Feature Extraction for Stress Detection in Electrodermal Activity

Erika Lutin

1,2 a

, Ryuga Hashimoto

, Walter De Raedt

and Chris Van Hoof

1,2,4 d

Electrical Engineering-ESAT, KU Leuven, Kasteelpark Arenberg 10, Heverlee, Belgium

Imec, Kapeldreef 75, Heverlee, Belgium

Dept. of Mechanical and Intelligent Systems Engineering, The University of Electro-Communications,

1-5-1 Chofugaoka, Chofu, Japan

Imec-Nl, OnePlanet Research Center, Stippeneng 2, Wageningen, The Netherlands

Keywords: Electrodermal Activity (EDA), Feature Extraction, Stress.

Abstract: Electrodermal activity (EDA) is a sensitive measure for changes in the sympathetic system, reflecting

emotional and cognitive states such as stress. There is, however, inconsistency in the recommendations on

which features to extract. In this study, we brought together different feature extraction methods: trough-to-

peak features, decomposition-based features, frequency features and time-frequency features. Regarding the

decomposition analysis, three different applications were used: Ledalab, cvxEDA and sparsEDA. A total of

forty-seven features was extracted from a previously collected dataset. This dataset included twenty

participants performing three different stress tasks. A Support Vector Machine (SVM) classifier was built in

a Leave-One-Subject-Out Cross Validation (LOOCV) set-up with feature selection within the LOOCV loop.

Three features were consistently selected over all participants: 1) the number of responses in the driver

function generated by Ledalab and 2) by sparsEDA and 3) a time-frequency feature, previously described as

TVSymp. The classifier obtained an accuracy of 88.52%, a sensitivity of 72.50% and a specificity of 93.65%.

This research shows that EDA can be successfully used in stress detection, without the addition of any other

physiological signals. The classifier, built with the most recent feature extraction methods in literature, was

found to outperform previous classification attempts.

1 INTRODUCTION

Prolonged exposure to psychological stress in daily

life has been associated to various diseases such as

depression and cardiovascular disease (Cohen et al.,

2007). To prevent these adverse consequences, there

is a strong need for correct quantification of personal

stress (Epel et al., 2018). The body’s response to

stress can be quantified by measuring activation of

the autonomic nervous system (ANS). The ANS

consists of the sympathetic, parasympathetic and

enteric branches. The main function of the

sympathetic nervous system (SNS) is to initiate a

rapid response in case of a dangerous or threatening

situation, the so-called “fight or flight” response

(Lovallo, 2005). As part of this response, the sweat

glands become activated (Critchley, 2002), which

causes a change in the electrical properties of the skin:

https://orcid.org/0000-0002-9254-7374

https://orcid.org/0000-0002-3189-1235

https://orcid.org/0000-0002-7117-7976

https://orcid.org/0000-0002-4645-3326

the skin becomes more conductive. This change in

electrical properties is referred to as Skin

Conductance Response (SCR) or Galvanic Skin

Response (GSR) (Boucsein, 2012). Many researchers

have demonstrated a high correlation between

Electrodermal activity (EDA) and cognitive and

emotional processes (Critchley, 2002). However, in

comparison to other physiological signals, such as an

Electrocardiogram (ECG) (Camm et al., 1996), there

are few guidelines on which features to extract.

Recent reviews indicate multiple approaches but do

not recommend a specific one (Topoglu et al., 2019;

Posada-Quintero & Chon, 2020).

Electrodermal activity generally consists of two

components. The first component is the tonic

component, referred to as Skin Conductance Level

(SCL). It varies slowly and changes only slightly

within tens of seconds to minutes. The second

Lutin, E., Hashimoto, R., De Raedt, W. and Van Hoof, C.

Feature Extraction for Stress Detection in Electrodermal Activity.

DOI: 10.5220/0010244601770185

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 177-185

ISBN: 978-989-758-490-9

177

component is the phasic component, also referred to

as the Skin Conductance Response (SCR). The phasic

component represents a rapid change with a very

short time to onset, usually between 1-5 seconds after

the onset of a stimulus. If a response occurs in the

absence of a stimulus, it is referred to as a non-

specific SCR (NS.SCR) (Boucsein, 2012).

Regarding feature extraction in EDA, research has

been focussed on the extraction of time domain

features. Within the time domain, the focus has been

on the evaluation of single SCRs, rather than patterns

of multiple SCRs. Every SCR shows a characteristic

course, which can be parameterized by features such

as latency time, rise time, peak amplitude and

recovery time (Boucsein, 2012). However, the

evaluation of a single response becomes difficult in

case of overlapping responses or superimposition. In

the last decade, new analysis applications have been

developed, which can handle superimposed SCRs

(Benedek & Kaernbach, 2010b; Ghaderyan &

Abbasi, 2016; Greco et al., 2016; Hernando-Gallego

et al., 2018). These applications model EDA as a

result of discrete bursts of sudomotor nerve activity,

mathematically referred to as the driver function. By

deconvolving the EDA signal into the driver function,

and then convolving the peaks in the driver function

with an SCR shape, these applications decompose an

EDA signal in pure SCRs, independently from SCL

and previous SCRs. Multiple features can be easily

extracted from both the resulting SCL and SCR signal

(Boucsein, 2012; Topoglu et al., 2019). Although

analysis of EDA in the frequency domain has been

described as less relevant (Boucsein, 2012), Posada et

al. (2016a & 2016b) proposed two new frequency-

related features, including one time-variant feature,

which can be used to detect stress.

The purpose of the current study is to extract both

traditionally studied features and recently developed

features and discuss their performance with respect to

stress detection.

2 METHODS

Data Collection

EDA data was collected in a previous study by Smets

et al., (2016). The dataset consisted of twenty healthy

participants (ten males and ten females, mean age =

40 years

േ

10 years), who did not suffer from any

mental or physical disease. The participants were

asked to perform three stress tasks, during which

EDA was recorded at the fingertip with the NeXus 10

– MK II (sampling rate = 32 Hz). Each of the three

stress tasks was carried out for two minutes. The first

task was the Stroop color-word test (Van Der Elst et

al., 2006). During this task, color words are displayed

in a different color as the words represent. For

example, when the word red is presented in blue, the

participant must answer the printed color (blue in this

example). The second task was an arithmetic test in

which the participant had to countdown from 1081 by

subtracting 7 in a serial manner. The final task

included a stress talk, in which participants were

asked to talk about past stressful or emotionally

negative events. In addition to these stress tasks,

participants were instructed to count from zero to

hundred out loud to control for the physiological

response caused by vocalization. The counting task

was performed before the Stroop color-word test and

after the stress talk. All tasks were separated by a

resting period of two minutes. The experiment was

conducted in a quiet and controlled laboratory.

Feature Extraction

In this study, multiple methods for features extraction

were performed: trough-to-peak features,

decomposition-based features, frequency features

and time-frequency features. Whereas time domain

analysis is usually performed in small windows

(about 10s) around a single Skin Conductance

Response (SCR) (Lim et al., 1997), frequency domain

analysis with inclusion of slow changing responses

requires longer processing windows. Therefore, a

window of 64s was selected. The windows were

acquired with 32s overlap from the start point of each

task to an integer multiple of 64 seconds. Pre-

processing, i.e. filtering and downsampling, differed

among the different feature sets, as it was performed

according to the pre-processing procedures in the

original research describing the feature set to be

extracted.

2.2.1 Trough-to-Peak Features

In the first feature extraction method, SCR onset and

latency were estimated based on the trough-to-peak

analysis (Boucsein, 2012). Regarding this analysis,

the procedure by Healey (2000) was adopted. The

data was first pre-processed using a low-pass filter

with a cut-off of 4 Hz. Thereafter, a threshold was

applied on the derivative of the filtered EDA signal.

Crossing the threshold indicated a new response if

this happened more than one second away from other

responses. For every new response, the onset (trough)

and peak were determined as the zero-crossings of the

derivative preceding and following the response

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

178

respectively. Given the onset and peak values, four

features could be extracted: the number of SCRs, the

summed magnitude of the SCRs, the summed

durations of the SCRs and the summed area under the

SCRs.

2.2.2 Decomposition-based Features

The second set of features was extracted following

the decomposition of the EDA signal into its tonic,

phasic and driver components. Three different

MATLAB applications were used: Ledalab (Benedek

& Kaernbach, 2010b), cvxEDA (Greco et al., 2016)

and sparsEDA (Hernando-Gallego et al., 2018).

The first application, Ledalab, provides two

analysis methods to calculate the components:

Discrete deconvolution Analysis (DDA) (Benedek &

Kaernbach, 2010a) and Continuous Deconvolution

Analysis (CDA) (Benedek & Kaernbach, 2010b).

Whereas DDA applies a strictly nonnegative

deconvolution, CDA merely tries to minimize

negativity and therefore facilitates a more robust

analysis. In this study, CDA was chosen as analysis

strategy. The second application, cvxEDA, performs

a nonnegative deconvolution by solving a convex

optimization approach problem (Greco et al., 2016).

The application parameters were defined as follows:

τ_0= 4.0, τ_1= 0.7, α = 0.0008, γ = 0.01. Lastly,

sparsEDA also performs a non-negative

deconvolution, though with specific focus on the

sparse nature of the driver components (Hernando-

Gallego et al., 2018). The application parameters

were defined as follows: ε = 0.0001, K_max = 40

iterations, N

min

= 5/4 fs = 10 samples, ρ = 0.025.

These applications were selected above others as

they do not require an a priori specification of events,

whereas PsPM, another application by Bach et al.,

(2011) relies on information of the presented stimuli

(Kelsey et al., 2018).

Following the pre-processing procedures

described in Hernando-Gallego et al., (2018), the

EDA data was first downsampled to 8 Hz. All three

decomposition methods were applied on the complete

signal, resulting in a continuous tonic, phasic and

driver component for every MATLAB application.

The sparsity of the driver components in both Ledalab

and cvxEDA were increased post-extraction by

applying thresholds of 0.1 and 0.01 respectively.

From each of the components, statistical features, i.e.

the mean, maximum, minimum and standard

deviation, were extracted in windows of 64 seconds,

resulting in a set of twelve features per application.

Four additional features were extracted regarding the

driver component: firstly, the number of responses,

i.e. the number of non-zero elements after

thresholding and secondly, the number of (inactive)

intervals between responses, together with their mean

and maximum length.

2.2.3 Frequency-based Features

The third set of features was derived from the work

of Posada-Quintero et al. (2016a), who expressed a

specific interest in sympathetic tone. The sympathetic

tone has been described by time domain features such

as the SCL and the NS.SCRs. These time domain

features are, however, highly variable between

persons. Therefore, Posada et al. proposed a new

frequency-domain approach. Starting from the low

frequency (LF) range for heart rate variability, they

tested whether stressful tasks resulted in a spectral

peak between 0.045 Hz and 0.15 Hz. Their results

indicated a broader range of 0.045Hz to 0.25 Hz.

Based on these results, they described a new feature

called EDASymp, which represents the spectral

power in this specific frequency band, calculated in

windows of two minutes using Welch’s periodogram

(Blackman window of 128 points) with 50% data

overlap. EDASympn is the normalized adaptation in

which the power of the frequency band is divided by

the total power. In this study, we examined the

occurrence of a spectral peak during stress tasks and

calculated continuous features based on EDASymp

and EDASympn.

Preprocessing was performed in accordance with

Posada-Quintero et al., (2016a). First, an 8th order

Chebyshev Type I low-pass filter (0.8 Hz) was

applied, followed by down-sampling to 2 Hz and

finally, an 8th order Butterworth high-pass filter (0.01

Hz).

The spectral peak was examined, per task, in the

following frequency ranges: VLF = 0-0.045 Hz, LF =

0.045-0.15 Hz, HF1 = 0.15-0.25 Hz, HF2 = 0.25-0.4

Hz and VHF = 0.4-0.5 Hz (Posada-Quintero et al.,

2016a). The PSD was obtained via Welch’s

periodogram as described above.

The continuous frequency domain features,

relative to EDASymp and EDASympn, were

calculated for every 64 seconds window within the

tasks. A Blackman window of 128 samples was

applied to each window. The PSD was obtained via

the Fast Fourier Transform (FFT).

2.2.4 Time-frequency-based Features

The final feature was also derived from Posada-

Quintero et al. In addition to invariant frequency-

domain analysis, Posada-Quintero et al., (2016b)

proposed a time-variant index of sympathetic tone,

Feature Extraction for Stress Detection in Electrodermal Activity

179

called TVSymp. The time-frequency representation

(TFR) of EDA was computed using the variable

frequency complex demodulation (VFCDM), a time-

frequency spectral (TFS) analysis technique that

provides accurate amplitude estimates and one of the

highest time-frequency resolutions (Wang et al.,

2006). The components comprising the frequency

power in the range from 0.08 to 0.24 Hz were used to

compute TVSymp.

In this study, the raw EDA was first downsampled

from 32Hz to 2Hz and thereafter high-pass filtered

using a 2nd-order Butterworth filter (0.01 Hz).

TVSymp was calculated by adapting the procedures

described in Wang et al., (2006) and Posada-Quintero

et al., (2016b). The resulting TVSymp was averaged

per window of 64 seconds.

Classification

A Support Vector Machine (SVM) with a Radial

Basis Function (RBF) kernel was used to classify rest

(baseline, relaxing and counting) and stress (Stroop

color-word test, arithmetic task, and stress talk). All

features were standardized before training the

classifier. As the sample size was rather limited,

Leave-one-subject-out cross-validation (LOOCV)

was used to evaluate the performan54ce of the

classifier. Feature selection was performed within the

LOOCV loop resulting in 20 sets of selected features.

All features were ranked according to their

correlation (point-biserial correlation) with the binary

target variable (Hall, 1999) indicating the occurrence

of either a resting task or a stress task. Only features

with a correlation over 0.5 were retained. The

remaining features were compared based on their

correlation among each other. If features had a

correlation over 0.85, only the feature with the

highest target correlation was retained. Accuracy,

sensitivity, specificity, F1-score and precision were

calculated and averaged over all 20 cross-validation

folds.

3 RESULTS

3.1 Feature Extraction

Figure 1 displays the tonic component of one

participant as computed by the three applications:

Ledalab, cvxEDA and sparsEDA. Similarly, the

phasic components are shown in Figure 2, and the

driver components in Figure 3. The light-yellow areas

represent the counting tasks, the light-red areas

represent the stress tasks. Figure 2 only displays the

phasic components as computed by Ledalab and

cvxEDA, since SparsEDA does not provide tools for

the calculation of the phasic component. Ledalab and

cvxEDA show a comparable decomposition, whereas

sparsEDA illustrates a different approach. The driver

function by sparsEDA is much sparser than the ones

by Ledalab and cvxEDA. The tonic component, based

on this sparse driver function, however, deviates from

the true tonic component as it crosses the raw data

Figure 1: Tonic component of one participant. Light yellow

areas indicate a counting task, light red areas a stress task.

Figure 2: Phasic component of one participant. Light

yellow areas indicate a counting task, light red areas a stress

task.

Figure 3: Driver component of one participant. Light

yellow areas indicate a counting task, light red areas a stress

task.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

180

Table 1 summarizes the average calculation time

for each application on the same machine, together

with the standard deviation. The results show that

sparsEDA has the lowest computation time, and

Ledalab the highest.

Table 1: Computation time per application.

Application Time

Ledalab 11.16 ±3.28

cvxEDA 2.40 ±0.45

sparsEDA 0.53 ±0.13

The results on the presence of a stress-related

spectral peak are reported in Table 2. Five frequency

ranges were examined: VLF (0-0.045 Hz), LF (0.045-

0.15 Hz), HF1 (0.15-0.25 Hz), HF2(0.25-0.4 Hz) and

VHF (0.4-0.5 Hz). The distribution of spectral energy

is presented per task. Table 3 shows the values of the

adapted TVSymp per task. The Stroop color-word

test resulted on average in the highest value.

3.2 Classification

Table 4 presents the model performance of the SVM

classifier. Table 5 shows the features selected in the

LOOCV procedure.

In the group of selected features, features of both

Ledalab and sparsEDA appeared. Post-classification

analysis showed that limiting the decomposition

features to features originating from either Ledalab or

sparsEDA, lowered sensitivity but increased specifity

further (SVM-spars; accuracy = 88.65%, sensitivity =

68.33%, specificity = 95.10%, F1 = 71.36%, SVM-

Led; accuracy = 88.40%, sensitivity = 70.83%,

specificity = 94.10%, F1=72.22%). When the same

model was built with decomposition features

originating only from cvxEDA, both sensitivity and

specificity were lowered (SVM-cvx; accuracy =

86.67%, sensitivity = 69.17%, specificity = 93.27%,

F1=70.21%)

Table 5: Features selected in LOOCV classification.

Feature Votes

TVSymp (~ Posada-Quintero et al.) 20

Number of responses - Driver Ledalab 20

Number of responses - Driver sparsEDA 20

Summed duration of SCRs (~ Healey et al.) 4

4 DISCUSSION

The purpose of the current study was to extract both

traditionally studied features and recently developed

features and discuss their behavior as well as their

performance in a simple classifier.

The first feature extraction method corresponded

to the traditionally applied trough-to-peak analysis. In

this analysis, SCRs are simply extracted based on

zero-crossings of the derivative, while maintaining a

minimal spacing between SCRs of 1 second. In four

folds of the LOOCV, the summed duration of the

SCRs was retained as an important feature. This

finding agrees with the wide use of these features and

Table 2: Percentage of energy within the five frequency ranges presented per task.

Range Baseline (%) Relax (%) Count (%) Stroop (%) Arithm. (%) Talk (%)

VLF 70.27 ± 26.23 83.48 ± 17.04 59.32 ± 20.33 53.50 ± 20.82 53.59 ± 19.23 61.27 ± 19.10

LF 31.30 ± 23.09 17.68 ± 14.35 39.03 ±15.31 36.41 ± 15.45 42.06 ± 16.25 40.01 ± 16.04

HF1 5.43 ± 6.90 4.12 ± 5.80 10.42 ± 6.95 18.46 ± 15.17 13.27 ± 8.71 8.42 ± 7.11

HF2 0.66 ± 1.28 0.42 ± 0.76 1.33 ± 1.56 1.69 ± 1.85 1.80 ± 1.70 1.11 ± 0.92

VHF 0.11 ± 0.19 0.12 ± 0.24 0.49 ± 0.73 0.85 ± 1.10 0.61 ± 0.72 0.39 ± 0.48

Table 3: TVSymp results per task.

Baseline Relax Count Stroop Arithmetic Talk

TVSymp

0.59 ± 0.51 0.54 ± 0.16 1.03 ± 0.34 1.69 ± 0.46 1.62 ±0.46 1.51 ± 0.51

Table 4: Model performance.

Accuracy (%) Sensitivity (%) Specificity (%) F1 (%) Precision (%)

SVM 88.52 ± 10.99 72.50 ± 30.24 93.65 ± 9.60 72.84 ± 28.96 81.20 ± 26.37

Feature Extraction for Stress Detection in Electrodermal Activity

181

the high accuracy (96%) in driver stress detection

originally obtained by Healey (2000), but is contrary

to the findings of Shukla et al., (2019), who reported

low weighted occurence of these features in

comparison to others in an emotion recognition task.

Although this method is easy to implement and

could be used in a real-time setting, it is limited in the

analysis of successive SCRs. They will most likely

appear as one response (Benedek & Kaernbach,

2010b), which will lead to an overall underestimation

of the number of responses.

To overcome the limitation of superimposed

SCRs, we included the decomposition of the EDA

signal as a second feature extraction method. We

compared three decomposition methods. SparsEDA

gave rise to the sparsest driver, which was claimed by

the original authors to improve interpretability while

reducing computation cost at the same time

(Hernando-Gallego et al., 2018). This is, however, at

the cost of an accurate tonic component as can be seen

in Figure 1. Moreover, Amin & Faghih (2019)

reported sparsEDA as oversparsifying the driver

function. Ledalab and cvxEDA both gave rise to a

more accurate tonic component, though at a much

slower rate. In this study, it was argued that the

sparsity and thus the interpretability of the driver

functions of Ledalab and cvxEDA could easily be

improved by introducing a threshold. Figure 3

illustrates that, nevertheless, sparsEDA remained the

sparsest. Two decomposition-related features were

retained in the LOOCV. Surprisingly, these were the

number of driver responses of both sparsEDA and

Ledalab. This suggests there is information in both

sparsity and continuity. Limiting the classification to

one single decomposition method showed that the

sparisity of sparsEDA results in a higher specificity

as it decreases the presence of drivers in resting

periods, while the continuity of Ledalab results in a

more robust sensitivity as it increases the presence of

drivers in stress periods. Ledalab might be

interchanged with cvxEDA to reduce time, though at

cost of accuracy.

The third feature extraction method was linked to

the frequency domain. We first tried to replicate the

results of Posada-Quintero et al., (2016a), who

investigated the spectral energy in multiple frequency

ranges during four conditions: baseline, postural

stimulation, a cold pressor and a Stroop color-word

test. Only the latter corresponded to the protocol of

the current study. Table 2 shows for all tasks the

distribution in spectral energy. In this study,

participants were fully at rest during the relaxation

task and not during the baseline. Therefore, the results

of relaxing resemble the results of the baseline in

Posada-Quintero et al., (2016a) (VLF between 79.2%

and 87.3%, LF between 8.1% and 14.6%, HF1

between 1.2% and 3.7%) more strongly. The

distribution of spectral energy within the Stroop

color-word test is also comparable to the one of

Posada-Quintero et al., (2016a). The results show

again the same trend (VLF 51.6%, LF 32.9%, HF1

10.7% in Posada-Quintero et al., 2016a). As

described by Posada-Quintero et al., (2016a), during

a stressor, the spectral energy in the VLF frequency

range goes down, whereas in the other ranges the

energy goes up. Remarkably, this trend is present in a

comparable magnitude within the counting tasks as

within the stress talk task. As the LF and HF1 ranges

were confirmed as most interesting ranges,

EDASymp and EDASympn were calculated

according to Posada-Quintero et al., (2016a). They

were not retained as important features during the

feature selection.

The final feature to be extracted was an adaptation

of TVSymp, a new time-frequency-based feature.

Again, the presented results (Table 3) show the same

trend as the data of Posada-Quintero et al. (2016b),

i.e. the values of TVSymp go up in case of a stressor.

The absolute numbers are comparable as well

(baseline: ~ 0.2-0.5, stressor: ~ 1.4-1.6) . TVSymp

was confirmed to be a highly sensitive index for stress

detection in EDA, as it was selected in all folds of the

cross validation. This in line with the conclusion of

Ghaderyan & Abbasi (2016), who reported time-

frequency domain features as highly performant in a

mental workload classification task. Whereas in

TVSymp the time-frequency representation is

obtained by VFCDM, Ghaderyan & Abbasi (2016)

obtained it via wavelets. In addition to time-

frequency features, Ghaderyan & Abbasi (2016)

extracted decomposition features via cepstral

analysis. These were found to perform equally well as

the wavelet features. The latter is contrary to the work

of Shukla et al. (2019) who reported that cepstrum-

based features outperformed wavelet-related features.

Classification using an SVM resulted in 88.52%

accuracy, 72.50% sensitivity and 93.65% specificity.

These results closely resemble the performance of an

earlier machine learning exercise on this dataset

which gave 72.0% sensitivity and 93.4% specificity

(Smets et al., 2016) while this exercise included heart

rate features on top of EDA features.

Previous studies have focused largely on datasets

including multiple physiological signals. Only few

studies have used EDA as a sole predictor for stress.

Kurniawan et al., (2013) obtained 80.72% accuracy

in classifying the Stroop color-word test from

recovery with statistical features and trough-to-peak

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

182

features. Liu & Du (2018) got an average accuracy of

81.82% in a three-level stress detection task on

driving stress with statistical features. While the

results of Kurniawan et al. (2013) and Liu & Du

(2018) present a lower accuracy as reported in this

study, Zangróniz et al., (2017) achieved an accuracy,

sensitivity and specificity as high as 89.18%, 93.90%

and 85.36% respectively using statistical features

(directly derived from the EDA signal, i.e. without

decomposition) and morphological features. The best

performing parameter in a single-parameter classifier

was the spectral power in bandwidth 0.2 Hz to 0.3 Hz,

this parameter outperformed the parameter based on

the spectral power in bandwidth 0.1 Hz to 0.2 Hz.

This result differs from the bands suggested by

Posada-Quintero et al., (2016a), which were

confirmed in this study. Nevertheless, EDASymp was

not retained in any of the feature selection folds.

A possible explanation for the difference in

performance might be the experimental design of

Zangróniz et al., (2017), as it was rather different

from the one described in this paper. In the study by

Zangróniz et al., (2017) calmness and distress were

elicited by pictures of the International Affective

Picture System (IAPS). This design might have

favoured a well-balanced dataset without interference

from vocalization, which is known to effect EDA

(Levenson, 2014). Grimley et al., (2019)

demonstrated that 36-78% of stress responses

involving vocalizations are solely attributed to

vocalizations. This is also apparent from the results

presented in this study: the counting task shows

highly increased EDA (Figures 1 – 3, Tables 2-3),

while it is assumed to be free of stress. Since daily life

measurements will often include vocalizations it is

important to take this into account when building a

classifier.

There are some limitations to this research.

Firstly, the current study included twenty

participants, which is rather limited for the multitude

of features examined in this work. In future work,

EDA data of more participants should be included.

Secondly, the current study included EDA data

collected at the fingertip in a controlled environment.

However, for future purposes, it would be more

relevant to include data collected at the wrist in an

ambulatory setting. This type of data would require

more intensive pre-processing related to artefact

removal.

Lastly, signal transformations such as

deconvolution and VFCDM were performed prior to

windowing, on the complete signal. This is consistent

with prior research in which these transformations

were performed on complete tasks or experiments

(Bobade & Vani, 2020; Murugappan et al., 2020;

Posada-Quintero & Bolkhovsky, 2019; Posada-

Quintero et al., 2018). However, for future purposes

such as continuous or real-time feature extraction, the

effect of performing these transformations within

windows as short as 64 seconds should be explored.

The application of windows prior to transformation

will abrupt ongoing responses which might introduce

noise into the feature extraction. In light of this

potential noise introduction, different window sizes

should be explored.

5 CONCLUSION

The aim of the present study was to assess different

feature extraction methods for stress detection in

EDA. An SVM classifier was built in a Leave-One-

Subject-Out Cross Validation (LOOCV) set-up with

feature selection within the LOOCV loop.

Decomposition-derived features and time-frequency

features were found to be most relevant. The resulting

classifier obtained an accuracy of 88.52%, a

sensitivity of 72.50% and a specificity of 93.65%.

Therefore, by including novel features, we could

outperform an earlier classification attempt.

The research shows that EDA can be successfully

used as a sole predictor for stress when using the most

recent features in literature. In future research, the

presented work should be repeated in a dataset

collected in ambulatory settings. In addition, a

continuous mode of feature extraction should be

envisioned, which requires signal transformations

such as decomposition to be performed in windows

instead of complete signals. Related to this, different

window sizes need to be explored.

ACKNOWLEDGEMENTS

This work was supported by a PhD fellowship from

the Research Foundation - Flanders (FWO) awarded

to EL (1SB4719N).

REFERENCES

Amin, M. R., & Faghih, R. T. (2019). Sparse deconvolution

of electrodermal activity via continuous-time system

identification. IEEE Transactions on Biomedical

Engineering, 66(9), 2585–2595.

Bach, D. R., Daunizeau, J., Kuelzow, N., Friston, K. J., &

Dolan, R. J. (2011). Dynamic causal modelling of

spontaneous fluctuations in skin conductance.

Feature Extraction for Stress Detection in Electrodermal Activity

183

Psychophysiology, 48(2), 252–257.

https://doi.org/10.1111/j.1469-8986.2010.01052.x

Benedek, M., & Kaernbach, C. (2010a). Decomposition of

skin conductance data by means of nonnegative

deconvolution. Psychophysiology, 47(4), 647–658.

Benedek, M., & Kaernbach, C. (2010b). A continuous

measure of phasic electrodermal activity. Journal of

Neuroscience Methods, 190(1), 80–91.

Bobade, P., & Vani, M. (2020). Stress Detection with

Machine Learning and Deep Learning using

Multimodal Physiological Data. In 2020 Second

International Conference on Inventive Research in

Computing Applications (ICIRCA) (pp. 51–57). IEEE.

Boucsein, W. (2012). Electrodermal activity. Springer

Science & Business Media.

Camm, A. J., Malik, M., Bigger, J. T., Breithardt, G.,

Cerutti, S., Cohen, R. J., … Kleiger, R. E. (1996). Heart

rate variability: standards of measurement,

physiological interpretation and clinical use. Task

Force of the European Society of Cardiology and the

North American Society of Pacing and

Electrophysiology.

Cohen, S., Janicki-Deverts, D., & Miller, G. E. (2007).

Psychological Stress and Disease. JAMA, 298(14),

1685–1687. https://doi.org/10.1001/jama.298.14.1685

Critchley, H. D. (2002). Electrodermal responses: What

happens in the brain. Neuroscientist, 8(2), 132–142.

https://doi.org/10.1177/107385840200800209

Epel, E. S., Crosswell, A. D., Mayer, S. E., Prather, A. A.,

Slavich, G. M., Puterman, E., & Mendes, W. B. (2018).

More than a feeling: A unified view of stress

measurement for population science. Frontiers in

Neuroendocrinology, 49(March), 146–169.

https://doi.org/10.1016/j.yfrne.2018.03.001

Ghaderyan, P., & Abbasi, A. (2016). An efficient automatic

workload estimation method based on electrodermal

activity using pattern classifier combinations.

International Journal of Psychophysiology, 110, 91–

101.

Greco, A., Valenza, G., Lanata, A., Scilingo, E. P., & Citi,

L. (2016). cvxEDA: A convex optimization approach to

electrodermal activity processing. IEEE Transactions

on Biomedical Engineering, 63(4), 797–804.

Grimley, S. J., Ko, C. M., Morrell, H. E. R., Grace, F.,

Bañuelos, M. S., Bautista, B. R., … Gurning, J. (2018).

The Need for a Neutral Speaking Period in

Psychosocial Stress Testing. Journal of

Psychophysiology.

Hall, M. A. (1999). Correlation-based feature selection for

machine learning.

Healey, J. A. (2000). Wearable and automotive systems for

affect recognition from physiology. Massachusetts

Institute of Technology.

Hernando-Gallego, F., Luengo, D., & Artés-Rodríguez, A.

(2017). Feature extraction of galvanic skin responses by

nonnegative sparse deconvolution. IEEE Journal of

Biomedical and Health Informatics, 22(5), 1385–1394.

Kelsey, M., Akcakaya, M., Kleckner, I. R., Palumbo, R. V.,

Barrett, L. F., Quigley, K. S., & Goodwin, M. S. (2018).

Applications of sparse recovery and dictionary learning

to enhance analysis of ambulatory electrodermal

activity data. Biomedical Signal Processing and

Control, 40, 58–70.

Kurniawan, H., Maslov, A. V, & Pechenizkiy, M. (2013).

Stress detection from speech and galvanic skin response

signals. In Proceedings of the 26th IEEE International

Symposium on Computer-Based Medical Systems (pp.

209–214). IEEE.

Levenson, R. W. (2014). The autonomic nervous system

and emotion. Emotion Review, 6(2), 100–112.

https://doi.org/10.1177/1754073913512003

Lim, C. L., Rennie, C., Barry, R. J., Bahramali, H., Lazzaro,

I., Manor, B., & Gordon, E. (1997). Decomposing skin

conductance into tonic and phasic components.

International Journal of Psychophysiology, 25(2), 97–

109. https://doi.org/https://doi.org/10.1016/S0167-

8760(96)00713-1

Liu, Y., & Du, S. (2018). Psychological stress level

detection based on electrodermal activity. Behavioural

Brain Research, 341, 50–53.

Lovallo, W. R. (2015). Stress and Health: Biological and

Psychological Interactions. SAGE Publications.

Retrieved from

https://books.google.be/books?id=kXtZDwAAQBAJ

Murugappan, R., Bosco, J. J., Eswaran, K., Vijay, P., &

Vijayaraghavan, V. (2020). User Independent Human

Stress Detection. In 2020 IEEE 10th International

Conference on Intelligent Systems (IS) (pp. 490–497).

IEEE.

Posada-Quintero, H. F., & Bolkhovsky, J. B. (2019).

Machine learning models for the identification of

cognitive tasks using autonomic reactions from heart

rate variability and electrodermal activity. Behavioral

Sciences, 9(4), 45.

Posada-Quintero, H. F., & Chon, K. H. (2020). Innovations

in electrodermal activity data collection and signal

processing: A systematic review. Sensors, 20(2), 479.

Posada-Quintero, H. F., Florian, J. P., Orjuela-Cañón, A.

D., Aljama-Corrales, T., Charleston-Villalobos, S., &

Chon, K. H. (2016a). Power spectral density analysis of

electrodermal activity for sympathetic function

assessment. Annals of Biomedical Engineering, 44(10),

3124–3135.

Posada-Quintero, H. F., Florian, J. P., Orjuela-Cañón, A.

D., & Chon, K. H. (2018). Electrodermal activity is

sensitive to cognitive stress under water. Frontiers in

Physiology, 8, 1128.

Posada-Quintero, H. F., Florian, J. P., Orjuela-Cañón, Á.

D., & Chon, K. H. (2016b). Highly sensitive index of

sympathetic activity based on time-frequency spectral

analysis of electrodermal activity. American Journal of

Physiology-Regulatory, Integrative and Comparative

Physiology, 311(3), R582–R591.

Shukla, J., Barreda-Angeles, M., Oliver, J., Nandi, G. C., &

Puig, D. (2019). Feature extraction and selection for

emotion recognition from electrodermal activity. IEEE

Transactions on Affective Computing.

Smets, E., Casale, P., Großekathöfer, U., Lamichhane, B.,

De Raedt, W., Bogaerts, K., … Van Hoof, C. (2015).

Comparison of machine learning techniques for

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

184

psychophysiological stress detection. In International

Symposium on Pervasive Computing Paradigms for

Mental Health (pp. 13–22). Springer.

Topoglu, Y., Watson, J., Suri, R., & Ayaz, H. (2019).

Electrodermal Activity in Ambulatory Settings: A

Narrative Review of Literature. In International

Conference on Applied Human Factors and

Ergonomics (pp. 91–102). Springer.

Van Der Elst, W., Van Boxtel, M. P. J., Van Breukelen, G.

J. P., & Jolles, J. (2006). The stroop color-word test:

Influence of age, sex, and education; and normative

data for a large sample across the adult age range.

Assessment, 13(1), 62–79.

https://doi.org/10.1177/1073191105283427

Wang, H., Siu, K., Ju, K., & Chon, K. H. (2006). A High

Resolution Approach to Estimating Time-Frequency

Spectra and Their Amplitudes. Annals of Biomedical

Engineering, 34(2), 326–338.

https://doi.org/10.1007/s10439-005-9035-y

Zangróniz, R., Martínez-Rodrigo, A., Pastor, J. M., López,

M. T., & Fernández-Caballero, A. (2017).

Electrodermal activity sensor for Classification of Calm

/ Distress Condition, 1–14.

https://doi.org/10.3390/s17102324.

Feature Extraction for Stress Detection in Electrodermal Activity

185