnaris muscle (Channel 1), Extensor Digitorum
of 8 channels with pre-amplifiers, 24 bit ADC, variable Communis muscle (Channel 2), Extensor Carpi Radialis
sampling rate (250 SPS – 32K SPS), low power consumption Longus muscle (Channel 3), Flexor Farpi Radialis
(0.75MW/channel) and SPI interface which is shown in Fig 2. muscle, (Channel 4), and Biceps Brachii muscle (Channel
Before each input, it owns a 2nd order passive low pass filter 5), [6].
of 1kHz cut-off frequency and Schottky diode protection d) Each activity has been performed five times after training
circuit for over voltage protection shown in Fig 3. Schottky the subject.
diode has high switching speed, this circuit can clip the over e) The EMG data has been collected as the volunteers
voltage with respect to the supply. Further Teensy 3.2 module performed seven upper-limb actions HO (Hand open), HC
was used to decode SPI data. (Hand closed), WE (Wrist extension), WF (Wrist
A application has been created in Processing software flexion), SG (Soft gripping), MG (Medium gripping) and
which shows real time data acquisition, change in the number HG (Hard gripping) shown in Fig. 5.
of input channels and sampling rate is also possible through f) A band-pass filter of 20–500 Hz bandwidth and an
this GUI shown in Fig 4. amplifier with 12x gain has been used with sampling
frequency 4000 Hz.
g) Then overlapped windowing technique has been applied
for segmentation of EMG signal, window size has been
taken 300ms with window shift of 0.032ms [10].

C. Feature Extraction
Because of the various artifacts and noises detected among
EMG signals, required information remains merged inside the
raw EMG signals. However, if these amalgamate signals are
used as an input in EMG classification, the efficiency of the
classifier decreases. To boost the performance of the
classification process, scientists have been using different types
of EMG features as an input to the classifier. To achieve
optimum classification performance, the properties of EMG
feature space (e.g., robustness, Maximum Class separatability,
and the computational complexity) should be taken into
Fig. 3. Protection circuit

2466
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

consideration [11]. In this paper two feature groups, which are III. EMG PATTERN CLASSIFICATION
defined in the time domain and frequency domain, have been An efficient means of classifying electromyography (EMG)
considered. Total ten features, six-time domain and four signal patterns has been the interest of several researchers in
frequency domain features examined are shown in Table 1. the modern era. There are several sorts of classifiers, which are
mostly used for different EMG applications, such as an
Artificial Neural Network (ANN), Decision Tree, fuzzy
classifier, Linear Discriminant Analysis (LDA), Random
Forest and Support Vector Machines (SVM). The raw EMG
signal is mapped as a feature vector in the feature extraction
process, which is applied as an input to the classifier. Because
raw EMG signals directly feed to the classifier, they are not
practical due to the randomness of the EMG signal. The
success of the electromyogram classification system highly
depends on the quality of the selected and extracted features
[11]. Feature extraction step in the classification system
increase information density of the signal [12-14].
Fig 5. Various arm activities It is better to classify data with different classifiers instead
of a single classifier to understand which classifier is
TABLE I. FEATURES WITH THEIR MATHMATICAL appropriate for the machine learning. Four types of classifiers
REPRESENTATION
have been used for classification: Decision Tree, Linear
ே Discriminant Analysis (LDA), Random Forest, Support Vector
 ൌ ෍ ȁš୧ ȁ Machines (SVM) for comparative analysis and classification of
Integrated
EMG ௜ୀ଴
Where xi denotes the EMG signal in a segment ‘i’
activities. In all types of classification 70% of data is taken as a
and ‘N’ represents the length of the EMG signal training data set and 30% is taken as testing data set.
Mean ே
ͳ III. RESULT
Absolute  ൌ ෍ ™୧ ȁš୧ ȁ
Value 
௜ୀ଴ This section presents experimental results of developed
ே system. Figure 6-12 shows the waveform for all activities (HC,
Root Mean ͳ
 ൌ ඩ ෍ ‫ݔ‬௜ଶ HO, WE, WF, SG, MG, HG) of subject1 for Carpi Radialis
Square ܰ
௜ୀ଴ Longus muscle (Channel 3).
ேିଵ
Waveform
ܹ‫ ܮ‬ൌ ෍ ȁš୧ାଵ െ š୧ ȁ
Length
௜ୀଵ
ேିଵ

ܼ‫ ܥ‬ൌ ෍ሾ•‰ሺš୧ ൈ  š୧ାଵ ሻ ‫ ת‬ȁš୧ାଵ െ š୧ ȁ စ 

Zero
௜ୀଵ
Crossing ͳǡ ݂݅‫ ݔ‬စ ‫݈݀݋݄ݏ݁ݎ݄ݐ‬
‹‰ሺšሻ ൌ ൜
Ͳǡ ‫݁ݏ݅ݓݎ݄݁ݐ݋‬
Simple ே

Square  ൌ ෍ š୧ଶ
Integral ௜ୀ଴
Fig. 6. Hand close actvity EMG waveform for
channel1
ெ ெ

 ൌ ෍ ˆ୨ ୨ Ȁ ෍ ୨
௝ୀଵ ௝ୀଵ
Mean
Where ‘fj’ is the frequency of the spectrum at
Frequency
frequency bin j, ‘Pj’ is the EMG power spectrum at
frequency bin ‘j’, and M is the length of the
frequency bin

ெ஽ி ெ ெ
Median ͳ
Frequency ෍ ୨ ൌ ෍ ୨ ൌ ෍ ୨
ʹ
௝ୀଵ ௝ୀெ஽ி ௝ୀଵ Fig. 7. Hand open actvity EMG waveform for
Peak channel1
frequency
ܲ‫ ܨܭ‬ൌ ƒš൫ܲ௝ ൯ ݆ ൌ ͳǡʹ ǥ ǥ ǥ Ǥ Ǥ ‫ܯ‬Ǥ
power
Mean ெ
frequency
  ൌ ෍ ୨ Ȁ 
Power
௝ୀଵ

2467
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

In classification process, the accuracy has been calculated

for different classifiers which shown in table II. The accuracies
of different classifiers have been compared, which shows that
the Random Forest classifier is giving the best classification
accuracy 92.61%, which is higher than other classifiers. Here,
confusion matrices of Random Forest and SVM classifier are
presented in Table III and IV, the accuracies and classes mis-
interpretation can be analyzed .

Fig. 8. Wrist extension actvity EMG waveform for TABLE II. CLASSIFICATION ACCURACIES FOR
channel1 VARIOUS CLASSIFIERS
S.NO. Classifier Accuracy %
1. Decision Tree 43.64%
2. Random Forest 92.61%
3. Support Vector Machines (SVM) 81.66%
4. Linear Discriminant Analysis (LDA) 67.14%

TABLE III. CONFUSION MATRIX FOR RANDOM FOREST

CLASSIFIER
Fig. 9. Wrist flexion activity EMG waveform for
channel1 Classes 1 2 3 4 5 6 7

1 672 9 1 0 6 17 23

2 4 728 4 3 2 6 11
3 3 5 659 5 3 2 1
4 6 2 8 501 3 9 6

5 6 5 5 7 535 17 20

6 22 11 3 5 30 520 13
Fig. 10. Soft gripping actvity EMG waveform for
channel1 7 10 8 0 2 15 16 565

TABLE IV. CONFUSION MATRIX FOR SVM CLASSIFIER

Classes 1 2 3 4 5 6 7

1 633 14 2 3 12 17 37

2 6 732 5 0 4 6 5

3 15 4 641 4 10 2 3
Fig. 11. Medium gripping actvity EMG waveform for
4 5 11 6 498 3 9 9
channel1
5 5 18 7 2 586 17 35

6 27 22 5 4 67 520 19

7 26 33 0 5 21 16 506

Random forest and SVM classifiers show the highest

accuracy of classification, activity 6th is misinterpreted with
activity 1st and 5th maximum because these activities are a little
bit similar to each other and activity 1st misinterpreted with
Fig. 12. Hard gripping actvity EMG waveform for activity 7th also, however these misinterpretations have very
channel1 less accuracy. To overcome this misinterpretation problem
more training data should be acquired from subjects.

2468
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

TABLE IV. SESNITIVITY AND AND PRECISION FOR RANDOM REFERENCES

FOREST CLASSIFIERS
[1] A. Boxtel, "Optimal signal bandwidth for the recording of surface EMG
S.NO. Class Sensitivity (%) Precision (%) activity of facial, jaw, oral, and neck muscles," Psychophysiology, vol.
38, no. 1, pp. 22–34, Jan. 2001.
1 1 92.30 92.94
[2] R. Merletti, D. Farina, and M. Gazzoni, "The linear electrode array: A
2 2 96.04 94.79 useful tool with many applications," Journal of electromyography and
kinesiology : official journal of the International Society of
Electrophysiological Kinesiology., vol. 13, no. 1, pp. 37–47, Dec. 2002.
3 3 97.19 96.91
[3] E. A. Clancy and N. Hogan, "Multiple site electromyograph amplitude
4 4 93.64 95.79 estimation," IEEE Transactions on Biomedical Engineering, vol. 42, no.
2, pp. 203–211, Feb. 1995.
5 5 90.21 90.06 [4] N. Hogan and R. W. Mann, "Myoelectric signal processing: Optimal
estimation applied to Electromyography - part II: Experimental
6 6 86.09 88.58 demonstration of optimal Myoprocessor performance," IEEE
Transactions on Biomedical Engineering, vol. BME-27, no. 7, pp. 396–
7 7 91.72 88.41 410, Jul. 1980.
[5] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, "Feature
reduction and selection for EMG signal classification," Expert Systems
with Applications: An International Journal, vol. 39, no. 8, pp. 7420–
TABLE IV. SESNITIVITY AND AND PRECISION FOR SVM 7431, Jan. 2012.
CLASSIFIERS [6] H. Ghapanchizadeh, S. A. Ahmad, and A. J. Ishak, "Developing
multichannel surface EMG acquisition system by using instrument
S.NO. Class Sensitivity (%) Precision (%) opamp INA2141," pp. 258–263, Apr. 2016.
[7] [Online].http://www.ti.com/lit/ds/symlink/ads1298.pdf.
1 1 88.16 88.28 [8] [Online].https://www.biopac.com/wpcontent/uploads/BSL-PRO-3_7-
2 2 96.56 87.76 Manual.pdf.
[9] G. Biagetti, P. Crippa, A. Curzi, S. Orcioni, C. Turchetti, “Analysis of
3 3 90.42 96.24 the EMG Signal During Cyclic Movements Using Multicomponent AM-
FM Decomposition” IEEE Journal of Biomedical and Health
4 4 92.05 96.51 Informatics, vol. 19, no. 5, pp. 1672–1681, Sept. 2015
5 5 87.46 83.35 [10] Rubana H. Chowdhury , Mamun B.,Rea , Mohd Alauddin Bin Mohd Ali
et al., “Surface Electromyography Signal Processing and Classification
6 6 78.31 88.58 Techniques” Sensors, 13, 12431-12466, 2013.
[11] F. H. Y. Chan, Y.-S. Yang, F. K. Lam, Y.-T. Zhang, and P. A. Parker,
7 7 83.36 82.41 "Fuzzy EMG classification for prosthesis control," IEEE Transactions
on Rehabilitation Engineering, vol. 8, no. 3, pp. 305–311, Sep. 2000.
[12] E. Scheme and K. Englehart, "Electromyogram pattern recognition for
control of powered upper-limb prostheses: State of the art and
IV. CONCLUSION challenges for clinical use," Journal of rehabilitation research and
This study demonstrates a low cost EMG data acquisition development., vol. 48, no. 6, pp. 643–59, Sep. 2011.
system and its utilization . This system provides flexibility to [13] H.-J. Yeom, H.-D. Park, Y.-H. Chang, Y.-C. Park, et al., "Stimulus
the user for online visualization of signal during activity and artifact suppression using the stimulation synchronous Adaptive impulse
correlated filter for surface EMG application," Journal of Electrical
also to export the data in desired format for signal processing. Engineering and Technology, vol. 7, no. 3, pp. 451–458, 2012.
The power consumption is low as compared to instrumentation [14] I. Batzianoulis, S. El-Khoury, S. Micera, and A. Billard, "EMG-Based
amplifiers based acquisition systems, as in those systems, analysis of the upper limb motion," pp. 49–50, Feb. 2015. [Online].
separate amplifier IC is used for each channel, where as in this Available: http://dl.acm.org/citation.cfm?id=2701997. Accessed: Aug. 9,
system only one IC is used for 8 channels. The developed 8 2016.
channel device can be extended to 16 channel by replicating
the same circuit with some minor changes in interfacing
device. To validate the utility of developed device arm activity
recognition using EMG data is presented and further various
classifiers are compared.
ACKNOWLEDGMENT
The authors are thankful to Biomedical Research Lab, EIED at
Thapar University Patiala, India to provide facilities for this
research.

2469