The Body’s Automatic Stress Detection Process Relies on

  • Journal List
  • Sensors (Basel)
  • v.19(8); 2019 Apr
  • PMC6515276

Sensors (Basel).
2019 Apr; 19(eight): 1849.

Continuous Stress Detection Using Article of clothing Sensors in Real Life: Algorithmic Programming Contest Case Study

Received 2019 Mar 2; Accustomed 2019 Apr 16.

Abstract

The negative effects of mental stress on human being wellness has been known for decades. High-level stress must be detected at early stages to prevent these negative effects. After the emergence of wearable devices that could be part of our lives, researchers have started detecting extreme stress of individuals with them during daily routines. Initial experiments were performed in laboratory environments and recently a number of works took a pace outside the laboratory surroundings to the real-life. We adult an automated stress detection system using physiological signals obtained from unobtrusive smart wearable devices which tin can be carried during the daily life routines of individuals. This system has modality-specific artifact removal and feature extraction methods for existent-life conditions. We farther tested our system in a real-life setting with collected physiological data from 21 participants of an algorithmic programming contest for ix days. This event had lectures, contests too equally gratis time. Past using heart activity, skin conductance and accelerometer signals, we successfully discriminated contest stress, relatively higher cerebral load (lecture) and relaxed fourth dimension activities past using different machine learning methods.

Keywords:

stress recognition, machine learning, wearable sensors, smartwatch, photoplethysmography, electrodermal activeness, daily life psychophysiological data, heart charge per unit variability

1. Introduction

Daily life stress is an important problem of our modernistic society. It is a growing issue and it has become an unavoidable part of our daily lives. Psychological stress types can be listed every bit astute and chronic [1]. Acute stress is more prevalent than chronic stress. American Psychological Clan noted that the causes of acute stress are pressure from contempo past and nearly future [ii]. Able-bodied challenges, examination taking, or anxiety when meeting new people can induce acute stress. On the other hand, long-standing pressures and demands as a consequence of socioeconomic conditions, difficulties in interpersonal relationships, or an unsatisfying career tin trigger chronic stress [2]. If chronic stress is not handled properly, it could result in serious health issues [3]. Since symptoms of acute stress are more apparent than chronic stress symptoms, acute stress is more widely investigated.

After musculoskeletal illnesses, which also could exist stress-related in some cases [four], stress is one of the most significant wellness bug in the world. The effect of stress on man health depends on the stress type. Emotional distress, muscular ache and tension, back pain, headache, heartburn, digestive tract issues, and overarousal tin be named as the effects of acute stress [5]. Overarousal can cause eye attacks, arrhythmias, and even sudden death for people with eye conditions [6]. Effects of the chronic stress on human wellness are akin to those of acute stress even so it can damage physical conditions more. Possible causes of the chronic stress tin can be listed equally hypertension and coronary disease [six,vii], irritable bowel syndrome, gastroesophageal reflux affliction [eight], generalized anxiety disorder, and depression [ix]. The above-mentioned stress-related diseases also affect the economy by increasing absenteeism, staff turnover [x], presenteeism, and tardiness. These issues decrease the product and increment the piece of work-related costs. Public surveys [11] unveiled that at least half of the European workers are subjected to stress at work. Furthermore, at least half of the lost working days in the business sector are causeless to be acquired past piece of work-related stress and psycho-social risks [12].

Researchers plant out that stress should be handled when the symptoms first come out to avoid the long-term consequences. In other words, stress must be discovered in early stages to refrain from more damages and impede information technology from being chronic. The higher up-mentioned damages of stress on homo health and detriments to social life and economy have forced researchers to come up with an automated stress monitoring scheme which exploits smart article of clothing devices and advanced affective computing algorithms. This scheme can be applied in automobiles, airplanes, factories, and offices, at job interviews and daily life environments. This scheme tin can further compute social stress stages during meetings or mutual intercommunication. The platonic scheme should exist applicable to daily life, i.east., it should use unobtrusive sensors and devices which users tin article of clothing easily in their daily routines.

In this work, we adult an automatic stress level detection scheme that uses physiological signals from wrist-worn devices. Our scheme can also exist applied to daily life of individuals. In real-life settings, movements of individuals are unrestricted and artifacts occur because of that. In society for our organization to be applicable in these settings, we practical several novel artifact detection and removal strategies. These antiquity detection algorithms are developed for specific sensors and their performances are scientifically proven. We further extracted features from heart activity, skin conductance, and accelerometer signals with our tools. From these features, we classified the stress level of an individual by employing auto learning algorithms. To test our organisation in real-life settings, we collected physiological signals of participants in an algorithmic programming summer army camp via smart wrist-worn habiliment devices. This military camp was composed of lecture, contest, and complimentary time sessions. We nerveless data for nine days from 21 participants. After the data were collected, we obtained promising results for detecting stress with these habiliment devices in real life scenarios. Our work addresses five prominent enquiry issues:

  • The comparison of stress detection model performances with dissimilar wearable devices;

  • The influence of the interpolation, aggregation window sizes and artifact detection threshold percentages;

  • Change in the functioning of the stress detection scheme with known context labels and the subjective reports as the ground truths;

  • The discriminative effect of each sensor modality; and

  • The performance of person-specific and full general models.

The structure of the rest of the paper is every bit follows: In Section 2, the related work for stress detection is provided. Real-life data collection bug are addressed in Section 3. In Section 4, our stress detection scheme is explained. Data drove result and our experiment design are presented in Section 5. In Section half-dozen, we present experimental results and discussion. The decision of the study and futurity work are given in Section 7.

ii. Related Piece of work

The early stress detection research was performed in the laboratory environments, while the current research continues on real-life environments (see
Table 1). Electrodermal action (EDA), centre activeness (60 minutes) and accelerometer are the most widely used physiological signals for the detection of stress levels. As shown in
Table one, EDA and HR combination has the best performances in the laboratory environments. Proposals with accuracies higher than 95% employ this combination every bit the physiological signals. Linear Discriminant Analysis (LDA), Support Vector Car (SVM), g Nearest Neighbors (kNN) and Fuzzy Logic classifiers are the best performing machine learning (ML) algorithms. An 89% accurateness was achieved in four-class stress classification by using EEG signals in [13]. However, current EEG (Electroencephalogram) measuring devices are obtrusive for individuals and they are non applicable to daily life routines.


Table 1

Stress detection experiments in controlled laboratory environments.

Commodity Stress Signal Stress Examination Method # of Classes Accuracy % Applicative in Daily Life?
[14] (2012) HRV Stress in the traffic Minimum Altitude Classifier 3 (Depression, Medium, High) 90 Yes
[15] (2011) EDA, PPG Hyperventilation and Talk Prep Fuzzy Logic 2 (S, R) 99 Yes
[16] (2013) Oral communication TSST SVM 2 (South,R) 72 Yes
[17] (2011) ECG, EMG, EDA Arithmetic, Puzzle, Retention Tasks Bayes, kNN, LSD 2 (South, R) eighty Yes
[18] (2016) PPG, EDA, Respiration, Thermal Cam Lie Detection DecisionTree 2 (Southward, R) 73 Yep
[13] (2016) EEG Arithmetic Task SVM 4 (Neutral, Medium, Low, Loftier) 89 No
[xix] (2015) Body Movements Arithmetic Task SVM 2 (S, R) 77 No
[20] (2016) Body Movements, EMG, EDA, Respiration Arithmetic Task SVM ii (Stress, Relax) 85 No
[21] (2017) Facial Cues Social Exposure and Stressful Media (IAPS) kNN, SVM, Naive Bayes three (Neutral, Relax, Stressed) 91.68 No
[22] (2014) Pupil Bore IAPS DecisionTree 2 (Stress, Relaxed) 90 No
[23] (2017) EDA, PPG, Speech communication, Accelerometer TSST Adaboost 2 (Stress, Relax) 94 Yes
[24] (2015) EDA, Accelerometer, Bluetooth Logistic Regression 2 (Stressed, Unstressed) 91 Aye
[25] (2012) Temperature, Oestrus Flux, EDA, Respiration, Accelerometer Arithmetic Task, Cold Pressor and loud Sounds Naive Bayes 2 (Stress, Relaxed) 82 No
[26] (2017) ECG, GSR, Respiration, Blood Pressure, Blood Oximeter Ice examination and IAPS SVM, kNN 2 (Stressed, Relax) 95.eight Yes
[27] (2014) EEG, ECG, EMG, EOG Mental and Retentivity Task ANN iii (Relaxed, Mental, Fatigue) 80 No
[28] (2015) Facial Blood Period SCWT Multiple Regression 2 (Due south, R) 88.6 No
[29] (2015) EDA Fail Scenarios LDA 2 (S, R) 98.88 Yes
[xxx] (2016) Homo Gaze, Mouse Click Arithmetic Task Random Woods 2 (Due south, R) 66 No
[31] (2018) BioRadar Mental Arithmetic Task Multilayer Perceptron 2 (Due south,R) 0.94 No
[32] (2016) Mobile Awarding Usage Blueprint-Physical Activity-Light Sensor-Screen Events Real Life SVM, ANN, kNN 2 (S, R) 70 Yes
[33] (2016) BVP-Pare Temperature-EDA-RR-Heart Charge per unit (Without Context Info) Existent Life Random Forest ii (S, R) 76 Yes
[34] (2017) Hr-IBI-HRV-EDA- Temperature Real Life Weka Toolkit 2 class (South, R) 70 Yep
[35] (2018) Phone usage data for different application categories Existent Life HMM with MPM 2 (S, R) 68 Yes

Almost all of the studies in
Tabular array i
employed a ii-class stress level classification. However, the stress upward to a sure level might be harmless. After a sure limit, the stress level should be detected and precautions taken. To this end, stress detection resolution must be increased (precision of detected stress levels should be increased) and multi-level stress detection systems with high classification functioning must exist adult. These schemes should further take advantage of multimodality to increase accuracies as the laboratory enquiry suggest the benefits (see
Tabular array i).

It is recognized that the stress level that subjects endure in this environment is dissimilar from real life stress [1]. Information technology is as well demonstrated that subjects are reluctant to wear obtrusive instruments for measurement and they are not comfortable with these devices. For these reasons, stress measurement research has taken a step outside the lab with the aim of developing an unobtrusive multi-level stress detection system for daily life. Since smartphones and wearable devices have become an integral function of our lives in our modern society, they are chosen equally the instruments for stress detection in daily lives research.

Afterwards laboratory environments, stress level detection research has been conducted in restricted and semi-restricted environments such every bit office, automobile and university campus. Part and workplace are amidst the environments which increase the stress levels the most. The stress level in the office environs is monitored past using EDA, ECG (Electrocardiogram) and Accelerometer [36,37,38]. Peculiarly in crowded cities, the stress levels of individuals increase in traffic jams. At that place are a number of studies in the automobile environments in the literature. Most of the studies used DriveDB database [39]. This database consists of ECG, EDA, EMG (Electromyogram) and respiratory sensor information collected from 24 drivers in Boston. Researchers applied automobile learning algorithms to this database and EDA-ECG signal combination and SVM-kNN classifiers accomplished the best performance [forty] in this environment. Campus environments are semi-restricted environments and the well-nigh similar surround to unrestricted daily life environment. Therefore, nomenclature performances are lower when compared with restricted laboratory, function and automobile environments. ECG betoken and the conclusion tree classifier have achieved the highest classification accuracy in ii-form classification in a campus environment [41]. Most works have only used features extracted from the smartphones [42,43,44]. Smart habiliment devices are not used in the campus environment in well-nigh of the works.

The stress detection research has taken a step to the unrestricted existent life since the ultimate aim is to detect stress levels of individuals in their daily routines. However, researchers should come up up with solutions to new problems ascend when taking a pace exterior the laboratory (see Section three). The stress level recognition performances of existent-life schemes are lower than restricted environments and laboratory environments [32] (2016), ref. [45] (2015), ref. [34] (2017) and [35] (2018). The listed works have classification accuracies around 70% and lxxx%. Low reliability of self-report answers, the unknown context of participants and unrestricted movements of subjects could be the main reasons. Furthermore, the devices used in real-life studies are non-obtrusive but their information quality is not comparable with their laboratory counterparts. There are a lot of smart unobtrusive wearable devices for daily life usage. Even so, their data quality and effect on stress level detection performance are non investigated thoroughly. Another open enquiry question for real-life studies is the unknown context and low reliability of self-reports. Gjoreski et al. [34] employed activity recognition to increment the knowledge regarding context and improve their recognition performance. The effect of context and questionnaires to the performance of stress recognition systems should exist investigated comprehensively. Lastly, to eliminate the negative effects of unconstrained movements, antiquity detection and removal algorithms specific for each sensor must be adult and used.

3. Preprocessing and Characteristic Extraction: Problems and Possible Solutions in Real Life

Existent-life data collection brings new problems that are not encountered in laboratory data collection. In a lab experiment, methods for data collection are less error-prone and relatively easy. In real life scenarios, new parameters that can cause new research issues are added to the organization. For example, the maximum runtime of the devices is express due to their express battery. Wrong placement of devices, loosely worn equipment, charging of instruments, unconstrained movement of subjects and issues with the ground truth collection should exist taken into business relationship.

three.one. Problems Related to the Movement and Improper Placement of Devices

Today’southward off-the-shelf wearable devices provide u.s.a. with high-quality data standards [46]. However, certain conditions must exist satisfied for loftier-quality information acquisition. Electrodes should be properly placed obeying the instructions of the equipment, wristbands must be tightly worn and body movements should exist express. Otherwise, signals are contaminated by noise, loosely worn devices, and body movements [12]. To remove the dissonance, some signal processing techniques must be applied. Every problem creates multiple options for researchers. To requite an illustration, if a subject wears the device loosely, and for some period the information could not be acquired, the researcher may opt to ignore this time period or interpolate the data. Another example would be the pick of handling data artifacts due to unconstrained movement of a subject in daily life. To clean the data, in that location are several filters such as Kalman filters, Butterworth depression-laissez passer filters, median filters, Wiener filters, and wavelet decomposition [12]. For removing the artifacts, to the lowest degree mean squares, regression analysis, contained component (ICA) and chief component analysis (PCA) could exist employed [12].

3.2. Information Fusion from Multifariousness of Sensors

To increment the success of stress measurement systems, researchers tend to collect multimodal information. The integration of multimodal data imposes challenges. Synchronization must be employed between dissimilar information types by using timestamps. When to integrate the information (earlier classification or during processing) and missing data from some modalities are other challenges.

3.iii. Selection of Non-Obtrusive Devices

To collect data during the daily life of individuals, stress measurement devices should be non-obtrusive. People should article of clothing these devices without beingness uncomfortable in their daily routines, during sleeping, meetings and everyday activities. Obtrusiveness may even lead to extra stress on participants. The ideal system should collect the information without the user even being aware of information technology.

3.four. Limited Runtime Due to Battery

Limited runtime is another meaning problem when collecting data from participants in real life. If the maximum runtime of a device is around iii–4 h (such equally the case of Samsung Gear S1, S2, and S3), researchers or users have to charge the device several times for a whole solar day of information collection. This causes gaps in the collected information and increases the amount of endeavor for recharging and restarting sessions. The imposed challenges on researchers would exist increasing the battery lives of devices by reducing power consumption (i.e., disabling some sensors, duty cycling devices, and decreasing brightness).

Read:   Which Phrase Best Describes the Function of the Atp Molecule

iii.5. Ground Truth Collection

In laboratory experiments, researchers know the ground truth such as relaxed, baseline, and stressed considering they designed the experiment timeline. Still, in existent life information drove, to mensurate the success of stress detection schemes, the ground truth from subjects must exist nerveless. To this end, researchers usually employ some surveys (Perceived Stress Scale, Stress Self-Rating Calibration, NASA-TLX, The State-Trait Anxiety Inventory, Self Cess Manikin and Positive and Negative Affect Schedule questionnaires)periodically during a day. Researchers have to collect the surveys from each participant and redistribute new ones when the time comes. This job can be automatic by developing a mobile survey app and collecting answers periodically through pop-upward surveys. In our instance, we have the context information during the data drove, such every bit if they are in a lecture, a contest or free time.

iv. Proposed Arrangement Description

In plethysmography, volumetric changes of organs are measured from the skin illuminated by the light emitted from a pulse oximeter PPG [47]. PPG sensors in our devices are used to measure the center activity past measuring blood menses during the middle’s pumping deportment. Heart activity betoken is composed of dissimilar peaks and valleys. R peak is the about prominent i, which is used to calculate heart rate variability. PPG provides the RR interval by measuring the elapsing between ii consecutive R peaks which tin can also be chosen as Interbeat Interval (IBI).

EDA also known as Galvanic Skin Response (GSR), is the change of electrical properties of peel. Under emotional arousal and stress, trunk sweats and skin conductance increases. EDA is 1 of the all-time and widely used discriminative point along with the eye charge per unit signal for measuring stress [48]. Mean aamplitude, standard departure, minimum and maximum values, RMS, the filibuster between applied stimuli and response, number of peaks, peak meridian, ascent time, recovery time, the position of maximum and minimum features were used in the literature to measure out the stress levels of the user [49].

In this study, we developed a multi-level stress detection system, which employed heart activity information from the PPG sensor, skin conductance data from the EDA sensor and accelerometer and temperature information. Our EDA preprocessing tool uses accelerometer and temperature signals to clean the artifacts in this signal. We further extracted features from the accelerometer sensor but temperature information were not used for characteristic extraction. The increment in the centre rate and electrodermal activeness levels can be seen in
Figure 1. Preprocessing and feature extraction tools for each modality were developed. For each sensor, modality-specific tools were applied to eliminate artifacts, cleaning signals and extracting features. After the characteristic extraction, the virtually successful machine learning algorithms in the literature were applied to the physiological information for the classification task. Our system is compatible with different smart wrist-worn wearables in spite of the fact that they have unlike platforms and sensors. System diagrams for Samsung Gear South family devices and Empatica E4 devices are shown in
Figure 2. Note that all parameters for artifact detection and preprocessing algorithms are universal and person contained.


Recorded physiological signals before and after the start of the stimuli. The increment in EDA point level and number of peaks and irregularities and sudden increases in HRV can be seen in this figure.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g002.jpg

The cake diagram of the stress level detection system for Samsung Gear Due south and S2 and Empatica E4. Since the sensors and platforms are different, please notation that EDA and temperature signals are just bachelor for E4.

4.1. Electrodermal Activity Signal Preprocessing and Feature Extraction Tools

four.ane.ane. Preprocessing and Artifact Removal

Electrodermal Activeness bespeak is afflicted past increased physical activity and temperature changes. In these situations, obtained signal is contaminated and should be filtered. To this cease, we employed the EDA Explorer tool from Taylor et al. [l]. The artifacts in the EDA indicate is manually labeled by the experts to train a auto learning model. Past applying the SVM (Support Vector Motorcar) classifier with the accelerometer and temperature information, this tool achieves 95% accurateness on detecting artifacts in the EDA signals (encounter
Figure 3
and
Figure 4). We added batch processing feature to this tool. If a data segment is detected every bit an antiquity segment, it is excluded in the characteristic extraction process. By this style, we eliminated false peaks caused by increased temperature or physical activeness when extracting features.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g003.jpg

The case filtered EDA indicate according to changes in the accelerometer indicate. Annotation that red components were deleted because of the high activity intensity.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g004.jpg

Activity intensity is shown past using the accelerometer sensor X, Y, and Z components corresponding to the example EDA signal in
Figure 3
Note that this case was recorded during a highly intensive activity.

4.1.2. Characteristic Extraction

After cleaning artifacts from the signals, features were extracted. EDA indicate has two components: phasic and tonic. Nosotros decomposed the EDA indicate by applying the
cvxEDA
tool [51] on the EDA bespeak, which makes use of a convex optimization approach to decompose the EDA signal. Skin Conductance Level (Tonic) component includes more long-term ho-hum changes, whereas phasic components include faster (event-related) changes. When evaluating the hateful, standard deviation, and percentile features, researchers use the tonic component because they do non desire to overestimate these long-term changes with consequence-related fast changes. The phasic part is subtracted and features are calculated. On the other mitt, some peak related features such equally peaks per 100 south, superlative aamplitude, and potent peaks (peaks that are more 1


μ

Siemens) per 100 s are calculated from the phasic element. An example of a decomposed signal is shown in
Figure v. After that, we extracted seven features from the EDA signal: mean, standard deviation, superlative, strong height, 20th percentile, 80th percentile and quartile deviation (75th percentile–25th percentile).


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g005.jpg

Decomposed EDA Signal from Empatica E4 wristband by applying cvxEDA tool.

4.ii. Center Activity Point Preprocessing and Feature Extraction Tools

four.2.1. Preprocessing and Artifact Removal

The heart rate activity signal is too sensitive to the movement of the subjects and loosely worn wrist devices. To cope with these bug and make clean the artifacts from the indicate, our enquiry grouping adult a preprocessing tool in MATLAB. With this tool, we employed an artifact detection percent threshold between the information and the local average. In the literature, this threshold is by and large prepare every bit xx% [38] and nosotros as well used this threshold. After nosotros detected the artifacts in the center activity point, a user can cull to remove and apply some additional constraints or supersede them with shape preserving cubic spline interpolation after removal (see
Figure 6). If the artifact data points are removed and non-interpolated, new rules can be ready on the remaining healthy information. A minimum corporeality of sequent information samples and minimum consecutive fourth dimension rules can exist fix to evaluate the remaining segments. These rules are used to exclude interrupted (with holes of removed data) small amount of consecutive data in the feature extraction process. We applied removal with extra exclusion rules and removal and interpolation separately and observed their effect on the functioning of our system.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g006.jpg

Gaps due to movement and loosely worn wristband from PPG (Photoplethysmography) data (Left) are filled with cubic interpolation function (Right).

The tool also has a batch processing feature. Length of the local hateful, the pct of artifact detection threshold, minimum consecutive time and data sample constraints can exist altered with parameters.

four.2.2. Characteristic Extraction

For feature extraction, nosotros used MATLAB born tools forth with Marcus Vollmer’s HRV toolbox [52] along with our preprocessing tool. The employed time domain features are the hateful value of the heart rate (Mean 60 minutes), the standard deviation of inter-beat interval (IBI), hateful value of the inter-beat (RR) intervals (Mean RR), root mean square of successive difference of the RR intervals (RMSSD), the percentage of the number of successive RR intervals varying more than than 50 ms from the previous interval (pNN50), the full number of RR intervals divided by the peak of the histogram of all RR intervals measured on a scale with bins of 1/128 s (HRV triangular index), and triangular interpolation of RR interval histogram (TINN).

We as well applied Fast Fourier Transform (FFT) and Lomb–Scargle periodogram [53] and the post-obit frequency domain features are calculated: low frequency ability (LF), loftier frequency power (HF), very depression frequency power (VLF), prevalent low frequency (pLF), prevalent loftier frequency (pHF), the ratio of LF to HF (LF/HF), (From Lomb–Scargle) LF, HF, and LF/HF. Definitions of these features are given in
Table 2.


Tabular array two

Heart charge per unit variability features and their definitions.

Feature Description
Mean RR Mean value of the inter-crush (RR) intervals
STD RR Standard divergence of the inter-beat interval
RMSSD Root mean square of successive difference of the RR intervals
pNN50 Percentage of the number of successive RR intervals varying more than 50 ms
from the previous interval
HRV triangular alphabetize Full number of RR intervals divided by the height of the histogram of all RR intervals
measured on a calibration with bins of 1/128 southward
TINN Triangular interpolation of RR interval histogram
LF Power in low-frequency ring (0.04–0.15 Hz)
HF Power in high-frequency ring (0.15–0.4 Hz)
LF/HF Ratio of LF-to-HF
pLF Prevalent low-frequency oscillation of heart rate
pHF Prevalent high-frequency oscillation of heart rate
VLF Power in very low-frequency ring (0.00–0.04 Hz)
SDSD Related standard divergence of successive RR interval differences

4.three. Accelerometer Processing and Feature Extraction

Body and caput movements can be used to detect the emotions and arousal level [54]. The accelerometer sensor records three-centrality acceleration with gravity. We employed the accelerometer modality in two means. Firstly, to find artifacts in the EDA information, accelerometer data were used forth with the temperature data. Secondly, we used these information for characteristic extraction. The mean value is calculated for each window. The energy of the signal is as well calculated with FFT.

4.4. Machine Learning Tools

For the classification of the data, we employed the Weka toolkit [55]. For preprocessing of features, we practical numeric to nominal transformation to the class column. Since our dataset is unbalanced in terms of membership of class instances, we added instances from the minority class and removed the samples from the bulk class to overcome the class imbalance problem. Therefore, we prevented classifiers from biasing towards the class with more instances. In this report, we evaluated the performance of half dozen well-known classifiers.

  1. Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA)

  2. PCA and Support Vector Motorcar with radial kernel (SVM)

  3. K-Nearest Neighbours (due north
    = i) (kNN)

  4. Logistic Regression

  5. Random Woods (RF with 100 trees)

  6. Multilayer Perceptron

10-fold cross-validation was applied. A three-class classification system was developed. The parameters of the classifiers were selected from the stress level detection studies in the literature.

5. Description of the Data Collection Event: Algorithmic Programming Summer Camp

To examination and evaluate our system in real-life settings, we conducted a information drove experiment in the INZVA algorithmic programming contest summer military camp, which is organized each year in Istanbul, Turkey [56]. This event is similar to the International Collegiate Programming Contest (ICPC) [57]. A photograph from the information collection setup is shown in
Figure 7. Algorithmic programming camp is designed for loftier-school and university students to ameliorate their programming skills and this contest will induce stress on the participating students.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g007.jpg

A view of smartwatches and wristbands after data extraction, charged and ready to use.

The algorithmic programming contest is conducted in iii levels, expert, advanced and foundation. 80-4 students with unlike levels of expertise gathered to participate in this algorithmic programming contest. Algorithmic programming competition camp was held for nine days. The physiological indicate and questionnaire data were collected from the 21 participants at the foundation level. Of these 21 people, 18 were men and 3 female person and the average age of the attendees was 20.

There were three types of sessions such as the grooming, the contest and the gratis day. The program was scheduled to be held from 10:00 to 17:00 for the whole nine days. The data collection schedule presented in
Figure 8
was repeated for the get-go 8 days. In those days, attendees had training classes with professors from the field of computer science and computer engineering science from high-ranking universities in Turkey for two hours. The participants entered daily trouble-solving contests in which the questions were derived from the aforementioned days of training lectures. In all sessions, raw NASA-TLX questionnaires were collected from the users. The frustration question of this questionnaire was used to measure the perceived stress levels of the individuals. On the endmost day, the participants entered the final contest and they should solve challenging questions, which were asked from all the topics covered in the total 8 days of training. Participants had to solve challenging questions within a time limit and the scores of the participants were projected on a wall, which created extra stress for the participants. To collect more points, participants should solve more than questions in a shorter fourth dimension period than their opponents. As a result, the mental demand likewise as the temporal demand was increased, which encouraged the participants to gain more points in a shorter time and to achieve a higher position in the final ranking. When a participant solved a question, a balloon was fastened to his/her table.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g008.jpg

The daily schedule and information drove process during the algorithmic programming contest.

5.1. Data Collection Procedure

Nosotros informed the volunteer participants about the purpose and the procedure of the study. The data collection procedure and all the interventions in this research fully run into the 1964 Annunciation of Helsinki [58] and before whatever data acquisition, all participants volunteering to accept office in the study start signed the informed consent documents.

Twenty-1 foundation level subjects were selected because that they would attend the entire week and with approximately the same conditions. Tutorials and farther guidelines were presented to all of them concerning how to use the devices and how to fill in the questionnaires. Three study investigators checked if the devices were worn properly and running correctly. For data extraction, collecting the forms and bombardment recharge procedures, which were being administered past our team, a schedule was set up and participants followed this schedule regularly.

A unique number was assigned to each participant and to each device during the study. Study investigators ensured that the participants wear the device with the correct number. After the data collection, the relations between the participant names and the numbers are anonymized.

5.ii. Ideals

The process of the methodology used in this written report was canonical by the Institutional Review Board for Research with Man Subjects of Boğaziçi University with the approval number 2018/16. Prior to the information acquisition, each participant received a consent grade, which explains the experimental process and its benefits and implications to both the social club and the subject field. The procedure was also explained vocally to the subject area. The data collection procedure and all of the interventions in this enquiry fully run across the 1964 Announcement of Helsinki [58]. All of the data are stored anonymously.

5.3. Types of Wearable Devices Used for Data Acquisition

Two Samsung Gear S1, ten Samsung Gear S2, four Samsung Gear S3 smartwatches and iv Empatica E4 wristbands were used to gather data in this result. The maximum runtime of the devices when they are fully charged varies. While Empatica E4s can collect information for over 48 h, Samsung smartwatches can collect data for at most 4 h when all the sensors are active. All these devices are off-the-shelf and they provided us with the power to access the raw information. Empatica E4 is solely built for inquiry and it provides the software for accessing data; however, nosotros had to develop an awarding for data conquering for Samsung devices that allows the choice of sensors to be used for data drove. Nonetheless, for this research, we gathered data from all sensors. While data collected from Samsung devices were collected straight by Wi-Fi, Empatica E4 information were offset sent to the cloud.

Read:   Annie Traveled 5 Times the Sum

While Empatica E4 devices have four sensors, namely 3D accelerometer (ACC), photoplethysmography (PPG), electrodermal activity (EDA) and the peel temperature (ST), Samsung Gear watches lack the EDA sensor just instead are equipped with Gyro and Barometer sensors. In this study, we used PPG, EDA and ACC sensors.

vi. Discussion of Experimental Results

We adult a iii-grade stress detection arrangement. The system can differentiate the stress level of the free twenty-four hours, lecture and competition sessions. It can further differentiate three levels of perceived stress (come across Section 6.5). Besides from Department half dozen.v, nosotros used the context label as the ground truth and we chosen the measured stress as physiological stress. Free day is enumerated equally 0, lecture is assigned 1 and contest is assigned 2 labels. We causeless that the stress levels of most of the subjects would be college in contest, medium in lecture and lower in the free fourth dimension with this context labels. In Section 6.5, the frustration scale of raw NASA-TLX is further used equally a ground truth and performance of the arrangement is compared under these 2 conditions. With the latter ground truth, perceived stress level of individuals from cocky-reports was measured.

vi.1. Effect of Unlike Physiological Modalities

Multi-modality of whatsoever stress detection scheme is proven to improve the accuracy and performance of the systems. However, the effects of each modality and their combinations are different when the functioning is taken into consideration. We examined the result of each modality. The heart action alone, heart activity and accelerometer combination, center activity and electrodermal activeness combination and all of the modalities together were investigated, equally presented in
Table three. We achieved the highest stress–activeness level detection accuracies when MLP was applied to the features from all the modalities. On the other hand, centre activity and electrodermal activeness combination for stress level detection achieved the best results with the logistic regression classifier.


Tabular array iii

Stress detection accuracies with different ML algorithms: three-course classification. On the left side, stress recognition results that only used HR and EDA signals are presented. On the right side, context information with accelerometer data is also added. The highest accuracy in every column is emphasized with bold.

Algorithm Stress Only Stress with Context
HR EDA HR + EDA Hr + EDA + ACC HR + ACC EDA + ACC
PCA + LDA 49.01 52.94 62.seventy 82.35 72.50 80.39
PCA + SVM (radial) 80.39 62.74 84.31 82.35 86.27 80.39
kNN 82.35 84.31 86.27 80.39 84.31 80.39
Logistic Regression 84.21 60.78 92.15 90.19 86.27 78.43
Random Forest 86.27 80.39 86.27 86.27 90.19 84.31
Multilayer Perception 86.27 68.62 ninety.nineteen 92.15 90.xix 82.35

We divided the operation evaluation into two categories. In the first category, heart activity and electrodermal activity signals were used to notice physiological stress levels. In the 2d category, we added accelerometer signal to these signals. With the addition of the data from the accelerometer sensor, information regarding the activity and the context of individuals were also evaluated in this category. To this extent, we chosen this category as “Stress with Context”. In our case, since we knew the context for all times, adding of accelerometer features to the feature vector might be little and these features increased the performance of our system. Nevertheless, we added these features to prove that context data is crucial in daily life studies; when information technology is completely unknown, adding them will likewise increase the performance of those systems. When we combined HR and EDA signals, the accurateness was higher than either signal alone in nigh all cases (in RF, it was equal to HR). Nosotros can infer from that using multiple modalities increases the performance of the stress level detection schemes.

The detailed stress with context detection accuracy results, f-measure, precision and recall values are presented in
Table iv,
Table v
and
Table 6.
Table 4
presents the classification accuracy results obtained from HR, EDA and ACC signals of the Empatica E4 device. The multilayer perceptron algorithm achieved the best classification accuracy of 92.fifteen%.


Tabular array 4

Stress with context classification accuracy, f-Measure, precision and recall values with different ML algorithms: three-course. 60 minutes + EDA +ACC for Empatica E4.

Algorithm HR + EDA + ACC (Empatica E4)
Accuracy f-Measure Precision Call back
PCA + LDA 82.35 82.xx 82.60 82.xl
PCA + SVM (radial) 82.35 82.50 83.30 82.40
kNN lxxx.39 80.40 eighty.80 lxxx.xl
Logistic Regression 90.19 xc.10 90.20 90.20
Random Forest 86.27 86.20 86.20 86.thirty
Multilayer Perceptron 92.xv 92.20 92.30 92.twenty


Table v

Stress with context classification accuracy, f-Measure, precision and recall values with different ML algorithms: three-class. Hour +ACC for Empatica E4.

Algorithm HR + ACC (Empatica E4)
Accuracy f-Measure Precision Recall
PCA + LDA 72.54 71.60 71.lxxx 72.five
PCA + SVM (radial) 86.27 86.20 86.xc 86.30
kNN 84.31 84.x 84.60 84.xxx
Logistic Regression 86.27 86.20 86.90 86.30
Random Woods 88.25 88.00 88.ten 88.xx
Multilayer Perceptron 92.19 90.xxx 91.40 90.twenty


Table 6

Stress with context nomenclature accuracy, f-Measure, precision and think values with different ML algorithms: iii-class. HR +ACC for all devices.

Algorithm Hour + ACC (All Devices)
Accuracy f-Measure Precision Remember
PCA + LDA 59.12 59.80 60.10 59.60
PCA + SVM (radial) 76.99 77.10 77.30 77.00
kNN 87.32 87.20 87.xxx 87.thirty
Logistic Regression 65.25 65.00 65.00 65.30
Random Forest 88.26 88.20 88.20 88.thirty
Multilayer Perceptron 83.09 83.00 83.20 83.10

In
Table v, nosotros go out the EDA betoken nerveless from Empatica E4 out, since Samsung Galaxy Gear devices do not have EDA sensors. In
Tabular array 6, we demonstrate the results from the data collected from eighteen participants for 32 h (nine days) by using 4 Empatica E4 and xiv Samsung Gear devices (All Devices). Note that we too nerveless data with the Samsung Gear S3 classic smartwatch from three participants. However, we did not use these data since part Samsung SDK no longer provides the RR interval of raw data. The Multilayer Perceptron algorithm accomplished the best consequence (92.nineteen%) from HR and ACC signals collected using Empatica E4, whereas the Random Forest algorithm gave the best classification accuracy (88.26%) with the Hour and ACC data collected from all devices.

6.2. Event of Device Type

Nosotros compared the result of using Empatica E4 and Samsung Gear South-S2 devices (Combination of Samsung Gear S and S2 devices) relative to each other. Samsung Gear Southward-S2 devices are a commercial type, relatively cheaper smartwatches. On the other hand, Empatica E4 is a more precise, relatively more expensive research device. Nosotros compared the nomenclature accuracies and data quality on both of these devices.

In the literature, RR intervals that differ more than than 20% of the local average are removed [38]. This is called the antiquity detection with a percent threshold. We changed the value of this threshold from ten% to 25% and observed the amount of remaining clean data. As shown in
Figure ix, Empatica E4 devices have approximately 25% more remaining information for all of the dissimilar artifact detection percentage thresholds. We deduced that the quality of RR intervals of Empatica E4 devices is higher than those of the Samsung Gear Southward-S2 devices. We farther investigated the effect of data quality on stress level nomenclature accuracies. We observed that nomenclature accuracies obtained from the data collected with Empatica E4 were higher than those from Samsung devices with all classifiers, as shown in
Tabular array 7
and
Tabular array 8. From these results, we tin can observe that the information quality has a significant effect on the stress level classification accuracy.


An external file that holds a picture, illustration, etc.
Object name is sensors-19-01849-g009.jpg

Percentage of the remaining information (for both device types) later on the artifacts are removed versus different pct thresholds of artifact detection.


Table 7

Consequence of the used device to three-class stress with context nomenclature accuracy when heart activity and accelerometer data are used together.

Algorithm Empatica E4 Samsung Gear S-S2 All Devices
PCA + LDA 88.88 72.sixty 59.12
PCA + SVM (rad) 92.06 78.60 76.91
kNN 87.30 85.30 87.30
Logistic Regression 90.47 83.30 65.25
Random Forest 90.40 88.threescore 88.30
Multilayer Perception 95.23 87.30 83.ten


Table viii

Effect of the device used to iii-course stress level classification accuracy when only heart activity signal is used (without context).

Algorithm Empatica E4 Samsung Gear S-S2 All Devices
PCA + LDA 65.07 55.33 52.58
PCA + SVM (rad) 90.xl 73.33 62.60
kNN 88.88 82.00 82.15
Logistic Regression 84.90 66.66 66.66
Random Forest 87.30 84.67 82.62
Multilayer Perception 88.88 78.00 71.36

half dozen.3. Effect of Artifact Detection Per centum Threshold, Interpolation and Aggregation Window on Accuracy

Physiological signals are sensitive to the movements of the subjects. Specially the quality of the heart charge per unit information declines very drastically in the example of intense concrete activities. Nosotros applied a few preprocessing techniques and filters to remove the contamination of the center rate information. We investigated the furnishings of artifact detection percentage threshold, interpolation and the assemblage window length. Artifact detection percentage threshold is the minimum percentage difference between a data betoken and the local average to evaluate the data point equally an artifact. If the value of the artifact correction percentage threshold increases, the filter loosens, i.eastward., the number of detected artifacts decreases. Furthermore, the aggregation window is the information segment in which features are extracted and averaged for the whole session to get the features of the session.

We applied artifact correction percentage thresholds from x% to 25% and investigated the stress level classification accuracies, every bit shown in
Tabular array 9. We were unable to observe a design when we practical different classifiers and changed the antiquity detection per centum thresholds. We tin infer that irresolute this threshold does not have a clear effect on classification accuracy.


Tabular array nine

Classification accuracies vs. irresolute pct based artifact detection and filtering rules.

Algorithm 10% fifteen% twenty% 25%
PCA + LDA 64.28 62.38 59.62 63.80
PCA + SVM(rad) 80.95 78.57 77.00 79.52
kNN 87.61 86.66 87.32 85.20
Logistic Regression 73.80 61.xc 66.25 66.19
Random Forest 89.00 88.09 89.26 82.60
Multilayer Perception eighty.00 78.57 83.09 eighty.95

Nosotros further examined the effect of assemblage window on the stress level classification accurateness (see
Table ten). Nosotros changed the length of the assemblage window from ii min to 20 min. We observed that the behavior changed for each ML algorithm. Researchers should consider the ML algorithm and its performance of different assemblage window sizes when deciding the optimum window length. Gjoreski et al. found that the aggregation window lengths betwixt 10 min and 17.5 min have better accuracy in full general [34], which is similar to our results.


Table 10

Outcome of the length of the aggregation window on classification accuracies.

Algorithm Aggregation Window Size (s)
120 300 600 1200
PCA + LDA 59.62 62.24 54.14 63.02
PCA + SVM (radial) 76.99 77.94 77.27 83.33
kNN 87.32 83.30 88.38 85.41
Logistic Regression 65.25 69.threescore 72.22 76.16
Random Forest 88.26 86.76 87.87 84.14
Multilayer Perception 83.09 86.76 81.81 88.54

Equally mentioned above, we provide a selector in our centre rate preprocessing tool that decides whether to interpolate the removed artifacts or remove and apply some minimum consecutive rules. Minimum consecutive rules could be either the minimum required number of samples or the time interval for a segment to extract features. We further investigated the effect of removal and interpolation to the classification accuracies. In
Table 11, nosotros can run across that applying interpolation achieved higher performance than filtering for some car learning methods (removal and minimum consecutive filter) and lower results for other algorithms. This decision depends on the practical ML method.


Tabular array 11

Classification accuracies of Empatica E4 when removed inter-trounce interval artifacts are replaced with interpolation vs. when they are removed.

Algorithm Filtering Interpolation
PCA + LDA 72.72 fifty.75
PCA + SVM 89.39 89.39
kNN 95.45 97.72
Logistic Regression 83.33 89.39
Random Forest 95.45 93.93
Multilayer Perception 89.39 95.45

half-dozen.4. Person Contained and Dependent Models

We adult two dissimilar stress detection systems. The first one is the general (person-contained) model. In this model, the collected data from all of the participants are divided into training and exam segments without considering the participant labels. By employing 10-fold cross-validation, the accuracy of the system is determined independently from whatever individual’southward information. The second model is the person dependent model. In this model, the information collected from different participants are divided. Afterward this division, preparation and test partitions are divided for each person and models are developed for each participant. The classification accuracy is calculated for each individual and an average of all accuracies of the participants is presented. Since everyone has a particular stress beliefs and person-specific models take just individuals information into consideration when developing models, these models are expected to accept a college operation. We nowadays the accuracy results in
Table 12. We observed that person specific stress detection models had higher nomenclature accuracies than general models, equally expected. Furthermore, we achieved the highest classification accurateness on person-specific models with Empatica E4 devices when the Random Forest algorithm was applied (97.92%) to features from all signals. With all algorithms, HR, EDA and ACC signal combination with Empatica E4 devices had higher accuracy than with all devices in person-specific models. These results demonstrate that stress level detection schemes should give more weight to the private’southward information than data from other people when edifice models.


Table 12

Nomenclature accuracies of general and person-specific models.

Algorithm General Person Specific
60 minutes + EDA + ACC-E4 60 minutes + ACC-All HR + EDA + ACC-E4 60 minutes + ACC-All
PCA + LDA 82.35 59.12 95.83 87.60
PCA + SVM (radial) 82.35 76.99 93.75 85.98
kNN 80.39 87.30 95.83 89.91
Logistic Regression 90.nineteen 65.25 95.83 90.17
Random Forest 86.27 88.twenty 97.92 90.17
MLP 92.15 83.twenty 95.83 91.54

6.5. Measuring the Perceived and Physiological Stress

Equally tin can be seen in the literature, in the aforementioned context, the perceived stress and physiological stress of individuals can be different. We investigated the effect of two different basis truth drove methods in this subsection. The first one is the known context equally the ground truth. In our case, the contest context was assumed to induce stress, the lecture context was causeless to gives some cognitive load and a lower amount of stress and free time was causeless to be relaxed sessions. In this method, the ground truth is enumerated from the known context as: gratuitous day, 1; lecture, 2; and contest, 3. When we examined the physiological signals, we could differentiate these three levels with loftier classification accuracies (see
Table thirteen). Perceived stress of individuals was also measured. Nosotros asked the following question to the participants to learn their stress level:

How irritated, stressed, and annoyed versus content, relaxed, and complacent did you lot feel during the task?


Table 13

Classification accuracies comparison of subjective report and known context. On the left, known context information (Complimentary,1; Lecture, 2; Competition, 3) was used equally class labels. On the correct, subjective ground truths are used equally class labels.

Algorithm Accurateness Wrt. Known Context Accurateness Wrt. Subjective Basis Truth
Hr + ACC-All Hr + EDA + ACC-E4 HR + ACC-All 60 minutes + EDA + ACC-E4
PCA + LDA 59.12 82.35 54.46 50.98
PCA + SVM (radial) 76.99 82.35 69.01 72.55
kNN 87.xxx eighty.39 85.44 78.44
Logistic Regression 65.25 90.19 57.27 78.43
Random Woods 88.20 86.27 86.38 76.47
MLP 83.20 92.15 80.28 68.62

The answers were on a scale of 0–100 with 5-point increments. We determined the stress level as ane if the reply was between 0 and thirty. Stress level was assigned to 2 if the answer was 35–75. The highest stress level (3) was assigned if the respond was at least eighty.

Read:   Blood Type a Would Give Which of the Following Results

When we look at the classification accuracies of the perceived stress level, for all ML algorithms and signal combinations, they were lower than the corresponding physiological stress level. This is because the perceived stress is subjective, depending on the individual. The survey answers were also prone to error and this might exist another reason for the decrease in the stress level detection accuracies. The correlation between the known context and perceived stress labels was computed to be 0.356, which is a moderate relation. The relation between the perceived stress and physiological stress is not investigated thoroughly in the literature. Development of personal perceived stress level models and filtering out the outlier survey answers might increment the performance of the classifiers.

7. Conclusions

We developed a stress detection scheme to be used in real life. Since our system employs unobtrusive wearable devices, it can hands exist used in the daily life of individuals. It can track the stress levels in existent-time and intervene if an extreme level of stress is detected. After the detection, some stress management methods can besides be offered to convalesce the high level of stress. We nerveless data from participants of an algorithmic programming competition to evaluate the operation of our system. We obtained labeled sessions for 21 subjects for nine days. Nosotros described the difficulties of existent-life data collection, which do not occur in laboratory environments. After describing our algorithms, we presented the results. For three-grade stress level detection, we obtained 90.40% accuracy by using Empatica E4 devices with high information quality, whereas the accuracy with Samsung S devices was 84.67%. We tin deduce from the results that the data quality of the devices increases the stress level classification accuracies. When compared with other real-life studies in
Table 1, our system has higher accuracies fifty-fifty for the results obtained with the three-grade classification. Later examining the upshot of different preprocessing methods and parameters, we can infer that their effect depends on the chosen ML algorithm. Researchers should select these methods past considering their performance with selected ML algorithms. Furthermore, person-specific models have always higher classification accuracies than full general models. We accomplished maximum 97.92% accurateness for 3-level stress detection with our person-specific models. On the other hand, nosotros obtained a maximum 88.20% classification accuracy with our general models. When physiological data from each person are sufficient for developing person-specific models, they should be applied. Otherwise, people should be clustered co-ordinate to their stress behaviors and models can be developed for clusters to increase the operation of general models. The best performing classifiers were the Random Woods and the Multilayer Perceptron algorithms. These classifiers outperformed other algorithms in almost cases. Another significant finding is that the combination of modalities increases the performance of our organisation. When we combined center activity with electrodermal activeness, we obtained 92.15% maximum 3-level classification accuracy, whereas this was 86.27% when these modalities were used separately. Finally, we observed that the perceived stress level classification results in lower accuracies than physiological stress level nomenclature. At that place were up to fifteen% decrease when compared with physiological stress level nomenclature accuracies. The possible causes of this subtract could be listed equally subjectivity and fallibility of cocky-report answers and the possible difference of physiological and psychological responses of individuals to the same stressor. Every bit a futurity piece of work, we plan to record data with an increased number of loftier-quality Empatica E4 devices. Nosotros further plan to develop personalized perceived stress level models from basis truth surveys and remove outlier answers to increase the perceived stress level classification accuracies.

Acknowledgments

We would like to show our gratitude to INZVA for providing us the opportunity for the data collection in their summertime camp.

Writer Contributions

Y.South.C., N.C. and D.E. contributed equally to this work in design, implementation, field study, information analysis and writing the manuscript. C.Due east. provided invaluable feedback and technical guidance to translate the design and details of the field study. C.E. also performed comprehensive disquisitional editing to increase the overall quality of the manuscript.

Funding

This work was supported by AffecTech: Personal Technologies for Affective Wellness, Innovative Training Network funded by the H2020 People Programme nether Marie Skłodowska-Curie grant agreement No. 722022. This work was supported by the Turkish Directorate of Strategy and Upkeep under the TAM Project number DPT2007K120610.

Conflicts of Involvement

The authors declare no conflict of involvement.

References

1.
Picard R.W. Automating the Recognition of Stress and Emotion: From Lab to Real-World Bear upon.

IEEE MultiMedia.
2016;23:3–7. doi: 10.1109/MMUL.2016.38.

[CrossRef] [Google Scholar]

ii.
England M.J., Liverman C.T., Schultz A.M., Strawbridge L.M. Epilepsy across the spectrum: Promoting health and understanding: A summary of the Plant of Medicine report.

Epilepsy Behav.
2012;25:266–276. doi: 10.1016/j.yebeh.2012.06.016.

[PMC free commodity]
[PubMed] [CrossRef] [Google Scholar]

three.
Ryvlin P., Nashef 50., Lhatoo South.D., Bateman L.M., Bird J., Bleasel A., Boon P., Crespel A., Dworetzky B.A., Høgenhaven H., et al. Incidence and mechanisms of cardiorespiratory arrests in epilepsy monitoring units (MORTEMUS): A retrospective study.

Lancet Neurol.
2013;12:966–977. doi: 10.1016/S1474-4422(13)70214-X.

[PubMed] [CrossRef] [Google Scholar]

4.
Colligan T.W., Higgins E.Chiliad. Workplace stress: Etiology and consequences.

J. Workplace Behav. Health.
2006;21:89–97. doi: ten.1300/J490v21n02_07.

[CrossRef] [Google Scholar]

five.
American Psychology Association .
Stress: The Dissimilar Kinds of Stress.
American Psychology Clan; Washington, DC, USA: 2019.
[Google Scholar]

6.
Krantz D.S., Whittaker K.S., Sheps D.S.
Heart and Listen: Evolution of Cardiac Psychology.
American Psychological Association; Washington, DC, USA: 2011. Psychosocial risk factors for coronary heart illness: Pathophysiologic mechanisms.
[Google Scholar]

vii.
Pickering T.G. Mental stress as a causal cistron in the development of hypertension and cardiovascular affliction.

Curr. Hypertens. Rep.
2001;3:249–254. doi: 10.1007/s11906-001-0047-1.

[PubMed] [CrossRef] [Google Scholar]

8.
Mönnikes H., Tebbe J., Hildebrandt M., Arck P., Osmanoglou Due east., Rose K., Klapp B., Wiedenmann B., Heymann-Mönnikes I. Function of stress in functional gastrointestinal disorders.

Dig. Dis.
2001;nineteen:201–211. doi: x.1159/000050681.

[PubMed] [CrossRef] [Google Scholar]

9.
Herbert J. Fortnightly review: Stress, the encephalon, and mental illness.

BMJ.
1997;315:530–535. doi: x.1136/bmj.315.7107.530.

[PMC free commodity]
[PubMed] [CrossRef] [Google Scholar]

10.
Milczarek M., Elke Schneider Eastward.G.
OSH in Figures, Stress at Work, Fact and Figures.
European Agency for Safety and Health at Work; Bilbao, Spain: 2009.
[Google Scholar]

eleven.
European Bureau for Rubber and Wellness at Work .
European Opinion Poll on Occupational Condom and Health.
European Agency for Safety and Health at Work; Bilbao, Spain: 2013. [CrossRef] [Google Scholar]

12.
Alberdi A., Aztiria A., Basarab A. Towards an automatic early stress recognition system for office environments based on multimodal measurements: A review.

J. Biomed. Inform.
2016;59:49–75. doi: x.1016/j.jbi.2015.11.007.

[PubMed] [CrossRef] [Google Scholar]

14.
Costin R., Rotariu C., Pasarica A. Mental stress detection using heart rate variability and morphologic variability of EeG signals; Proceedings of the 2012 International Conference and Exposition on Electrical and Ability Technology; Iaşi, Romania. 25–27 October 2012; pp. 591–596.
[Google Scholar]

xv.
De Santos Sierra A., Avila C.S., Casanova J.1000., del Pozo G.B. A Stress-Detection System Based on Physiological Signals and Fuzzy Logic.

IEEE Trans. Ind. Electron.
2011;58:4857–4865. doi: 10.1109/TIE.2010.2103538.

[CrossRef] [Google Scholar]

16.
Soury M., Devillers L. Stress Detection from Audio on Multiple Window Assay Size in a Public Speaking Task; Proceedings of the 2013 Humaine Clan Conference on Melancholia Computing and Intelligent Interaction; Geneva, Switzerland. 2–5 September 2013; pp. 529–533.
[Google Scholar]

17.
Wijsman J., Grundlehner B., Liu H., Hermens H., Penders J. Towards mental stress detection using wearable physiological sensors; Proceedings of the 2011 Annual International Conference of the IEEE Engineering science in Medicine and Biology Gild; Boston, MA, U.s.. 30 Baronial–3 September 2011; pp. 1798–1801. [PubMed] [Google Scholar]

18.
Abouelenien M., Burzo M., Mihalcea R. Human Acute Stress Detection via Integration of Physiological Signals and Thermal Imaging; Proceedings of the 9th ACM International Conference on PErvasive Technologies Related to Assistive Environments; Corfu Isle, Greece. 29 June–1 July 2016; New York, NY, USA: ACM; 2016. pp. 32:i–32:8.
[Google Scholar]

xix.
Aigrain J., Dubuisson S., Detyniecki One thousand., Chetouani M. Person-specific behavioural features for automated stress detection; Proceedings of the 2015 11th IEEE International Conference and Workshops on Automatic Confront and Gesture Recognition (FG); Ljubljana, Slovenia. 4–8 May 2015; pp. i–6.
[Google Scholar]

20.
Aigrain J., Spodenkiewicz 1000., Dubuisson S., Detyniecki M., Cohen D., Chetouani Thou. Multimodal stress detection from multiple assessments.

IEEE Trans. Touch on. Comput.
2016;9:491–506. doi: 10.1109/TAFFC.2016.2631594.

[CrossRef] [Google Scholar]

21.
Giannakakis G., Pediaditis G., Manousos D., Kazantzaki East., Chiarugi F., Simos P.G., Marias G., Tsiknakis M. Stress and feet detection using facial cues from videos.

Biomed. Point Process. Control.
2017;31:89–101. doi: 10.1016/j.bspc.2016.06.020.

[CrossRef] [Google Scholar]

22.
Baltaci S., Gokcay D. Role of student dilation and facial temperature features in stress detection; Proceedings of the 2014 22nd Signal Processing and Communications Applications Conference (SIU); Trabzon, Turkey. 23–25 April 2014; pp. 1259–1262.
[Google Scholar]

23.
Mozos O.Thou., Sandulescu V., Andrews S., Ellis D., Bellotto N., Dobrescu R., Ferrandez J.M. Stress Detection Using Clothing Physiological and Sociometric Sensors.

Int. J. Neural Syst.
2017;27:1650041. doi: 10.1142/S0129065716500416.

[PubMed] [CrossRef] [Google Scholar]

24.
Zubair G., Yoon C., Kim H., Kim J., Kim J. Smart Vesture Band for Stress Detection; Proceedings of the 2015 5th International Briefing on IT Convergence and Security (ICITCS); Kuala Lumpur, Malaysia. 24–27 Baronial 2015; pp. 1–four.
[Google Scholar]

25.
Hong J.H., Ramos J., Dey A.K. Understanding Physiological Responses to Stressors During Concrete Activity; Proceedings of the 2012 ACM Briefing on Ubiquitous Computing; Pittsburgh, PA, The states. 5–viii September 2012; New York, NY, USA: ACM; 2012. pp. 270–279.
[Google Scholar]

26.
Akmandor A.O., Jha N.One thousand. Keep the Stress Abroad with SoDA: Stress Detection and Alleviation System.

IEEE Trans. Multi-Scale Comput. Syst.
2017;3:269–282. doi: 10.1109/TMSCS.2017.2703613.

[CrossRef] [Google Scholar]

27.
Akhonda M.A.B.S., Islam S.K.F., Khan A.S., Ahmed F., Rahman G.One thousand. Stress detection of estimator user in office like working environment using neural network; Proceedings of the 2014 17th International Conference on Computer and Information Technology (ICCIT); Dhaka, Bangladesh. 22–23 December 2014; pp. 174–179.
[Google Scholar]

28.
Mohd Yard.H., Kashima K., Sato K., Watanabe M. Mental stress recognition based on non-invasive and non-contact measurement from stereo thermal and visible sensors.

Int. J. Affect. Eng.
2015;fourteen:9–17. doi: 10.5057/ijae.fourteen.nine.

[CrossRef] [Google Scholar]

29.
Liapis A., Katsanos C., Sotiropoulos D., Xenos Yard., Karousos N. Stress Recognition in Human being-computer Interaction Using Physiological and Self-reported Information: A Study of Gender Differences; Proceedings of the 19th Panhellenic Briefing on Information science; Athens, Greece. 1–3 October 2015; New York, NY, USA: ACM; 2015. pp. 323–328.
[Google Scholar]

xxx.
Huang G.X., Li J., Ngai G., Leong H.Five. StressClick: Sensing Stress from Gaze-Click Patterns; Proceedings of the 2016 ACM on Multimedia Briefing; Amsterdam, The Netherlands. 15–nineteen Oct 2016; New York, NY, The states: ACM; 2016. pp. 1395–1404.
[Google Scholar]

31.
Fernández J.R.G., Anishchenko L. Mental stress detection using bioradar respiratory signals.

Biomed. Signal Procedure. Control.
2018;43:244–249. doi: 10.1016/j.bspc.2018.03.006.

[CrossRef] [Google Scholar]

32.
Ciman M., Wac K. Individuals’ stress cess using human-smartphone interaction assay.

IEEE Trans. Touch on. Comput.
2016;9:51–65. doi: 10.1109/TAFFC.2016.2592504.

[CrossRef] [Google Scholar]

33.
Gjoreski K., Gjoreski H., Luštrek One thousand., Gams M. Continuous Stress Detection Using a Wrist Device: In Laboratory and Real Life; Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct; Heidelberg, Germany. 12–sixteen September 2016; New York, NY, USA: ACM; 2016. pp. 1185–1193.
[Google Scholar]

34.
Gjoreski M., Luštrek K., Gams M., Gjoreski H. Monitoring stress with a wrist device using context.

J. Biomed. Inform.
2017;73:159–170. doi: 10.1016/j.jbi.2017.08.006.

[PubMed] [CrossRef] [Google Scholar]

35.
Vildjiounaite Eastward., Kallio J., Kyllönen V., Nieminen M., Mäntyjärvi J., Gimel’farb M. Unobtrusive stress detection on the basis of smartphone usage data.

Pers. Ubiquitous Comput.
2018;22:671–688. doi: 10.1007/s00779-017-1108-z.

[CrossRef] [Google Scholar]

36.
Hernandez J., Morris R.R., Picard R.Due west.
Melancholia Computing and Intelligent Interaction—Volume Part I, Proceedings of the fourth International Conference on Affective Calculating and Intelligent Interaction, Memphis, TN, USA, 9–12 Oct 2011.
Springer; Berlin/Heidelberg, Germnay: 2011. Call Middle Stress Recognition with Person-specific Models; pp. 125–134.
[Google Scholar]

37.
Garcia-Ceja E., Osmani Five., Mayora O. Automatic Stress Detection in Working Environments From Smartphones’ Accelerometer Information: A Get-go Stride.

IEEE J. Biomed. Health Inform.
2016;20:1053–1060.

[PubMed] [Google Scholar]

38.
Cinaz B., Arnrich B., Marca R., Tröster M. Monitoring of Mental Workload Levels During an Everyday Life Part-work Scenario.

Pers. Ubiquitous Comput.
2013;17:229–239. doi: x.1007/s00779-011-0466-1.

[CrossRef] [Google Scholar]

39.
Healey J.A., Picard R.W. Detecting stress during real-world driving tasks using physiological sensors.

IEEE Trans. Intell. Transp. Syst.
2005;6:156–166. doi: 10.1109/TITS.2005.848368.

[CrossRef] [Google Scholar]

40.
Chen L.l., Zhao Y., Ye P.F., Zhang J., Zou J.Z. Detecting driving stress in physiological signals based on multimodal characteristic analysis and kernel classifiers.

Expert Syst. Appl.
2017;85:279–291. doi: x.1016/j.eswa.2017.01.040.

[CrossRef] [Google Scholar]

41.
Castaldo R., Xu Due west., Melillo P., Pecchia 50., Santamaria L., James C. Detection of mental stress due to oral academic examination via ultra-brusque-term HRV assay; Proceedings of the 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC); Orlando, FL, USA. 16–twenty August 2016; pp. 3805–3808. [PubMed] [Google Scholar]

42.
Gjoreski M., Gjoreski H., Lutrek Grand., Gams Thou. Automatic Detection of Perceived Stress in Campus Students Using Smartphones; Proceedings of the 2015 International Conference on Intelligent Environments; Prague, Czech Republic. 15–17 July 2015; pp. 132–135.
[Google Scholar]

43.
Bogomolov A., Lepri B., Ferron M., Pianesi F., Pentland A.South. Pervasive stress recognition for sustainable living; Proceedings of the 2014 IEEE International Conference on Pervasive Computing and Communication Workshops (PERCOM WORKSHOPS); Budapest, Hungary. 24–28 March 2014; pp. 345–350.
[Google Scholar]

44.
Bauer 1000., Lukowicz P. Can smartphones detect stress-related changes in the behaviour of individuals?; Proceedings of the 2012 IEEE International Conference on Pervasive Computing and Communications Workshops; Lugano, Switzerland. nineteen–23 March 2012; pp. 423–426.
[Google Scholar]

45.
Sysoev M., Kos A., PogaăźNik 1000. Noninvasive Stress Recognition Considering the Current Activeness.

Pers. Ubiquitous Comput.
2015;19:1045–1052. doi: 10.1007/s00779-015-0885-5.

[CrossRef] [Google Scholar]

46.
Wyatt J., Liu J. Basic concepts in medical informatics.

J. Epidemiol. Customs Wellness.
2002;56:808–812. doi: 10.1136/jech.56.11.808.

[PMC free article]
[PubMed] [CrossRef] [Google Scholar]

47.
Shelley K., Shelley S. Pulse oximeter waveform: photoelectric plethysmography. In: Lake C., Hines R., Blitt C., editors.
Clinical Monitoring.
WB Saunders Company; Philadelphia, PA, USA: 2001. pp. 420–428.
[Google Scholar]

48.
Sharma North., Gedeon T. Objective measures, sensors and computational techniques for stress recognition and classification: A survey.

Comput. Methods Programs Biomed.
2012;108:1287–1301. doi: 10.1016/j.cmpb.2012.07.003.

[PubMed] [CrossRef] [Google Scholar]

49.
Can Y.S., Arnrich B., Ersoy C. Stress detection in daily life scenarios using smart phones and clothing sensors: A survey.

J. Biomed. Inform.
2019;92:103139. doi: 10.1016/j.jbi.2019.103139.

[PubMed] [CrossRef] [Google Scholar]

50.
Taylor Due south., Jaques N., Chen Due west., Fedor S., Sano A., Picard R. Automated identification of artifacts in electrodermal activity data; Proceedings of the 2015 37th Almanac International Briefing of the IEEE Engineering in Medicine and Biology Social club (EMBC); Milan, Italian republic. 25–29 August 2015; pp. 1934–1937.
[PMC free article]
[PubMed] [Google Scholar]

51.
Greco A., Valenza K., Lanata A., Scilingo E.P., Citi L. cvxEDA: A Convex Optimization Approach to Electrodermal Activeness Processing.

IEEE Trans. Biomed. Eng.
2016;63:797–804. doi: 10.1109/TBME.2015.2474131.

[PubMed] [CrossRef] [Google Scholar]

52.
Vollmer M.
MarcusVollmer/HRV Toolbox.
GitHub; San Francisco, CA, USA: 2019.
[Google Scholar]

53.
Lomb N.R. Least-squares frequency assay of unequally spaced data.

Astrophys. Space Sci.
1976;39:447–462. doi: 10.1007/BF00648343.

[CrossRef] [Google Scholar]

54.
Ekman P. Differential communication of bear on by head and body cues.

J. Pers. Soc. Psychol.
1965;two:726–735. doi: 10.1037/h0022736.

[PubMed] [CrossRef] [Google Scholar]

55.
Eibe F., Hall Yard., Witten I.
The WEKA Workbench. Online Appendix for Information Mining: Practical Machine Learning Tools and Techniques.
Morgan Kaufmann; Burlington, MA, USA: 2016.
[Google Scholar]

57.
International Collegiate Programming Contest. [(accessed on six January 2019)];
Available online: https://icpc.baylor.edu/

58.
World Medical Clan World Medical Association Declaration of Helsinki. Ethical principles for medical research involving human subjects.

Bull. World Health Organ.
2001;79:373.

[PMC free commodity]
[PubMed] [Google Scholar]


Articles from
Sensors (Basel, Switzerland)
are provided here courtesy of
Multidisciplinary Digital Publishing Institute (MDPI)


The Body’s Automatic Stress Detection Process Relies on

Source: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6515276/

Check Also

Which Phrase Best Describes the Context of a Speech

Which Phrase Best Describes the Context of a Speech. C the amount of prove. D …