Predictive Analysis of Stress Based on Vital

Physiological Metrics
Athar Sayed Gaurav Sahu Shardul Gore
Mukesh Patel School of Technology Mukesh Patel School of Technology Mukesh Patel School of Technology
Management & Engineering, Management & Engineering, Management & Engineering,
(MPSTME), Shri Vile Parle Kelavani (MPSTME), Shri Vile Parle Kelavani (MPSTME), Shri Vile Parle Kelavani
Mandal's Narsee Monjee Institute of Mandal's Narsee Monjee Institute of Mandal's Narsee Monjee Institute of
Management Studies, Mumbai, India Management Studies, Mumbai, India Management Studies, Mumbai, India
sayedathar242@[Link] gauravsahu7045@[Link] shardulggore@[Link]

Kshamta Mathur Harikrishnan R Rashmi Patel

Mukesh Patel School of Technology Symbiosis Institute of Technology, Mukesh Patel School of Technology
Management & Engineering, Pune Campus, Symbiosis International Management & Engineering,
(MPSTME), Shri Vile Parle Kelavani Deemed University, India. (MPSTME), Shri Vile Parle Kelavani
Mandal's Narsee Monjee Institute of harikrishnan.r@[Link] Mandal's Narsee Monjee Institute of
Management Studies, Mumbai, India Management Studies, Mumbai, India
[Link]@[Link] [Link]@[Link]

Abstract— Stress comes from various sources, stress is a normal thing, prolonged exposure to stress
including personal struggles, professional expectations, can cause side effects on mental health which can
and academic constraints. It is a significant aspect of a become a serious issue later. Specifically, stress in
person's life. Stress in the workplace affects many workplaces and schools can lead to anxiety, burnout,
people, including corporate professionals and students,
and other mental health issues, making it crucial to
and it can have serious negative effects on one's
physical and emotional well-being. Older stress
identify and manage stress effectively.
assessment methods are biased since they are frequently
subjective or rely on self-reports. The physiological The earlier stress assessment techniques depended
health metrics such as the body temperature, the heart on surveys and practical evaluations, which are
rate, blood volume pulse , & electrodermal activity of a subject to bias and human error. However, the
person are used in this research to suggest a new modern method eliminated these biases and time-
method of stress identification and treatment. We consuming practical evaluations and included the
provide an unbiased and trustworthy evaluation of stress ability to provide real-time data for stress detection,
levels by using real-time data from wearable sensors of such as body heat , blood, the volume of blood
different students. We examine the relationship between pulse,the heart rate, and electrodermal activity. With
stress and these health metrics using machine learning
the aid of contemporary wearable sensors, they
techniques, providing a prediction model for stress
identification. The findings show that, in comparison to
provide vital information in real-time, completely
conventional techniques, our model can precisely and eliminating the need for subject methods.
promptly detect stress events with very high and
optimum performance, greatly enhancing the timeliness In this study, we offer an implementable strategy.
of stress monitoring. Future research will concentrate With the use of wearable technology that was used to
on providing user suggestions based on their stress level measure the students' health metrics, we offer a
and providing stress-reduction strategies based on real- method that uses the data provided to identify stress
time data feedback. levels and manage it depending on an individual's
health metrics. This stress-related work offers a
Keywords— Stress Prediction, Machine Learning, Dataset, thorough method for putting a prediction model into
Physiological Metrics, real-time data
practice using machine learning techniques. When
I. INTRODUCTION compared to the conventional techniques, our
method's performance is efficient for stress
Stress is an unavoidable aspect of human life, there monitoring and its accuracy in identifying an
are various factors like academics, Career pressure, individual's stress episodes is remarkable.
stock trading, handling business and countless other
reasons that can cause stress. Even though having
 II. LITERATURE SURVEY stress detection the only thing that it requires is two
important signal metrics they are the heart rate and
Stress can be explored and investigated using a the skin conductance and the system that they have
variety of medical techniques in the medical designed is very implementable in real-time
[Link] order to monitor and measure stress applications.
related health, the authors of [1] came up with a new
algorithm to evaluate respiratory rate using a One week before to the test and while using the
photoplethysmography sensor input. The algorithm internet, the writers of [9] calculate the mental stress
also mentions that Low-quality signals and loss of of the pupils. The goal is to examine stress in college
signals from the sensor can also be handled. The students at various stages of their life. It's common to
algorithm is a combination of different processes overlook the impact that recruiting stress or test
such as signal conditioning and preprocessing, pressure have on students.[10] Using data from a
window selection, modified Welch filtering and person's voice and/or GSR, the authors of this
postprocessing of signal, thus achieving high research have examined how classification
accuracy and making the algorithm resistant to noise. algorithms may be applied to automatically identify
[2] The purpose of this study was to use data from a times of acute stress.
wrist-worn watch to present relevant information and
statistics about stress in healthy people. Participants [11] This research presents an approach that
in the study were evaluated using a series of considers the range of stress assessments while
standardized psychological stress induction protocols examining the outcomes of multimodal stress
during a 60-minute laboratory-based testing session detection. Using a socially assessed mental
using a quasi-experimental repeated measures arithmetic exam, we initially gathered physiological,
approach. depth, and video data from 25 participants under
stressful conditions. As a second phase, we have
[3]In this paper a new multimodal SWELL obtained three distinct stress assessments: one from a
knowledge work (SWELL-KW) dataset for user physiology specialist, one from an outside observer,
modelling and stress research is described in this and one from ourselves. Lastly, we use a
publication.[4] In this research the authors have used classification task to extract 101 behavioural and
a unique stress detection dataset produced in a physiological variables and test their predictive
hospital's natural working environment and that is potential for the three assessments that were
presented in this research. also the biometric gathered.
information of nurses during the COVID-19
pandemic is collected in this dataset. [12] In this paper the authors have reflected the
possibility of unsupervised learning using Self-
In[5]authors have gathered information on a Organizing Maps (SOM) for stress detection. The
population of traumatized teenagers receiving labels based on perceived stress levels are no longer
residential care, the Empatica E4 wristband's required in unsupervised learning environments.
accuracy and predictive value were to be verified in Initially, a controlled stress experiment was carried
comparison to the VU University Ambulatory out in which the phases of tension and relaxation
Monitoring System (VU-AMS). [6] The authors of were switched. Test subjects' electrocardiograms
this study have proposed a method for evaluating (ECGs) and skin conductance’s (SCs) were recorded.
people's stress levels that makes use of wearable A training set of SC and ECG features was then used
chest straps and smartphone characteristics. to construct the SOM's structure. This study
proposes a feature-level fusion method based on long
The impact of stress on various physiological short-term memory and one-dimensional
reactions is examined by the authors of [7]. The convolutional neural networks to extract temporal
effectiveness, resilience, and consistency of and spatial features from ECG, EMG, electrical
biosignal data aspects are highlighted in relation to activity, temperature, accelerator, and response data,
the current level of knowledge in stress detection. In respectively [13]. In order to determine effect and
order to derive precise stress correlates, it also stress, these characteristics are then combined in a
investigated multimodal biosignal analysis and summation way.
modelling techniques. The purpose of this study is to
present a thorough analysis of the biosignal patterns [14]The authors investigated the automatic
that arise under stress and to offer trustworthy, identification of psychological stress in people
useful recommendations for improving stress through social media. By examining the relationships
detection.[8] The authors of this paper came up with between users' stress and their tweeting content,
a innovative approach that can precisely perform the social engagement, and behaviour patterns using
actual online microblog data. [15]In this paper the C. Software Requirements-
authors have proposed an evaluation based on
correlation based time series feature technique using 1. Visual Studio Code
the stress predict [Link] uses a package with the It is a very popular IDE among developers
help of that package, the algorithm computes a list of since it provides fast performance and
1400+ features of the heart beat rate, the respiration excellent debugging capabilities and it's
metrics. quite a popular choice for machine-learning
tasks.

2. Flask
It is an important web development
III. PROPOSED WORK framework that is available as an extension
A. Objective of the work
in many IDEs.
1) Collect and preprocess raw data for stress D. System Implementation
analysis.
2) Engineer relevant features and train various The designed system monitors and detects stress
classification models. levels in real-time by using data from wearable
3) Optimise models through hyperparameter devices. It tracks key health metrics such as heart
tuning and evaluate their performance. rate, skin response ,the body temperature, & the
4) Select the best-performing model and create breathing rate, all of which are important factors to
its pickle file for further use. measure stress. These metrics are analysed
5) Develop a user interface (UI) to predict continuously to assess a person’s stress level
stress levels based on user-input health parameters. accurately. The accumulated data is then cleaned
and processed before being used to train various
B. Detailed Problem Statement machine learning models, the best-performing model
being chosen based on its accuracy. The selected
Stress is an inevitable thing in human life. It affects model is then saved and integrated into the
the mind and body both. Prolonged stress can lead to application to predict stress levels. The model’s
many health problems that in the long term become settings are fine-tuned during training to improve its
very dangerous like heart attacks, Blood pressure performance, ensuring the predictions are as accurate
spikes, anxiety attacks and depression Traditional as possible. We get the reading of stress on the UI in
methods of identifying stress are self-reported the range of 0-100 scale.
surveys and questionnaires. This is often subjective
and may lack precision. With the advent of wearable
technology and advances in machine learning,
(Predictive Modeling) We can use physiological data
to identify real-time stress.

Wearable devices like smartwatches and fitness

bands are popularly used to monitor important
physiological signals related to stress. including
heart rate variability , skin conductance & the
breathing patterns These physiological data points
use machine learning algorithms to detect stress in
real-time. This creates the potential for personal
stress management. Prediction models can be trained
to recognize patterns that indicate increased stress. It Fig. 1. Stress-O-Meter Input Page
helps individuals take preventive measures before
stress escalates. Serious health problems, such as
breathing exercises or pauses or when this In Fig. 1, we take input from the user, specifically
technology improves It is considered a revolution in their health parameters, which are collected through
stress management. Make people in various the wearable device. The parameters include
environments More accessible and proactive. Electrodermal Activity, the heart rate, Blood Volume
Pulse, & temperature. These health metrics are very
important for assessing stress levels. The same
features that the model was trained on are then used
to predict stress. Once the user enters all the required
inputs and hits the "Predict" button, the system
processes the data and transfers the user to a new
window. In this new window, the system uses the
entered health parameters and then predicts the
user’s stress level in real-time.

In Fig. 2, the Stress-O-Meter page is displayed,

showing the predicted pressure level of the person.
Based on the entered information supplied by means
of the person, the gadget calculates and presents the
stress level, providing a clear indication of the user’s
modern-day emotional and bodily country. Stress-O-
Meter gives the visual illustration of expected strain, Fig. 3. Workflow for Stress-O-Meter
allowing the person to apprehend their stress level
and take appropriate moves if wished. The web page
serves as an interactive interface that now not most A. Data Collection
effective displays the results but may also provide
insights into how the inputs have an impact on strain Starting from the data collection which is the first
predictions, assisting users manage and display their process in this project and undoubtedly the most
properly-being efficiently important one. The obtained dataset from wearable
devices include various health metrics such as the
blood volume pulse, Electro dermal activity,
temperature, accelerometer data & the heart rate.
Apart from this the additional data that are obtained
normal from wearable devices include oxygen
saturation , skin conductance and the breathing rate.
These factors are also very important in determining
the stress of an individual , as they represent the
responses that are commonly associated with stress

B. Data Preprocessing & Feature Engineering

After the data collection process, the next step was

data preprocessing. The dataset contained missing
Fig. 2. Stress-O-Meter Predicting Stress values for features such as EDA and heart rate,
which were handled through imputation. Features
IV. EXPERIMENTATION AND RESULTS like accelerometer data, which were not significant
contributors to stress determination, were dropped.
Through detailed analysis, we identified the most
The experimental setup system for predicting and
important factors that contribute to stress and can
developing the Stress-O-Meter machine is described
help accurately predict stress levels.
extensively in this component. It includes numerous
important steps, inclusive of function engineering,
records amassing, facts pre-processing, model
training, and assessment. The system's typical
workflow is displayed above which explains the
whole workflow from the start to the result in a
sequential manner.

Table 1. Dataset of Health metrics

 This tuning process is important in optimizing the
model's performance and improving its overall
accuracy. By carefully adjusting the hyperparameters
We were able to increase the predictive ability of the
model and obtain the best results.

Fig. 4. GPU consumption stats during GPU Intensive

Model Training

For the dataset shown in Table 1, it represents the

processed data obtained after performing data In Fig.4. To make the training process quicker we
preprocessing and feature engineering. Based on this configured the GPU (graphics processing unit)
dataset, classification models were trained, and resources so that instead of the model running and
hyperparameter tuning was performed to optimize consuming CPU resources it could run on the GPU
accuracy directly we configured the GPU with the necessary
libraries and dependencies such as CUDA and
cuDNN, both of which are developed by NVIDIA.
These libraries are designed to support GPU-
intensive tasks. This makes it ideal for accelerating
C. Model Training and Evaluation deep learning and machine learning workflows.
Using GPUs significantly reduces the training time
of these models. This allows for faster model
Table 2. Performance Evaluation of Classification experimentation and evaluation. Otherwise, utilizing
Models a CPU was a very time-consuming task to do.

In Table 2 we present the accuracy metrics for

different classification models. This is to ensure
consistency and objectivity in the assessment
process. All models were trained using the same Fig. 5. Classification report of the Algorithm
preprocessed dataset. Models considered for
comparison include XGBoost, AdaBoost and many In Fig.4 we consider the highest-performing model
others. All models except Logistic Regression were (XgBoost) and generated its classification report
hyperparameter adjusted. which contains all the metrics related to the overall
performance of the model along with that we
generated the confusion matrix which is very crucial & Cochran, K. (2022). A multimodal sensor
in comparison between the predicted values with dataset used to continuously identify stress among
the actual values and also it provides enhanced hospital nurses. 9(1), Scientific Data, 255.
visual of the model's evaluation metrics
rather than a single accuracy score. [5] De Looff, P., Rosada, C., Scholte, R. H.,
Popma, A., Nijhof, K. S., Schuurmans, A. A., &
Otten, R. (2020). A comparison with
CONCLUSION electrocardiography (ECG) demonstrates the
validity of the Empatica E4 wristband for
Stress is an inevitable occurrence in life, and the
measuring heart rate variability (HRV)
causes can vary greatly from person to person. This
parameters. Medical Systems Journal, 44, 1–11.
project can assist in monitoring stress levels and
improving one's quality of life. Anyone who can [6] Tröster, G., Arnrich, B., and Muaremi, A.
make the necessary changes can utilize the deployed (2013). Towards quantifying stress during the
system. working and when sleeping using wearable
technology and smartphones. 1, 172-183;
This research uses a variety of existing machine
BioNanoScience, 3.
learning approaches to help produce an intelligent
stress predicting system. It determines the best [7] Tsiknakis, M., Simantiraki, O., Giannakakis,
machine learning model for stress prediction after a G., Grigoriadis, D., and Giannakakis, K. (2019).
protracted process of data collecting, feature Review of biosignals for detecting psychological
engineering, and model training. stress. IEEE Affective Computing Transactions,
13(1), 440-460.
In light of this, we also incorporate an intuitive user
interface that allows users to input their health [8] Casanova, J. G., de Santos Sierra, A., Ávila, C.
metrics and determine their stress level, thereby S., & Del Pozo, G. B. (2011). a system for
facilitating meaningful health management. detecting stress that uses fuzzy logic and
physiological signals. IEEE Industrial Electronics
Future research will concentrate on recommending
Transactions, 58(10), 4857-4865.
stress-reduction strategies based on real-time data
feedback and user stress levels. [9] Banga, A., and R. Ahuja (2019). Machine
learning algorithms for detecting mental stress in
college students. 349–353 in Procedia Computer
V. REFERENCES
Science, 152.
[1] Iqbal, T., Ganly, S., Wijns, W., Elahi, A., & [10] Maslov, A. V., Kurniawan, H., & Pechenizkiy,
Shahzad, A. (2022). algorithm for estimating M. (2013, June). Speech and galvanic skin reaction
respiratory rates based on photoplethysmography signals are used to detect stress. IEEE International
for use in health monitoring applications. 42(2), Symposium on Computer-Based Medical Systems,
242-252; Journal of Medical and Biological Proceedings, pp. 209-214. IEEE.
Engineering.
[11] Cohen, D., Dubuisson, S., Detyniecki, M.,
[2] Glynn, N., Killilea, J., Walsh, J., Simpkin, A. Aigrain, J., Spodenkiewicz, M., & Chetouani, M.
J., Simpkin, T., Roshan, D., & Shahzad, A. (2016). Multiple assessment-based multimodal stress
(2022). A pilot research and stress-predict dataset detection. IEEE Affective Computing Transactions,
for wearable sensor-based stress monitoring. 8135 9(4), 491-506.
in Sensors, 22(21).
[12] Helic, D., Bogaerts, K., Van Diest, I., De Raedt,
[3] Neerincx, M. A., Verberne, S., Koldijk, S., W., Van Hoof, C., Smets, E., and Huysmans, D.
Sappelli, M., & Kraaij, W. (2014, November). (2018). Unsupervised learning for the identification
Research on stress and user modeling using the of mental stress. Biomedical engineering systems
Swell Knowledge Work dataset. Pages 291-298 of and technologies, 11th International Joint
Proceedings of the 16th International Conference Conference Proceedings (Vol. 4, pp. 26-35).
on Multimodal Interaction.
[13] Feng, M., Li, M., He, C., Fang, T., & Liu, J.
[4] Hosseini, S., Borst, C. W., Katragadda, S., (2024). Affect and stress detection using LSTM and
Gottumukkala, R., Ashkar, Z., Bhupatiraju, R. T., 1DCNN feature fusion. 512–520 in Computer
Methods in Biomechanics and Biomedical
Engineering, 27(4).

[14] Huang, J., Xue, Y., Li, Q., Lin, H., Jia, J., Guo,
Q., Xue, Y.,... & Feng, L. (2014, November). Deep
neural networks are used to detect psychological
stress at the user level from social media. Pages 507–
516 of Proceedings of the 22nd ACM International
Conference on Multimedia.

[15] Elahi, A., Wijns, W., Amin, B., Shahzad, A., &
Iqbal, T. (2023). Automatic feature selection from
heart rate and breathing rate time data has improved
stress classification. 2950 in Applied Sciences,
13(5).