Assessment of Pain using Optimized Feature Set from Corrugator EMG

Full text

Assessment of Pain using Optimized Feature Set
from Corrugator EMG
Pallab Das
'HSDUWPHQWRI$SSOLHG3K\VLFV
University of Calcutta
Kolkata, India
daspallab36@gmail.com
Kausik Sen
'HSDUWPHQWRI$SSOLHG3K\VLFV
University of Calcutta
Kolkata, India
kausiksen1993@gmail.com
Jhilik Bhattacharyya
'HSDUWPHQWRI$SSOLHG3K\VLFV
University of Calcutta
Kolkata, India
jhilikbhattacharyya96@gmail.com
Saurabh Pal
'HSDUWPHQWRI$SSOLHG3K\VLFV
University of Calcutta
Kolkata, India
spal76@gmail.com
Abstract—Pain is one of the most complex sensation of
human physiology. Till now, physicians use subjective scores
for measuring pain of any individual and doctors need to
completely depend on patient’s response for assessment of
pain. Although, these methods are not always effective in the
medical field, when the subjects are non-cooperative or unable
to response. Hence, subject’s response independent pain
recognition systems are utmost important. Noxious stimulus
excites Sympathetic Nervous System (SNS), which is related to
changes in neuro-somatic biosignals and facial expression. In
this present work, EMG of corrugator muscle which is
pertaining to pain sensitiveness is analyzed. Considering non-
linear & non-stationary nature of the EMG signal stimulated
through pain, Empirical Mode Decomposition technique is
appli ed on EMG for its data adaptive nature. Taguchi Method
of feature optimization is applied on the IMFs for ranking of
features according to their significance. Classification of
different nociception levels with ‘no pain’ was performed
employing linear SVM algorithm, using all extracted features
as well as the most significant features. Appreciable increase in
classification accuracy is noti ced with optimized set of features.
Keywords—EMG, corrugator , EMD, Optimizatio n, SVM,
Pain assessment
I. INTRODUCTION
Pain sensation is one the most commonly felt signal by
human. Inception of pain can occur due internal or external
noxious stimulus which causes actual or potential tissue
damage[1]. Receptors in tissues, known as Nociceptor[2]
gets activated specifically by these painful stimuli.
Thesenociception sensing neurons are localized in the skin,
joints, viscera, and muscles etc[2]. The noxious information
is transduced by the receptors into an electrical signal and
transmitted from the periphery to the central nervous system
along axons[2].These harmful stimulus causes acute pain. If
pain persists for long period it is considered as chronic pain.
So, estimation of pain is critical for clinical practices as it
indicates the severity of the injury. Till now, examining
physician must rely on the patient’s qualitative description
about the location, quality and intensity of the pain sensation.
It is possible to quantify pain with the help of the Visual
Analogue Scale (VAS), Numeric Rating Scale (NRS) or the
Verbal Rating Scale (VRS). Visual Analogue Scale[3] is a
straight line with two extreme points ’no pain’ and ‘worst
possible pain’. Numeric Rating Scale[4] is a 11-point rating
scale, where subjects are asked for scale his/her pain between
the scale of 0-10, where 0 indicates no pain and 10 denotes
maximum pain. However, these subjective methods for pain
mea sur em ent only work wh en the pat ient is sufficiently alert
and co-operative, assessment of pain completely depends on
response of patient which is not always possible in the
medical field. Drowsiness and oblivion can puzzle patients to
provide proper data for pain assessment. Thus, patient’s
response independent pain assessment technique can be
appreciated.
As mentioned earlier, pain is associated with sympathetic
nervous system (SNS)[5]. SNS can control triggering of
nociceptors and thus curbs peripheral inflammation. Rising
in sympathetic activity can alters the common bio-potential
levels of ECG, EEG, EMG, electrodermal activities etc.
Thus measuring of pain can be studied as a viable variation
of different bio-signals.
Now a days, there is a growing interest of doctors &
researchers on bio-signal based analysis of pain stages
around the globe. Different stages of heat pain were realized
using biosignals such as EMG, ECG, GSR, EEG etc[6].
Change of facial expression is recorded using video signals
during various pain stages[6]. However, video based pain
recognition requires complex and costly setup. Thus, there is
an interest to acquire change in facial expression during
different pain stages by using EMG signals of from
zygomaticus and corrugator muscles. Painful heat stimuli
was elicited to subjects that alteration ofbio-potential level
of EMG signal of corrugator muscle with reasonable
accuracy at pain tolerance (P4) condition with respect to ‘no
pain’ condition is observed[7]. To emphasize this concept,
computer-based pain recognition systems are developed that
incorporates the correlation between pain sensation & the
temporal and spatial variation of basic bio-signals.
In this present work, nociceptive level of pain is perceived
from the encapsulated electromyogram (EMG) signal from
corrugator muscle (see Fig. 1)of the subjects exposed with
heat stimuli using BioVid heat pain database[7] (explained
in next section).
Proceedin
g
s of 2020 IEEE A
pp
lied Si
g
nal Processin
g
Conference
(
ASPCON
)
ISBN: 978-1-7281-6882-1 PART: CFP20P52-ART
349
,(((
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on December 10,2020 at 16:41:21 UTC from IEEE Xplore. Restrictions apply.

Fig. 1. Position of muscle[8].
Variation of EMG being a non-stationary phenomena, thus a
data adaptive technique like Empirical Mode Decomposition
(EMD) is used for feature extraction. Various levels of
nociception were realized in two ways. Firstly, using energy
levels of all extracted features. Secondly, feature
optimization technique using Taguchi Method is used to
detect the most significant features. Linear SVM algorithm
is applied to compare various nociceptive levels with ‘no
pain’ condition. Superiority of the Taguchi based method is
observed from the results. The detailed explanations are
provided in the imminent sections.
II. BIOVID HEAT PAIN DATABSE
The database consists of 90 healthy subjects. They were in
resting condition during the experiment and induced heat
pain by a Medoc Pathway thermal stimulator[7] at right
hand (see Fig. 2). Heat stimuli were applied to the subjects
in 4 different pain levels (P1, P2, P3, P4). The baseline (no
pain) was 32°C. P1 is the threshold temperature level. The
maximum temperature which was used for tolerance heat
pain stimulus (P4) is 50.50C. All the temperature levels
were evenly distributed. The intermediate temperature
levels (P2 and P3) were determined as follows:
1
P)
3
1
P
4
P
(
2
P

(1)
1
P)2)
3
1
P
4
P
((
3
Pu

(2)
Every pain stages including no pain was applied 20 times
making 100 responses. Each stimulus was applied for 4
seconds. The pauses between the stimuli were 8–12
seconds.
Bio signals such as Galvanic skin response (GSR), EMG,
ECG, EEG and video signals were recorded using Nexus
32 amplifier[7].
Fig. 2. Location of heat stimulus.
III. METHODOLOGY
Due to induction of pain facial expression changes, this alters
the biopotential of facial muscles. EMG y(k) of corrugator
muscle was taken from BioVid heat pain database.
Fig. 3. Block diagram of proposed method.
Variation of EMG is a non-stationary phenomena. Thus,
data adaptive technique like Empirical Mode
Decomposition (EMD) is applied to analyze the EMG
signal.
Empirical Mode Decomposition[11] is an adaptive, data-
driven technique to analyze complex, non-linear, non-
stationary signals. It is based on local decomposition of data
into a into a sum of oscillatory functions, namely ‘intrinsic
mode functions’ (IMFs) that represents frequency modulated
components and zero mean of amplitude. IMFs should have
two basic features- (1) It has same number of extr ema and
zero-crossings or differ at most by one. (2) Its envelopes are
symmetric with respect to local zero mean[12].
Actual signal can be reorganized by adding the IMFs that
can be mathematically explained as,
)()()(
1
kskIMFkx
n
N
n
n
 ¦
(3)
Wher e, ‘IMFn’ ar e the intrinsic mode functions and ‘Sn’
is denoted as final residue. Decomposition ends when the
value of normalized standard difference (NSD) is within a
predefined threshold, which is usually considered as 0.02,
and expressed [13] as,
¦


M
jj
ii
kd
kdkd
NSD
1
2
2
1
)(
)()(
(4)
Heat
stimulus
applying
position
Corrugator
muscle
Proceedings of 2020 IEEE Applied Signal Processing Conference (ASPCON)
ISBN: 978-1-7281-6882-1 PART: CFP20P52-ART
350
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on December 10,2020 at 16:41:21 UTC from IEEE Xplore. Restrictions apply.

Here, i = 0 to n.
Energy of each IMFs were calculated usingthe formula,
>@
2
1
10
)(log10
¦
N
k
nIMF
kIMFE
(5)
Actual EMG signals with their resulting IMFs (IMF1-IMF
7)
of a subject are given below (see Fig. 4).
NP
P4
Original signal
Original signal
IMF 1
IMF 1
IMF2
IMF2
IMF3
IMF 3
IMF4
IMF 4
IMF5
IMF5
IMF6
IMF6
IMF 7
IMF7
Fig. 4. Original EMG signals and resulting IMFs (1-7)) of
NP and P4.
Here, horizontal axis represents the no. of samples and
vertical axis represents signals.
The energy of IMFs were normalized between 0 to 1 using
Equ. 6
ᇱൌെ݁
୫୧୬
݁୫ୟ୶ െ݁
୫୧୬
 ሺ͸ሻ
where e is the data under test and ᇱis the normalized
value. Firstly, these normalized energies of IMFs are used
for classification. Secondly, feature optimization was
performed to obtain the most impactful IMF energies.
Taguchi Method (TM) is a feature optimization[9] method
developed by Dr. Genichi Taguchi. It identifies the most
significant features and reduce number of features which
are irrelevant, noisy and contain redundant dataset, to
obtain the optimum results of the process and make the
data processing smoother and faster. Orthogonal arrays are
used as per the process requirement. Signal to Noise Ratio
(SNR) is then calculated to determine the robustness of the
features[15]. The feature whose SNR value is lowest have
highest impact and whose SNR value is highest have
lowest impact. Now they are assigned with some values
according to their rank. Similarly, it is performed for all
subjects and finally by adding the assigned values of a
feature for all subjects and by arranging them, the most
impactful features can be extracted. The feature which gets
maximum marks is the most impactful feature. In this
present work, IMF 5 & 6 are estimated as most significant
features.
Support vector machine (SVM) is one type of supervised
learning model in machine learning, used for data
classification & regression analysis[14]. It is introduced by
Boser, Gyon & Vapnik in 1992. The goal of an SVM is to
develop a predictive model based on given training
samples and try to recognize characteristics pattern in such
sample, after learning phase this model can be applied to a
testing data set that can be used to classify unknown input
data into a category. However, in SVM feature space is
divided into two subspaces of two different classes. These
two classes are separated by a hyperplane. There may be
0500 1000 1500 2000 2500 3000
-80
-60
-40
-20
0
20
40
60
80
100
120
0500 1000 1500 2000 2500 3000
-25
-20
-15
-10
-5
0
5
10
15
20
25
0500 1000 1500 2000 2500 3000
-60
-40
-20
0
20
40
60
0500 1000 1500 2000 2500 3000
-25
-20
-15
-10
-5
0
5
10
15
20
25
0500 1000 1500 2000 2500 3000
-40
-30
-20
-10
0
10
20
30
40
0500 1000 1500 2000 2500 3000
-20
-15
-10
-5
0
5
10
15
20
0500 1000 1500 2000 2500 3000
-25
-20
-15
-10
-5
0
5
10
15
20
25
0500 1000 1500 2000 2500 3000
-10
-8
-6
-4
-2
0
2
4
6
8
10
0500 1000 1500 2000 2500 3000
-15
-10
-5
0
5
10
15
0500 1000 1500 2000 2500 3000
-6
-4
-2
0
2
4
6
0500 1000 1500 2000 2500 3000
-4
-3
-2
-1
0
1
2
3
4
0500 1000 1500 2000 2500 3000
-4
-3
-2
-1
0
1
2
3
4
5
0500 1000 1500 2000 2500 3000
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
0500 1000 1500 2000 2500 3000
-2
-1.5
-1
-0.5
0
0.5
1
1.5
0500 1000 1500 2000 2500 3000
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
0500 1000 1500 2000 2500 3000
-1.5
-1
-0.5
0
0.5
1
1.5
Proceedings of 2020 IEEE Applied Signal Processing Conference (ASPCON)
ISBN: 978-1-7281-6882-1 PART: CFP20P52-ART
351
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on December 10,2020 at 16:41:21 UTC from IEEE Xplore. Restrictions apply.

several hyperplan es exist but good separation between
them is achieved by the hyperplane that has the largest
distance to the nearest training-data point of any class. This
is known as margin. Nearest training-data points are called
as support vectors. The larger the margin the lower the
generalization error of the classifier. Thus the hyperplane
serves as the decision function. SVM can be efficiently
perform both linear & nonlinear classification. But the
simplest formulation of SVM is the linear one, where the
hyperplane lies on the space of the input feature vector
x[14]. In this case the hypothesis space is a subset of all
hyperplanes & the linear feature function in the form:
f(x) = wڄx+b (7)
where x is the input/unknown feature vector, w is the
weight vector, represents the orientation of hyperplane and
b is the bias term represents the position of hyperplane in
D-dimensional space.
IV. RESULTS AND DISCUSSION
T-test result shows, no changes in EMG signal during ‘no
pain’ and ‘P1’ stimulus. Moreover, from the mean test also
same inference was established between the two mentioned
stages of EMG, which suggest that subjects have shown
hardly any changes in expression for the lowest stimulus.
Thus, classification between ‘no pain’(NP) and ‘P1’is not
performed in this work.
TABLE I. CLASSIFICATION PERFORMANCE WITH AND WITHOUT
FEATURE OPTIMIZATION USING TM AND LINEAR SVM
Binary
Classif
ication
Linear SVM Classificatio n
Accuracy (%) Specificity (%)
Sensitivity (%)
without
optimiz a
tion
(IMF 1
to 7)
using
optimiz
ation
(I
MF 5
& 6)
without
optimiz a
tion(IMF
1 to 7)
using
optimiz
ation (I
MF 5 &
6)
without
optimiz a
tion
(IM
F 1 to 7)
using
optim
izatio
n
(IMF
5 &
6)
NP vs
P2
58.7
62.7
67.2
41.3
54.8
58.7
NP vs
P3
59.6
64.3
63.5
46.4
55.4
54.2
NP vs
P4
66.7
68.25
74.6
61.9
58.7
74.6
Although, higher pain stages (P2, P3, and P4) were
compared with respect to ‘no pain’. Binary classification was
performed by linear SVM algorithm using the energy levels
of IMFs. From the results it can be observed that accuracy of
the classifier is low for lower stimulus. From r esults it is
again observed that the classification accuracy is improved
when feature optimization is applied using TM compared to
without feature optimization. This justifies the idea of
finding the most significant IMFs. Confusion matrix is a
table that is often used to describe the performance of the
classifier on a set of test data for which the true values are
known. Accuracy, sensitivity and specificity of classifier or
various pain stages are obtained from the confusion matrix
are listed in TABLE I are defined as follows:
Accuracy = σ்௥௨௘ ௉௢௦௜௧௜௩௘ ାσ்௥௨௘ே௘௚௔௧௜௩௘
σ்௢௧௔௟௦௨௕௝௘௖௧௦
ൌ ܶܲ ൅ ܶܰ
ܶܲ ൅ ܨܲ ൅ ܶܰ ൅ ܨܰሺͺሻ
Sensitivity = ே௨௠௕௘௥௢௙்௥௨௘௉௢௦௜௧௜௩௘ ௗ௘௖௜௦௜௢௡௦
ே௨௠௕௘௥௢௙௦௨௕௝௘௖௧௦௪௜௧௛ௗ௜௦௘௔௦௘
ൌ ܶܲ
ܶܲ ൅ ܨܰ  ሺͻሻ
Specificity = ே௨௠௕௘௥௢௙்௥௨௘ே௘௚௔௧௜௩௘ ௗ௘௖௜௦௜௢௡௦
ே௨௠௕௘௥௢௙௦௨௕௝௘௖௧௦௪௜௧௛௢௨௧ௗ௜௦௘௔௦௘
ൌ ܶܰ
ܶܰ ൅ ܨܲ ሺͳͲሻ
False positive rate (FPR)ൌ୊୔
୊୔ା୘୒ ൈ ͳͲͲሺͳͳሻ
Here, TP or True Positive is when a test correctly identifies
a positive result, TN or True Negative is when a test
correctly identifies a negative result, FP or False Positive is
when a test falsely or incorrectly identifies a positive result
and FN or False Negative is when a test falsely or
incorrectly identifies a negative result.
Graphical representation of the parameters is provided in
Fig. 5-6.
Fig
. 5(a). NP vs P4 without
feature
optimization
.
Fig
. 5(b). NP vs P4
with feature
optimizati on using TM
.
Fig. 5. Confusion matrix.
Fig
. 6(a). NP vs
P4 without feature
optimization
.
Fig
.6
(b). NP vs P4 with feature
optimizati on using TM
.
Fig. 6. ROC curves.
The Receiver Operating Characteristic (ROC) curve is a
graphical plot that illustrates the diagnostic ability of
binary classification system at various thresolds settings.
It tells us how the proposed model is capable of
Proceedings of 2020 IEEE Applied Signal Processing Conference (ASPCON)
ISBN: 978-1-7281-6882-1 PART: CFP20P52-ART
352
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on December 10,2020 at 16:41:21 UTC from IEEE Xplore. Restrictions apply.

distingushing between classes. The ROC curve is plotted
with TPR against FPR, where TPR is on Y
-
axis and FPR
is on the X
-axis.
The proposed work is compared with previous works
(TABLE II) based on the classification accuracy. The
superiority of the proposed method is observed.
TABLE II. COMPARISON TABLE
Classifica
tion
Proposed
Method
Markus K¨achele et al [7]
Ehsan
Othman et
al.[16]
Parameters used
EMG
Corrugato r
(IMF 5 &
6)
EMG
(zygomaticus)
EMG
(trapezius)
ECG
GSR
Video si gnal
NP vs P2
62.7
--------
--------
------
------
65.8
NP vs P3
64.3
--------
--------
------
------
NP vs P4
68.25
67.80
61.30
63.50
67
V. CONCLUSION
An approach was framed to classify pain, based on
energy of IMFs of EMG signal using optimizing features set.
Linear SVM seems to be less accurate during lower stimulus,
However, for higher stimulus it exhibits better accuracy. The,
study manifests the possibility of estimating pain severity
using EMG signal with association of data adaptive analysis.
Classification accuracy was further increased with optimized
set of features. The proposed method showed better
classification rate compared to some of the previous works
which classified pain; based on biosignal as well as video
signal. Video signal-based analysis requires complex
computation and costly set up. From the comparative study it
can be remarked that pain analysis by EMG; sometimes can
be a better alternative. Further, other classifiers can also be
tried for prospective results. Relation of other bio-signals like
ECG, EEG, respiration etc. with pain can also be explored
for further studies.
ACKNOWLEDGMENT
Authors acknowledge their gratitude to BioVid heat pain
database (http://www.iikt.ovgu.de/BioVid) a dministrators for
permitting to use the database.
REFERENCES
[1] K. Hanoch Kumar, P. Elavarasi, “Definition of pain and classification
of pain disorder s”,Advanced Clinical & Research Insights, vol.3,
pg.87-90, 2016
[2] Charlotte E Steeds, “The anatomy and physiology of pain”, Basic
Science, 2009.
[3] M. Haefeli and A. Elfering, “Pain assessment,” European Spine
Journal, vol. 15, no. SUPPL. 1. Springer, pp. S17–S24, 01-Jan. 2006,
doi: 10.1007/s00586-005-1044-x.
[4] A. Williamson and B. Hoggart, “Pain: a review of three commonly
used painrating scales,” J. Clin. Nurs., vol. 14, no. 7, pp. 798–804,
Aug. 2005, doi: 10.1111/j.1365-2702.2005.01121.x.
[5] T. Schlereth and F. Birklein, “The sympathetic nervous system and
pain,” NeuroMolecular Med., vol. 10, no. 3, pp. 141–147, Sep. 2008,
doi: 10.1007/s12017-007-8018-6.
[6] K. Markus, P. Werner, and A. Al-hamadi , “Bi o-Visual Fusion for
Person-Independent Recognition of Pain Intensity,” pp. 220–230,
2015, doi: 10.1007/978-3-319-20248-8.
[7] S. Walter et al., “The biovid heat pain database: Data for the
advancement and systematic validation of an automated pain
recognition,” in 2013 IEEE International Conference on Cybernetics,
CYBCONF 2013, 2013, pp. 128–131, doi:
10.1109/CYBConf.2013.6617456.
[8] https://en.wikipedia.org/wiki/Corrugator_supercilii_muscle.
[9] K. Ja mes Mathai, Kshiti Agnihotri, “Optimization techniques for
feature selection in classification”, in International Journal of
Engineering Development and Research, vol. 5, no. 3, ISSN: 2321-
9939, 2017Jeffrey E Janis et al, “The anato my of t he corrugator
supercilli muscle: Part II. Supraorbital nerve branching patterns”, in
Plastic and Reconstructive Surgery, vol. 121, no. 1, pp. 233-40, Feb.
2008.
[10] N. E. Hua ng et al., “The empirical mode decomposition and the
Hilbert spectrum for nonlinear and non-stationary time series
analysis,” Proc. R. Soc. A Math. Phys. Eng. Sci., vol. 454, no. 1971,
pp. 903–995, 1998, doi: 10.1098/rspa.1998.0193.
[11] S. Pal and M. Mitra, “Empirical mode decomposition based ECG
enhancement and QRS detection,” Comput. Biol. Med., vol. 42, no. 1,
pp. 83–92, Jan. 2012, doi: 10.1016/j.compbiomed.2011.10.012.
[12] S. Pal, U. Maji, and M. Mitra, “Characterizing Atrial Fibrillation in
Empirical Mode Decomposition Domain,” J. Med. Biol. Eng., vol. 36,
no. 5, pp. 693–703, Oct. 2016, doi: 10.1007/s40846-016-0168-2.
[14] T. Evgeniou and M. Pontil, ‘’ Workshop on support vector machines:
theory and applications,”doi:10.1007/3-540-44673-7_12, Feb. 2001.
[15] Ng Chin Fei, Nik Mizamzul Mehat, and Shahrul Kamaruddin, “
Practical Applications of Taguchi Method for Optimization of
Processing Parameters for Plastic Injection Moulding: A
Retrospective Review”in ISRN Industrial Engineering vol 13,pp. 1 -
11, Article ID 462174.
[16] E. Othman, P. Werner, F. Saxen, A. Al-Hamadi, and S. Walt er,
“Cross-da tabase evaluation of pain recognition from facial video,” in
International Symposium on Image and Signal Processing and
Analysis, ISPA, 2019, vol. 2019-Septe, pp. 181–186, doi:
10.1109/ISPA.2019.8868562.
Proceedings of 2020 IEEE Applied Signal Processing Conference (ASPCON)
ISBN: 978-1-7281-6882-1 PART: CFP20P52-ART
353
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on December 10,2020 at 16:41:21 UTC from IEEE Xplore. Restrictions apply.