Feature Selection Framework for XGBoost Based

on Electrodermal Activity in Stress Detection

Cheng-Ping Hsieh1 , Yi-Ta Chen2 , Win-Ken Beh2 , and An-Yeu (Andy) Wu2 , Fellow, IEEE
1
Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan
2
Graduate Institute of Electronics Engineering, National Taiwan University, Taipei, Taiwan
1
b04901020@ntu.edu.tw, 2 {edan, kane, andywu}@access.ee.ntu.edu.tw

Abstract—Since stress has a strong influence on human’s

health, it is necessary to automatically detect stress in our
daily life. In this paper, we aim to improve the performance
and obtain the dominant features in stress detection based
on Electrodermal Activity (EDA). Compared to the methods
in Wearable Stress and Affect Dataset (WESAD), we propose
several enhancements to get higher f1-scores, including less over-
lapped signal segmentation, more signal processing features, and
extreme gradient boosting classification algorithm (XGBoost).
Furthermore, we select dominant features according to their
importance in classifier and correlation among other features
while keeping high performance. Experiment results show that
with 9 dominant features in XGBoost, we can achieve 92.38%
(+17.87%) and 89.92% (+14.58%) f1-scores compared to WESAD
on chest- and wrist-based EDA signal respectively. The features
we choose suggest that the magnitude of low frequency and
the complexity of high frequency EDA signal contain the most
significant information in stress detection.
Index Terms—Stress detection, feature selection, electrodermal
activity(EDA), signal processing, extreme gradient boosting
Fig. 1. Overall view of the processing flow. Black text: original methods used
I. I NTRODUCTION in WESAD [5]. Red text: several modifications and methods we proposed.
There is a strong link between the stress and diseases while (baseline, stressed, amusement), was the newest publicly avail-
pro-longed stress can cause disorders on the psychological able benchmark to the best of our knowledge. The baseline
and physiological functions [1]. Therefore, automated stress and amusement states can be combined to a non-stress class,
detection is particularly significant, which can enable people to which was the first time to compare stress with other affect
better manage their health. Lots of studies have been proposed stimuli. The dataset had collected EDA signals from chest and
to identify stress, such as facial expression or conversation wrist sensor and built a stress detection model with a total of
speech [2]-[3]. However, these methods may arise privacy- 14 features and 133,000 segments.
intrusive problems while human inclines to hide their true However, the works based on their achievement still have
feelings. In contrast, physiological measurements collecting room for improvement with some modifications, as shown in
from wearable devices cannot be controlled by conscious and Fig. 1. In this paper, we first modify the overlapped proportion
have similar properties among different subjects. of segments considering the influence of over-fitting [8]. After
Among all the physiological signals, Electrodermal Activity then, lots of features in time, frequency, entropy and wavelet
(EDA) is the most sensitive to stress level due to the high domain will be extracted for better analysis on EDA signal [9].
correlation between EDA and sympathetic nervous system To strengthen the classification algorithms, we introduce XG-
[4]. Moreover, it can be obtained by several sensors on the Boost [10] as our learning model, which has been supported
skin, e.g. fingers, foot, wrist or chest [5]-[6], which is easily helpful in emotion recognition [11].
integrated with a smart device. EDA signal consists of two According to the above enhancement, we can improve the
components: a tonic (skin conductance level, SCL) represents performance compared to previous results. However, in order
a slowly varying baseline conductivity; a phasic (skin conduc- to build an effective and efficient stress detection system, we
tance response, SCR) shows a fast varying reaction to specific have to reduce the load of models with the same performance.
arousing stimulus and can be visible as bursts or peaks. We propose feature selection methods with feature importance
Although several datasets have applied EDA signal for stress and correlation to acquire the most dominant features related
detection [6]-[7], previous experiments commonly regarded to stress and verify these features with various analysis.
stress condition as neutral and stressed state. WESAD [5], There are two main contributions in our study. 1) We have
a multimodal dataset containing three different affect states modified several part of WESAD data processing flow. The
This research was supported in part by the Ministry of Science and
Technology of Taiwan (MOST 107-2622-8-002-009-TA, MOST 108-2622-
8-002-012-TA).
experiment result suggested that the modified processing flow
is more effective for stress detection. 2) We have proposed
two feature selection methods considering feature importance
and correlation. The experiment results suggest our methods
is capable of selecting useful, complementary features. By
elimination most of redundant features, our system yield the
same higher classification results with less computation effort.
II. R ELATED W ORK
In recent years, stress detection based on psychological
sensors has been a hot research topic. Several studies had
Fig. 2. WESAD experiment setup on EDA signal. Grey: baseline. Blue: stress.
worked hard only using EDA signals. Setz et al. [12] discrim- Pink: amusement. White: transient.
inated stress from cognitive load in a laboratory environment.
They extracted 16 features in time domain and analyzed each iteration based on a kernel estimator. With RFE, we can
their performance with three common learning algorithms, find the optimal features based on a specific size to achieve the
e.g. linear discriminant analysis (LDA), k-nearest neighbors best result. However, RFE may take more computation costs
(KNN), and support vector machine (SVM). Hindra et al. [13] during training, so we aim to propose more efficient feature
conducted experiments also in controlled laboratory settings selection methods for XGBoost in this paper.
and used 9 statistical features into different classification mod-
els, e.g. k-means, gaussian mixture model (GMM), decision III. DATA P ROCESSING
tree (DT), and SVM. Both of the above research investigated In our research, we first utilise the “chest” and “wrist”
simple features with complex models to acquire higher per- EDA raw signal from WESAD. Next, following the classic
formance, which is lack of other features in different signal data processing flow, the signal will be split into segments
representation. Since EDA may be varied from individual or after undergoing some preprocessing techniques such as down-
environment, it is necessary to increase the diversity of signal sampling and filtering to remove the artifacts or noise. After
features for a more comprehensive detection system. then, more signal information will be acquired from EDA
According to the feature set in WESAD [5], 14 features are decomposition into tonic and phasic components with peak
calculated to characterize stress responses in Table I. Besides detection. Further details of each step are introduced as below.
the basic statistical features of EDA signal (e.g. mean, standard
deviation, min, max), they had extracted more features consid- A. Dataset
ering tonic and phasic decomposition. However, there are still WESAD [5] (WEarable Stress and Affect Dataset) includes
more signal processing techniques can be applied. Healey et data of subjects experiencing both an emotional and a stress
al. [6] focused on the information from peak response, Posada- stimulus. The data collection was conducted with 15 subjects
Quintero et al. [14] proposed power spectral density analysis (3 female) in a laboratory setting. As shown in Fig. 2, each
in different frequency bands, and Sara et al., [15] suggested subject experienced three main affect conditions: “baseline”
that wavelet transforms for EDA have been successfully used (neutral reading task), “amusement” (watching a set of funny
in several noise reduction. Furthermore, Visnovcova et al. video clips), and “stress” (being exposed to the TSST [17]).
[16] indicated that entropy features would increase obviously The dataset recorded physiological and motion data from
during stress tests and decrease during recovery periods. both a chest- and a wrist-worn device. The sensor modali-
Considering different signal processing methods, EDA can ties included: electrodermal activity (EDA), electrocardiogram
be extracted into various useful signal features. Due to high (ECG), blood volume pulse (BVP), electromyogram (EMG),
dimension of the features, it is necessary to select and verify respiration (RESP), skin temperature (TEMP), and 3-axis
the most dominant ones for stress detection. Recursive Feature accelerometer (ACC). In our study, we combine baseline and
Elimination (RFE) has been widely used for feature selection, amusement states as a non-stress class to compare with stress
which is a recursive method to remove redundant features in class in a binary classification task. The amusement state was
proposed to be a positive stimuli for non-stress class, which
TABLE I is different from only containing baseline state. Among the
WESAD F EATURE E XTRACTION [5]
physiological signals in WESAD, we focus on EDA signal.
Feature Description
µEDA , σEDA mean and std of the EDA B. Pre-processing
minEDA , maxEDA min and max of the EDA
∂EDA , rangeEDA slope and range of the EDA As dataset description, EDA raw signal of the chest- and
µtonic , σtonic , σphasic mean and std of the tonic/phasic wrist-worn device was recorded at sampling rate 700Hz and 4
corr(tonic, time) correlation between tonic and time Hz respectively. The raw signals may contain lots of noise due
numberphasic number of peaks
Pamp Ptime sum of phasic startle magnitudes and to power lines interferences, movements, faulty connections
phasic , phasic response duration etc. Therefore, we downsample the chest-based EDA signal
R
phasic area under the responses into 4Hz like wrist signal and apply a 2Hz low-pass filter [9]
 be fed into a XGBoost machine learning model and output
prediction of stress states, which XGBoost is the first attempt
to utilize in stress detection to our best knowledge. Finally, we
propose an feature selection method using feature importance
and correlation in XGBoost to determine the most dominant
features for stress detection based on Electrodermal Activity.

A. Proposed Signal Segmentation

The choices of an appropriate window size and overlapped
proportion are crucial and depend on several aspects, such as
the classification task or the sensor modalities. According to
WESAD, segmentation of the preprocessed EDA signals were
Fig. 3. Data processing flow of EDA signal. Preprocessing and decomposition done using a 60 seconds sliding window with a highly over-
in the specific segmentation data. Green line: tonic component. Light Blue lapped shift of 0.25 seconds, which would generate 133,000
line: phasic component. Red point: peak. yellow point: onset.
segments. However, lots of similar data may lead to over-fitting
to all EDA signal since physiological plausible changes in the when training a classification model. Therefore, we simply
EDA signal are commonly in the low-frequency domain. increase the shift size from 0.25 seconds to 30 seconds, half of
each segment size for less overlapped segments. Following our
C. Segmentation proposed segmentation methods, we extract the same feature
The data considered in this paper amount to approximately set in WESAD and use the same classification algorithm to
36 minutes per subject. Among these long time experiments, compare the effects of signal segmentation.
we implement segmentation to extract different data samples.
We segment the samples with 60 seconds window size and B. Proposed Feature Extraction
overlap the window in 30 seconds, shifting in half proportion Different from only fewer features in the previous studies,
of each segments. Total 1041 segments are generated with 15 we aim to analyze the EDA signal with various signal pro-
subjects. Out of these segments, 70.1% belongs to the non- cessing methods to enhance the diversity of features. Features
stress class and 29.9% represents the stress condition. contain 36 time domain, 15 frequency domain, 12 entropy
domain, and 21 wavelet domain for both tonic and phasic
D. Decomposition component. Additionally, 13 extra time domain features about
Unlike the complicated EDA decomposition method from peak responses for phasic component are proposed. Total
WESAD, a 0.2 Hz low-pass filter [18] can divide the signal 84 tonic and 97 phasic proposed features are extracted as
into a smooth and slightly changed tonic components while described in succeeding sections and shown in Table II.
the residual signal is a high variation phasic component. After 1) Time domain: Previous studies [19] have used 12 tradi-
separating the tonic and phasic part, we can view these two tional time domain features for basic extraction while first dif-
component as two different decomposition domain and hence ference and second difference of signal with the same method
additional features can be computed. were also calculated, which can provide more information
about the variation of EDA signal. In addition to the statistical
E. Peak Detection features of signal, specific features along with peak responses
A series of rising responses is shown in Fig. 3 along with are extracted for stimuli analysis, e.g. amplitudes, intervals,
the marks indicating the red peaks and yellow onsets. The and duration time etc.
algorithm [6] detects the peaks and onsets by finding the 2) Frequency domain: To map stress response apart from
occurrence of the local maximum (peak) between two consec- time domain features, various frequency domain features have
utive local minimum (onset) in phasic component. Detected been proposed such as power spectrum density (PSD) of EDA
peaks with an amplitude smaller than a critical threshold of signal [14]. Since the power within different frequency bands
the maximum value are excluded in the segment. By this can extract more detailed features, low and high frequency
procedure, we can take into account more information on the component are computed in tonic (0-0.1Hz & 0.1-0.2Hz)
level of peak features. and phasic (0.2-1Hz & 1-2Hz), respectively. After then, the
statistical frequency analysis of the raw signal also provides
IV. P ROPOSED F EATURE S ELECTION F RAMEWORK the information between power and frequency band.
There are four proposed methods in our research. First, we 3) Entropy domain: Entropy describes the randomness, un-
reduce the overlapped proportion of segmentation and extract certainty and messiness of a system and more entropy-domain
the same features set in WESAD. Second, in order to find features have been used in analyzing the physiological signal.
more information from the EDA signal, EDA signals from First, we analyze the symbolic information entropy, which is
tonic and phasic are analyzed in various signal representation coarse graining of time series into symbolic sequences with
as proposed features, e.g. time, frequency, entropy, and wavelet alphabet to classify the dynamic changes. Next, approximate
domain. After then, lots of features in each domain will entropy reflects the probability that two sequences of length
 TABLE II where θ is the parameters of the model, l is the training loss
P ROPOSED F EATURE E XTRACTION function that measures the difference between the prediction
Domain Type Feature yˆi and the ground truth yi , respectively. Ω is the additional
mean, standard deviation, skewness, kurtosis regularized term penalizing the complexity of the model to
raw mean of 1st /2nd diff. avoid over-fitting.
Time
1st diff. mean of absolute 1st /2nd diff.
(T.)(P.)
2nd diff. mean of negative diff., ratio of negative diff.
# local minimum, # local maximum D. Proposed Feature Selection for XGBoost
mean/max/min peaks amplitude from onsets
Time peak
mean/max/min peaks raw value For XGBoost, good features will be picked as node in
mean/max/min peaks phasic value the trees and feature size has little influence on the model
(P.) responses
# peaks, interval of peaks
rising time, recovery time performance, which we can extract a great amount of features
Frequency L.F.
to improve accuracy during training. However, lots of features
mean/max/min/ratio/sum of PSD. may reduce the efficiency and increase computational cost for
(T.)(P.) H.F.
the future application. Therefore, we propose some efficient
Frequency mean/max/min/median of frequency
(T.)(P.)
raw
interquartile range of frequency
feature selection method in this section to realize fewer
computation by finding the most dominant features.
approximate entropy
Entropy 1) Feature Importance: Since decision tree is typically
raw symbolic information entropy
(T.)(P.)
refined composite multi-scale entropy 1-10 chosen as the predictor, the importance of each feature can
L.F. mean, standard deviation, skewness, kurtosis be calculated by counting how many times a feature is used
Wavelet
M.F. root mean square
(T.)(P.) to split the data across all trees, which is helpful for selecting
H.F. approximate/information entropy
T.=tonic P.=phasic diff.=difference #=number of the dominant features in stress detection.
L.F./M.F./H.F.=low/medium/high frequency PSD.=power spectral density We apply a mechanism to gradually increase the feature size
are similar within a tolerance, even if the length of sequence from the most significant features until approaching the same
increases by one [16]. After then, we calculate the refined performance with all features. However, similar features may
composite multi-scale entropy [11] from scale 1-10 which is not only reduce their importance due to the less influence on
an adaptation of multi-scale entropy to resolve the problems the model performance but also limit our mechanism not to
of undefined value. select useful features with different contribution. In order to
4) Wavelet domain: Wavelet Transforms have been suc- find dissimilar features along feature importance, we have to
cessfully used in several noise reduction application [15] obtain the correlation among all features.
because of their good time-frequency localization. A wavelet 2) Feature Correlation: Correlation-based feature selection
transform decomposes a signal into coefficients at multiple method has been used widely for machine learning. We reduce
scales. In our case, Daubechies-3(db3) mother wavelet is used the similar features based on Pearson's linear correlation coef-
obtaining for wavelet coefficients [9]. Both tonic and phasic ficient [20], which is a simple and efficient approach. Among
components of EDA signal are decomposed up to level 5. all features, the features with correlation coefficient absolute
Various statistical and entropy features are extracted from high, value higher than 0.9 were grouped and only one selected as
medium, and low-frequency wavelet coefficients. representative. After then, we can choose dissimilar important
features while keeping model performance and consider these
C. Extreme Gradient Boosting (XGBoost)
features as the most dominant features in stress detection.
In addition to the common machine learning algorithms
used in WESAD [5], we implement XGBoost as our clas- V. E XPERIMENT S ETUP AND R ESULTS
sification model to predict stress class because of its ability
to deal with high-dimensional features or imbalanced data. We conduct two sets of experiments: classification and
XGBoost is an ensemble technique based on gradient boosting, dominant features analysis. We improve the performance and
consisting of sets of predictors (e.g. decision trees) to become show the effective enhancements with only few dominant
a stronger model. The output of the model can be described features on the classification results. The dominant features
in the form: analysis is to verify how the important features we select are
K
X useful in stress detection, which we can use an unsupervised
yˆi = fk (xi ), fk ∈ F, (1)
classification methods with dominant features to distinguish
k=1
stress and non-stress easily.
where K is the total number of predictors, fk for kth predictor In our experiment, 15 subjects are selected and 1041 sam-
is a function in the functional space F . We use training feature ples are generated after data processing. All EDA signals
xi to predict a target variable yi . In the training, a specific loss (chest and wrist) are processed into different signal representa-
function for XGBoost which is optimized at each iteration of tion, such as time, frequency, entropy and wavelet domain. We
gradient boosting is proposed as: normalize the data samples to [-1,1] among different subjects
L(θ) =
X
l(yˆi , yi ) +
X
Ω(fk ), (2) and different features before feeding them into the machine
i k
learning model.
 TABLE III
P ROPOSED S EGMENTATION P ERFORMANCE

RFDT AB LDA kNN XGB

Chest
WESAD [5] 69.88 73.63 71.97 74.51 66.64 NA
our segmentation 85.69 88.78 84.02 88.01 88.18 89.29
Wrist
WESAD [5] 70.95 70.88 75.34 69.86 68.30 NA
our segmentation 87.17 87.24 86.70 85.59 88.91 87.67
DT=Decision Tree RF=Random Forest AB=AdaBoost XGB=XGBoost
LDA=Linear Discriminant Analysis kNN=k-Nearest Neighbour

TABLE IV
P ROPOSED F EATURE E XTRACTION /S ELECTION P ERFORMANCE
Fig. 4. F1 score and computation time along different feature size for chest-
WESAD [5] Time Frequency Entropy Wavelet Fusion based signal in fusion modality. The performance was compared with different
Chest feature selection methods. Blue line: proposed selection f1 score. Red line:
Extraction 89.29 (14) 92.35 (85) 87.87 (30) 78.87 (24) 91.44 (42) 92.27 (195)
Correlation 89.91 (9) 89.74 (37) 87.45 (15) 74.61 (9) 89.89 (29) 91.53 (94) RFE f1 score. Black bar: proposed selection computation time. Grey bar: RFE
Importance 90.08 (5) 89.92 (9) 88.53(5) 76.51 (4) 90.31 (10) 92.38 (9) computation time.
Wrist
Extraction 87.67 (14) 89.73 (85) 84.95 (30) 78.60 (24) 88.42 (42) 89.48 (195)
Correlation 87.02 (9) 88.47 (37) 83.11 (15) 78.36 (9) 89.62 (29) 88.94 (94) sensitive to stress level on time scale and suitable for time-
Importance 88.40 (5) 88.76 (9) 83.70 (5) 77.95 (4) 89.92 (10) 88.78 (9)
Extraction=Extracted features split into time, frequency, entropy, wavelet
frequency analysis.
Correlation=Dominant features without high correlation
Importance=Dominant features important for XGBoost without high correlation
3) Proposed Feature Selection Performance: With lots of
signal processing features, we proposed two selection methods
A. Classification Results to select the dominant features based on their importance
We have three goals for the classification experiment, in- and correlation. Furthermore, we compare our methods with
cluding performance improvement with less overlapped seg- Recursive Feature Elimination (RFE) considering F1-score and
mentation, various signal processing features, and selection computation time, as shown in Fig. 4. The results suggest our
for dominant features. We use XGBoost as our experiment methods with enough feature information can have comparable
model and apply grid search for model parameters tuning. The F1 score to RFE. However, it is considerably less computation
classification performance is evaluated in terms of mean F1- time for our methods during training. By removing correlated
score, which is the harmonic mean of precision and recall. We features and selecting important features for XGBoost, our
employ leave-one-subject-out as our cross-validation scheme, methods can have more efficient performance.
where the classification models are trained using all data but After feature selection for XGBoost, we can obtain the dom-
a previously unseen subject which is used in testing. inant features with similar high classification scores comparing
1) Proposed Segmentation Performance: The stress detec- to all features, as shown in Table IV. Correlation selection can
tion performance with different segmentation methods com- reduce more than half the original features while importance
pared to WESAD is shown in Table III. The extracted features selection can acquire the most dominant features in XGBoost,
and the five machine learning algorithms except XGBoost e.g. features in fusion modality can reduce from 195 features
are all applied the same as WESAD. The results show that to 9 features. The best performance results on XGBoost are
segmentation can have a great influence on model performance 92.38% and 89.92% using chest- and wrist-based EDA signal
with +15% improvement on the testing set, which we suggest respectively. The values indicate that XGBoost is able to
that lots of segments may lead to over-fitting on the similar identify the stress and non-stress class with the proposed
training samples. dominant features. Most importantly, our selection methods
Comparing the performance of the employed algorithms, have the ability to select the features useful for stress detection.
it is obvious that XGBoost has higher classification scores
B. Dominant Feature Analysis
than other ensemble-based or common-used models on both
chest- and wrist-based EDA signals. Therefore, we carry out Dominant features in terms of the feature selection of
more experiments on XGBoost to explore its ability for stress fusion modality are shown in Table V and the features we
detection. propose are highlighted in boldface. The tonic component
2) Proposed Feature Extraction Performance: As for our captures the mean level and the signal range while the phasic
enhanced results on XGBoost, we analyze more features on component emphasizes the entropy and difference. The result
different feature extraction domain as shown in Table IV. shows that the low frequency magnitude and the high fre-
Slight improvements compared to WESAD features are +3% quency complexity of EDA signal contain the most significant
and +2% for chest- and wrist-based signal, respectively, which information in stress detection. Moreover, the property of these
indicate that additional signal processing features of EDA sig- features indicate the positive correlation with stress level,
nal are useful for stress detection. Regarding more information which suggests that stress can lead to high level and high
on different extraction domains, the results suggest that good disorder electrodermal activity.
features are derived more from time and wavelet domain. In addition to the discussion on the information of features,
These processing features show that the EDA signal wave is we take the dominant features as a nine-dimensional vector
 TABLE V [4] W. Boucsein, Electrodermal activity. New York: Plenum, 1992.
D OMINANT F EATURES OF F USION M ODALITY [5] P. Schmidt, A. Reiss, R. Duerichen, C. Marberger, and K. Van Laer-
hoven, “Introducing wesad, a multimodal dataset for wearable stress and
No. Dominant features
affect detection,” in Proceedings of the 2018 on International Conference
1 mean of the tonic
on Multimodal Interaction, 2018, pp. 400–408.
2 number of local maximum in 1st difference of phasic
[6] J. Healey, R. W. Picard et al., “Detecting stress during real-world driving
3 approximate entropy of high-frequency wavelet of tonic
tasks using physiological sensors,” IEEE Transactions on Intelligent
4 information entropy of high frequency wavelet of tonic
Transportation Systems, vol. 6, no. 2, pp. 156–166, 2005.
5 1-2Hz spectral power of phasic
[7] M. Gjoreski, H. Gjoreski, M. Luštrek, and M. Gams, “Continuous stress
6 sum of phasic startle magnitudes
detection using a wrist device: in laboratory and real life,” in Proceedings
7 standard deviation of phasic
of the 2016 ACM International Joint Conference on Pervasive and
8 area under the phasic responses
Ubiquitous Computing: Adjunct, 2016, pp. 1185–1193.
9 range of the EDA
[8] A. Anusha, J. Jose, S. Preejith, J. Jayaraj, and S. Mohanasankar,
“Physiological signal based work stress detection using unobtrusive
sensors,” Biomedical Physics & Engineering Express, vol. 4, no. 6, p.
065001, 2018.
[9] V. Sharma, N. R. Prakash, and P. Kalra, “Audio-video emotional
response mapping based upon electrodermal activity,” Biomedical Signal
Processing and Control, vol. 47, pp. 324–333, 2019.
[10] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,”
in Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, 2016, pp. 785–794.
[11] K. Tung, P.-K. Liu, Y.-C. Chuang, S.-H. Wang, and A.-Y. A. Wu,
“Entropy-assisted multi-modal emotion recognition framework based on
physiological signals,” in 2018 IEEE-EMBS Conference on Biomedical
Engineering and Sciences (IECBES), 2018, pp. 22–26.
[12] C. Setz, B. Arnrich, J. Schumm, R. La Marca, G. Tröster, and U. Ehlert,
“Discriminating stress from cognitive load using a wearable eda device,”
IEEE Transactions on Information Technology in Biomedicine, vol. 14,
Fig. 5. t-SNE visualization for dominant features based on EDA signal. Red: no. 2, pp. 410–417, 2010.
non-stress. Blue: stress. [13] H. Kurniawan, A. V. Maslov, and M. Pechenizkiy, “Stress detection from
speech and galvanic skin response signals,” in Proceedings of the 26th
among all data segments. These vectors are projected on a two- IEEE International Symposium on Computer-Based Medical Systems,
2013, pp. 209–214.
dimensional space by t-SNE [21], which we can distinguish [14] H. F. Posada-Quintero, J. P. Florian, A. D. Orjuela-Cañón, T. Aljama-
the data distribution easily, as shown in Fig. 5. The visualiza- Corrales, S. Charleston-Villalobos, and K. H. Chon, “Power spectral
tion result vindicates the ability of the dominant features to density analysis of electrodermal activity for sympathetic function
assessment,” Annals of Biomedical Engineering, vol. 44, no. 10, pp.
detect the non-stress class and stress class while the non-stress 3124–3135, 2016.
class contains baseline and amusement affect states. [15] S. Taylor, N. Jaques, W. Chen, S. Fedor, A. Sano, and R. Picard,
“Automatic identification of artifacts in electrodermal activity data,” in
VI. C ONCLUSION 2015 37th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), 2015, pp. 1934–1937.
In this paper, we propose an enhanced framework for stress [16] Z. Visnovcova, M. Mestanik, M. Gala, A. Mestanikova, and I. Ton-
detection based on EDA signal. The proposed improvements hajzerova, “The complexity of electrodermal activity is altered in mental
for higher performance include less overlapped segmentation, cognitive stressors,” Computers in Biology and Medicine, vol. 79, pp.
123–129, 2016.
more signal processing features, and extreme gradient boosting [17] C. Kirschbaum, K.-M. Pirke, and D. H. Hellhammer, “The trier social
classifier. Moreover, we obtain the most dominant features stress test–a tool for investigating psychobiological stress responses in
related to stress based on the selection of feature importance a laboratory setting,” Neuropsychobiology, vol. 28, no. 1-2, pp. 76–81,
1993.
and correlation. The analysis of dominant feature suggests [18] J. Kim and E. André, “Emotion recognition based on physiological
that the magnitude of low frequency and the complexity of changes in music listening,” IEEE Transactions on Pattern Analysis and
high frequency of EDA signal are useful for stress detection. Machine Intelligence, vol. 30, no. 12, pp. 2067–2083, 2008.
[19] R. Zangróniz, A. Martı́nez-Rodrigo, J. Pastor, M. López, and
According to our studies, we can acquire 92.38% and 89.92% A. Fernández-Caballero, “Electrodermal activity sensor for classification
f1-scores with 9 dominant features in XGBoost classifier based of calm/distress condition,” Sensors, vol. 17, no. 10, p. 2324, 2017.
on chest- and wrist-based EDA signals, which is the state of [20] J. Biesiada and W. Duch, “Feature selection for high-dimensional data-
a pearson redundancy based filter,” in Computer recognition systems 2.
the art on WESAD using EDA for stress detection. Springer, 2007, pp. 242–249.
[21] J. Birjandtalab, D. Cogan, M. B. Pouyan, and M. Nourani, “A non-
R EFERENCES eeg biosignals dataset for assessment and visualization of neurological
[1] G. P. Chrousos and P. W. Gold, “The concepts of stress and stress status,” in 2016 IEEE International Workshop on Signal Processing
system disorders: overview of physical and behavioral homeostasis,” Systems (SiPS), 2016, pp. 110–114.
The Journal of the American Medical Association, vol. 267, no. 9, pp.
1244–1252, 1992.
[2] H. Gao, A. Yüce, and J.-P. Thiran, “Detecting emotional stress from
facial expressions for driving safety,” in 2014 IEEE International Con-
ference on Image Processing (ICIP), 2014, pp. 5961–5965.
[3] H. Lu, D. Frauendorfer, M. Rabbi, M. S. Mast, G. T. Chittaranjan, A. T.
Campbell, D. Gatica-Perez, and T. Choudhury, “Stresssense: Detecting
stress in unconstrained acoustic environments using smartphones,” in
Proceedings of the 2012 ACM Conference on Ubiquitous Computing,
2012, pp. 351–360.