Automatic Pain Intensity Estimation based on Electrocardiogram and

Demographic Factors

Stefanos Gkikas

, Chariklia Chatzaki

, Elisavet Pavlidou

, Foteini Verigou

Kyriakos Kalkanis

and Manolis Tsiknakis

Department of Electrical and Computer Engineering, Hellenic Mediterranean University,

Estavromenos, 71410, Heraklion, Greece

Keywords:

Pain Recognition, ECG, Machine Learning, Age, Gender.

Abstract:

Automatic pain intensity estimation possess signiﬁcant importance for reliable and complete pain manage-

ment. The accurate and continuous monitoring is essential in order to attain objective insight about the con-

dition of the patient. In this work, we elaborate physiological signals in order to estimate the pain intensity

and investigate the impact of demographic factors. Speciﬁcally, we exploit electrocardiography signals, adopt-

ing the Pan-Tompkins algorithm to extract important features and apply well-validated classiﬁcation methods,

while we explore the correlation of gender and age with the pain manifestation.

1 INTRODUCTION

Pain according to the most accepted deﬁnition is ”an

unpleasant sensory and emotional experience associ-

ated with actual or potential tissue damage, or de-

scribed in terms of such damage” (Merskey et al.,

1979). The two main types of pain are acute and

chronic, where their main difference is related to the

duration; in the case of acute pain the sensation is

momentary, whereas in chronic lasts longer than a

few months. Pain is a major issue in humans’ life-

time, since every day people of all ages experience

pain, either due to an accident or due to an illness or

even during a treatment. A large percentage of peo-

ple who have been treated for serious illnesses, even

if they have overcome them, may suffer from persis-

tent pain (Lynch, 2011). Since pain is a situation with

physical and psychological dimensions, it affects peo-

ple in a major degree provoking a plethora of daily

life challenges, and especially in chronic pain condi-

tion, it often leads to depression, sleep problems, and

anorexia (Lopez-Martinez and Picard, 2017). Pain

constitutes an incumbrance in health care systems,

since more than 50% of those who are in a hospital

are experience the phenomenon of pain (Cordell et al.,

https://orcid.org/0000-0002-4123-1302

https://orcid.org/0000-0003-1312-8401

https://orcid.org/0000-0003-3744-9648

https://orcid.org/0000-0002-3428-158X

https://orcid.org/0000-0003-2292-8806

https://orcid.org/0000-0001-8454-1450

2002). Large resources of medical and nursing staff

are consumed for the therapy of chronic pain, exceed-

ing the cost of diseases such as cardio-vascular and

neoplasms (Gaskin and Richard, 2012). A consid-

erable body of research indicates divergence on pain

expression and sensation among individuals with dif-

ferent gender or age. In the study of Bartley and

Fillingim (2013) conducting a psychological review

research, concluded that females experience greater

discomfort and pain in more areas of the body than

males, and generally are more sensitive. Similarly, in

the research (Toomey, 2008) also discovered that fe-

males had lower thresholds evaluating equal stimuli

as more intense and painful compared to men, while

Hadjistavropoulos and Craig (2002) refereed that el-

ders present important alterations in pain perception,

compared to younger people.

Self-report is the common approach in clinical

practice used for determining the presence and the

severity of pain, by rating scales or questionnaires.

This process is time consuming and challenging es-

pecially for patients with communicational abnormal-

ities, intellectual disabilities, people with serious ill-

ness or infants, and moreover, the continuous pain

monitoring is unfeasible with the absence of computer

systems (Werner et al., 2014). Adequate and objec-

tive pain assessment is required, in order to provide

the necessary care to people who are suffering, since

inappropriate management of pain can lead to addi-

tional serious health problems.

The process of recognizing pain, is based on

analysis of behavioural and physiological responses;

Gkikas, S., Chatzaki, C., Pavlidou, E., Verigou, F., Kalkanis, K. and Tsiknakis, M.

Automatic Pain Intensity Estimation based on Electrocardiogram and Demographic Factors.

DOI: 10.5220/0010971700003188

In Proceedings of the 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2022), pages 155-162

ISBN: 978-989-758-566-1; ISSN: 2184-4984

155

the behavioural reaction is related with facial ex-

pressions, vocalizations and body-head movements,

while the physiological are the interactions from the

neural structures which lead to the sympathetic out-

ﬂow being affected and increased, something that

can be observed in the physiological signals (Stewart

and Panickar, 2013), such as electroencephalography

(EEG), electrodermal activity (EDA), electromyogra-

phy (EMG) and electrocardiography (ECG).

This study elaborates a pain estimation procedure,

exploiting ECG signals and investigates the differ-

ences on various demographic groups, related to gen-

der and age. In particular, we scrutinize the varia-

tion in pain manifestation among males and females,

we examine the pain diversities across age groups,

and furthermore we combine the factor of age with

the gender. The remaining of this paper is organ-

ised as follows: in Section 2 we present the related

work on automatic pain estimation, Section 3 de-

scribes the methodology for the feature extraction ap-

proach and the machine learning methods, in Section

4 we present the conducted experiments and results

and the paper is concluded in Section 5.

2 RELATED WORK

A signiﬁcant number of published research efforts

(Wang et al., 2020; Mamontov et al., 2019), were

based on biosignals to interpret the pain sensation,

since in many cases the video modalities e.g. facial

expression are relatively difﬁcult to elaborate, espe-

cially in clinical settings where the patient rotates or

due to facial occlusions from medical support devices

or the patient’s hands. Additionally, there are circum-

stances where the individuals pretend facial expres-

sion in order to elicit personal interest (Rohling et al.,

1995). The correlation among pain and physiolog-

ical signals like heart rate variability, electrodermal

activity, brain function, and respiration rate have been

examined in plethora of studies (Werner et al., 2019).

According to a study done by Chu et al. (2017)

linear discriminant analysis was utilized based on

data of six healthy participants, to categorize the pain

into ﬁve distinct levels, based on ECG and EDA

employing Genetic Algorithms and Principal Com-

ponent Analysis. K

achele et al. (2016) exploited

EDA, ECG and EMG, extracting several features

such as skewness, standard deviation and QRS com-

plexes. Applying Random Forest (RF) conducted

multi-classiﬁcation on ﬁve pain levels, while Lopez-

Martinez and Picard (2017) designed a multitask neu-

ral network for personalized pain recognition using

ECG and EDA signals. Similarly, in their work

Amirian et al. (2016) explored the continuous and

discrete pain estimation exploiting ECG, EMG and

EDA. Time domain (e.g. Willison amplitude), fre-

quency domain and entropy features (e.g. Shannon)

were extracted and a Radial Basis Function (RBF)

Neural Network was designed achieving accepted

performances.

Recently, Deep Learning (DL) approaches have

been studied for their application in pain estimation,

after their success in different scientiﬁc areas. The re-

search of Thiam et al. (2019) investigated DL mod-

els for pain categorization. They applied 1D Con-

volutional Neural Networks (CNN) on galvanic skin

response signals improving the binary classiﬁcation

between no pain and the highest pain intensity. Yu

et al. (2020) studied three levels of pain based on

EEG, developing a framework that consisted of sev-

eral convolutional modules, each one related to differ-

ent frequency range, while Lopez-Martinez and Pi-

card (2018) proposed a Recurrent Neural Network

(RNN) for determining the severity of the pain, based

on the deconvolved signals, extracting tonic-phasic

components from EDA and R peaks as well as inter-

beat intervlas (IBIs) from ECG.

Furthermore, to the best of our knowledge there

exist two studies which investigate demographic fac-

tors. In the study of Hinduja et al. (2020) the au-

thors exploited several physiological signals (e.g. di-

astolic blood pressure, respiration rate, EDA), facial

action units and the combination of them as well, re-

vealing the existence of signiﬁcant differences among

men and women in pain sense, on both unimodal

and multimodal approaches. Similarly, Subramaniam

and Dass (2021) conducted experiments through ECG

and EDA adopting Convolutional Neural Networks

(CNN) disclosed performance variations between the

gender.

3 METHODOLOGY

The employed pain database, the electrocardiogra-

phy processing algorithm, as well as the extraction

method of features and classiﬁcation algorithms will

be described in this section.

3.1 BioVid Heat Pain Dataset

In this study we utilized the publicly available

”BioVid Heat Pain Database” (Walter et al., 2013),

which comprises facial videos, and biosignals (ECG,

EMG, EDA) from 87 subjects (44 males and 43 fe-

males, age 20-65 ), and currently is the only pub-

licly available dataset which includes the subjects’

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

156

age and gender. Data were collected by subjecting

heat stimulus on the right arm by a thermode. Be-

fore the data recording was started, for each subject

the pain threshold (the temperature for which the par-

ticipant’s sensing changes from heat to pain) and pain

tolerance (the temperature at which the pain becomes

intolerable) were determined. The speciﬁc thresh-

olds utilized as the temperatures for lowest and high-

est pain levels, and addition two intermediate levels

were included, resulting to 5 pain conditions: No pain

(NP), mild pain (P1), moderate pain (P2), severe pain

(P3), very severe pain (P4). The participants were

stimulated 20 times for every intensity, thus gener-

ating 100 samples for each of the four modalities.

We employed Part A of the BioVid, speciﬁcally the

pre-processed ECG samples with a Butterworth band-

pass ﬁlter (87 × 100 = 8700).

3.2 ECG Signal Processing and Analysis

An ECG signal reﬂects the electrical activity of the

heart over a period of time. A normal ECG consist

of a series of waves, namely P, Q, R, S, T and some-

times U. These waves and their intervals give impor-

tant information regarding the heart’s function. The P

wave signs atrial depolarization. The QRS complex

represents ventricular depolarization and contraction,

while the T wave represents repolarization of ventri-

cles. Therefore, each heartbeat is represented by the

PQRST complex (see Figure 1). The accurate and re-

liable detection of the R wave in the QRS complex is

of high importance since it is the most prevalent peak

in the complex. By accurately detecting the R wave

we can compute the heart rate (HR) and the heart rate

variability (HRV), which is related with the time inter-

val between consecutive R waves, called R-R interval

or Interbeat interval.

One the most widely used real time QRS detec-

tion algorithm is the Pan-Tompkins Algorithm (Pan

and Tompkins, 1985). Over the past decades the orig-

inal Pan-Tompkins algorithm and several modiﬁca-

tions have been evaluated, with the results to support

its efﬁciency even in noisy and low-quality data (Far-

iha et al., 2020; Liu et al., 2017). The performance

of the Pan-Tompkins Algorithm has been repeatedly

evaluated and therefore considered to be one of the

state-of-the-art algorithms for QRS detection, used

also for performance comparison in new approaches

(Zhao et al., 2021). In this study we adopted the

original Pan-Tompkins Algorithm for the detection of

the QRS complex. The integration of the algorithm

emerged in two stages: the preprocessing and the de-

cision. The preprocessing is an essential procedure

to prepare the ECG, removing the noise and artefacts,

QRS

Complex

Interval

QT Interval

Figure 1: The PQRST complex.

Low Pass FilterECG High Pass Filter Differentiation

Adaptive Thresholds

Moving Window

Integration

Squaring

QRS

Complex

Band-Pass Filter

Low Pass FilterECG High Pass Filter

Differentiation

Adaptive

Thresholds

Moving Window

Integration

Squaring

QRS

Band-Pass Filter

Figure 2: The ﬂow diagram of the pre-processing procedure

of the Pan-Tompkins algorithm.

smoothing the signal, and increasing the QRS slope.

The ﬂow diagram of the preprocessing procedure of

the Pan-Tompkins algorithm is shown on Figure 2.

3.3 Feature Extraction

The succeeding phase is the extraction of speciﬁc fea-

tures based on the inter-beat intervals (IBIs). In our

work, the mean of IBIs, the root mean square of suc-

cessive differences (RMSSD), the standard deviation

of IBIs (SDNN), the slope of the linear regression of

IBIs, the ratio of SDNN to RMSSD, and the heart beat

rate, were calculated as follows:

1. Mean of IBIs

µ =

∑

i=1

(RR

i+1

− RR

) (1)

where RR are consecutive R peaks.

2. Root mean square of successive differences

RMSSD =

n − 1

n−1

∑

i=1

(RR

i+1

− RR

)

(2)

3. Standard deviation of IBIs

SDNN =

n − 1

∑

i=1

(RR

− µ)

(3)

Automatic Pain Intensity Estimation based on Electrocardiogram and Demographic Factors

157

4. Slope of the linear regression of IBIs

Ax = A

b (4)

based on the least-square approximation, where b

is the vector of RR peak intervals and A is the cor-

responding time series.

5. Ratio of SDNN to RMSSD

SR =

SDNN

RMSSD

(5)

6. Heart beat rate

HR =

60 · FS

(6)

where FS is the frequency of ECG recording and

is equal to 512 Hz. Figure 3 shows the raw ECG

signal and the applied algorithm’s steps as well.

3.4 Classiﬁcation Methods

For the classiﬁcation phase, three well known clas-

siﬁers were deployed: Linear Discriminant Analysis

(LDA), the Support Vector Machine (SVM) with lin-

ear kernel, and the SVM with Radial Basis Function

(RBF) kernel. Furthermore, all the conducted exper-

iments repeated threefold with identical settings, uti-

lizing a particular classiﬁer in every repetition in order

to compare their performances, founded on the leave-

one-subject-out (LOSO) cross validation, employing

all the available subjects and ECG samples, and as

evaluation performance we adopted the metric of ac-

curacy.

1. Linear Discriminant Analysis

P(X|y = k) =

exp



−

(X −µ

)

−1

(X −µ

)



(2π)

d/2

|Σ

1/2

(7)

where P is the probability density function of fea-

tures X given the target y and class k.

2. SVM with linear kernel

K(x

) = x

(8)

where x

, x

are features from two distinct classes.

3. SVM with Radial Basis Function (RBF) kernel

K(x

) = exp

−

||x

− x

2σ

(9)

where σ is the width of the kernel.

4 EXPERIMENTS & RESULTS

Utilizing the aforementioned classiﬁcation algo-

rithms, we conducted several experiments with the

objective of pain recognition and it’s relation with

demographic factors. The classiﬁcation tasks were

based on the pain conditions and implemented in a

multi-class classiﬁcation manner, as well as binary

classiﬁcation. Speciﬁcally ﬁve distinct experiments

were performed: (1) multi-class pain classiﬁcation,

(2) NP vs P1, (3) NP vs P2, (4) NP vs P3, (5) NP vs

P4. In (1) the purpose is to classify an ECG signal

in one of the ﬁve pain conditions, while in (2)-(5) to

classify it in one of the two pain conditions, i.e. no

pain and the corresponding pain level. In addition,

taking into consideration the gender and the age of

subjects, we developed four different schemes; (1) the

basic scheme where we employed the whole dataset,

(2) the gender scheme where the data were divided

based on the gender of subjects i.e. males-females, (3)

the age scheme based on the age of subjects, creating

three groups i.e. ’20-35’, ’36-50’, ’51-65’, and ﬁnally

(4) the gender-age scheme was in accordance to gen-

der and age combined, creating six different subjects

groups i.e. ’males 20-35’, ’females 20-35’, ’males 36-

50’, ’females 36-50’, ’males 51-65’, ’females 51-65’.

The best classiﬁcation results are presented in Figures

4-5 for every corresponding task and the utilized clas-

siﬁcation method, while the Tables 1-5 enclose every

individual experiment.

Table 1 presents the results utilizing the whole

dataset, where for the multi-class pain classiﬁcation

we achieved 23.79%, and the performance scores in-

creased as much as the pain intensity raise, reaching

58.62% on NP vs P4, indicating the challenges to de-

tect the low magnitude of pain severity. Regarding the

classiﬁcation algorithms, the SVM (linear) performed

ameliorative, besides the last task related to the higher

level of pain, while the SVM (RBF) exhibited inferi-

orly. On the gender scheme (see Table 2), we observe

differences among males and females. Females in to-

tal presented 1.12% variation from males, where in

task NP vs P4 attained 60.69% accuracy over 56.07%,

where the 4.62% increase indicates that women are

more sensitive in higher pain levels than men. Curi-

ously, in NP vs P1 and NP vs P2 the males outper-

formed by 1.16% and 1.78%, respectively, and sim-

ilarly to the ﬁrst scheme, the SVM (linear) obtained

greater results in most of the tasks. Figure 4 depict

the gender differences, based on the classiﬁcation ac-

curacy.

On the age scheme (see Table 3), the group ’20-

35’ presented 25.06% in multi-level classiﬁcation

over 23.27% and 22.35% from the groups ’36-50’and

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

158

0 500 1000 1500 2000 2500

Time (ms)

Raw Signal

0 500 1000 1500 2000 2500

Time (ms)

Band Pass Filtered

0 500 1000 1500 2000 2500

Time (ms)

Derivative Filtered

0 500 1000 1500 2000 2500

Time (ms)

Squared

500 1000 1500 2000 2500

Time (ms)

Moving Window Averaged

Signal

QRS

Noise Level

Signal Level

Adaptive Threshold

Figure 3: The signal pre-processing with Pan-Tompkins algorithm.

Table 1: Classiﬁcation results of the Basic Scheme, re-

ported on % accuracy.

Group

Task

Algorithm

LDA SVM LN SVM RBF

All

MC 23.72 23.79 22.77

NP vs P1 50.97 52.38 49.97

NP vs P2 52.55 52.78 52.70

NP vs P3 55.20 55.37 53.87

NP vs P4 58.62 58.39 57.41

MC: multi-classiﬁcation NP: no pain P1: mild pain P2: moderate pain

P3: severe pain P4: very severe pain LDA: Linear Discriminant Analysis

LN:Linear RBF: Radial Basis Function

Table 2: Classiﬁcation results of the Gender Scheme, re-

ported on % accuracy.

Group

Task

Algorithm

LDA SVM LN SVM RBF

Males

MC 22.13 22.25 20.70

NP vs P1 51.53 52.61 47.72

NP vs P2 53.12 53.69 52.15

NP vs P3 54.94 54.71 51.36

NP vs P4 55.28 56.07 51.36

Females

MC 25.11 24.41 23.41

NP vs P1 50.23 51.45 49.06

NP vs P2 51.62 51.86 51.91

NP vs P3 55.98 55.87 55.29

NP vs P4 60.17 60.69 59.82

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Basic Gender Age Gender-Age

Schemes (max scores)

Basic

Gender

Age

Gender-Age

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Gender (max scores)

Males

Females

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Males Females

Age

20-35

36-50

51-65

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

20-35 36-50 51-65

Gender-Age

Males 20-35

Females 20-35

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

Males 20-35 Females 20-35 Males 36-50 Females 36-50 Males 51-65

Figure 4: Classiﬁcation results on the Gender Scheme.

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Basic Gender Age Gender-Age

Schemes (max scores)

Basic

Gender

Age

Gender-Age

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Gender (max scores)

Males

Females

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Males Females

Age

20-35

36-50

51-65

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

20-35 36-50 51-65

Gender-Age

Males 20-35

Females 20-35

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

Males 20-35 Females 20-35 Males 36-50 Females 36-50 Males 51-65

Figure 5: Classiﬁcation results on the Age Scheme.

’51-65’ respectively, revealing that age is a factor

which inﬂuences the pain sensation. Especially, in

NP vs P4 demonstrated a difference nearly 9% be-

tween the youngest and the oldest group. Analo-

gously to the gender scheme, in low pain intensi-

ties there exist slight differences among the groups,

which increased as the pain escalates. Speciﬁcally,

the variance (σ

) between the three groups in NP

vs P1 was 1.38% while in the remaining tasks was

Automatic Pain Intensity Estimation based on Electrocardiogram and Demographic Factors

159

2.44%, 6.35% and 20.42% respectively, revealing that

in order to identify the differences of pain perception

among the groups, is essential for pain to increase to a

high intensity. On the basis of classiﬁcation accuracy,

the group ’20-35’ described with the highest sensitiv-

ity followed by the group ’36-50’ and ’51-65’. Ad-

ditionally, relating to the classiﬁcation methods, the

SVM (RBF) in group ’51-65’ performed superior in

almost every task, while in group ’20-35’ completely

underperformed, indicating that is the ﬁnest choice

for difﬁcult separable classes. In Figure 5 illustrated

the results of Age scheme.

In the last scheme, we studied in a more precisely

manner the subjects, in order to obtain a better insight

about the correlation of pain and the demographic fac-

tors of gender and age. In Tables 4-5 we observe that

in the multi-class pain classiﬁcation the higher accu-

racy achieved by ’females 20-35’ with 24.80% accu-

racy, while in NP vs P1 the ’females 51-65’ reached

55.38%, disclosing again that the female gender char-

acterized by a elevated sensitivity. Likewise, the high-

est performance in NP vs P2 achieved by the ’females

51-65’ followed by the ’males 51-65’, although in NP

vs P3 the ’females 36-50’ exceeded by 3.5% from the

second best group ’males 20-35’. Finally, in the last

task of NP vs P4 the ’females 20-35’ outperformed

attaining 67%, where the minimal performance ob-

served on ’males 51-65’ with 54.50%, where the par-

ticular groups are the uttermost and the minor sensi-

tive group respectively. We report that in some cases

the classiﬁcation accuracy is lower than others, de-

spite the fact the pain level is increased (e.g. ’Fe-

males 36-50’). An explanation could be related to

the recording procedure of biosignals, where the sub-

jects may get accustomed to the stimulation thence-

forwards.

In Figure 6 we visualize the classiﬁcation perfor-

mances of the six groups in the Gender-Age scheme.

Additionally, in Table 6 we compare ours accom-

plished results with related studies which utlized the

ECG signals from BioVid database and followed the

same evaluation protocol, in order to have objective

and fair comparison. We are able to achieve the best

classiﬁcation performance in the multi-class setting,

as well as in NP vs P1 and NP vs P2. In the the re-

maining binary classiﬁcation tasks we obtain accept-

able results.

5 CONCLUSION

Automatic pain intensity estimation possess great

value in effective pain management. This paper stud-

ied the ECG biosignals, utilizing the Pan-Tompkins

Table 3: Classiﬁcation results of the Age Scheme, reported

on % accuracy.

Group

Task

Algorithm

LDA SVM LN SVM RBF

20-35

MC 25.06 24.73 21.96

NP vs P1 52.83 52.83 49.90

NP vs P2 54.33 53.75 52.75

NP vs P3 55.58 56.16 54.66

NP vs P4 63.83 63.41 60.75

36-50

MC 23.27 22.06 23.03

NP vs P1 50.34 48.36 50.68

NP vs P2 49.13 51.20 50.17

NP vs P3 58.10 58.70 58.27

NP vs P4 58.10 57.75 55.94

51-65

MC 21.89 22.07 22.35

NP vs P1 52.23 51.87 52.58

NP vs P2 52.14 51.69 52.76

NP vs P3 53.66 53.39 54.10

NP vs P4 54.46 54.19 54.91

Table 4: Classiﬁcation results of the Gender-Age Scheme

(Males), reported on % accuracy.

Group

Task

Algorithm

LDA SVM LN SVM RBF

Males 20-35

MC 23.13 23.20 18.73

NP vs P1 52.50 52.83 45.83

NP vs P2 54.00 53.50 53.16

NP vs P3 56.33 56.50 54.83

NP vs P4 60.00 59.00 53.66

Males 36-50

MC 23.21 22.21 20.92

NP vs P1 50.53 50.53 46.42

NP vs P2 50.00 51.78 47.50

NP vs P3 54.64 56.25 47.32

NP vs P4 55.53 56.25 51.96

Males 51-65

MC 20.06 21.60 19.60

NP vs P1 52.66 51.66 50.66

NP vs P2 54.00 54.66 51.50

NP vs P3 53.00 54.66 51.50

NP vs P4 53.33 54.50 49.83

algorithm for the detection of QRS complexes extract-

ing features related to the inter-beat intervals. Addi-

tionally, we experimented with three machine learn-

ing methods, comparing them in tasks of multi-class

and binary pain classiﬁcation of different pain inten-

sities. Furthermore, we scrutinized the effect of gen-

der and age in pain manifestation, revealing that they

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

160

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Basic Gender Age Gender-Age

Schemes (max scores)

Basic

Gender

Age

Gender-Age

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Gender (max scores)

Males

Females

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Males Females

Age

20-35

36-50

51-65

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

20-35 36-50 51-65

Gender-Age

Males 20-35

Females 20-35

Males 36-50

Females 36-50

Males 51-65

Females 51-65

BL vs P1

BL vs P2

BL vs P3

BL vs P4

BL vs P1

BL vs P2

BL vs P3

BL vs P4

Males 20-35 Females 20-35 Males 36-50 Females 36-50 Males 51-65 Females 51-65

Figure 6: Classiﬁcation results on the Gender-Age Scheme.

Table 5: Classiﬁcation results of the Gender-Age Scheme

(Females), reported on % accuracy.

Group

Task

Algorithm

LDA SVM LN SVM RBF

Females 20-35

MC 24.73 24.80 23.26

NP vs P1 49.83 51.50 52.00

NP vs P2 54.50 53.66 46.50

NP vs P3 53.50 52.83 49.00

NP vs P4 65.83 67.00 62.16

Females 36-50

MC 23.06 22.73 21.93

NP vs P1 48.16 49.33 48.33

NP vs P2 48.66 49.83 47.83

NP vs P3 57.50 60.00 55.00

NP vs P4 59.00 58.83 56.16

Females 51-65

MC 21.23 21.84 23.92

NP vs P1 48.84 49.80 55.38

NP vs P2 51.15 48.65 55.96

NP vs P3 53.07 53.07 50.96

NP vs P4 52.69 55.00 56.34

are major factors directly related to pain perception.

The conducted experiments exhibited great variation

among the genders where the males presented lower

sensitivity, especially in high pain intensities. Regard-

ing to age factor, signiﬁcant variations demonstrated

as well, disclosed that as long the age increases the

pain sensation is diminished, and as consequent the

increased risk for further injury. In particular demo-

graphic groups the difference is over 12%, indicating

the divergence of pain sensation among people. We

suggest that clinical pain assessment tools, need to be

speciﬁcally designed for certain groups related to the

particular demographic factors considering the unique

pain manifestation’ characteristics. Furthermore, we

indicate to researchers who will involve in the cre-

ation of new pain databases, the necessity for the inte-

gration of demographic factors, as well as information

Table 6: Comparison of studies which utilized BioVid,

ECG signals and LOSO cross validation, reported on %

accuracy.

Method Task Results

Lopez-Martinez and

Picard (2018)

NP vs P4 57.69

Werner et al. (2014)

NP vs P1 48.70

NP vs P2 51.60

NP vs P3 56.50

NP vs P4 62.00

Thiam et al. (2019)

MC 23.23

NP vs P1 49.71

NP vs P2 50.72

NP vs P3 52.87

NP vs P4 57.04

Ours

MC 23.79

NP vs P1 52.38

NP vs P2 52.78

NP vs P3 55.37

NP vs P4 58.62

about the social context and the psychological condi-

tions of subjects. Our future work will explore the

utilization of the remaining biosignals of BioVid, ei-

ther in unimodal or in a multimodal approach as well.

REFERENCES

Amirian, M., K

achele, M., and Schwenker, F. (2016). Using

radial basis function neural networks for continuous

and discrete pain estimation from bio-physiological

signals. In Artiﬁcial Neural Networks in Pattern

Recognition, pages 269–284. Springer Verlag.

Bartley, E. J. and Fillingim, R. B. (2013). Sex differences

in pain: a brief review of clinical and experimental

ﬁndings. British journal of anaesthesia, 111(1):52–

58.

Chu, Y., Zhao, X., Han, J., and Su, Y. (2017). Physiological

Automatic Pain Intensity Estimation based on Electrocardiogram and Demographic Factors

161

signal-based method for measurement of pain inten-

sity. Frontiers in Neuroscience, 11(MAY):279.

Cordell, W. H., Keene, K. K., Giles, B. K., Jones, J. B.,

Jones, J. H., and Brizendine, E. J. (2002). The high

prevalence of pain in emergency medical care. Amer-

ican Journal of Emergency Medicine, 20(3):165–169.

Fariha, M. A. Z., Ikeura, R., Hayakawa, S., and Tsutsumi, S.

(2020). Analysis of Pan-Tompkins Algorithm Perfor-

mance with Noisy ECG Signals. Journal of Physics:

Conference Series, 1532(1):12022.

Gaskin, D. J. and Richard, P. (2012). The Economic Costs

of Pain in the United States. The Journal of Pain,

13(8):715–724.

Hadjistavropoulos, T. and Craig, K. D. (2002). A theoretical

framework for understanding self-report and observa-

tional measures of pain: A communications model.

Behaviour Research and Therapy, 40(5):551–570.

Hinduja, S., Canavan, S., and Kaur, G. (2020). Multi-

modal Fusion of Physiological Signals and Facial Ac-

tion Units for Pain Recognition. In Struc V., G.-F. F.,

editor, IEEE International Conference on Automatic

Face and Gesture Recognition, pages 577–581. Insti-

tute of Electrical and Electronics Engineers Inc.

achele, M., Thiam, P., Amirian, M., Schwenker, F.,

and Palm, G. (2016). Methods for Person-Centered

Continuous Pain Intensity Assessment from Bio-

Physiological Channels. IEEE Journal on Selected

Topics in Signal Processing, 10(5):854–864.

Liu, F., Wei, S., Li, Y., Jiang, X., Zhang, Z., Zhang, L., and

Liu, C. (2017). The accuracy on the common Pan-

Tompkins based QRS detection methods through low-

quality electrocardiogram database. Journal of Medi-

cal Imaging and Health Informatics, 7(5):1039–1043.

Lopez-Martinez, D. and Picard, R. (2017). Multi-task neu-

ral networks for personalized pain recognition from

physiological signals. In International Conference on

Affective Computing and Intelligent Interaction Work-

shops and Demos, pages 181–184, San Antonio. Insti-

tute of Electrical and Electronics Engineers Inc.

Lopez-Martinez, D. and Picard, R. (2018). Continuous Pain

Intensity Estimation from Autonomic Signals with

Recurrent Neural Networks. In International Confer-

ence of the IEEE Engineering in Medicine and Biol-

ogy Society., volume 2018, pages 5624–5627.

Lynch, M. E. (2011). The need for a Canadian pain strategy.

Pain Research and Management, 16(2):77–80.

Mamontov, D., Polonskaia, I., Skorokhod, A., Semenkin,

E., Kessler, V., and Schwenker, F. (2019). Evo-

lutionary algorithms for the design of neural net-

work classiﬁers for the classiﬁcation of pain intensity.

Multimodal Pattern Recognition of Social Signals in

Human-Computer-Interaction, 11377 LNAI:84–100.

Merskey, H., Albe-Fessard, D. G., Bonica, J. J., Carmen, A.,

Dubner, R., Kerr, F. W. L., and Pagni, C. A. (1979).

Editorial: The need of a taxonomy. Pain, 6(3):247–

252.

Pan, J. and Tompkins, W. J. (1985). A Real-Time QRS De-

tection Algorithm. IEEE Transactions on Biomedical

Engineering, BME-32(3):230–236.

Rohling, M. L., Binder, L. M., and Langhinrichsen-

Rohling, J. (1995). Money Matters: A Meta-

Analytic Review of the Association Between Finan-

cial Compensation and the Experience and Treatment

of Chronic Pain. Health Psychology, 14(6):537–547.

Stewart, G. and Panickar, A. (2013). Role of the sympa-

thetic nervous system in pain.

Subramaniam, S. D. and Dass, B. (2021). Automated Noci-

ceptive Pain Assessment Using Physiological Signals

and a Hybrid Deep Learning Network. IEEE Sensors

Journal, 21(3):3335–3343.

Thiam, P., Bellmann, P., Kestler, H. A., and Schwenker, F.

(2019). Exploring deep physiological models for no-

ciceptive pain recognition. Sensors, 19(20):4503.

Toomey, M. (2008). Gender differences in pain: does X =

Y? AANA journal, 76(5):355–359.

Walter, S., Gruss, S., Ehleiter, H., Tan, J., Traue, H. C.,

Crawcour, S., Werner, P., Al-Hamadi, A., Andrade,

A. O., and Da Silva, G. M. (2013). The biovid heat

pain database: Data for the advancement and system-

atic validation of an automated pain recognition. In

2013 IEEE International Conference on Cybernetics,

pages 128–131.

Wang, J., Wei, M., Zhang, L., Huang, G., Liang, Z., Li,

L., and Zhang, Z. (2020). An Autoencoder-based

Approach to Predict Subjective Pain Perception from

High-density Evoked EEG Potentials. In International

Conference of the IEEE Engineering in Medicine Bi-

ology Society, pages 1507–1511. Institute of Electrical

and Electronics Engineers Inc.

Werner, P., Al-Hamadi, A., Niese, R., Walter, S., Gruss,

S., and Traue, H. C. (2014). Automatic pain recog-

nition from video and biomedical signals. In Inter-

national Conference on Pattern Recognition, pages

4582–4587. Institute of Electrical and Electronics En-

gineers Inc.

Werner, P., Lopez-Martinez, D., Walter, S., Al-Hamadi, A.,

Gruss, S., and Picard, R. (2019). Automatic Recogni-

tion Methods Supporting Pain Assessment: A Survey.

IEEE Transactions on Affective Computing.

Yu, M., Sun, Y., Zhu, B., Zhu, L., Lin, Y., Tang, X., Guo,

Y., Sun, G., and Dong, M. (2020). Diverse frequency

band-based convolutional neural networks for tonic

cold pain assessment using EEG. Neurocomputing,

378:270–282.

Zhao, K., Li, Y., Wang, G., Pu, Y., and Lian, Y. (2021).

A robust QRS detection and accurate R-peak identiﬁ-

cation algorithm for wearable ECG sensors. Science

China Information Sciences, 64(8):182401.

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

162