Proceedings of NILES2024:

6th Novel Intelligent and Leading Emerging Sciences Conference

$ 'HHS /HDUQLQJ $SSURDFK 8VLQJ :(6$' 'DWD IRU

0XOWL &ODVV &ODVVLILFDWLRQ ZLWK :HDUDEOH 6HQVRUV
Roaa Mubarak
Mohamed Abd Al Aleem Nancy M. Salem
Electronics and Communication
Biomedical Engineering Department Biomedical Engineering Department
Engineering Department
Faculty of Engineering, Helwan University Faculty of Engineering, Helwan University
Faculty of Engineering, Helwan University
Helwan, Cairo Helwan, Cairo
Helwan, Cairo
[email protected] [email protected]
[email protected]
Ibrahim Sadek
Biomedical Engineering Department
Faculty of Engineering, Helwan University
Helwan, Cairo
[email protected]
2024 6th Novel Intelligent and Leading Emerging Sciences Conference (NILES) | 979-8-3503-7851-1/24/$31.00 ©2024 IEEE | DOI: 10.1109/NILES63360.2024.10753228

Abstract² Stress is a common part of life, but chronic or Acute stress is typically triggered by daily short-term
intense stress can severely impact safety and disrupt daily stressors, which our bodies can usually handle. However,
activities. Early detection of mental stress can prevent many excessive or prolonged stress can cause physiological issues
related health issues. Stress causes significant changes in bio- like elevated blood pressure and a heightened risk of mental
signals, which can be used to identify stress levels. This paper health concerns such as anxiety [7]. To mitigate severe mental
proposes various machine learning and deep learning health challenges, it is essential to consistently or periodically
techniques for detecting stress using multimodal datasets from track physiological and biochemical markers associated with
wearable physiological and motion sensors, aiming to prevent the sympathetic nervous system (SNS) and hypothalamic-
stress-related health problems. Data from sensor modalities
pituitary-adrenal (HPA) axis. Current stress level
such as three-axis acceleration (ACC), electrocardiogram
(ECG), blood volume pulse (BVP), body temperature (TEMP),
measurements rely on subjective methods like clinical
respiration (RESP), electromyogram (EMG), and electrodermal interviews and self-report questionnaires, which are time-
activity (EDA) were taken from the WESAD dataset, covering consuming and require trained clinicians, limiting their
amusement, neutral, and stress states. The best results for the accessibility [8,9]. These methods may also yield unreliable
Adam optimizer were achieved with a batch size of 16, yielding results due to personal bias, potentially affecting clinical
an accuracy of 90.26%, while the best results for the SGD decisions. Consequently, there is growing interest in wearable
optimizer were obtained with a batch size of 16, yielding an sensors that can continuously collect real-time bio-signals
accuracy of 87.84%. related to mental status [10].
Keywords²WESAD, Stress, DNN, Early Detection, Wearable This study proposes a feed-forward DNN incorporating
Sensors, COVID-19 Dropout layers between the hidden layers for effective stress
detection. Utilizing the publicly available WESAD dataset,
I. INTRODUCTION this research explores and compares the performance of two
Close to two billion individuals worldwide experience different optimizers, namely Stochastic Gradient Descent
mental health challenges, characterized by impaired brain (SGD) and Adam, across various patch sizes and a 20%
function and these issues can involve unusual perceptions, Dropout layers. The analysis focuses on determining the
altered thoughts, irregular emotions, atypical behavior, and optimal combination for stress detection, employing three-
disrupted social interactions. The global cost of managing level classification (No-Stress/Medium/High).
these issues is expected to reach $16 trillion by 2030 [1].
Despite their significance, Psychological health care has II. RELATED WORK
historically been undervalued compared to physical With significant advancements interest in wearable
treatments, with many countries' services under-resourced, a technology is growing, focusing on continuously tracking
situation worsened by the COVID-19 pandemic [2,3]. This stress through various physiological signals [11,12]. [13]
strain has notably impacted Individuals affected by COVID- Studies have explored the links mmeasuring the distinction
19, healthcare frontline workers, people with pre-existing between suffering and stress the role of wearable sensors and
mental health conditions, and the broader population, leading diagnostic implants in this context. These sensors can detect
to a growing interest in developing accessible sensors for physiological signals such as BP, EDA, HR, MA, BR, SC.
continuous mental status monitoring [4,5]. The goal of researchers is to create wearable health service
Stress, triggered by physical, cognitive, or emotional systems that assess stress and pain through the analysis of
challenges, is managed by the interconnected autonomic wearable sensors in healthcare.
nervous system (ANS) and hypothalamic-pituitary-adrenal The study [14] compares ML techniques for accuracy in 3-
(HPA) axis [6]. The ANS, with its sympathetic (fight-or- class and 2-class. Additionally, a feed-forward deep learning
flight) and parasympathetic (rest-and-digest) branches, network is proposed for these classifications, achieving
regulates body functions. During stress, the sympathetic accuracies of up to 81.65% and 93.20% with ML, and up to
system activates, increasing heart rate, respiration, and sweat 84.32% and 95.21% with DL, respectively.
production to meet the challenge. The HPA axis releases Previous research explored stress assessment through
cortisol, further supporting the stress response through various social media sentiment analysis using ML and BERT for tweet
physiological changes [6]. classification [15]. Their model identified emotions (joy,

Authorized licensed use limited to: UNIVERSITY OF HERTFORDSHIRE. Downloaded on February 20,2025 at 12:36:38 UTC from IEEE Xplore. Restrictions apply.
979-8-3503-7851-1/24/$31.00 ©2024 IEEE 194
 Proceedings of NILES2024:
6th Novel Intelligent and Leading Emerging Sciences Conference

sadness, neutrality, anger, fear) with 94% accuracy, In this study, we utilized the WESAD dataset, a publicly
highlighting the potential of NLP for stress detection. available resource, to implement a feed-forward deep neural
In [16], two DNNs were developed: MLP and a 1D-CNN. network (DNN) for distinguishing between different stress
These networks analyzed wrist and chest-worn physiological levels. We compared the performance of two optimizers, SGD
data for two tasks. For binary classification of stressed vs. non- and Adam, across various patch sizes. Additionally, we
stressed states, average accuracies were 99.65% (MLP) and experimented with Dropout layers at 20% to enhance the
99.80% (1D CNN). For three-class classification (stressed, model's generalization by preventing overfitting. The
baseline, amused), accuracies were 98.38% (MLP) and inclusion of Dropout layers helped mitigate the risk of
99.55% (1D CNN).
overfitting by randomly deactivating a fraction of neurons
In [17], a convolutional neural network with multi-level
during training, thereby improving the robustness of our
deep learning capabilities was introduced. The model was
evaluated using the WESAD dataset for mental health, model.
demonstrating superior performance compared to state-of-the- The next section details our methodology, providing an in-
art methods with an outstanding accuracy of 87.7%. depth look at the techniques and configurations used in our
In [18], a 1D - CNN was developed to categorize arousal, analysis. This comprehensive approach allowed us to identify
valence, and liking based on peripheral physiological signals, the optimal settings for effective stress detection.
such as EMG, BVP, hEOG, vEOG, SKT, and RSP. By
leveraging time and frequency domain features, the CNN III. METHODOLOGY
network obtained accuracies of 77.03% for arousal, 74.68% This study used the WESAD dataset, made publicly
for liking, and 68.75% for valence. available in [27]. It includes movement and physiological data
In [19], researchers developed a novel framework called from fifteen participants, recorded with the Raspbian Advance
MFBPST-3D-DRIFT was developed for emotion recognition chest apparatus and the Empatica E4 arm sensor during
based on EEG signals, employing a parallel spatial-temporal activities like baseline, preparation, amusement, stress,
3D deep residual learning approach. This framework meditation, and recovery. Detailed sensor setup, positioning,
harnessed various EEG frequency bands (delta, theta, alpha, methodology, and data collection processes are in [28]. The
beta, gamma) to construct a 3D feature representation, which Raspbian measured heart rate, ACC, RESP, EMG, and TEMP
was subsequently trained using a 3D deep residual CNN at 700 Hz. The Empatica E4 recorded TEMP, RESP, ACC,
model. It achieved an impressive accuracy of 96.67% on the and EDA at frequencies of 8 Hz, 16 Hz, 32 Hz, and 64 Hz.
SEED dataset, which includes positive, neutral, and negative The primary reason for employing a DNN in this study
emotions, and 88.21% on the SEED-IV dataset, which was its parameter-sharing capability, allowing a few filters to
comprises happy, fear, sad, and neutral emotions. efficiently extract features from the entire input. This DNN
In [20], research concentrated on recognizing emotions was designed to process physiological signals from eight
through EEG records introduced the Spatial-Temporal sensors, including ECG, EDA, EMG, RESP, TEMP, and a 3-
Information Learning Network (STILL), achieving 68.31% axis ACC. Each ACC axis was treated as a separate input,
accuracy for arousal and 67.52% for valence classification. In resulting in eight signals for the DNN network. Data from
[21], an attention-LRCN model was developed to mitigate each sensor were segmented into 5-second windows,
motion artifacts in photo plethysmography (PPG) data. This providing simultaneous inputs for the DNN, which was
model combined a baseline network with an attention configured to perform a three-class classification.
mechanism and employed frequency domain features from Dropout layers were incorporated between each hidden
PPG, resulting in an accuracy of 97.11%. layer to prevent overfitting and enhance generalization. The
Bio-signals such as EEG, ECG, and PPG are widely used DNN consists of five hidden layers with node capacities of 64,
To evaluate physical and mental conditions such as stress, 128, 256, 128, and 64, respectively, all using the ReLU
focus, and emotions. Open datasets like AMIGOS [22], activation function, while the output layer, configured for
DREAMER [23], SWELL [24], SEED [25], DEAP [26], and three-level stress classification, uses the softmax function.
WESAD [27] are curated for research with experiments This hierarchical design with Dropout layers enables nuanced
designed to induce stress and capture bio-signals. These feature extraction. The dataset was randomly scrambled and
datasets include eexperiments are meticulously crafted to split into training and testing subsets with a 70:30 ratio,
ensuring the model's robustness and generalizability by
elicit stress and collect bio-signals. AMIGOS and SEED used
allowing it to learn from diverse samples and validating
video clips to elicit emotions, gathering EEG and ECG data,
performance on unseen data.
while DEAP uutilized music videos to stimulate emotions Two optimization techniques, Adam and SGD, were
over various arousal-valence quadrants while simultaneously HYDOXDWHG IRU WKHLU LPSDFW RQ WKH PRGHO¶V FODVVLILFDWLRQ
peripheral signals and recording EEG. performance. During training, the model underwent 40
The WESAD dataset gathered peripheral signals using iterations per epoch, and the influence of varying batch sizes
video clips and the Trier Social Stress Test. SWELL exposed (16 to 512) on performance was thoroughly investigated. This
participants to office stressors like time pressure, and analysis provided valuable insights into optimizing the
collecting ECG data during tasks including delivering training process for DNN models in stress classification.
presentations and writing a report. These datasets provide The primary metrics utilized to evaluate the deep neural
comprehensive physiological data, making them invaluable network for each optimizer included accuracy, precision,
for understanding human stress and emotional responses. recall, and F1-score, as shown in Equations 1 to 4,
They have been adapted for a variety of studies due to their respectively.
்௥௨௘ ௣௢௦௜௧௜௩௘ା்௥௨௘ ௡௘௚௔௧௜௩௘
diverse stimuli and robust experimental designs. These ‫ ݕܿܽݎܿܿܣ‬ൌ (1)
்௢௧௔௟ ௣௢௣௨௟௔௧௜௢௡
resources are critical for advancing research in emotion ்௥௨௘ ௣௢௦௜௧௜௩௘
recognition and stress detection. ܲ‫ ݊݋݅ݏ݅ܿ݁ݎ‬ൌ (2)
்௥௨௘ ௣௢௦௜௧௜௩௘ାி௔௟௦௘ ௣௢௦௜௧௜௩௘

Authorized licensed use limited to: UNIVERSITY OF HERTFORDSHIRE. Downloaded on February 20,2025 at 12:36:38 UTC from IEEE Xplore. Restrictions apply.
195
 Proceedings of NILES2024:
6th Novel Intelligent and Leading Emerging Sciences Conference

்௥௨௘ ௣௢௦௜௧௜௩௘
ܴ݈݈݁ܿܽ ൌ (3)
்௥௨௘ ௣௢௦௜௧௜௩௘ାி௔௟௦௘ ௡௘௚௔௧௜௩௘
௉௥௘௖௜௦௜௢௡ ൈோ௘௖௔௟௟
‫ ͳܨ‬െ ܵܿ‫ ݁ݎ݋‬ൌ ʹ ൈ (4)
௉௥௘௖௜௦௜௢௡ାோ௘௖௔௟௟

IV. RESULTS
The effectiveness of the feed-forward Neural Network
was thoroughly evaluated in terms of its performance metrics
for the three-class classification task. This assessment
included two specific optimization methods: The Adam Fig. 1. Confusion matrices for DNN performance with Adam optimizer (16
optimizer and the SGD optimizer, applied under different epochs).
batch size settings. This comparative study aimed to uncover
how optimization techniques and batch size variations TABLE 2 PERFORMANCE METRICS OF THE DNN MODEL USING THE SGD
OPTIMIZER ACROSS DIFFERENT BATCH SIZES
influence the overall performance and efficiency of DNN
models in stress classification. Additionally, Dropout layers No. No. F1-
Optimizer ACC. Precision Recall
Class Epochs Score
were strategically incorporated between each hidden layer to
16 0.878 0.874 0.910 0.878
enhance model robustness and prevent overfitting during 32 0.851 0.844 0.896 0.851
training. 3 64 0.819 0.805 0.882 0.819
SGD
In our initial analysis, we began by evaluating the Class 128 0.774 0.746 0.864 0.774
performance metrics of the feed-forward DNN model using 256 0.700 0.664 0.725 0.700
the Adam optimizer and 20% Dropout layers across different 512 0.772 0.754 0.816 0.772
numbers of epochs. Our assessment concentrated on key
metrics such as F1-Score, precision, recall, and accuracy, as The performance of the DNN model using the SGD
outlined in Table 1. optimizer showed noticeable variations across different batch
sizes. The highest accuracy was achieved with a batch size of
TABLE 1 PERFORMANCE METRICS OF THE DNN MODEL USING THE ADAM 16, reaching 87.84%. However, as the batch size increased, a
OPTIMIZER ACROSS DIFFERENT BATCH SIZES decline in performance metrics was observed, with the lowest
No. No. F1- accuracy of 70.05% at a batch size of 256. The model with a
Optimizer ACC. Precision Recall
Class Epochs Score batch size of 16 also demonstrated a high F1-Score of 0.875
16 0.902 0.900 0.924 0.903 and precision of 0.911. Overall, the results indicate that
32 0.871 0.867 0.907 0.872 smaller batch sizes yielded better performance in stress
3 64 0.847 0.839 0.895 0.847
Adam
Class 128 0.871 0.866 0.906 0.871
classification tasks with the SGD optimizer.
256 0.835 0.825 0.889 0.836 To visually represent and clarify the best result of the
512 0.862 0.857 0.902 0.863 DNN when using the SGD optimizer, Fig. 2 presents the
confusion matrices. The optimal outcome was achieved with
From Table 1 it is clear that the Adam optimizer 16 epochs, demonstrating the model's performance under
demonstrated strong performance across all batch sizes in the these conditions.
stress classification task, with accuracies ranging from 83.5%
to 90.2%. The highest accuracy and F1-score 0.900 were
achieved with a batch size of 16, highlighting the model's
effective stress detection. Precision and recall were also
highest with a batch size of 16, at 0.9245 and 0.9030,
respectively. Smaller batch sizes yielded the best results,
suggesting better generalization and faster convergence.
Overall, the Adam optimizer proved robust and efficient,
particularly with smaller batch sizes, for classifying stress Fig.2. Confusion matrices for DNN performance with Adam optimizer (16
using physiological signals. epochs).
To visually represent and clarify the best result of the
DNN when using the Adam optimizer, Fig. 1 presents the DISCUSSION
confusion matrices. The optimal outcome was achieved with Our study explored stress classification using DNNs with
16 epochs, demonstrating the model's performance under Dropout layers, finding the Adam optimizer most effective
these conditions. for classifying three stress levels after 16 epochs. This
In the following analysis, we will examine the research examines how the optimizer, number of epochs, and
performance of the DNN with a 20% Dropout layer using the Dropout layers affect DL model efficiency, using chest patch
SGD optimizer. This evaluation involved a thorough signals for the first time.
investigation of different epoch settings, with performance A limitation of this study is that the model, trained on lab-
metrics such as Accuracy, F1-Score, Precision, and Recall collected data, may not generalize well to real-life conditions
carefully recorded and detailed in Table 2 for in-depth due to its inability to handle sensor noise in everyday
comparison and analysis. environments. In Table 3, a review will be conducted based
on the results from previous research.
TABLE 3REVIEW OF RESULTS BASED ON PREVIOUS RESEARCH FINDINGS

Ref. ML/DL Model ACC.

Authorized licensed use limited to: UNIVERSITY OF HERTFORDSHIRE. Downloaded on February 20,2025 at 12:36:38 UTC from IEEE Xplore. Restrictions apply.
196
 Proceedings of NILES2024:
6th Novel Intelligent and Leading Emerging Sciences Conference

[29] ML RF 0.856 [14] Bobade, P.; Vani, M., "Stress detection with machine learning and deep
[30] ML Fuzzy 0.765 learning using multimodal physiological data," in *Proceedings of the
[31] DL 1D-CNN 0.705 2020 Second International Conference on Inventive Research in
Computing Applications (ICIRCA)*, Coimbatore, India, 15-17 July
CONCLUSION 2020, pp. 51-57.
[15] T. Nijhawan, G. Attigeri, and T. Ananthakrishna, "Stress detection
This study demonstrated the effectiveness of using a DNN using natural language processing and machine learning over social
with Dropout layers and various optimization techniques for interactions," J. Big Data, vol. 9, p. 33, 2022.
stress classification. By leveraging physiological signals [16] R. Li and Z. Liu, "Stress detection using deep neural networks," BMC
from multiple sensors, the DNN was able to perform three- Med. Inform. Decis. Mak., vol. 20, p. 285, 2020.
class stress classification with high accuracy. The analysis [17] A. Kumar, K. Sharma, and A. Sharma, "Hierarchical deep neural
network for mental stress state detection using IoT based biomarkers,"
highlighted that the performance of the model significantly Pattern Recognition Letters, vol. 145, pp. 81-87, 2021.
increased with a smaller number of epochs, especially when [18] S. Vijayakumar, R. Flynn, P. Corcoran, and N. Murray, "CNN-based
using the Adam optimizer. The use of publicly available Emotion Recognition from Multimodal Peripheral Physiological
datasets such as WESAD allowed for rigorous training and Signals," in Proceedings of the IMX'22: ACM International
Conference on Interactive Media Experiences, Aveiro, Portugal, 22-24
validation, ensuring the robustness and generalizability of the June 2022.
model. These findings underscore the potential of DNNs in [19] M. Miao, L. Zheng, B. Xu, Z. Yang, and W. Hu, "A multiple frequency
enhancing stress detection and contribute to the broader field bands parallel spatial-temporal 3D deep residual learning framework
of physiological signal analysis for mental health assessment. for EEG-based emotion recognition," Biomed. Signal Process. Control,
vol. 79, p. 104141, 2023.
REFERENCES [20] Y. Tang, Y. Wang, X. Zhang, and Z. Wang, "STILN: A novel spatial-
temporal information learning network for EEG-based emotion
[1] Mental Health Wkly., "Report: Mental illness will cost the world $16
trillion (USD) by 2030," vol. 28, pp. 7-8, 2018. recognition," Biomedical Signal Processing and Control, vol. 85, p.
104999, 2023.
[2] J. Alonso, M. Codony, V. Kovess, M.C. Angermeyer, S.J. Katz, J.M.
[21] J. Choi, J. S. Lee, M. Ryu, G. Hwang, G. Hwang, and S. J. Lee,
Haro, G. De Girolamo, R. De Graaf, K. Demyttenaere, G. Vilagut, et
"Attention-LRCN: Long-term Recurrent Convolutional Network for
al., "Population level of unmet need for mental healthcare in Europe,"
Br. J. Psychiatry, vol. 190, pp. 299-306, 2007. Stress Detection from Photoplethysmography," in Proceedings of the
2022 IEEE International Symposium on Medical Measurements and
[3] S. Rathod, N. Pinninti, M. Irfan, P. Gorczynski, P. Rathod, L. Gega, Applications (MeMeA), Messina, Italy, 22-24 June 2022, pp. 1-6. New
and F. Naeem, "Mental Health Service Provision in Low- and Middle- York, NY, USA: IEEE, 2022.
Income Countries," Health Serv. Insights, vol. 10, pp.
1178632917694350, 2017. [22] J.A. Miranda-Correa, M.K. Abadi, N. Sebe, and I. Patras, "Amigos: A
dataset for affect, personality and mood research on individuals and
[4] M. Luo, L.X. Guo, M.Z. Yu, W.Y. Jiang, and H.Y. Wang, "The groups," IEEE Trans. Affect. Comput., vol. 12, pp. 479-493, 2018.
psychological and mental impact of coronavirus disease 2019
[23] S. Katsigiannis and N. Ramzan, "DREAMER: A database for emotion
(COVID-19) on medical staff and general public²A systematic review
and meta-analysis," Psychiatry Res., vol. 291, pp. 113190, 2020. recognition through EEG and ECG signals from wireless low-cost off-
the-shelf devices," IEEE J. Biomed. Health Inform., vol. 22, pp. 98-
[5] K.Y. Pan, A.A.L. Kok, M. Eikelenboom, M. Horsfall, F. Jorg, R.A. 107, 2017.
Luteijn, D. Rhebergen, P. van Oppen, E.J. Giltay, and B.W.J.H.
[24] S. Koldijk, M. Sappelli, S. Verberne, M.A. Neerincx, and W. Kraaij,
Penninx, "The mental health impact of the COVID-19 pandemic on
"The swell knowledge work dataset for stress and user modeling
people with and without depressive, anxiety, or obsessive-compulsive
research," in Proc. 16th Int. Conf. Multimodal Interact., Istanbul,
disorders: A longitudinal study of three Dutch case-control cohorts,"
Turkey, 2014, pp. 291-298.
Lancet Psychiatry, vol. 8, pp. 121-129, 2021.
[25] W.-L. Zheng and B.-L. Lu, "Investigating critical frequency bands and
[6] S. Rotenberg and J. J. McGrath, "Interrelation between autonomic and
channels for EEG-based emotion recognition with deep neural
HPA axis activity in children and adolescents," Biol. Psychol., vol.
networks," IEEE Trans. Auton. Ment. Dev., vol. 7, pp. 162-175, 2015.
117, pp. 16-25, 2016.
[26] S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T.
[7] S. Gedam and S. Paul, "A Review on Mental Stress Detection Using
Ebrahimi, T. Pun, A. Nijholt, and I. Patras, "Deap: A database for
Wearable Sensors and Machine Learning Techniques," IEEE Access,
vol. 9, pp. 84045-84066, 2021. emotion analysis using physiological signals," IEEE Trans. Affect.
Comput., vol. 3, pp. 18-31, 2011.
[8] T. Richter, B. Fishbain, G. Richter-Levin, and H. Okon-Singer,
[27] P. Schmidt, A. Reiss, R. Duerichen, C. Marberger, and K. Van
"Machine Learning-Based Behavioral Diagnostic Tools for
Laerhoven, "Introducing WESAD, a multimodal dataset for wearable
Depression: Advances, Challenges, and Future Directions," J. Pers.
Med., vol. 11, p. 957, 2021. stress and affect detection," in Proc. 20th ACM Int. Conf. Multimodal
Interact., Boulder, CO, USA, 16±20 Oct. 2018, pp. 400-408.
[9] W. M. L. Finlay and E. Lyons, "Methodological issues in interviewing
[28] X. Long, B. Yin, and R.M. Aarts, "Single-accelerometer-based daily
and using self-report questionnaires with people with mental
retardation," Psychol. Assess., vol. 13, pp. 319-335, 2001. physical activity classification," in Proc. Int. Conf. IEEE Eng. Med.
Biol. Soc., Minneapolis, MN, USA, 2009, pp. 6107-6110. doi:
[10] M. Sheikh, M. Qassem, and P. A. Kyriacou, "Wearable, 10.1109/IEMBS.2009.5334925.
Environmental, and Smartphone-Based Passive Sensing for Mental
Health Monitoring," Front. Digit. Health, vol. 3, p. 662811, 2021. [29] Y. S. Can, N. Chalabianloo, D. Ekiz, J. Fernandez-Alvarez, G. Riva,
and C. Ersoy, "Personal Stress-Level Clustering and Decision-Level
[11] M. Namavari, J. Lippoth, S. Knight, A. A. Jamali, M. Hedayati, R. J. Smoothing to Enhance the Performance of Ambulatory Stress
Spiteri, and S. Syed-Abdul, "Photoplethysmography Enabled Detection With Smartwatches," IEEE Access, vol. 8, pp. 38146±
Wearable Devices and Stress Detection: A Scoping Review," J. Pers. 38163, 2020.
Med., vol. 12, p. 1792, 2022.
[30] S. Pourmohammadi and A. Maleki, "Continuous mental stress level
[12] T. Iqbal, A. J. Simpkin, D. Roshan, N. Glynn, J. Killilea, J. Walsh, G. assessment using electrocardiogram and electromyogram signals,"
Molloy, S. Ganly, H. Ryman, E. Cohen, et al., "Stress Monitoring *Biomed. Signal Process. Control*, vol. 68, p. 102694, 2021.
Using Wearable Sensors: A Pilot Study and Stress-Predict Dataset,"
Sensors, vol. 22, p. 8135, 2022. [31] $ /LVRZVND 6 :LON DQG 0 3HOHJ &DWFKLQJ SDWLHQW¶V DWWHQWLRQ DW
the right time to help them undergo behavioural change: Stress
[13] J. Chen, M. Abbod, and J. S. Shieh, "Pain and Stress Detection Using classification experiment from blood volume pulse," in *Artificial
Wearable Sensors and Devices²A Review," Sensors, vol. 21, p. 1030, Intelligence in Medicine*, ser. Lecture Notes in Computer Science,
2021 Cham, Switzerland: Springer International Publishing, 2021, pp. 72±
82.

Authorized licensed use limited to: UNIVERSITY OF HERTFORDSHIRE. Downloaded on February 20,2025 at 12:36:38 UTC from IEEE Xplore. Restrictions apply.
197