Design of a real time portable low-cost multi-channel surface

electromyography system to aid neuromuscular disorder and post

stroke rehabilitation patients
Vinay Chandrasekhar1 , Vikas Vazhayil2 , and Madhav Rao1

Abstract— Surface and needle-based electromyography usually observed at 50-150 Hz frequency spectrum band.
signals are used as diagnostic markers for detecting Surface EMG is widely accepted in medical fraternity due to
neuromuscular disorders. Existing systems that are used to its hassle free application of electrodes on the patients under
acquire these signals are usually expensive and invasive in
practice. A novel 8 channel surface EMG (sEMG) acquisition investigation, and minimum turn around time to determine
system is designed and developed to acquire signals for the signals. The motor functionality of muscles can be as-
various upper limb movements in order to evaluate the certained using the sEMG signal, as the amplitude and other
motor impairment. The real time sEMG signals are generated characteristics of the signal measured are directly correlated
from the muscle fibre movements, originated solely from with the muscle activity [2], [3]. Hence sEMG is commonly
the upper limb physical actions. Intuitively, sEMG signals
characterize different actions performed by the upper limb, used to evaluate motor functionality of post stroke patients
which is considered apt for assessing the improvement for and assess the degree of impairment [4], [5]. Accurate motor
post stroke patients undergoing routine physical therapy impairment assessments could help physicians understand
activities. The system is designed and assembled in a view the type and severity of the impairment, which will help in
to make it affordable and modular for easier proliferation, better planning of physiotherapy. Periodic motor assessments
and extendable to motor classifying applications. The system
was validated by recording realtime sEMG data using six would help in gauging the motor improvement and helps
differential electrodes for various finger and wrist actions. guide further course of rehabilitative measures [6].
The signals are filtered and processed to develop a machine Commercially available sEMG system allows multi-
learning (ML) model to classify upper limb actions, and channel and raw-data collection, and are usually either
other electronic systems are designed in the portable form expensive, bulky, or sparsely available, leading to highly
around the patch electrodes. A classifier was trained to
predict each action and the accuracy of the classifier was inaccessible to the patients under need. A two channel
assessed across different usage of channels. The accuracy low cost EMG signal acquisition and processing circuit is
of the classifier was improved by optimizing the number of described in [7], [8], however picking all possible muscle
electrodes as well as the spatial position of these electrodes. signals from two channels is not adequate to character-
The sEMG circuit designed has the capacity to characterize ize upper limb actions, hence multi-channel sEMG signal
wrists, and finger movements. The improvement observed
in the sEMG signals should benefit the physiotherapists to acquisition and processing system in a compact form is
plan further protocols in the prescribed rehabilitation program. required. An ultra low power microcontroller was considered
for implementing EMG signal processing algorithms and
Clinical relevance— A portable and low-cost sEMG system evaluate muscle signals as reported in [9], hence a similar
allows patients to have easy access to motor functionality system for multi-channel, with additional characterization of
assessments as well as aid physiotherapeutic exercises. Ready
access to such a system will not only allow physicians to upper limb actions is needed.
perform motor impairment studies but also help to quantify and In the past, correlating muscles for specific activities are
gauge the rehabilitation progress through periodic and frequent reported [10], [11], [12], however the finer movements of
assessments of the motor system. the upper limb are still not characterized efficiently, hence
a multichannel sEMG signal with spatial positioning of
I. INTRODUCTION electrodes for picking specific muscle movements needs
Extremely low electric currents are generated from the thorough investigation. Characterization of finer movements
muscle fibres due to the physical movement of human body will assist in recovery of partially deformed parts of the upper
driving the muscle movements [1]. These low currents are limb, especially for patients suffering from specific parts of
observed to have a pattern for different movements, and are fingers, or wrist. A wearable system with dry electrodes to
commonly measured non-invasively by placing electrodes measure EMG signals was also investigated in the past for
over the skin. The signal acquired by the electrodes are measuring the agility index [13]. Again the system lacked the
known as surface Electromyography (sEMG) signals that characterization of finer movements, hence the system may
ranges from µV to mV at the surface of the skin, and is not be directly applicable for rehabilitation of post stroke and
neuromuscular disorder patients.
*This work was supported by EHRC at IIIT-Bangalore EMG signal is inherently stochastic in nature and contains
1 International Institute of Information Technology, Bangalore, Bangalore-
560100, India [email protected] noise, hence special signal processing technique with prior
2 NIMHANS, Bangalore-560029, India feature extraction is generally a preferred method [14].

978-1-7281-1990-8/20/$31.00 ©2020 IEEE 4138

 In the design, bipolar electrode architecture was preferred
over the unipolar architecture due to the following reasons.
Bipolar architecture allows the strategic electrode placement
on various muscle groups with different spatial configura-
tions, allowing the choice of acquiring differential signal
between two distinct points along the length of a muscle
fibres. In addition, bipolar system favourably rejects common
mode noise better than the unipolar configuration. Single
use gel adhesive electrodes (3M Red Dot Sticky Gel 2560)
were used in the design as they are affordable in the order
of 10 cents per electrode, and are widely available. These
electrodes provide a very high signal quality while requiring
minimal to no skin preparation. Alligator clips connected
Fig. 1: Block diagram of the EMG signal acquisition system these electrodes to the front end circuitry. Ideally the wires
carrying signal from the electrodes to the circuitry should be
shielded to reduce noise. The wires were designed to be short
in length and the differential pairs were twisted together to
avoid noise.
The analog circuitry was designed to accommodate eight
simultaneous differential channels. The analog system was
designed with an intention to use minimal circuitry. An in-
strumentation amplifier was used as the first stage to amplify
the differential signal, that rejects common mode noise, while
preserving and amplifying differential signal. Hence a differ-
Fig. 2: Circuit diagram of a single channel EMG amplifier ential amplifier with a high common mode rejection ratio was
chosen. A natural choice would be to use an instrumentation
amplifier (Fig 2 U1 INA106U/2K5) such as the INA106
The paper proposes not only a circuit for EMG signal series, which meets the required specifications. The gain of
acquisition, but also a real time classifier technique for this instrumentation amplifier is fixed. This signal was further
characterizing upper limb movements. Classifier techniques amplified in a second stage (Fig 2 U2A LM358DT) which
in the form of Linear Discriminant Analysis (LDA) and has a variable gain controlled by a potentiometer and passed
Support Vector Machine (SVM) were applied for similar to an active high pass filter (30Hz cutoff frequency) circuit
upper limb movements [15] and accuracy for five channels as shown in the Figure 2. The buffered output (Fig 2 U2B:
EMG system were reported earlier. However the spatial LM358DT) of the filter is directly digitized and stored using
arrangement of patch electrodes towards higher accuracy are an ADC. The instrumentation amplifiers and the active filters
not studied in the past, hence an optimized number, and are housed on a PCB. The overall setup with 8 channel
spatial position of electrodes to cover all upper limb actions electrodes are shown in the Figure 3(a). The PCB was
is much needed and critical for a robust system design. In fabricated using a additive PCB printer, where conductive
this paper, an 8-channel sEMG system is proposed, with real ink was deposited onto the PCB, then appropriate SMD
time classification capability, and an optimized placement components were reflowed to get the printed circuit as shown
of electrodes to acquire signals and characterize upper limb in the Figure 3(b). The filtered analog signal was then used
actions are presented. The device is envisioned as a sensory as an input to the analog to digital converters in a suitable
system applicable towards domestic rehabilitation assessment microprocessor such as the Atmel SAM3X8E ARM Cortex-
for post stroke patients, considered an alternative to the M3 board. The reference of the ADC was adjusted to convert
supervised assessment currently practiced. The physicians complete analog signal swing to digital. The microprocessor
can modify the prescribed set of therapy based on the sEMG sampled all 8 channels data at the preset sampling rate, and
signal assessment report, and chart a plan towards complete transferred data to a system where it is stored and processed
recovery of upper limb. further. A snapshot showing ARM cortex microprocessor
and potentiometer to adjust ADC level is shown in the
II. DESIGN Figure 3(c).
The sEMG signal acquisition system was designed to be The EMG signal is dominant between 50 Hz to 150 Hz.
modular, with 8 separate channels, and the overall fabrication The system samples the signal at 20kHz frequency, which
of the system was made cost effective. Major components of more than satisfies the requirements of EMG recording. The
the system are: the electrodes in the form of patch connected acquired signal is processed to remove any DC component
to front end analog design for signal amplification and and noise using digital filters. Two major filters were de-
conditioning, followed by signal digitization and last stage signed: a notch filter to remove any AC power noise around
consisting of storage, and analysis, as shown in the Figure 1. 50 Hz and a band-pass filter from 30Hz to about 300Hz

4139
 (a) (b)
Fig. 4: Schematic representing placement of six electrodes
with (a) three electrodes on the posterior side of the arm,
Fig. 3: Prototype showing (a) sEMG signal acquisition and (b) three electrodes patched on the anterior side. Image
system with electrodes, (b) Image of sEMG printed circuit redrawn from [16]
board, (c) Image of microprocessor and ADC setup that is
interfaced to the acquisition system.

to eliminate other signal noise. In this processed signal,

any EMG signal would appear as an oscillating wave with
changing amplitude. For further analysis, the processed EMG
signal is passed through a moving window RMS calculator,
to estimate energy of the signal at discrete time points. The
moving window RMS calculator depicts the strength of the
EMG signal for different upper limb actions performed over
time.
III. E XPERIMENTAL A NALYSIS AND R ESULTS
The overall sEMG setup was applied to record wrist
movement, and flexion action of the fingers. Six channels
from the eight designed channels were used and mounted
radially on the arm as shown in the Figure 4. Six pairs
of electrodes completely covered the circumference of the Fig. 5: EMG signal response for individual finger move-
forearm, where the electro-muscular activity was detected. ments, using two channel system.
On posterior side of the arm, the muscle density is less when
compared to the anterior side, hence only two electrodes
were patched on the sides, that provided adequate signal
level. The individual fingers of upper limb were actuated to the difference in the magnitude of signal recorded at
for flexion action, and signal response was experimentally each electrode. As observed in the Figure 5, the amplitude
acquired by the EMG acquisition system. signature is used to estimate the origin of the muscle set
signal, and characterize the performed action. Similar change
A. Principle of operation in signal level is detected for all six channel based EMG
Figure 5 shows RMS response of the sEMG signal for system when placed on arm. The change in signal level
two channels over time, characterizing the flexion action of is detected in each of the six electrodes when each of the
individual fingers. Both channels show a sharp rise in the fingers were actuated. If each action is repeated exactly by
signal level, attributing to the movement of the individual the person, the performed action is actuated using a set of
fingers. The amplitude of the signal is different for each of muscles fibres and the amplitude of signal picked up by the
the channel, which is fundamentally related to the source electrodes has a unique fingerprint. However, every time an
of the EMG signal generated due to the muscle fibres at action is performed, the generated EMG signal is different,
different locations, away from the placed electrodes. The due to various factors including difference in force applied,
generation of the EMG signals at varying locations leads sets of muscle fibres used, and muscle fatigue [17]. Hence to

4140
overcome this issue, the absolute value of the EMG signal is achieve a higher testing accuracy. In order to investigate the
seldom used. Instead, the EMG data is normalized and then same, all combinations of electrode choices were considered
used for further analysis. in the experiments. For example, in one such case, electrodes
1, 3, and 5 as a combination is used to generate features
B. Feature extraction and calculate accuracy. Further among the chosen electrode
The proposed method when applied for every action, RMS subset, the choice of the electrode used for normalization is
waveform for all six channels were generated, with maxi- also a variable. The result of one such experiment with 4
mum value of RMS from the six channels were considered target finger actions is shown in Figure 6, where each data
as absolute valued feature vector. A fixed-length moving point (blue star) represents a unique variable set of electrode
rectangular window was used for calculating the RMS value. subset and normalizer choice, which resulted in certain
The length was determined by trail and error to maximise accuracy. The orange line shows the maximum possible
the RMS energy difference between actuation and relaxation. accuracy achieved for respective number of electrodes used.
Vector consisting of six RMS values were further normalized
and reduced to vector of length five, by considering the ratio
of RMS of each of five channels to one chosen channel RMS
value. Vector of length five is utilized in the classifier to
predict the upper limb actions. The choice of one of these
RMS values as the normalization factor is also a variable,
which heavily affects the classifier accuracy.
C. Action classification and model accuracy

Fig. 7: EMG signal classification accuracy for N fingers

versus the number of electrodes used

D. Optimal number of electrodes

Figure 7 contains the data of model accuracy for various
target classification sets as the number of electrodes, used
Fig. 6: EMG signal classification accuracy for 4 fingers were changed. If one requires to only differentiate between
vs. the number of electrodes used. (*, *, ..) represents the two finger actions, then two electrodes was adequate to
electrode combination corresponding to that accuracy. serve the purpose as more number of electrodes does not
change the accuracy of the model. Similarly if one required
to differentiate between 5 finger classes, than 5 electrodes
sEMG signal was collected for all the actions including usage was recommended. In general it was observed that the
finger flexion and the wrist movements from an individual. number of electrodes required is same as the distinct number
In total 331 actions were captured and the features were of classes one targets to predict.
extracted and normalized. The data was divided into 80%
training and 20% testing classes. The training features were E. Optimal spatial placement of electrodes
then used to train support vector machine (SVM [18]) based From the Figure 6, the highest data point (blue star)
classifiers. Several permutations and combinations were tried for each electrode subset number represents an optimal
to get the optimal feature creation and find the maximum combination of electrode choice which yields highest model
testing accuracy with minimum number of electrode usage. accuracy. It was observed that this choice of electrode subset
Model for different target classes were investigated. One such followed a pattern. It was found that electrode #3 was always
case is the prediction model for 4 finger classes, in which top choice or within 5% accuracy of the top choice, followed
data pertaining to only those 4 finger movements were used in order by the electrodes #4, #2, #5, #1 and then #6. The
to calculate the accuracy. Next, all possible variable states experimental results indicated that if one wanted to use the
were tried to attain maximum accuracy. The classifier model system with 3 sets of differential electrodes, it is optimal
was also investigated for different number of electrodes being to place them at positions of #3, #4, and #2. This is also
used. It was felt that using only a subset of electrodes may biologically intuitive as most of the flexor muscles lie directly

4141
beneath those areas as also shown graphically in Figure [2] V. Glaser, A. Holobar, and D. Zazula, “Real-time motor unit identifi-
4(b). The closer the electrode to the target muscle group, cation from high-density surface emg,” IEEE Transactions on Neural
Systems and Rehabilitation Engineering, vol. 21, no. 6, pp. 949–958,
the higher signal quality is achieved hence allowing a better Nov 2013.
classification accuracy. [3] O. W. Samuel, X. Li, P. Fang, and G. Li, “Examining the effect of
subjects’ mobility on upper-limb motion identification based on emg-
F. Realtime action classification pattern recognition,” in 2016 Asia-Pacific Conference on Intelligent
Robot Systems (ACIRS), July 2016, pp. 137–141.
The proposed system was designed to render a realtime [4] M. Chen, X. Zhang, and P. Zhou, “A novel validation approach for
classifier. The model was trained on an individual and then high-density surface emg decomposition in motor neuron disease,”
tested in real time. The system constantly measures EMG IEEE Transactions on Neural Systems and Rehabilitation Engineering,
vol. 26, no. 6, pp. 1161–1168, June 2018.
signals, and detects possible muscle actions when it sees an [5] X. Li, A. Holobar, M. Gazzoni, R. Merletti, W. Z. Rymer, and P. Zhou,
increase in EMG signal energy corresponding to an increase “Examination of poststroke alteration in motor unit firing behavior
in RMS value of the signal. As soon as an action is detected, using high-density surface emg decomposition,” IEEE Transactions on
Biomedical Engineering, vol. 62, no. 5, pp. 1242–1252, May 2015.
EMG data is collected and using a classifier the target action [6] O. W. Samuel, M. G. Asogbon, Y. Geng, A. H. Al-Timemy, S. Pirb-
is predicted. The EMG data from start and end of the hulal, N. Ji, S. Chen, P. Fang, and G. Li, “Intelligent emg pattern
action is used to generate a feature vector which is used recognition control method for upper-limb multifunctional prostheses:
Advances, current challenges, and future prospects,” IEEE Access,
as input to the classifiers. It was generally found that the vol. 7, pp. 10 150–10 165, 2019.
system collected data for about a hundred milliseconds and [7] T. S. Poo and K. Sundaraj, “Design and development of a low cost
generate a prediction, within a span of tens of milliseconds. emg signal acquisition system using surface emg electrode,” in 2010
IEEE Asia Pacific Conference on Circuits and Systems, Dec 2010, pp.
The reduced latency of the realtime classifier system can be 24–27.
further utilized to various other upper limb actuation systems [8] M. S. Çelik and Eminoğlu, “Fuzzy logic based classification for
including prosthetic hand and wearable exoskeleton devices. multifunctional upper limb prostheses,” in 2016 Medical Technologies
National Congress (TIPTEKNO), Oct 2016, pp. 1–4.
G. Overall cost of the system [9] T. Roland, W. Baumgartner, S. Amsuess, and M. F. Russold, “Signal
evaluation of capacitive emg for upper limb prostheses control using
It was observed that the overall cost of this system an ultra-low-power microcontroller,” in 2016 IEEE EMBS Conference
was around USD 200. The majority of the cost being the on Biomedical Engineering and Sciences (IECBES), Dec 2016, pp.
317–320.
instrumentation amplifier ICs, that can be replaced with [10] Costa, M. Itkonen, H. Yamasaki, F. S. Alnajjar, and S. Shimoda,
more affordable options. The cost of the entire system “Importance of muscle selection for emg signal analysis during
could be further reduced by employing bulk manufacturing upper limb rehabilitation of stroke patients,” in 2017 39th Annual
International Conference of the IEEE Engineering in Medicine and
techniques. The simplicity and modular nature of the design Biology Society (EMBC), July 2017, pp. 2510–2513.
allows for employing cheaper manufacturing techniques as [11] B. N. Cahyadi, I. Zunaidi, S. A. Bakar, W. Khairunizam, S. H. Majid,
well as allowing swapping of modules. Z. M. Razlan, M. N. Muhammad, M. N. Rudzuan, and W. A. Mustafa,
“Upper limb muscle strength analysis for movement sequence based
IV. CONCLUSION on maximum voluntary contraction using emg signal,” in 2018 In-
ternational Conference on Computational Approach in Smart Systems
A novel portable low-cost eight channel sEMG signal Design and Applications (ICASSDA), Aug 2018, pp. 1–5.
acquisition system was designed, fabricated, and tested as a [12] R. Merletti and D. Farina, Techniques for Information Extraction from
the Surface EMG Signalhigh-Density Surface EMG, 2016, pp. 126–
proof of concept. To validate the system, the sEMG signals 157.
for single finger actuation and wrist actuation was acquired [13] E. J. Pino, Y. Arias, and P. Aqueveque, “Wearable emg shirt for upper
multiple times. The collected data was processed and was limb training,” in 2018 40th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC), July
used to train an SVM model with relative RMS voltage peaks 2018, pp. 4406–4409.
as multiple features. The acquisition system characterizes [14] M. S. Hazam Majid, W. Khairunizam, A. B. Shahriman, I. Zunaidi,
the five finger flexion movements with an accuracy of 85% B. N. Sahyudi, and M. R. Zuradzman, “Emg feature extractions
for upper-limb functional movement during rehabilitation,” in 2018
in real time. From the data collected and analysed it was International Conference on Intelligent Informatics and Biomedical
concluded that the optimal number of electrodes to be used Sciences (ICIIBMS), vol. 3, Oct 2018, pp. 314–320.
is roughly equal to the number of target classification classes. [15] A. Zhang, N. Gao, L. Wang, and Q. Li, “Combined influence of
classifiers, window lengths and number of channels on emg pattern
Further, the most optimal placement of these set of electrodes recognition for upper limb movement classification,” in 2018 11th
is as close as possible to the muscle group of interest. The International Congress on Image and Signal Processing, BioMedical
system was thus shown to be capable of identifying the Engineering and Informatics (CISP-BMEI), Oct 2018, pp. 1–5.
[16] A. F. Currier, in (Ed.) The Foundation Library (New York, NY: The
action performed and the intensity of the action performed. Educational Society), 1911.
The developed sEMG signal acquisition circuit automatically [17] M. D. F. Ma’as, Masitoh, A. Z. U. Azmi, and Suprijanto, “Real-time
collected signal over time from an individual, that will pro- muscle fatigue monitoring based on median frequency of electromyo-
graphy signal,” in 2017 5th International Conference on Instrumenta-
vide insight into the rehabilitation progress of the individual. tion, Control, and Automation (ICA), Aug 2017, pp. 135–139.
Thus avoiding tedious manual evaluation of the individual [18] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
and provide personalized exercise and physiotherapy plans. O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Van-
derplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and
R EFERENCES E. Duchesnay, “Scikit-learn: Machine learning in Python,” Journal
of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.
[1] T. Koshio, S. Sakurazawa, M. Toda, J. Akita, K. Kondo, and Y. Naka-
mura, “Identification of surface and deep layer muscles activity by
surface emg,” in 2012 Proceedings of SICE Annual Conference (SICE),
Aug 2012, pp. 1816–1821.

4142