=Paper=
{{Paper
|id=Vol-3706/Paper16
|storemode=property
|title=Evaluating Mental Well-being through Wearable Sensors Utilizing Machine Learning
|pdfUrl=https://ceur-ws.org/Vol-3706/Paper16.pdf
|volume=Vol-3706
|authors=Aarti,Swathi Gowroju,Raju Pal,Pushpendra Kumar Rajput
|dblpUrl=https://dblp.org/rec/conf/icaids/AartiGPR23
}}
==Evaluating Mental Well-being through Wearable Sensors Utilizing Machine Learning==
https://ceur-ws.org/Vol-3706/Paper16.pdf
Evaluating Mental Well-being through Wearable
Sensors Utilizing Machine Learning
Aarti1,† , Swathi Gowroju2,∗,† , Raju Pal3,† and Pushpendra Kumar Rajput4,†
1
Lovely Professional University, Punjab, INDIA
2
Sreyas Institute of Engineering and Technology, Hyderabad, Telangana, INDIA
3
Jaypee Institute of Information Technology,Noida, Uttar Pradesh, INDIA
4
Sharda University, Gr. Noida, INDIA
Abstract
Stress is known to play a significant role in the development of serious medical illnesses such diabetes,
hypertension, and cardiovascular disease. Numerous studies on the viability of using several physiological
markers to detect stress have been conducted in light of the increased emphasis on wearable health
monitoring. The goal of this study is to classify individuals based on physiological information, using
the easily accessible WESAD (Wearable Stress and Affect Detection) dataset. The main objective is to
use this dataset to create algorithms that can forecast stress based on physiological markers. Using
the Synthetic Minority Oversampling Technique (SMOTE), a model is developed in this research to
improve the precision of stress level detection. SMOTE is designed to balance out dataset imbalances by
oversampling the minority class. This study used the SMOTE approach to efficiently balance the dataset
groups due to the uneven nature of the data.
Keywords
Artificial intelligence, machine learning, stress detection, physiological signal, mental health, WESAD
1. Introduction
Real-time physiological data gathering for stress evaluation has become easier in recent years
because of the development of wearable sensors and physiological monitoring technology.
Physiological signals are useful markers of stress levels because they shed light on how the body
reacts to stress. These signals provide a clear indicator of physiological arousal and reactivity
since they are a direct reflection of the autonomic nervous system’s activation and the release of
stress hormones. Numerous physiological signals, including the electroencephalogram (EEG),
electrocardiogram (ECG), heart rate variability (HRV), skin conductance (SC), electromyography
(EMG), and respiration rate, can be continuously monitored thanks to these technologies. Using
these physiological signs, there has been an increase in interest in the subject of automatic
mental stress detection. This interest results from the requirement to create approaches and
tools that can precisely extract features from diverse physiological data, process those signals
using machine learning methods, and then analyze those signals. In order to identify stress, a
ACI’23: Workshop on Advances in Computational Intelligence at ICAIDS 2023, December 29-30, 2023, Hyderabad, India
∗
Corresponding author.
†
These authors contributed equally.
Envelope-Open aarti.1208@gmail.com ( Aarti); swathigowroju@sreyas.ac.in (S. Gowroju); raju3131.pal@gmail.com (R. Pal);
pushpendrarajputs@gmail.com (P. K. Rajput)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
200
lot of people now use the WESAD (Wearable Sensor and Affect Detection) dataset. WESAD
dataset was utilized by numerous studies to examine stress detection [1, 2, 3, 4, 5].The WESAD
dataset includes numerous physiological parameters that were taken from 15 participants under
varied circumstances. An Empatica E4 bracelet and a RespiBAN device worn on the chest were
used to record these measurements. Blood volume pulse (BVP), electrodermal activity (EDA),
skin temperature, and accelerometer data were all recorded using the Empatica E4 wristband.
The BVP signal, which is produced from a photoplethysmography (PPG) sensor, was used to
extract information on the interval between beats and heart rate. The RespiBAN gadget, on the
other hand, captured data from the accelerometer, ECG signal, EDA, electromyography (EMG),
respiration rate, temperature, and other sensors. In the WESAD dataset, this combination of
sensors and measurements enables a thorough evaluation of physiological reactions during
various activities and situations.
The following main reasons are the motivations behind the use of the WESAD dataset for
research:
• Multimodal data collection: The WESAD dataset contains information from a number of
different modalities, including electrocardiography (ECG), electrodermal activity (EDA),
electromyography (EMG), respiration, and temperature.
• Comparability and benchmarking: The availability of a standardized dataset like WESAD
makes it possible to benchmark and compare against existing techniques.
• Open accessibility: Researchers can freely access the WESAD dataset.
2. Related Work
The performance of K-nearest neighbors (KNN) models used to classify the WESAD dataset
is evaluated in [1]. The K-fold cross-validation parameter and the total number of nearest
neighbors taken into consideration are two crucial factors that are changed to conduct the
evaluation. The researchers evaluate the effects of various cross-validation procedures on the
effectiveness of the models by altering this parameter. The task is addressed by F. D. Martino et
al. [2] proposed solution, which makes use of ensemble learners and recurrent neural networks
(RNNs). The research uses the Leave-One-Subject-Out (LOSO) cross-validation scheme to
evaluate each model’s ability to generalize in predicting individual stress scores.
Based on their significance to the classifier and their association with other features, Hsieh et
al. [4] stressed the selection of dominating features. For classification, the extreme gradient
boosting technique (XGBoost) was used. To manage variations in sensing settings, often known
as the noise context, N. Rashid et al. [5] suggested a system dubbed SELF-CARE (Selective Sensor
Fusion for Stress Detection). The study measures SELF-CARE performance using wearable
sensors that are worn on the wrist and chest. SELF-CARE obtains an accuracy of 86.34% using
wrist-based sensors and 86.19% using chest-based sensors for the 3-class stress classification
problem. Similar to this, SELF-CARE achieves an accuracy of 94.12% (wrist-based) and 93.68%
(chest-based) for the 2-class stress categorization problem. The WESAD dataset was used in [6]
to examine the effectiveness of six classifiers. Of the classifiers studied, the random forest
(RF) classifier demonstrated the best performance. Overall, the performance of the wrist-worn
201
sensors was worse to that of the sensors positioned on the chest. The wrist-worn sensors
performed at their peak level 95.54%.
The performance of the chest-worn sensors, however, was higher at 97.15%. A unique
technique for identifying mental stress is presented by S. Ghosh et al. [7], employing the
ECG and GSR, two widely utilized physiological signals. The adaptive synthetic minority
oversampling (ADASYN) technique was used to alleviate dataset imbalance and guarantee
uniform classifier training. The detection of mental stress was then carried out using a multi-
class random forest (RF) classifier. Using the WESAD dataset, the performance of the suggested
method was assessed. The outcomes show the usefulness of the suggested approach, with a
97.08% overall accuracy. In [8], PPG and EDA data were acquired from the Empatica E4 bracelet
and binary categorization was done using Random Forest, Support Vector Machine (SVM), and
Logistic Regression.
Certain machine learning approaches were expermented in Ayurveda based physical con-
stituent balancing detection and psychological impact og COVID-19 on hman being [9, 10, 11].
The study’s findings showed that the Random Forest model had the maximum stability and,
when including all the features, had an accuracy of 76.5%. A novel multimodal artificial intelli-
gence (AI)-based technique for stress detection and categorization as well as the recognition of
stress patterns across time is presented by Rahee W. et al. in their paper published in Science
Translational Medicine [12]. The proposed method achieves accuracy of 96.07% using the ANN
and SWELL-KW dataset.
3. Proposed Methodology:
The major goal of the study paper is to suggest a novel and promising method for recognizing
stress using CNN and encoding the multiple-variant time series dataset to GAF pictures following
accurate pre-processing, required transformation, and normalization of the dataset. When it
comes to the WESAD dataset, each subject’s chest data that has been captured was taken,
removed, and transformed to data frames with the chest sensor keys serving as the columns.
The labels were taken separately from the data frames. The labels list the stress level, which
ranges from ’0’ to ’3’. The proposed methodology is shown in Fig. 1.
In our method, we classify each person’s state every 0.1 seconds. PPG value (ppg), PPG
autocorrelation value (ppgau), HRV value (hrv), and EDA value (eda) are the four measurements
that make up each state. Therefore, we considered for each xt =ppgt, ppgaut, hrvt, edat as
a sample at time t as the feature vector, where t equals sampled at intervals of 0.1 seconds.
Each sample xt was tagged with the current condition of the subject. i.e., under stress or not.
Consequently, our dataset consisted of the measurements. With their matching label as D =,
the results that were achieved at each time interval. (Xt, Lt), where L = ”stressed, not stressed”
is the formula. The person’s condition was the activity that individual was engaged in at time t.
3.1. Dataset Description:
The WESAD dataset includes information gathered from people who were exposed to emotional
and stressful stimuli. This data collection was conducted in a controlled laboratory setting
with 15 volunteers, 3 of whom were female. Each participant underwent three basic effect
202
Figure 1: Proposed Methodology
conditions: ”baseline” (consisting of an impartial reading task), ”amusement” (where subjects
watched comedic video clips), and ”stress” (using the Trier Social Stress Test - TSST).This dataset
includes physiological and movement data that was obtained from a wrist- and chest-worn
device. Electrodermal activity (EDA), electrocardiogram (ECG), blood volume pulse (BVP),
electromyogram (EMG), respiration (RESP), skin temperature (TEMP), and an accelerometer
with three axes (ACC) are some of the sensor modalities. For the purposes of comparison with
the stress class in a binary classification task, we combine the baseline and amusement states
into a non-stress category in our particular inquiry.
3.2. Pre-processing and Feature Extraction:
With the exception of EMG-features, all physiological signal-based characteristics were calcu-
lated across a window size of 60 seconds.To recognize individual heartbeats, peak detection
techniques were applied to the raw ECG/BVP signals. By monitoring the time interval between
successive peaks, the heart rate (HR) was determined using these identified peaks as reference
points. The mean and standard deviation were also calculated from these heart rate readings in
addition to the HR calculation.
203
Figure 2: Total no of Samples in each Label
3.3. The sympathetic nervous system (SNS):
SNS has an impact on the EDA signal, making it more sensitive to states of increased arousal.
The raw EDA signal was initially passed through a 5 Hz low pass filter. A tonic segment (known
as skin conductance level, or SCL) and a phasic segment (known as skin conductance response,
or SCR), are also included in the unprocessed EDA signal. The SCR denotes a brief reaction to a
stimulus, whereas the SCL represents a steady change in baseline conductivity [13, 14].
The electromyogram (EMG) signal was processed using two different processing steps. A
highpass filter was used to remove the DC component in the original sequence. The filtered
signal was then divided into 5-second windows by segmentation. Both statistical characteristics
and frequency-domain features (such peak frequency) were calculated from these windows.
Additionally, seven frequency bands evenly spaced from 0 to 350 Hz were used to compute the
power spectral density (PSD) of the EMG signal.The raw EMG signal was subjected to a lowpass
filter application in the second processing chain, especially at 50 Hz. This signal after it had been
processed was then divided into 60-second segments for study. The respiratory (RESP) signal
was first subjected to a bandpass filter with cutoff frequencies of 0.1 and 0.35 Hz. Within the
desired frequency range, this filtering procedure assists in isolating pertinent respiratory data.
The minima and maxima sites in the filtered RESP signal were then found using a peak detector
method. To reduce variances, the data were standardized using the min-max normalization
method. The number of samples in each group is shown in Fig. 2. The Synthetic Minority
Over-sampling Technique (SMOTE) was used to alleviate dataset imbalance [15]. SMOTE is a
well-liked method for addressing class imbalance in datasets for machine learning. Traditional
204
machine learning methods may fail to perform well when faced with imbalanced datasets when
one class contains much less samples than the other because they may be biased towards the
dominant class.
3.4. Classification:
Linear Discriminate Analysis (LDA), Decision Tree, XGBoost, and Logistic Regression were the
four machine learning techniques used in the study. The main goal was binary classification,
with the intention of classifying people as either stressed or unstressed. The performance of
various classifiers on the WESAD dataset is shown in Table 1. Entropy, a measure of information
gain, was used as the criterion for evaluating splits in the instance of Decision Tree classifiers,
which had a maximum depth of 5. For XGBoost, the number of estimators (trees) was set to
50, the maximum depth of trees was set at 5, and the learning rate was set at 0.1. In order to
maximize the model’s convergence in Logistic Regression, the Newton-cg solver was used, with
a maximum iteration count set at 1000. The Singular Value Decomposition (svd) solver was
chosen for linear discriminant analysis. A 10-fold cross-validation strategy was used for all of
these models, enabling thorough performance analysis.
Table 1
Performance Comparison of different Classifiers
Techniques Precision Recall F1-Score Accuracy
Decision Tress 85.76% 96.62% 90.34% 92.27%
Xgboost 90.83% 97.14% 93.16% 94.22%
Logistic Regression 90.21% 90.36% 90.19% 90.36%
LDA 71.48% 84.20% 74.55% 71.60%
4. Results and Discussion
The following experiment was carried out on a DELL Inspiron 15 5518 laptop with an 11th
generation Intel Core processor, 16 GB of memory, and 8 GB of random access memory (RAM).
The whole source code for this experiment was produced using a Jupyter notebook on an
Ubuntu 22.04 1 LTS 64-bit computer. We also take note of the four evaluation metrics, accuracy,
precision, recall, and F1 score, used in the current work before calculating the training and
testing accuracies reached by the current image-encoding-based deep neural network. From
table 1 it is observed that XGBoost outperforms all the other classifiers with maximum accuracy
of 94.22%.
The experiment was run in the WESAD dataset to forecast a person’s degree of stress, which
ranged from 0 (Baseline) to 3 (Amusement). A promising training accuracy of 99.48% and a
testing accuracy of 94.77% were attained after training for roughly 100 epochs. The proposed
image-encoding-based deep neural network model for the WESAD dataset yielded the confusion
matrix, which is displayed in Fig. 3 with the predicted labels on the X-axis and the actual labels
205
Figure 3: Confusion matrix for the labeled data
of the data on the Y-axis. Table 2 presents the accuracy, precision, recall, and F1 score for each
stress identification label as well as the average of all the individual stress-wise performances for
the proposed model for the WESAD dataset. Fig. 4 show that as the model is trained for a larger
number of epochs, the classification accuracy increases while the graphs presented for loss
function substantially reduce. The results also show that, in comparison to other relevant works
without the use of encoding time series images, the accuracy is increased when a multivariate
time series dataset is encoded to its corresponding image.
We ran the study on the two benchmark datasets, calculated each class’s accuracy as well
as their F1 score, recall, and precision, and averaged them all to produce the results shown in
Table 4. Table 4 shows that the proposed image-encoding-based deep neural network achieves
classification accuracy for the WESAD and SWELL datasets of 94.77% and 99.39%, respectively.
The SWELL dataset’s data length is shorter than the WESAD dataset’s, and it required a lot
fewer epochs to train the model. Therefore, the plots of accuracy versus epoch size and loss
function versus epoch size are not required in that case because the difference in the number of
epochs is so small.
5. Future Work
By addressing these issues and enhancing the WESAD dataset, researchers would be able to
investigate more intricate research questions, create more precise models, and learn novel
affective computing and activity detection insights.
The WESAD dataset can be connected with a number of new trends and technologies to
206
Figure 4: Predicted stress graph
expand its capabilities and open up new research opportunities, including:Integrating the
WESAD dataset with wearable devices equipped with advanced sensors can provide richer
and more comprehensive physiological data. For instance, incorporating devices that capture
brain activity (e.g., electroencephalography, EEG) or eye-tracking technology would enable
researchers to study cognitive processes and visual attention alongside physiological responses.
The WESAD dataset can be used in conjunction with mobile sensing technology, such as
smartphones or smartwatches, to gather a wider variety of data and offer contextual information.
In addition to physiological signals, mobile sensing enables researchers to gather information on
location, activity, and social interactions. This provides a more comprehensive understanding
207
of affecting states and activity recognition in real-world environments.
Multi-modal analysis can be facilitated by combining the WESAD dataset with information
from other sources, such as social media, sentiment analysis, or natural language processing.
The accuracy and applicability of models would be increased by combining physiological data
with textual and contextual information to provide a more thorough understanding of affecting
states and activity patterns.
6. Conclusion
Different human physiological indicators have been studied in recent study to evaluate and
monitor degrees of physical and mental stress. Many of these signs have been used in wearable
sensor-driven devices on their own. However, the focus of this study is on utilizing physiological
cues for stress detection. This research has carefully assessed the effectiveness of several
classifiers through a meticulous comparative analysis. SMOTE was used to the problem of
imbalanced datasets in order to improve model accuracy, producing better outcomes. The
results particularly emphasize the XGBoost classifier’s outstanding performance, which led to
an impressive accuracy of 94.22%.
References
[1] D. Bajpai, L. He, Evaluating KNN performance on WESAD dataset, in: 2020 12th Interna-
tional Conference on Computational Intelligence and Communication Networks (CICN),
IEEE, 2020.
[2] Praveenkumar., T. Karthick, Automatic stress recognition system with deep learning using
multimodal psychological data, in: 2022 International Conference on Electronic Systems
and Intelligent Computing (ICESIC), IEEE, 2022.
[3] F. D. Martino, F. Delmastro, High-resolution physiological stress prediction models based
on ensemble learning and recurrent neural networks, in: 2020 IEEE Symposium on
Computers and Communications (ISCC), IEEE, 2020.
[4] C.-P. Hsieh, Y.-T. Chen, W.-K. Beh, A.-Y. A. Wu, Feature selection framework for XGBoost
based on electrodermal activity in stress detection, in: 2019 IEEE International Workshop
on Signal Processing Systems (SiPS), IEEE, 2019.
[5] N. Rashid, T. Mortlock, M. A. A. Faruque, Stress detection using context-aware sensor
fusion from wearable devices (2023). arXiv:2303.08215 .
[6] R. Gupta, A. Bhongade, T. K. Gandhi, Multimodal wearable sensors-based stress and
affective states prediction model, in: 2023 9th International Conference on Advanced
Computing and Communication Systems (ICACCS), IEEE, 2023.
[7] S. Ghosh, S. Mukhopadhyay, R. Gupta, A new physiology-based objective mental stress
detection technique with reduced feature set and class imbalanced dataset management,
in: 2021 IEEE International Conference on Technology, Research, and Innovation for
Betterment of Society (TRIBES), IEEE, 2021.
[8] S. Campanella, A. Altaleb, A. Belli, P. Pierleoni, L. Palma, A method for stress detection
208
using empatica E4 bracelet and machine-learning techniques, Sensors (Basel) 23 (2023)
3565.
[9] C. Gupta, D. Gaur, P. Agrawal, D. Virmani, Huda_covid human disposition analysis
during covid-19 using machine learning, International Journal of E-Health and Medical
Communications (IJEHMC) 13 (2021) 1–15.
[10] R. Kaur, V. Madaan, P. Agrawal, Fuzzy expert system to calculate the strength/immunity
of a human body, Indian Journal of Science and Technology (2016).
[11] K. Ranjit, M. Vishu, A. Prateek, S. S. Kumar, K. Amandeep, Fuzzy expert system for
identifying the physical constituents of a human body, Indian Journal of Science and
Technology (2016).
[12] R. Walambe, P. Nayak, A. Bhardwaj, K. Kotecha, Employing multimodal machine learning
for stress detection, J. Healthc. Eng. 2021 (2021) 1–12.
[13] S. Kim, W. Rhee, D. Choi, Y. J. Jang, Y. Yoon, Characterizing driver stress using physiological
and operational data from real-world electric vehicle driving experiment, Int. J. Automot.
Technol. 19 (2018) 895–906.
[14] W. Du, Application of improved SMOTE and XGBoost algorithm in the analysis of
psychological stress test for college students, J. Electr. Comput. Eng. 2022 (2022) 1–8.
[15] S. Gowroju, S. Kumar, Aarti, A. Ghimire, Deep neural network for accurate age group
prediction through pupil using the optimized UNet model, Math. Probl. Eng. 2022 (2022)
1–24.
[16] A. Ragav, N. H. Krishna, N. Narayanan, K. Thelly, V. Vijayaraghavan, Scalable deep
learning for stress and affect detection on resource-constrained devices, in: 2019 18th IEEE
International Conference On Machine Learning And Applications (ICMLA), IEEE, 2019.
209