International Journal of Engineering Science and Advanced Technology (IJESAT)

Vol 24 Issue 06, JUNE, 2024

MULTI CLASS STRESS DETECTION THROUGH HEART RATE

VARIABILITY USING LOGISTIC REGRESSION ALGORITHM
1
G;Kusuma Harinadh, [Link], Assistant Professor, Department of CSE, Universal College of Engineering And
Technology, Dokiparru, Perecharla(V), Guntur, Andhra Pradesh-522438.
2
[Link] Venkata Ramana, [Link], Department of CSE, Universal College of Engineering And Technology,
Dokiparru, Perecharla(V), Guntur, Andhra Pradesh-522438.

Abstract: This paper discuss stress as a natural human response to pressure, highlighting the risks associated with chronic
stress on mental and physiological health. It also introduces the use of heart rate variability (HRV) as a measure of stress
and describes a study investigating the role of HRV features in detecting stress using machine learning techniques,
particularly a Long Short Term Memory(LSTM), SVM, CNN and Decision Tree Algorithms. The study claims to achieve
high accuracy in multi-class stress detection (no stress, interruption stress, and time pressure stress) based on HRV
features, outperforming existing methods in the literature. Additionally, it demonstrates the effectiveness of essential HRV
features for stress detection using an analysis of variance technique.

1. INTRODUCTION because the ANS controls an individual's level of mental

Stress is necessary to preserve homeostasis when stress [8]. HRV is often extracted from ECG signals [9].
unpleasant stimuli lead to physical or mental imbalances. HRV is assessed by taking the difference in time between
Chronic stress causes an overactive sympathetic nervous two successive heartbeat peaks from an ECG signal. It is
system, which can result in abnormalities of the body, defined as the variance across periods between consecutive
mind, and behavior [1]. Subjective approaches are regular RR intervals, 1. Traditionally, fluctuations in both
frequently used to measure stress levels in order to extract RR intervals and instantaneous heart rate have been
perceptions of stress. Assessing stress levels using gathered referred to as HRVs [12]. Clinical circumstances and
heart rate viability (HRV) data can assist in eliminating specialist technological skills for data interpretation are
stress by tracking its impact on the autonomic nervous necessary in order to extract HRV from ECG
system (ANS) [2]. Anxiety disorder sufferers typically measurements. Heart rate readings can now be monitored
have a persistently lower resting heart rate than healthy and recorded using commercially accessible wearable or
individuals. HRV rises with relaxation and falls with stress, non-wearable Internet of medical things (IOMT) devices,
as [2] and [3] have shown. In fact, a slowing heart beats thanks to recent technological advancements in this field
with a greater HRV, and vice versa. Consequently, there is [17].
typically an inverse relationship between heart rate and
HRV [2], [3]. HRV changes throughout time in response to Many machine learning (ML) and deep learning (DL)
activity levels and the degree of stress associated with one's algorithms have been created in recent years for stress
job. prediction based on the analysis of ECG data (or HRV
Additionally, stress is typically linked to a negative features) [20], [21], [22], [23], [24], [25], [26], [27] (see
perception of an individual and is thought to be a more details in Sec. II). SWELL−KW is one of the two
subjective human emotion that may have an impact on most widely used publically available datasets for stress
one's physical and emotional health. It is defined as a detection, having been developed in [13] and [14].
biological and psychological response to environmental Nevertheless, no ML or DL study has yet attained ultra-
stimuli and biological or chemical agents that cause stress high accuracy for multi-class stress classification using the
in an organism [5], as well as internal or external stressors SWELL−KW dataset, particularly when it comes to multi-
[4]. Stress affects the ANS on a molecular level [6], which class stress level classification [15], [16]. Consequently,
controls the cardiovascular system using both sympathetic there is a research void in the area of creating innovative
and parasympathetic components. The sympathetic nervous machine learning models that can reach incredibly high
system in humans [7] functions similarly to the gas pedal in prediction accuracy. We have designed and developed a
an automobile. It triggers the fight-or-flight reaction, which one-dimensional convolution neural network (1D CNN)
gives the body more energy to react to harmful stimuli. On model for multi-class stress classification, and we show its
the other hand, the parasympathetic nervous system acts as superiority over the state-of-the-art models based on the
the body's brake. When a threat has passed, it relaxes the SWELL-KW dataset in terms of prediction accuracy. This
body, which promotes slumber and digestion. model is motivated by various existing applied ML and DL
Physiological measures including electrocardiogram based studies on HRV feature processing for stress level
(ECG), electromyogram (EMG), galvanic skin response classifications. More precisely, we have conducted
(GSR), heart rate, blood pressure, breath frequency, and research on stress detection with multi-layer perceptron
respiration rate can be used to measure mental stress (MLP) methods, which draw inspiration from the fully

ISSN No: 2250-3676 [Link] Page | 68

International Journal of Engineering Science and Advanced Technology (IJESAT)
Vol 24 Issue 06, JUNE, 2024
connected neural network (FCNN) architecture, as well as assessment for a long time, but they are not able to track
conventional machine learning algorithms. We have stress levels over time or in real time, which makes it
created a 1D CNN model in our work that is predicated on difficult to pinpoint the causes and stressful components of
the convolution technique. Because MLP accepts vectors the job. An efficient way around this restriction is to
as input and CNN accepts tensors in order to comprehend analyze physiological signals, which are continuously
spatial relationships, CNN minimizes the number of measurable using ambient or wearable sensors. Prior
training parameters. We have also introduced a feature research in this area has mostly concentrated on stress
reduction algorithm based on analysis of variance assessment using invasive wearable systems that are prone
(ANOVA) F-test, and while the accuracy achieved with to noise and performance-degrading aberrations. A
full features is nearly 100%, we show that it is possible to wearable and ambient hardware-software platform that can
achieve an accuracy score of 96.5% with less than half of detect human stress without interfering with regular work
the features available in the SWELL−KW dataset. During tasks is described in one of our recently published
the model training phase, this kind of feature extraction publications. It is also slightly prone to artifacts caused by
lessens the computational strain. The primary contributions motions. The system's low accuracy in identifying various
and innovation of this work can be summed up as follows: stress levels is one of its limitations. For this reason, the
study focused on enhancing the platform's software
With exceptional performance, we have created a novel 1D performance through the application of deep learning
CNN model that can detect multi-class stress state with techniques. Three neural networks were used for this
99.9% accuracy, a Precision, F1-score, and Recall score of purpose; the 1D-convolutional neural network performed
1.0, respectively, and a 99.9% Matthews correlation the best, identifying two degrees of stress with an accuracy
coefficient (MCC) score. We think this is the first study to of 95.38%, which is a significant improvement over the
obtain such a high accuracy score for classifying stress into results reported earlier.
many classes. 2) Stress Detection via Multimodal Machine Learning
Moreover, we disclose that the precise classification of Summary: The modern information age has made people
multi-class stress does not require the use of all 34 HRV more knowledge-focused in their lifestyles, which has
variables. Our model with selected top-ranked HRV resulted in sedentary jobs. Numerous medical conditions
features achieves excellent accuracy without sacrificing and mental illnesses have resulted from this. One of the
critical information, and it does so without requiring most important yet often overlooked components of the
resource-intensive computation, outperforming existing fast-paced environment we live in today is mental
classification models based on the SWELL-KW dataset. wellbeing. A person's daily activities and performance
We have performed feature optimization to select an might be hindered by mental health problems, which can
optimized feature set to train a 1D CNN classifier, and we also have a direct or indirect impact on other areas of
have achieved a performance score that beats them. This is human physiology. Finding a person's stress level and
how the rest of the paper is structured. We briefly explain stress pattern, which could result in major mental illnesses,
the framework for stress state categorization, dataset, and can be difficult and entail a number of variables. Accurate
data preprocessing in Sec. III after summarizing previous identification can be accomplished by combining these
work and highlighting the differences between our work several modalities that result from an individual's
and the current work in Sec. II. In Section IV, the created behavioral patterns, which can be caused by a variety of
CNN model is then given. The performance measures to causes. For this objective, certain methodologies are
assess the suggested classifier are then defined in Sec. V, recognized in the literature; nonetheless, for such
and the numerical results are shown in Sec. VI. Section VII multimodal fusion problems, very few machine learning-
provides more discussion. In Sec. VIII, the paper is finally based solutions are provided. This paper proposes a
ended. multimodal AI-based system to track an individual's stress
levels and work habits. We provide a technique that
2. LITERATURE SURVEY combines heterogeneous raw sensor data streams (e.g.,
posture, heart rate, computer interaction, and facial
1) A Deep Learning-Based Platform for Minimally expressions) to effectively identify stress brought on by
Invasive Multisensory Device Stress Detection in Workers workload. This information can be safely kept and
Summary: When it comes to assessing the stress levels of examined to identify and comprehend individual
employees who work in increasingly complex work behavioral patterns that cause weariness and mental strain.
environments, Industry 4.0's requirement for significant This work makes two contributions: first, it suggests a
human-machine interaction poses new difficulties. Without multimodal AI-based fusion technique to detect stress and
a question, stress at work has a big impact on people's its degree; second, it identifies a pattern of stress over time.
overall stress levels, which can have negative effects on On the test set, we were able to detect and classify stress
people's quality of life and long-term health problems. with 96.09% accuracy. Furthermore, by employing these
Psychological questionnaires have been used for stress modalities, we were able to lower the stress scale
prediction model loss to 0.036. The general public, in

ISSN No: 2250-3676 [Link] Page | 69

International Journal of Engineering Science and Advanced Technology (IJESAT)
Vol 24 Issue 06, JUNE, 2024
particular those with sedentary professions, may find it Additional research indicates that although occupational
useful to monitor and diagnose stress levels, particularly in stress and thermal comfort have an impact on HRV, their
light of the COVID-19 pandemic. effects may not overlap and that they can be differentiated
3) Wearable devices as feasible physiological markers for almost exactly from one another. These findings suggest
stress detection through HRV features that an autonomous, non-intrusive system that provides
Summary: One of the main factors contributing to car thermal comfort and forecasts work stress based on
crashes, which in turn cause a significant number of individuals' heart rate variability may be feasible to create.
fatalities and injuries annually, has been identified as
stress. Physiological measures can be used to quantify 3. EXISTING SYSTEM
stress, and in this study, the aspects that are often derived A thorough analysis of the HRV data quality pertaining to
by wearable technology will be the main focus. As a result, data obtained from ECG and IoMT devices, including the
heart rate variability will be the primary focus of the Motorola Droid, Elite HRV, H7, and Polar, can be found in
investigation (HRV). The purpose of this work is to [18]. When comparing the HRV values derived from
examine the function of features obtained from HRV as commercially available IoMT devices with measurements
stress markers. This is accomplished by trying various based on an ECG instrument, 23 studies found small
machine learning techniques, such as K-Nearest Neighbor, inaccuracies. Such a small-scale inaccuracy in HRV
and creating a strong predictive model that can precisely measurements is appropriate in practice because obtaining
categorize stress levels from ECG-derived HRV data, HRVs using portable IoMT devices is more convenient,
collected from drivers of motor vehicles. In addition, a economical, and doesn't require any laboratory or clinical
machine learning model that classifies stress using HRV equipment [18], [19]. However, there have been numerous
features generated from heart rate measurements received recent studies on the use of ML and DL algorithms to
from wearable devices will be developed using the models categorize stress using ECG data analysis [20], [21], [22],
that yielded the highest prediction power as a guide. We and [23]. Prioritizing binary (stress versus nonstress) and
show that HRV characteristics are useful indicators for multi-class stress classifications has been the main focus of
stress detection because the best machine learning model existing algorithms. For example, the authors in [4]
that was created had an 80% recall rate. Additionally, this distinguished between physiological states that are normal
study shows that HRV measures were significant for stress and stressed based on HRV data. The authors examined
detection, including the Root mean square differences of many machine learning (ML) techniques, including naive
successive NN intervals (RMSSD), Standard deviation of Bayes, random forest, knearest neighbor (KNN), support
the average NN intervals (SDNN), and Average of normal- vector machine (SVM), MLP, and gradient boosting, for
to-normal (AVNN) intervals (AVNN). The suggested the purpose of identifying stress. Their highest recall rating
approach can also be applied to any application where it's was eighty percent. Using the time-domain and frequency-
critical to track stress levels in a non-intrusive way, such as domain properties of HRV, the authors of a comparable
mental health, anxiety management, or physical comparison study [27] shown that SVM with radial basis
rehabilitation. function (RBF) produced an accuracy score of 83.33% and
4) In an office setting, thermal comfort and stress 66.66%, respectively. Moreover, the optimal temporal and
recognition frequency domain components in HRV have been chosen
Summary: Among the challenges that office workers deal using dimension reduction approaches [24]. In [25], the
with on a daily basis are work-related stress and discomfort authors used CNN to do binary classification, or stressed
from the heat. Office workers may feel stressed out at work versus not stressed, and they were able to acquire an
since their jobs require a lot of mental effort and are accuracy score of 98.4%.
lengthy. In addition, the methods used to provide thermal In a different study called StressClick [26], mouse-click
comfort today are energy-intensive and inefficient. In events—that is, the gaze-click pattern obtained from a
earlier research, we suggested a technique for providing commercial computer webcam and mouse—were used to
effective thermal comfort based on an individual's heart identify subjects as stressed or not stressed using a random
rate variability (HRV). However, because a person's HRV forest method. In [14], SVM based on the SWELL−KW
can also be impacted by work stress, this study looks into dataset was used to accomplish tasks for multi-class stress
the feasibility of differentiating between changes in HRV classification (e.g., no stress, interruption stress, and time
that result from work stress and changes that are caused by pressure stress). They were 90% accurate at their best.
thermal discomfort. We tested this on participants Additionally, for multi-class (amusement versus baseline
performing the Advanced Trail Making Test (ATMT) and versus stress) and binary (stress versus non-stress)
found that stress changes heart rate variability (HRV) and classifications, another publicly accessible dataset,
that there is a 100% accuracy difference between stressed WESAD, was employed in [27]. ML algorithms were able
and non-stressed respondents. Using the multimodal to reach up to 81.65% accuracy for three-class
SWELL knowledge work (SWELL-KW) stress dataset, we categorization in their investigations. The three-class stress
verified our technique and obtained comparable results classification task was completed with an accuracy level of
(average recall of 99.75% and accuracy of 99.25%). 84.32% by the authors, who also evaluated the

ISSN No: 2250-3676 [Link] Page | 70

International Journal of Engineering Science and Advanced Technology (IJESAT)
Vol 24 Issue 06, JUNE, 2024
effectiveness of deep learning algorithms. In addition, it is  To make the data fit the feature ranking algorithm,
noteworthy to mention that newer deep learning methods preprocessing is done. For feature ranking and
have emerged as effective tools for two-dimensional data selection in this work, ANOVA F-tests and
categorization tasks, such as genetic deep learning forward sequential feature selection are utilized,
convolutional neural networks (GDCNNs) [38], [39]. respectively.
However, further alterations or adaptations are needed in  A DL-based multi-class classifier that has been
order to apply GDCNN to 1D data, and this subject is constructed is subjected to training, testing, and
outside the purview of this work. validation using important features and
A novel double-check mapping function called DCMF, annotations labeled by medical specialists, such as
which is utilized to offer early attack detection at the no stress, interruption condition, and time
switch level, is never proposed by the system. This is one constraint.
of its [Link] following issues plague an
existing system: Three steps: 1) replace missing values; 2) SYSTEM ARCHITECTURE
encode categorical data; and 3) scale features.

Furthermore, it is worth mentioning that novel deep

learning techniques, such as genetic deep learning
convolutional neural networks (GDCNNs) [38], [39], have
appeared as a powerful tool for two-dimensional data
classification tasks. To apply GDCNN to 1D data,
however, comprehensive modifications or adaptations are
required and such a topic is beyond the scope of this paper.

DISADVANTAGES
The system never proposes a new double-check mapping
function called DCMF, which is used to provide early
attack detection at the switch level.
An existing system has the following problems
1) missing values replacement 
2) encoding categorical data and Fig 1: System Architecture
3) feature scaling. 5. ALGORITHMS
5.1 DECISION TREE CLASSIFIERS
4. PROPOSED SYSTEM Decision tree classifiers have proven effective in a wide
With exceptional performance, we have created a novel range of applications. The ability to extract descriptive
1D CNN model that can detect multi-class stress state decision-making knowledge from the provided data is their
with 99.9% accuracy, a Precision, F1-score, and Recall
key characteristic. Training sets can be used to create
score of 1.0, respectively, and a 99.9% Matthews
correlation coefficient (MCC) score. We think this is the decision trees. The following is the process for creating
first study to obtain such a high accuracy score for such a generation based on the set of objects (S), each of
classifying stress into many classes. Moreover, we which is a member of one of the classes C1, C2,..., Ck:
disclose that the precise classification of multi-class stress Step 1: The decision tree for S has a leaf labeled with this
does not require the use of all 34 HRV variables. In order class if every item in S is a member of the same class, such
to train a 1D CNN classifier, we undertook feature as Ci.
optimization. By doing so, we were able to choose an
Step 2: If not, let T represent a test with the possible results
improved feature set and outperform the competition in
terms of performance, using the SWELL-KW dataset as O1, O2,... On. The test divides S into subsets S1, S2,..., Sn
our base. Our model, which uses a subset of the highest- where each object in Si has result Oi for T. This is because
ranked HRV features, provides exceptional accuracy each object in S has a single outcome for T. T becomes the
without requiring computationally demanding processes. decision tree's root, and we create a subsidiary decision tree
for each outcome Oi by applying the same process
ADVANTAGES recursively to the set Si.
 Medical professionals label important features and
5.2 CLASSIFIER OF LOGISTIC REGRESSION
annotations (such as no stress, interruption
situation, and time pressure) to train, test, and The relationship between a set of independent
validate the developed DL-based multi-class (explanatory) variables and a categorical dependent
classifier. variable is examined using logistic regression analysis.

ISSN No: 2250-3676 [Link] Page | 71

International Journal of Engineering Science and Advanced Technology (IJESAT)
Vol 24 Issue 06, JUNE, 2024
When the dependent variable simply has two values, such optimization issue analytically, it is a discriminating
as 0 and 1, or Yes and No, the term logistic regression is approach that consistently yields the same optimum hyper
employed. When the dependent variable, such as married, plane parameter—in this case, in contrast to perceptions or
single, divorced, or widowed, has three or more distinct genetic algorithms (GAs), which are both often employed
values, the term multinomial logistic regression is typically in machine learning for categorization. The commencement
reserved for that situation. While the dependent variable's and termination criteria have a significant influence on the
data type differs from multiple regression's, the procedure's solutions for perceptions. Training yields precisely
practical application is comparable. When it comes to specified SVM model parameters for a given training set
examining categorical response variables, discriminant for a certain kernel that converts the data from the input
analysis and logistic regression are rivals. Compared to space to the feature space; in contrast, the perception and
discriminating analysis, many statisticians believe that GA classifier models vary with each training set. Only
logistic regression is more adaptable and suitable for minimizing mistake during training is the goal of GAs and
predicting the majority of scenarios. This is so because, perceptions, which translates into several hyper planes
unlike discriminant analysis, logistic regression does not fulfilling this need.
make the assumption that the independent variables are 5.4 RANDOM FOREST
normally distributed. On both categorical and numeric Random forests, also known as random decision forests,
independent variables, this program computes binary and are an ensemble learning technique that builds a large
multinomial logistic regression. Together with the number of decision trees during the training phase for
regression equation, quality of fit, odds ratios, confidence problems including regression, classification, and other
intervals, probability, and deviance are also reported. applications. The class that the majority of the trees choose
Complete residual analysis is carried out, including with is the random forest's output for classification problems.
diagnostic residual reports and charts. It has the ability to The mean or average prediction made by each individual
search for the optimal regression model with the fewest tree is returned for regression tasks. The tendency of
independent variables using an independent variable subset decision trees to overfit to their training set is compensated
selection search. It offers ROC curves and confidence for by random decision forests. Although they are less
intervals on expected values to assist in figuring out the accurate than gradient enhanced trees, random forests still
ideal cutoff point for categorization. It does this by perform better than choice trees in most cases. Their
automatically categorizing rows that are not utilized in the performance, however, may be impacted by the
study, allowing you to verify your results. peculiarities of the data. Tin Kam Ho[1] developed the first
5.3 SVM random decision forest algorithm in 1995 by utilizing the
A discriminate machine learning approach is used in random subspace technique, which is a means of putting
classification problems to create a discriminate function Eugene Kleinberg's "stochastic discrimination" approach to
that can properly predict labels for newly acquired classification into practice. Leo Breiman and Adele Cutler
instances, based on an independent and identically created an expansion of the algorithm and filed for a
distributed (iid) training dataset. A discriminating trademark for "Random Forests" in 2006; as of 2019,
classification function takes a data point x and assigns it to Minitab, Inc. is the owner of this [Link] order to
one of the several classes that are part of the classification create a set of decision trees with controlled variance, the
job, in contrast to generative machine learning techniques extension combines Breiman's "bagging" concept with
that need calculations of conditional probability random feature selection, which was initially presented by
distributions. Discriminate procedures are less effective Ho[1] and then separately by Amit and Geman[13].
than generative approaches, which are typically employed Businesses commonly utilize random forests as "blackbox"
when prediction requires outlier identification. This is models because they produce good predictions across a
especially true for multidimensional feature spaces and wide variety of data with minimal processing power.
situations where just posterior probabilities are required.
Discriminate approaches also need fewer training data and 6. RESULTS
processing resources. Finding the equation for a 6.1 Output Screens
multidimensional surface that optimally divides the various
classes in the feature space is the geometric equivalent of
learning a classifier. Because SVM solves the convex

ISSN No: 2250-3676 [Link] Page | 72

International Journal of Engineering Science and Advanced Technology (IJESAT)
Vol 24 Issue 06, JUNE, 2024

Fig 6.4 Prediction Result

Fig 6.1 Comparison of Accuracy graph in bar chart In above screen shows the prediction result for the stress
detection.
In above screen is the Comparison of Accuracy graph in
bar chart for different ml algorithms 7. CONCLUSION
Using HRV signals, we have created a novel 1D CNN
model for stress level categorization in this work. The
proposed model has been verified using the publically
available SWELL−KW dataset. We also used an ANOVA
feature selection method for dimension reduction in our
model. We show, by comprehensive training and
validation, that our model outperforms the state-of-the-art
models when all features are used in terms of key
performance measures, such as Accuracy, Precision,
Recall, F1-score, and MCC. Additionally, our method with
ANOVA feature reduction performs exceptionally well. In
order to make real-time stress detection a reality, we intend
Fig 6.2 Comparison of Accuracy graph in line chart
to continue researching the viability of refining the model
to fit it into edge devices in future work.
In above screen is the Comparison of Accuracy graph in
line chart for different ml algorithms
8. REFERENCES
[1] K. K. Kumar, S. G. B. Kumar, S. G. R. Rao and S. S. J.
Sydulu, "Safe and high secured ranked keyword searchover
an outsourced cloud data," 2017 International Conference
on Inventive Computing and Informatics (ICICI),
Coimbatore, India, 2017, pp. 20-25, doi:
10.1109/ICICI.2017.8365348.

[2] D. Muhajir, F. Mahananto, and N. A. Sani, ‘‘Stress

level measurements using heart rate variability analysis on
Android based application,’’ Proc. Comput. Sci., vol. 197,
pp. 189–197, Jan. 2022.
Fig 6.3 Comparison of Accuracy graph in pie chart
In above screen is the Comparison of Accuracy graph in [3] J. Held, A. Vîslă, C. Wolfer, N. Messerli-Bürgy, and C.
pie chart for different ml algorithms Flückiger, ‘‘Heart rate variability change during a stressful
cognitive task in individuals with anxiety and control
participants,’’ BMC Psychol., vol. 9, no. 1, p. 44, Mar.
2021.

[4] Kiran Kumar Kommineni, Ratna Babu Pilli, K.

Tejaswi, P. Venkata Siva, Attention-based Bayesian
inferential imagery captioning maker, Materials Today:

ISSN No: 2250-3676 [Link] Page | 73

International Journal of Engineering Science and Advanced Technology (IJESAT)
Vol 24 Issue 06, JUNE, 2024
Proceedings, 2023, , ISSN 2214-7853, Ambient Intell. Human. Comput., pp. 1–11, Aug. 2022,
[Link] doi: 10.1007/s12652- 022-04365-z.

[5] J. A. Miranda-Correa, M. K. Abadi, N. Sebe, and I. [16] R. Walambe, P. Nayak, A. Bhardwaj, and K. Kotecha,
Patras, ‘‘AMIGOS: A dataset for affect, personality and ‘‘Employing multimodal machine learning for stress
mood research on individuals and groups,’’ IEEE Trans. detection,’’ J. Healthcare Eng., vol. 2021, Oct. 2021, Art.
Affect. Comput., vol. 12, no. 2, pp. 479–493, Apr./Jun. no. 9356452.
2021.
[17] A. Ibaida, A. Abuadbba, and N. Chilamkurti,
[6] E. Won and Y.-K. Kim, ‘‘Stress, the autonomic nervous ‘‘Privacy-preserving compression model for efficient IoMT
system, and the immune-kynurenine pathway in the ECG sharing,’’ Comput. Commun., vol. 166, pp. 1–8, Jan.
etiology of depression,’’ Current Neuropharmacol., vol. 14, 2021.
no. 7, pp. 665–673, Aug. 2016.
[18] W. C. Dobbs, M. V. Fedewa, H. V. MacDonald, C. J.
[7] B. Olshansky, H. N. Sabbah, P. J. Hauptman, andW. S. Holmes, Z. S. Cicone, D. J. Plews, and M. R. Esco, ‘‘The
Colucci, ‘‘Parasympathetic nervous system and heart accuracy of acquiring heart rate variability from portable
failure: Pathophysiology and potential implications for devices: A systematic review and meta-analysis,’’ Sports
therapy,’’ Circulation, vol. 118, no. 8, pp. 863–871,Aug. Med., vol. 49, no. 3, pp. 417–435, Mar. 2019.
2008.
[19] C.-M. Chen, S. Anastasova, K. Zhang, B. G. Rosa, B.
[8] S. Goel, P. Tomar, and G. Kaur, ‘‘ECG feature P. L. Lo, H. E. Assender, and G.-Z. Yang, ‘‘Towards
extraction for stress recognition in automobile drivers,’’ wearable and flexible sensors and circuits integration for
Electron. J. Biol., vol. 12, no. 2, pp. 156–165, Mar. 2016. stress monitoring,’’ IEEE J. Biomed. Health Informat., vol.
24, no. 8, pp. 2208–2215, Aug. 2020.
[9] V. N. Hegde, R. Deekshit, and P. S. Satyanarayana, ‘‘A
review on ECG signal processing and HRV analysis,’’ J. [20] R. A. Rahman, K. Omar, S. A. M. Noah, M. S. N. M.
Med. Imag. Health Informat., vol. 3, no. 2, pp. 270–279, Danuri, and M. A. Al-Garadi, ‘‘Application of machine
Jun. 2013. learning methods in mental health detection: A systematic
review,’’ IEEE Access, vol. 8, pp. 183952–183964, 2020.
[10] M. Vollmer, ‘‘A robust, simple and reliable measure
of heart rate variability using relative RR intervals,’’ in [21] S. H. Jambukia, V. K. Dabhi, and H. B. Prajapati,
Proc. Comput. Cardiol. Conf. (CinC), Sep. 2015, pp. 609– ‘‘Application of machine learning methods in mental
612. health detection:Asystematic review,’’ in Proc. Int. Conf.
Adv. Comput. Eng. Appl., 2015, pp. 714–721.
[11] M. H. Kryger, T. Roth, andW. C. Dement, Principles
and Practice of Sleep Medicine, 5th ed. Amsterdam, The [22] S. Celin and K. Vasanth, ‘‘ECG signal classification
Netherlands: Elsevier, 2011. using various machine learning techniques,’’ J. Med. Syst.,
vol. 42, no. 12, p. 241, Oct. 2018.
[12] M. Malik, J. T. Bigger, A. J. Camm, R. E. Kleiger, A.
Malliani, A. J. Moss, and P. J. Schwartz, ‘‘Heart rate [23] A. Padha and A. Sahoo, ‘‘A parametrized quantum
variability: Standards of measurement, physiological LSTM model for continuous stress monitoring,’’ in Proc.
interpretation, and clinical use,’’ Eur. Heart J., vol. 17, no. 9th Int. Conf. Comput. Sustain. Global Develop.
3, pp. 354–381, Mar. 1996. (INDIACom), Mar. 2022, pp. 261–266.

[13] S. Koldijk, M. Sappelli, S. Verberne, M. A. Neerincx, [24] S. Sriramprakash, V. D. Prasanna, and O. V. R.

and W. Kraaij, ‘‘The SWELL knowledge work dataset for Murthy, ‘‘Stress detection in working people,’’ Proc.
stress and user modeling research,’’ in Proc. 16th Int. Conf. Comput. Sci., vol. 115, pp. 359–366, Dec. 2017.
Multimodal Interact., Nov. 2014, pp. 291–298.
[25] P. Sarkar and A. Etemad, ‘‘Self-supervised learning
[14] K. K. . Kommineni and A. . Prasad, “A Review on for ECG-based emotion recognition,’’ in Proc. IEEE Int.
Privacy and Security Improvement Mechanisms in Conf. Acoust., Speech Signal Process. (ICASSP), May
MANETs”, Int J Intell Syst Appl Eng, vol. 12, no. 2, pp. 2020, pp. 3217–3221.
90–99, Dec. 2023.

[15] M. Albaladejo-González, J. A. Ruipérez-Valiente, and

F. G. Mármol, ‘‘Evaluating different configurations of
machine learning models and their transfer learning
capabilities for stress detection using heart rate,’’ J.

ISSN No: 2250-3676 [Link] Page | 74