See discussions, stats, and author profiles for this publication at: [Link]

net/publication/357754986

Stress Analysis and Care Prediction System for Online Workers

Conference Paper · December 2021

DOI: 10.1109/ICAC54203.2021.9671106

CITATIONS READS

2 2,040

6 authors, including:

Pradeep Kumara Wijesekara Abeygunawardhana Suranjini Silva

Sri Lanka Institute of Information Technology Sri Lanka Institute of Information Technology
79 PUBLICATIONS 408 CITATIONS 7 PUBLICATIONS 21 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Pradeep Kumara Wijesekara Abeygunawardhana on 30 January 2022.

The user has requested enhancement of the downloaded file.

 Stress Analysis and Care Prediction System for
Online Workers
Amarasinghe A.A.S.M Malassri I.M.S. Weerasinghe K.C.N.
Department of Information Technology Department of Information Technology Department of Information Technology
Sri Lanka Institute of Information Sri Lanka Institute of Information Sri Lanka Institute of Information
Technology Technology Technology
Colombo, Sri Lanka Colombo, Sri Lanka Colombo, Sri Lanka
it17059428@[Link] it181473608@[Link] it180209224@[Link]
2021 3rd International Conference on Advancements in Computing (ICAC) | 978-1-6654-0862-2/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICAC54203.2021.9671106

Jayasingha I.B. Pradeep K. W. Abeygunawardhana Suranjini Silva

Department of Information Technology Department of Computer System Department of Computer System
Sri Lanka Institute of Information Engineering Engineering
Technology Sri Lanka Institute of Information Sri Lanka Institute of Information
Colombo, Sri Lanka Technology Technology
it181614344@[Link] Colombo, Sri Lanka Colombo, Sri Lanka
pradeep.a@[Link] suranjini.s@[Link]

Abstract— Working from home (WFH) online during the

covid-19 pandemic has caused increased stress level. Online Using a webcam while the user was working, System
workers/students have been affecting by the crisis according recognized stressful behaviors (sleeping while working,
to new researches. Natural response of body, to external and anxious behaviors, frequent yawning, and smoking while
internal stimuli is stress. Even though stress is a natural working) throughout a one-hour period as the first
occurrence, prolonged exposure while working Online to parameter. It also detects seven facial expressions of the
stressors can lead to serious health problems if any action will users. Following the identification of stressful behaviors
not be applied to control it. Our research has been conducted and expressions, the system counts them and calculates
deeply to identify the best parameters, which have connection their percentage value. The system delivers Preserved
with stress level of online workers. As a result of our research,
a desktop application has been created to identify the users
Stress Scale (PSS) [2] to the user for validation. Then add
stress level in real time. According to the results, our overall the two outputs together and show the user the ultimate
system was able to provide outputs with more than 70% stress level as a percentage number.
accuracy. It will give best predictions to avoid the health
Stress is effect for the person in cognitively while
problems. Our main goal is to provide best solution for the
online workers to have healthy lifestyles. Updates for the users
fatigue affect for person physically. However, severe stress
will be provided according to the feedback we will have in the feels for individual as physical fatigue, which called stress
future from the users. Our System will be a most valuable related fatigue condition. Keystroke dynamic is another
application in the future among online workers. ideal way to detect of stress related fatigue condition of a
person.
Keywords— stress detection, fatigue recognition, emotion
detection, behavior detection, keystroke detection, CNN, VGG16
We developed an IOT (Internet of Things) based device
I. INTRODUCTION with emended with computer mouse. In addition, use heart
People experience a lot of stress because of their rate variability to real-time stress predictions of an online
demanding employment working with electronic devices. worker using the extracted data from the device. The
Because of the pandemic situation, it has been increasing algorithm forecasts diseases that the user will be affected
day by day since people are working from home. Stress can by in the future based on recognized results. System
caused by situations such as repeated exposure to risk, short provides stress-relieving recommendations for the user
deadlines, difficult activities, exams, and repetitious tasks, based on the confirmed stress value. Our solution will help
which assigned by the companies and the universities. all the people who work with computers most of the time
Nearly one-third of workers in Europe and the US feel in their day-to-day life.
stressed at work according to the researches. Workplace
stress and anxiety can lead to decreased productivity and II. LITERATURE REVIEW
absenteeism, costing the economy an estimated 3% to 4%
Before planning our system, we researched about the
of GDP (Gross Domestic Product) [1].
past studies that are done based on all the parameters that
Most of the time, the employee is unaware that they have we selected. Most of the research have been done by
an issue with their working life. As a result, early detection physiological parameters. [3] It is difficult to get those
of mental health issues is particularly critical. Because of parameters from day by day from online workers.
our research, we developed a tool, which can help users to Therefore, we decided to use Machine learning and IOT
prevent health issues. “The online worker stress analysis (Internet of Things) technologies to make it easy for the
tool” assists online workers in identifying their stress levels users to detect their stress level.
while on the job and studies. Most of the past researches
According to the followed researches, they contain
have been conducting based on one or two parameters. In
about identifying the different stress behaviors [4]-[5].
our research, we selected four main parameters to detect
When they identified the stressed behaviors, but they have
stress level as well as fatigue level of the online worker.

978-1-6654-0862-2/21/$31.00 ©2021 IEEE

329

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on January 30,2022 at 00:51:45 UTC from IEEE Xplore. Restrictions apply.
 not predicted the user’s stress level, using that identified their research, and they were able to obtain subject
behaviors and they provided the stress releasing tips independent classification accuracy of up to 74 percent and
according to the identified stress level. Most of them subject dependent classification accuracy of up to 91 %.
considered about bellow variables.
Stress is effect for the person in cognitively while fatigue
 Stressed based diseases affect for person physically. If person suffers both of this
 Stressed behaviors condition in same time that cause of cognitive stress of
person, that is called stress related fatigue condition. There
In this proposed system, contain above functions, is high probability of doing unintentional mistakes [1] by
 Identify stressed behaviors (1) people who suffered these conditions. Therefore, prior
 Predict stress level (2) detection of stress, fatigue or stress related fatigue
 Display stress level in dashboard (3) condition is imported as well as important to make the
 Validate stress level using PSS (4) method of detection unobtrusive. Therefore, detection of
changes of key typing pattern (keystroke dynamics) and
 Suggest the tips for recover from stress (5)
Pattern changes in Mouse movements is another ideal way
TABLE I. RESEARCH GAP to detect above conditions of a person.
Research (1) (2) (3) (4) (5)
Research 01 [4] ✔ ✖ ✖ ✖ ✖ Many Researches use different techniques to detect
Research 02 [6] ✔ ✖ ✖ ✖ ✖ individual cognitive stress level using their typing patterns.
Javier H et el [11] tries to detect individual stress condition
Stress Analysis and Care ✔ ✔ ✔ ✔ ✔ by using pressure sensitive keyboard and mouse in
Prediction System
experimental room and giving transcript to do some tasks.
While doing that task they collected data using background
For many years, researchers have been researching app. They detect 79% of people’s typing pressure increased
automatic stress detection by collecting images of facial and 75% of people get more contact with mouse while they
expressions. From intrusive methods, such as saliva or are in stressful conditions. Suranga et el[12] calculate
blood testing, to less intrusive methods, such as picture individual stress while using in specific application
collecting of the people. (Microsoft word). keypress duration in selected key pairs
(diagraphs[12]) and keypress duration of selected three key
One of these studies [7] used a web camera installed
pairs (triagraph[12]) and rate of backspace and delete press
within a car's dashboard to capture and gather pictures of
As well as stress many researches interested in fatigue
the driver's face to detect stress. As per that research, the
detection using keystroke dynamic data, Robertas et al [13]
gathered pictures categorized into six facial expressions
detect fatigue by feeding 5 features generated by keypress
using Support Vector Machines (SVMs) trained on
and release times to the k-nearest neighbors (KNN). They
publicly gathered facial expression dataset. Following that,
used dataset provided which consist of data fifty-three
an algorithm counts the classifications within a certain time
people type specific text in fifteen times of a day. James D.
gap. The driver considered stressed if the number of anger
Hayes used custom modified keyboard and mouse to
and disgust facial expression count reaches a certain pre-
calculate fatigue on keypress pressure and mouse
assigned level. Not only were photos from public datasets
movements
used to train the best classifier, but also images of the
people posing for the stress categorization. As a result, the In the study [14] built a system to forecast stress using
models will be able to adapt to the way the individuals sensor data and personal data, by continuously tracking and
display their facial emotions. They were able to classify measuring data in a realistic context The platform is
stress with a 90.5 percent accuracy. primarily comprised of an Empatica E3 wristband, a stress
tracker iPhone agent (to store and transmit signals from the
In a study [8] created a method for stress detection that
wristband), mobile motion sensors, and a daily
gathers Remote Photo plethysmography (rPPG) signals
questionnaire. They visualized real-time and periodic
using the web camera of computers. After that, the rPPG
displays of rPPG signals with annotations using a web
data converted into a sinusoidal n wave, which represented
application. They used five healthy volunteers for the trial,
the heart rate. With 94.40 percent accuracy, an SVM
who worked 8–10 hours each week. They predicted stress
classifier produced the results.
using machine learning. With a f1-score of 0.91, the
A research team [9] created a system that uses video- random forest classifier outperformed the others.
recorded face signals to detect stress/anxiety emotional
states of the people. They retrieved the most significant III. METHODOLOGY
characteristics for stress categorization using a variety of
approaches such as Active Appearance Models, Optical Stress analysis system is designed for online workers.
Flow, and rPPG. A K-NN classifier achieved the greatest During their day today works system identifies behaviors,
classification accuracy of 87.72 percent. facial expressions, keystroke and heart rate of the user.
In the study [10], a method for identifying stress
indicators using Facial Action Units (FAUs) derived from
movies was proposed. They have used multiple basic
classifiers to conduct binary classification on FAUs
extracted from each video frames, which they gathered in

330

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on January 30,2022 at 00:51:45 UTC from IEEE Xplore. Restrictions apply.
 b) Training the Dataset
The architecture used to train the model is called a
Convolutional Neural Network (CNN). 256 layers were
employed to enhance the model's accuracy. Increased the
epochs to their maximum for the optimum accuracy. Then
develop the algorithm to split images to testing and training.
The model is trained using images from the training
category. Then saved the model for testing phase.

c) Testing the model

Based on the training data and created model, extract
stressful behaviors from a live video feed and deliver a
labeled output with high accuracy.

d) Validate Stressed Level

Fig. 1. System Diagram For one hour, the system recognized stressful behaviors and
sent them to a database. After one hour, the system gives the
This system contains four main sub topics. Methodology user the Preserved Stress Scale (PSS) [2]. The algorithm
of each component are described end to end in this section. then compares the number of detected stressful behaviors to
the PSS findings given by the user.
A. Monitor the behaviors of online workers and identify
their stressed behaviors e) Display Stress Level to User
According to finalized stress level, percentage value of level
is display to user in dashboard. Then display stressed
releasing tips for displayed stress level.

B. Detect stress level using facial expressions via

Webcam
The purpose of the function is to recognize and alert the
user when he exhibits stress-related facial expressions
using video/pictures. The convenience of webcams when
working with computers is a benefit of employing images
Fig. 2. Model developing flow for personal stress detection. The application will operate
in the background and monitor the user's facial expressions,
alerting him or her if he or she is displaying indications of
a) Dataset Creation stress. The modules that are using to build the system are
When making the dataset manually collect the video shown in Figure 3.
clips from online workers and split them to images using
video splitting algorithm. This dataset contains 18840
images under five categories (sleeping while working,
nervous behaviors, frequently yawning, smoking while
working, and normal behaviors). When creating this dataset
used 140 small video clips of twenty-nine male and twenty-
eight female online workers. In this function, use the web
camera to monitor the behaviors. Therefore, those videos
recorded from webcam when they are doing their works
through online. Manually gathered the images because
validation of the dataset. When use the global dataset it is
Fig. 3. System diagram for Facial Expression recognition
not 100% accurate for Asian country people like Sri Lanka.
The first module of the system will oversee taking
When creating dataset collected images are merged to
photos from the user’s computer/laptop webcam in real time
single array and then convert them to gray scale. When
and transferring them to the second module. The position of
convert images to gray scale can identify the features of the
the user's face will be determined in the second module,
image. Then gray scaled converted images divided to
using a feature selection approach, and cropped. After that,
relevant categories and label them properly. The purpose of
face will be divided in to 299x299 pixels and normalized by
the labeling is given proper input to the model.
dividing all pixel values by 255, resulting in values ranging
from 0 to1. The Emotion Classification module will
categorize the face and export a list of seven probability

331

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on January 30,2022 at 00:51:45 UTC from IEEE Xplore. Restrictions apply.
 values, one for each facial emotion, using a trained to the output of the fatigue and Stress model, System
classification model. Then system will record each detail of predict user is in Normal Stress or the user is in Stress
the records with the generated time and emotion type. Data Related Fatigue Condition.
will be stored in the database real-time. Then it will separate
stressed facial expressions from all the other the other facial
expressions. According to the latest results, it will calculate
the stressed facial expression rate per selected period.
Stress Rate = (Stressed Facial Expressions Count /All Facial
Count)/Time Period

C. Stress and Fatigue detection using keystroke dynamics

and mouse movements
Keystroke and mouse movement pattern is unique for
every person. That pattern may change while user faced an
external factor like stress or fatigue conditions. Fig. 4. Ways of extracting PH, PP1, PR1, RP1, RR1 and PP2, PR2, RP2,
Understanding that pattern change helps us to clarify user’s RR2 values
current stress condition is affect for their current fatigue
condition or not. To understand that pattern difference we
must generate individual weight files for each user.
In our app, we collect keypress and release times, mouse
movement speed (movement events per millisecond),
user’s mouse and keyboard inactivity times, time working
on an application and store that data in ‘.tsv’ file. Moreover,
we use Experience Sampling Method (ESM) [15] to take
user’s current conditions, which include

 Photographic Affect Meter (PAM) [5]

 Questionnaire of Feeling of Energy Level,
Pleasant Level and Stress
 Questionnaire of Feeling of Fatigue level
Fig. 5: Taking User condition using Experience Sampling Method
(ESM)
When data collected consecutive 3-day period, pre-
processed that data and remove outliers using MAD
(median absolute deviation). D. Using Heart Rate Sensor to Detect Mental Stress
The developed prototype uses heart rate variability to
System generates two separate feature sets for stress and determine whether a person is stressed. It can also aid in the
fatigue detection. System Extract 9 features from keypress detection of patterns of variations in a person's heart rate
and release data (key press and key release PH, 1st key while working on a computer. The heartbeat readings are
press and 2nd key press (PP1), 1st key press and 2nd key sent to the firebase, where they are filtered by a user's
release (PR1), 1st key release and 2nd key press (RP1), 1st network id so that readings for a specific person can be kept
key release and 2nd key release (RR1) [13] and 1st key track.
press and 3nd key press(PP2), 1st key press and 3nd key
release(PR2), 1st key release and 3nd key press(RP2), 1st For hardware implementation of the prototype, use Node
key release and 3nd key release(RR2)), Apart from that MCU (Machine Control Unit) and MAX30102 Pulse
mouse movement speed (in milliseconds), Number of key sensor. Node MCU has an ESP8266 Wi-Fi module and a
presses (per second) is extracted for fatigue detection. For Tensilica Xtensa LX106 32-bit microprocessor with
stress detection PH, PP, PR, RP, RR [13] is mainly used, 160KB RAM, which is significantly superior than the
apart from that total mouse movement distance, mean time Arduino UNO's 8-bit ATMega328P with 2KB RAM
of mouse button press, time of application used, total time (Fig.6). It connects to the internet through Wi-Fi and runs
of specific key typed is used After completing the feature micro python, a stripped- down version of Python. The
extraction, in fatigue detection divide dataset into three photodetectors, optical elements, and internal LEDs make
class according to time as Morning, Afternoon evening. up the MAX30102 sensor, which is a pulse oximetry
Then each class feed into Artificial Neural Network and module. A low-cost IC is utilized to measure our heart'sz
generate three different weight files that is used to predict electrical pulse action. It runs on a 1.8V power source and
user’s fatigue condition. For Stress detection, we used uses standard I2C connectivity to communicate with
SVM for classification. To train SVM we used data microcontrollers. This sensor operates at a high sample rate
generated by consecutive 3 days by the user. and has a high SNR. And small(16mmx21mm).

After generating models system automatically, start to

predict using generated models. Every 30-minute period
system input for both fatigue and stress models. According

332

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on January 30,2022 at 00:51:45 UTC from IEEE Xplore. Restrictions apply.
 stress occurrences might progressively place the person in
a stressed state. In a classifiers model, this might lead to the
incorrect conclusion that just the most recent incident
produced stress, while all earlier occurrences are ignored
Fig. 6. ESP8266 Wi-Fi module and MAX30102 Pulse sensor and stay concealed.
Sensor Source: Adapted from [16][17] Fig. 8. Stress prediction flow
One more technique we employed was a stress detection
Proposed device embedded with computer mouse that approach based on adaptive analysis. Multiple dynamical
monitor can monitor heart rate and blood oxygen level. features are used in this algorithm: Variational plot SD1,
Vibrational plot SD2, and Approximate Entropy (En).
Stress is detected using this approach if,

Fig. 9. Logical Equation

Fig. 7. Heart rate capturing device
We used Python to calculate Approximate Entropy
Method is proposed based on mm-wave sensor proven based on the method provided by the authors to recreate this
the most accurate in detecting the tiny mechanical process. The relaxing and game-playing parts were both
vibrations of the human thumbs caused by heartbeats, out scanned using a sliding window. We labeled the entire 01-
of all the many ways to electronically gather data about minute period as Tension if stress was identified in any
heart function. We proposed a method for detecting step, and REST if rest was detected in all stages of the full
heartbeats as separate occurrences using a constant portion [18].
heartbeat sensor and an artificial neural network (ANN) in
this study. The ANN took the raw wave signal as input to We calculated the accuracy, specificity, and sensitivity
make the technique computationally simple, while the for each of the three techniques to compare their results. A
output was little modified to ensure low latency operation true positive result was recorded if the method marked the
(less than 1 second). game playing part as STRESS, a true negative result was
recorded if the relax music part was marked as REST, a
Designers created a basic stress detection system that false positive result was recorded if the relax music part
simply utilize moment HRV characteristics. A justification was marked as STRESS, and a false negative result was
for omitting lot of positive is that they need significantly recorded if the game playing part was marked as REST
more computer resources to calculate than significant [18].
components, which is an objection worth considering in a
custom app. To detect stress, designers implemented IV. RESULT AND DISCUSSION
several the mean HR, pNN50, and RMSSD characteristics.
The Hrv wave was split into four equal portions by a sliding One hundred epochs were utilized to identify the
window. We experimented with several lengths of the behaviors in the stressed behaviors identification model
sliding window, and the minimum width that produced (sleeping while working, nervous behaviors, frequently
excellent results was 560 Update gaps with a shift of 20 RR yawning, smoking while working, and normal behaviors).
intervals in each step. To determine the optimal upper and It had an average accuracy of 93.63%.
lower limits for each HRV characteristic, designers utilized
a conventional warfare approach. Consequently, the
algorithm detects stress if the mean heart rate in the fourth
part increases by more than 5% compared to the first part,
and RMSSD and pNN50 measurements fall by much more
than 9% in the fourth part compared to the third part. It is Fig. 10. Results of stressed behavior identification
important to note that this method does not detect the
resting state. If no stress is detected, the rest state is used to Using the trained model 7 facial expressions were
calculate the accuracy, specificity, and sensitivity. The identified. Some of them are shown in below figure. We
subject's condition following the stressor is likewise created one model for local users and another model to
unknown. Instead of a single large incident, a sequence of global users. It was found in the research that facial
expressions are different country to country therefore, we
created two models.

Fig. 11. Results of facial recognition

In stress detection the highest accuracy the SVM

algorithm is archived is between 70%-80%. Accuracy

333

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on January 30,2022 at 00:51:45 UTC from IEEE Xplore. Restrictions apply.
 gradually decreased when more data is used therefore 3- REFERENCES
day period is ideal. [1] C. S. Dewa and J. S. Hoch, “Barriers to mental health service use
among workers with depression and work productivity,” J. Occup.
Environ. Med., vol. 57, no. 7, pp. 726–731, 2015.
[2] Chan, Sherilynn & La Greca, Annette. (2013). Perceived Stress
Scale (PSS).10.1007/978-1-4419-1005-9_773.
[3] M. Chauhan, S. V. Vora, and D. Dabhi, “Effective stress detection
using physiological parameters,” in 2017 International Conference
on Innovations in Information, Embedded and Communication
Systems (ICIIECS), 2017, pp. 1–6.
[4] B. Kim et al., “Interpretability beyond feature attribution:
Quantitative testing with Concept Activation Vectors (TCAV),”
arXiv [[Link]], pp. 2668–2677, 2017.
[5] John P. Pollak, Phil Adams, and Geri Gay. 2011. PAM: a
photographic affect meter for frequent, in situ measurement of
affect. Proceedings of the SIGCHI Conference on Human Factors in
Computing Systems. Association for Computing Machinery, New
Fig. 12. Result in Stress Detection using SVM
York, NY, USA, 725–734.
[6] Schneiderman, Neil & Ironson, Gail & Siegel, Scott. (2005). Stress
In Fatigue detection three model’s Accuracy given as and Health:Psychological, Behavioral, and Biological Determinants.
Morning-98.02percentage, Afternoon-96.76percentage, Annual review of clinical psychology. 1. 607-28.
and Evening-99.56percentage. After the model building, 10.1146/[Link].1.102803.144141.
our application starts to predict using those models. Input [7] “Gao D, et al. (2014),” [Link]. [Online].
Available:[Link]
forwarded to the stress and relevant fatigue model. ccessed: 13-Sep-2021].
According to the output of both models, we conclude user’s [8] M. A. El Maaoui, D. Feurer, O. Planchon, M. R. Boussema, and
stress condition. M.H. Snane, “Assessment of kite born DEM accuracy for gullies
measuring,”[Link].[Online].Available:
We obtained strongly correlated data segments once the [Link] [Accessed: 13-
synchronization operation was completed. The distribution Sep-2021].
of signal segment lengths shown in the graph below. Most [9] A. Pampouchidou et al., “Depression assessment by fusing high and
parts are between 3 and 18 minutes long, as may be seen. low level features from audio, video, and text,”in Proceedings of the
6th International Workshop on Audio/Visual Emotion
110 minutes (about 2 hours) is the longest section that is Challenge,2016.
highly linked with the gold standard data. The default [10] B. Kim et al., “Interpretability beyond feature
parameter choices are designed to reduce the number of attribution:Quantitative testing with Concept ActivationVectors
segments that are too short (less than five minutes). Most (TCAV),” arXiv [[Link]], pp. 2668–2677, 2017.
of the poor portions are under one minute in length, with [11] Hernandez, P. Paredes, A. Roseway and M. Czerwinski, Under
pressure: sensing stress of computer users. New York, NY, USA: In
only one exceeding 60 minutes. Proceedings of the SIGCHI Conference on Human Factors in
Computing Systems (CHI '14), 2014.
TABLE 2. CLASSIFICATION ACCURACY LEVELS
[12] S. D. W. Gunawardhane, P. M. De Silva, D. S. B. Kulathunga and
Classifier Accuracy S.M. K. D. Arunatileka, "Non invasive human stress detection using
key stroke dynamics and pattern variations," 2013
Gaussian Naïve Bayes 89.98% InternationalConference on Advances in ICT for Emerging Regions
K Nearest Neighbor 93.22% (ICTer), 2013, pp. 240-247, doi: 10.1109/ICTer.2013.6761185.
Decision Tree 88.12% [13] M. Ulinskas, R. Damaševičius, R. Maskeliūnas, and M. Woźniak,
Support Vector Machine (SVM) 87.96% “Recognition of human daytime fatigue using keystroke data,”
Neural Network (1 hidden layer) 96.67% Procedia Comput. Sci., vol. 130, pp. 947–952, 2018
[14] A. Ghosh, M. Danieli, and G. Riccardi, “Annotation and prediction
With the combination of each result above system, of stress and workload from physiological and inertial signals,” in
graphical statistics shown in the UI (User Interface). If the 2015 37th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC), 2015, pp. 1621–1624.
stress level of the user is high than the pre-defined [15] J. Hektner, M. Csikszentmihalyi and J. Schmidt, Experience
benchmark, system will show pop-up alert to the user with sampling method. Thousand Oaks [Calif.]: SAGE, 2007
tips which can be extremely useful to prevent the health [16] “NodeMCU ESP8266,” [Link]. [Online]. Available:
problems. [Link]
esp8266-pinout-features-and-datasheet. [Accessed: 19-Sep-2021]
V. CONCLUSION [17] “MAX30102 Sensor Module 1.8V-5V Pulse Oximeter Heart Rate
Detection,” [Link].[Online].Available:
The rise of online workers has been gradually [Link]
increasing due to covid-19 pandemic. It is important to [Link].[Accessed: 19-
have healthy lifestyle while working from home. As Sep-2021]
we discussed throughout the paper our application [18] Melillo P., Bracale M., Pecchia L. Nonlinear Heart Rate Variability
features for real-life stress detection. Case study: students
help to users to prevent stress. The accuracy levels and understress due to university examination. BioMedical Engineering.
outputs shown in Results and discussion section shows
efficiency of the system and how it will assist to provide
a great service.

334

Authorized licensed use limited to: SLIIT - Sri Lanka Institute of Information Technology. Downloaded on January 30,2022 at 00:51:45 UTC from IEEE Xplore. Restrictions apply.
View publication stats