See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/376244663

Automated Stress Level Detection for Hospital Nurses: A Single Triaxial

Wearable Accelerometer Sensor System Approach

Conference Paper · October 2023

DOI: 10.1109/CCE60043.2023.10332832

CITATIONS READS

6 92

2 authors, including:

Mohammad Sakib
University of Liberal Arts Bangladesh
14 PUBLICATIONS 20 CITATIONS

SEE PROFILE

All content following this page was uploaded by Mohammad Sakib on 01 January 2024.

The user has requested enhancement of the downloaded file.

2023 20th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE). Mexico City, Mexico. October 25-27, 2023

Automated Stress Level Detection for Hospital

Nurses: A Single Triaxial Wearable Accelerometer
Sensor System Approach
Mohammad Sakib, Student Member, IEEE Syeda Shanaz Pervez
Department of Electrical and Electronic Engineering Department of Electrical and Electronic Engineering
University of Liberal Arts Bangladesh University of Liberal Arts Bangladesh
Dhaka, Bangladesh Dhaka, Bangladesh
Sakib.eee@ulab.edu.bd shanaz.pervez.eee@ulab.edu.bd
(Corresponding Author)

Abstract—Advancements in wearable sensor technologies have and immediate remedies to alleviate stress and enhance overall
revolutionized healthcare, particularly in nurses stress level mental well-being [2]. Nurses' stress level detection involves
detection using accelerometer-based data. Accurate detection of the utilization of cutting-edge technology, encompassing smart
nurse stress throughout a day holds significant potential to
enhance healthcare outcomes and optimize nursing workflows. devices and advanced algorithms, to continuously monitor and
However, existing research in this domain has been limited by evaluate the physiological and behavioral metrics while
small data-sets and short sampling duration’s, leading to performing their professional duties [3].
constrained generalizability. To address these limitations
comprehensively, this paper presents a data on nurses’ mental The primary aim is to identify and detect signs of stress
health in a practical setting are gathered using a wearable during their duties. Recent studies have found a strong
accelerometer sensor. To record the physiological and behavioral relationship between a person’s physical and mental health
traits of the nurses, a single triaxial accelerometer is mounted on during the previous two decades. A cohort of assiduous
their heads. Multiple machine learning methods, such as Nearest
Neighbor Classifiers (k-NN), Neural Networks (NN), Support
researchers undertook a comprehensive meta-analysis of varied
Vector Machines (SVM), Naive Bayes (NB), Discriminant Analysis studies aimed at evaluating stress levels in the human
(DA), Decision Trees (DT), and Ensemble Classifiers, were used to population. Stress constitutes a significant concern within
assess and compare the accuracy of nurses’ stress levels. The contemporary society. The timely detection of stress is
Medium Gaussian-SVM study showed the best average paramount as it curtails deleterious effects and precludes its
performance of 80.4% and mean efficacy F-measure of 80.2%
across all cases. whereas the Coarse Gaussian-SVM evaluation deep-seated manifestation. Presently, a surge in interest
exhibited the lowest mean performance of 26.5%, accompanied by surrounds the development of automated approaches,
a mean F-measure of 26.1%. On a scale of 1–7, a questionnaire leveraging smart devices and computer algorithms, to discern
evaluated privacy issues. The results show our method’s potential stress [4]. The potential applications of such automated stress
for accurately determining nurses’ levels of stress, making it detection systems span diverse environments, including
helpful for implementation in the future.
nursing, workplaces, individuals coping with air travel phobias,
Keywords—Nurses Stress Detection, Accelerometer-Based
Data, Machine Learning, Wearable Sensor, Classification motorcycle riding, and vehicular operation.
Algorithm, Healthcare Outcomes, Better Manage Heath The main goal of the proposed study is to automatically
assess nurses’ stress levels using data from wearable single 3D
I. INTRODUCTION accelerometer sensors collected during their activities. This
In the contemporary era, the global problem of stress is on discernment holds the potential to significantly contribute to
the rise, notably affecting nurses who grapple with elevated stress management and proactively prevent stress-induced
workload and occupational stress, further contributing to the ailments, thus serving as a valuable safeguard. Furthermore,
worldwide scarcity of nursing professionals. Regrettably, no this study is expected to stimulate and inspire further
prior research has been undertaken to gauge the extent of exploration by researchers in this domain. In the ensuing
workload and occupational stress among nurses in Dhaka, segment of the investigation, relevant antecedent research is
Bangladesh [1]. Stress, in its negative form, can adversely cited. To accomplish the goals of this endeavor, a series of
impact a person’s life and well-being. However, when stress is systematic measures were pursued. These entailed acquiring a
constructive and manageable, it can encourage innovation and comprehensive comprehension of the dataset framework,
help people come up with responses to common problems. painstakingly purifying and transforming the data to conform
Although eradicating stress entirely may prove challenging, with machine learning classification methodologies, and
mastering the monitoring and regulation of its physical and methodically scrutinizing, constructing, and juxtaposing
emotional repercussions holds paramount importance. In light diverse classification models. In a related study conducted by
of this, the provision of additional mental health counselling Qian Zhang et al. [5], real-time stress monitoring models were
and support programs becomes imperative, offering practical employed, focusing on physiological attributes and targeting

979-8-3503-0676-7/23/$31.00 ©2023 IEEE

intensive care unit (ICU) nurses. Over a span of 12-hour shifts,
23 subjects' accelerometer data were continuously logged.
Notably, a variety of machine learning strategies, such as
eXtreme Gradient Boosting (XGBoost), were used to
accurately categorize stress levels.

II. RELATED WORK

Several studies have focused on stress conditions during
various activities such as nursing, working, and driving.
Researchers have explored various methodologies and datasets
to address the challenges associated with this domain. This
section presents a review of prominent works in the field of
automatic stress detection among nurses and other professions,
employing a similar approach. The contributions of these
studies are highlighted and discussed in Table I. Fig. 1. Stress Detection Device and Associated Methodology
Sozo Inoue et al. employed wearable sensors and machine
machine learning approaches are used. The outcomes are then
learning in a study to recognize nursing activities [6]. They
carefully analyzed and contrasted. For improved health and
utilized a sizable dataset of nurse activities in controlled
wellness, the proposed approach addresses the crucial issue of
settings. The approach involved leveraging prior knowledge of
nurses’ levels of stress.
daily activity patterns, incorporating steps like importance
sampling, likelihood and Bayesian estimation. III. METHODOLOGY
Seyedmajid Hosseini et al. have presented an exceptional
The present section provides a comprehensive description of
stress detection dataset from nurses working during the
the materials and procedure techniques employed for the
COVID-19 pandemic [7]. This dataset includes contextual
collection of participant data. We also concentrated on the pre-
details of stress-related events and physiological metrics like
processing of the data and the outcome analysis. The suitable
heart rate, skin temperature, and electrodermal activity.
data collection tools and investigation methods for assessing
Collected via recurrent smartphone surveys, this valuable
nurses’ levels of stress are shown in Fig. 1.
dataset sheds light on nurses' stress responses during
challenging circumstances. It holds promise for advancing A. Data Collection
stress detection research in healthcare, addressing the intricate For our research work, we used a 3D accelerometer sensor
nature of stress assessment influenced by diverse factors. called ADXL345 and an Arduino UNO for data collection. The
Hossein Hamidi Shishavan et al. developed a versatile sensor was calibrated from the beginning to guarantee
wristband for ongoing physiological monitoring [8]. Through accuracy. The precision and repeatability of the sensors are
24-hour trials with two working groups, the wristband's verified through calibration, which enhances the quality of the
effectiveness and reliability were demonstrated. Unobtrusively, data [11]. To begin, our research team selected some
it tracked vital signs like pulse transit time and activity levels government hospitals in the Sher-Bangla Nagar area of Dhaka,
during personal and professional activities. Bangladesh (hospital names were excluded to avoid bias). This
Ashlesha Akella et al. explored EEG signals and latent location allowed easy access to multiple hospitals. The nurse
biomarker representations [9]. They assessed three key stress directors, human resources, and hospital executives heard about
biomarkers using four classifiers, achieving 91% accuracy our study design and stress measuring methodology. We got the
compared to the original 83%. Emil Jovanov et al. introduced hospital compliance's permission after the nursing department
UAHealth, a wearable system for nurse stress monitoring [10]. expressed interest. The study participants were nurses working
It integrates personal monitors with iPhone smartphones via regular shifts, including morning and night shifts. In order to
Ant+ wireless tech, collecting real-time physiological and record physiological signals throughout their regular work
movement data. A stress index is computed, triggering a shifts with the least amount of disruption, accelerometer
personalized questionnaire when stress surpasses a set devices were deployed. Participants were free to use the system
threshold for 30 minutes, aiding stress source identification. as they wished throughout the study period. Over the course of
In summary, existing research has made significant approximately one week, data was gathered from a cohort of 12
progress in detecting nurses’ stress levels. However, there are nurses, encompassing both female and male individuals, who
opportunities for improvement in terms of dataset scale, were actively engaged in hospital duties, with ages ranging
generalization capabilities, and sensor modality selection. This from 27-50 years. Each participant provided their
study aims to overcome previous limitations by leveraging psychological data for 80-90 minutes between 03/06/2023 to
an extensive dataset and introducing new wearable sensor 10/06/2023. The data was acquired at a sampling frequency of
technology for automatic stress detection in nurses. For 100Hz [12]. The data was collected from nurses in various
precise stress level detection and categorization, a variety of departments, including Coronary Artery Disease (CAD),
 TABLE I
SUMMARY AND FINDINGS OF THE REVIEWED ARTICLE

References Apparatus Subject Setting Algorithm Classes Accuracy(%)

[5] Electrodermal activity, eye 23 Real eXtreme Gradient Boosting 3 88%
tracking, accelerometer (XGBoost)
data, and skin temperatures
[6] Big Dataset 100 Real Bayesian Network (BN) N/A 25.81%, 6.5% the
F-measure
[7] Heart Rate, skin 15 Lab N/A 2 N/A
temperature, EDA, and
accelerometer
[8] Wearable armband 12 Real & N/A N/A N/A
Lab
[9] Electroencephalogram 80 Lab SVM, KNN, AdaBoost, Linear 3 83%
(EEG) Discriminant Analysis, Ridge
Classifier, Random Forest
[10] Heart Monitor Sensor, 12 Lab N/A N/A N/A
Motion Monitor Sensor
[22] Raspberry Pi 3, 15 Real ANN, K-NN, SVM, and RF 2 90%
Electroencephalogram
(EEG)-SMT
[23] Photoplethysmography 40 Lab RF 2 94%
(PPG), Skin Temperature
(ST), EDA
[24] Speech, Video, and Bio- 25 Real Long Short-Term Memory (LSTM) 2 66.4%
Signal data and Feed-Forward Networks (FFN)
[25] Local processor, 15 Real SVM, K NN N/A 96.7%, 95.8%
Accelerometer

B. Sensor Signal Processing

Pre-processing is done before features are extracted from the
raw accelerometer sensor data. The first step in pre-processing
is to apply a high-pass filter [13], which is intended to eradicate
the Direct Current (DC) inherent in the signals with a 0.1 Hz
cut-off frequency. Subsequent to the removal of the direct
correct, the signals undergo normalization to effectively
address subject-specific variations. We have used six
Fig. 2. Placement of the system in the nurses’ head and the data collection
procedure normalization techniques based on the value. These techniques
include no normalization, RMS normalization (centered around
Hyper-tension (High Blood Pressure), Heavy Smoking, Sprains 1), median and mean-based normalization (adapting the range),
and Strains, Fractures, Pregnancy, etc. The device was placed on the max absolute value normalization, and normalization by mean
and Strains, Fractures, Pregnancy, etc. The device was placed on the of peaks. The final value ranges depend on the chosen
heads of both male and female nurses to collect data, which normalization method and the characteristics of the input data.
resulted in reliable and precise data (see Fig. 2). The study’s This normalization makes sure that all signals are scaled to the
main objective was to monitor behaviors and actions connected same range, enabling fair comparisons between various
to stress, however participants were not made aware of this to subjects [14]. The signals are then split into non-overlapping,
reduce potential biases or altered reactions. Each participant set time intervals known as “epochs”. Each epoch in this study
gave their willing consent for the data to be recorded. We also lasts for 5 seconds, and subsequent segments don’t cross over.
obtained subjective data on the participants level of mental Following this pre-processing step, the accelerometer data are
stress through a series of questions and responses. The shown in Fig. 3. It presents participant raw data, with the
questions and responses underwent a comprehensive horizontal axis representing time in seconds (from 0 to 6), and
evaluation by multiple hospital doctors and a distinguished the vertical axis showing acceleration values in meters per
psychologist. Their valuable insights and feedback were second squared (݉‫ି ݏ‬ଶ ). The data spans a duration of 90 minutes,
carefully considered, leading us to make necessary refinements which corresponds to (5.4 ൈ ͳͲହ ) data points, with a sampling
to our questions. This iterative process of consultation and frequency of 100 Hz. The acceleration values range between ± 2.5
modification ensured the thorough validation of our ݉‫ି ݏ‬ଶ Ǥ
questionnaire. On a 7- point scale, questions such “Are you
C. Feature Extraction
stressed?” and “Are there any patients in crisis??” were posed.
These subjective mental levels of stress were noted both during
Non-overlapping segmentation of the data was required for the
and after data collection in a questionnaire during their work feature extraction procedure. The following 13 distinct features
shifts and after collecting data. A score of 1-2 indicated low were chosen for this study: Mean Absolute Value (MAV),
stress, 3-5 indicated medium stress, and 6-7 indicated high Ratio of Max to RMS (RMR), Root Mean Square (RMS), Max
stress.
 except one, using the excluded subject's data for testing. The
final accuracy is determined by averaging all testing accuracies.
This approach enables extrapolation to older, previously
unexplored samples, as it ensures model independence from
specific subjects and enhances overall performance. The
present study utilized data from 12 participants for the purposes
of training, testing, and evaluating all categorization models. A
measure that combines precision and recall, the F1 score offers
an accurate assessment of the model’s performance.

்௉ା்ே
Accuracy = (1)
்௉ା்ேାி௉ାிே
Fig. 3. Participants’ accelerometer sensor system’s 3-axes of raw data signals ்௉
were collected
Sensitivity = (2)
௉
்ே
of Absolute (MMAV), Ratio of Mean to RMS (RMER), Specificity = (3)
ே
Spectrum energy (SE), Slope sign changes, Number of peaks, ்௉
Precision =  (4)
Waveform Length (WL), Average peaks range, Entropy of ்௉ାி௉
்௉
Spectrum (ES), Number of zero crossings, and Entropy [15]. Recall = (5)
்௉ାிே
Consequently, a total of 39 features were extracted from the ଶൈ௉௥௘௖௜௦௜௢௡ൈோ௘௖௔௟௟
accelerometer data of each participant, encompassing 13 F1 Score = (6)
௉௥௘௖௜௦௜௢௡ାோ௘௖௔௟௟
features from each axis. These features facilitated a
comprehensive analysis of stress patterns among the The confusion matrix is a fundamental metric for assessing
participants, leading to valuable insights and potential stress- model accuracy, comprising key values such as P (positive), N
reduction strategies. The inclusion of abundant accelerometer (negative), TP (true positives), TN (true negatives), FP (false
data from multiple axes enabled a detailed assessment of stress- positives), and FN (false negatives) [18]. In equations 1, 2, 3,
related behaviors and movements. 4, 5, and 6 P represents instances belonging to the positive
class, which the model aims to correctly detect. TP corresponds
D. Classification Algorithm to the count of positive class instances correctly identified by
Seven machine learning algorithms were used in this study, the model, while FP represents the cases where positive results
including K-NN, NN, SVM, NB, DA, DT, and Ensemble were falsely predicted. As opposed to P & TP, which allude to the
Classifiers. The effectiveness of each approach was evaluated. positive class, N & TN speak of the negative class. A fraction of the
Three classes—low stress, medium stress, and high stress—are model's performance may be used to construct Equation (1), which
represented by the labels 1, 2, and 3, respectively, in the dataset represents the model's accuracy [19]. However, accuracy isn't
utilized for the analysis. Twenty-nine classification models always enough to assess a model's effectiveness in its entirety.
based on the seven machine learning techniques were assessed Furthermore, the metrics of sensitivity, specificity, precision, and
to distinguish between these epochs. K-NN is a classification recall are commonly utilized measures. Equation (2) derives its
and regression algorithm that assigns labels to data points based score based on the proportion of accurate positive predictions,
on the most common class among their k-nearest neighbors. which is founded upon the total number of positive predictions.
Neural networks, inspired by the brain, consist of connected On the other hand, equation (3) defines the actual negative
nodes for tasks like pattern recognition. Naive Bayes is a prediction rate by dividing the true negative predictions by the
probabilistic algorithm assuming feature independence given total number of negative predictions. A high rate of prediction
class labels; it's used for text classification. The DA classifier is implied by precision. It is a percentage of all true positives in
separates classes by optimizing class differences and the model that are associated with all demanded positives. The
minimizing variance. Decision trees (DT), shown as branches rate of precision is seen in equation (4). The recall, also known
and nodes, are widely used for classification and regression. as the TP value, compares the precise total of positives in the
[26]. SVM is a supervised learning technique, produced the information to the total positives in the system states. The recall
best results for the classification problem out of all of these rate is represented by equation (5). The F1 evaluation, that is
methods. Due to the fact that SVM uses the ideal hyperplane to produced by the weighted average of the model's recall and
operate, it is very effective in high-dimensional spaces. SVM precision and is defined in equation (6) [20], is another useful
demonstrated strong generalization and speed when applied to metric for evaluating a model's performance.
the 2-dimensional data with dataset classes, making it a useful
method for this classification employment [16]. The IV. RESULT AND DISCUSSION
classification model in this work was trained using the
MATLAB 2023a Classification Learner tool from MathWorks To put the suggested framework into execution, we used a
Inc. The 10-fold and Leave- One-Subject-Out (LOSO) cross- dataset from 12 subjects in this study with a sampling rate of
validation approach was used to assess the performance of each 100Hz. Utilizing the LOSO approach, the dataset was split
classification model [17]. The LOSO cross-validation method between training and testing datasets. Table II displays the
involves iteratively training models using data from all subjects average results for accuracy, sensitivity, specificity, precision,
 TABLE II
THE PERFORMANCE OF THE THREE-CLASS CLASSIFICATION TASK

ML Approaches Classification Accuracy Sensitivity Specificity Precision Recall F-score

(%) (%) (%) (%) (%) (%)
Nearest Neighbor Classifiers (k-NN) Fine K-NN 76.0% 77.3% 75.3% 73.4% 71.5% 75.7%
Medium K-NN 78.5% 79.2% 77.1% 74.3% 72.3% 78.1%
Coarse K-NN 77.9% 78.3% 75.0% 73.7% 71.0% 77.3%
Cosine K-NN 78.4% 79.5% 76.3% 74.7% 72.4% 77.8%
Cubic K-NN 78.7% 79.2% 75.5% 75.3% 71.2% 78.1%
Weighted K-NN 77.7% 78.8% 76.1% 75.4% 72.0% 77.0%
Neural Network (NN) Narrow Neural Network (NNN) 75.2% 76.6% 73.9% 71.5% 71.4% 73.1%
Medium Neural Network (MNN) 74.3% 75.1% 72.1% 72.2% 70.9% 73.7%
Wide Neural Network (WNN) 71.1% 72.7% 70.6% 69.8% 68.4% 70.3%
Bilayered Neural Network (BNN) 76.8% 78.1% 74.3% 73.2% 71.2% 75.5%
Trilayered Neural Network (TNN) 75.7% 77.3% 73.5% 74.7% 69.7% 75.1%
Support Vector Machines (SVM) Linear-SVM 79.8% 80.1% 77.2% 78.2% 76.1% 78.9%
Quadratic-SVM 54.6% 57.3% 53.9% 55.3% 51.8% 54.3%
Cubic-SVM 55.0% 56.5% 54.3% 54.1% 51.2% 54.2%
Fine Gaussian-SVM 78.9% 79.3% 77.9% 76.2% 73.8% 78.5%
Medium Gaussian-SVM 80.4% 81.1% 79.9% 79.4% 78.4% 80.2%
Coarse Gaussian-SVM 26.5% 29.1% 25.0% 25.2% 23.1% 26.1%
Naïve Bayes (NB) Gaussian Naïve Bayes 72.1% 73.2% 71.9% 71.4% 69.8% 71.3%
Kernel Naïve Bayes 74.5% 76.1% 73.2% 72.3% 70.3% 74.0%
Discriminant Analysis (DA) Linear Discriminant (LD) 57.4% 60.1% 56.9% 56.0% 54.7% 56.5%
Quadratic Discriminant (QD) 59.6% 60.3% 58.7% 56.4% 53.6% 58.9%
Decision Tress (DT) Fine Tree (FT) 77.4% 78.2% 76.4% 73.0% 71.3% 77.1%
Medium Tree (MT) 77.8% 77.3% 76.1% 74.6% 72.5% 77.6%
Coarse Tree (CT) 79.3% 79.9% 78.3% 78.5% 76.8% 78.9%
Ensemble Classifiers Boosted Trees 77.8% 79.1% 78.9% 75.2% 73.7% 77.2%
Bagged Trees 76.1% 77.3% 75.1% 74.4% 73.2% 75.4%
Subspace Discriminant 75.3% 76.1% 76.6% 72.2% 71.5% 74.9%
Subspace k-NN 74.2% 75.9% 73.3% 71.8% 70.2% 73.7%
RUSBoosted Trees 77.3% 78.3% 71.8% 75.1% 68.7% 79.8%

recall, and F-score based on the data separately gathered from data training process. The CG-SVM’s comparable training
each of the 12 participants. Out of all the classifiers used, the accuracy was measured at 94.1%.
Coarse Gaussian-SVM classifier had the lowest accuracy, as During testing, the MG-SVM model was evaluated using
shown in Table II. Using the table to compare each classifier’s 761 (1 subject) epochs, and the result were presented in Table
performance, it is obvious that the Medium Gaussian-SVM V as a confusion matrix for all subjects. For the low stress class,
outperformed all other machine learning models. the model accurately classified 280 positive class epochs as true
The classification results for all subjects were presented in positives (TP), and (10 + 6 = 16) false negative class epochs
Table III. The MG-SVM model achieved promising performance (FN) and (3 + 8 = 11) false positive class epochs (FP) were
for all classes. The training phase of MG-SVM model was observed. Additionally, (245 + 4 + 205 = 454) the model
evaluated using 8314 (11 subjects) epochs. For the low stress correctly identified true negative class epochs as negative class
class, the model accurately classified 3081 positive class instances (TN). For the medium stress class, the model
epochs as true positives (TP), and (34 + 24 = 58) false negative achieved 245TP7FN, 10FP, and 499TN. Regarding the high
class epochs (FN) and (12 + 29 = 41) false positive class epochs stress class, the model obtained 205TP, 8FN, 10FP, and 538TN.
(FP) were observed. Additionally, (2918 + 9 + 2207 = 5134) the The MG-SVM achieved a testing accuracy of 80.4%, indicating
model correctly identified true negative class epochs as relatively good performance during testing. The confusion
negative class instances (TN). Similarly, for the medium stress matrix in Table V provides a comprehensive overview of the
class, the model achieved 2918TP, 21FN, 34FP, and 5341TN. model’s performance for all subjects during testing,
For the high stress class, the model attained 2207TP, 29FN, demonstrating a high number of correctly classified epochs
33FP, and 5341TN. The MG-SVM model demonstrated high with only a relatively small number of misclassifications.
training accuracy of 98.7%, indicating effective performance Conversely, Table VI presents the confusion matrix of CG-
during training [21]. However, in contrast, Table IV displays SVM, showcasing the lowest accuracy attained during the data
the confusion matrix of CG-SVM, which attained the lowest testing phase. The corresponding testing accuracy obtained by
accuracy compared to the other evaluated methods during the CG-SVM was measured at 26.5%.

TABLE III TABLE IV

CONFUSION-MATRIX FOR MG-SVM MODEL (TRAINING) CONFUSION-MATRIX FOR CG-SVM MODEL (TRAINING)

Predicted Class Predicted Class

Total = 8314 Low Medium High Total = 8314 Low Medium High
Ture Low 3081 34 24 Ture Low 2797 266 76
class Medium 12 2918 9 class Medium 39 2889 11
High 29 0 2207 High 111 1 2124
 TABLE V [6] Inoue, Sozo, Naonori Ueda, Yasunobu Nohara, and Naoki Nakashima.
CONFUSION-MATRIX FOR MG-SVM MODEL (TESTING) "Recognizing and understanding nursing activities for a whole day with a big
dataset." Journal of Information Processing 24, no. 6 (2016): 853-866..
Predicted Class [7] Hosseini, Seyedmajid, Raju Gottumukkala, Satya Katragadda, Ravi Teja
Total = 761 Low Medium High Bhupatiraju, Ziad Ashkar, Christoph W. Borst, and Kenneth Cochran. "A
Ture Low 280 10 6 multimodal sensor dataset for continuous stress detection of nurses in a
hospital." Scientific Data 9, no. 1 (2022): 255.
class Medium 3 245 4
[8] Shishavan, Hossein Hamidi, Jennifer Garza, Robert Henning, Martin
High 8 0 205 Cherniack, Liane Hirabayashi, Erika Scott, and Insoo Kim. "Continuous
physiological signal measurement over 24-hour periods to assess the impact of
TABLE VI work-related stress and workplace violence." Applied ergonomics 108 (2023):
CONFUSION-MATRIX FOR CG-SVM MODEL (TESTING) 103937.
[9] Akella, Ashlesha, Avinash Kumar Singh, Daniel Leong, Sara Lal,
Phillip Newton, Roderick Clifton-Bligh, Craig Steven Mclachlan et al.
Predicted Class
"Classifying multi-level stress responses from brain cortical EEG in nurses and
Total = 761 Low Medium High non-health professionals using machine learning auto encoder." IEEE Journal
Ture Low 182 98 27 of Translational Engineering in Health and Medicine 9 (2021): 1-9.
class Medium 18 208 14 [10] Jovanov, Emil, Karen Frith, Faye Anderson, Mladen Milosevic, and
High 45 1 168 Michael T. Shrove. "Real-time monitoring of occupational stress of nurses." In
2011 Annual International Conference of the IEEE Engineering in Medicine
V. C ONCLUSION and Biology Society, pp. 3640-3643. IEEE, 2011.
[11] Maag, Balz, Zimu Zhou, and Lothar Thiele.” A survey on sensor
An easily worn single 3D accelerometer sensor was used to calibration in air pollution monitoring deployments.” IEEE Internet of Things
automatically calculate the stress levels of nurses. The Journal 5, no. 6 (2018): 4857-4870.
classification accuracy was assessed using data filtering, [12] Gao, Ming, Yajie Liu, Yike Chen, Yimin Li, Zhongjie Ba, Xian Xu,
Jinsong Han, and Kui Ren. "Device-independent Smartphone Eavesdropping
normalization, feature extraction, segmentation, and 10-fold Jointly using Accelerometer and Gyroscope." IEEE Transactions on
cross- validation tests using the LOSO method. Numerous Dependable and Secure Computing (2022).
popular machine learning approaches, such as K-NN, NN, [13] Li, Tingfeng, Ruohui Wang, and Xueguang Qiao. ”Demodulation of
fiber Bragg grating accelerometer using in-line Sagnac interferometers.” IEEE
SVM, and NB, were successful in detecting stress. The most Sensors Journal 22.5 (2022): 4077-4082.
precise categorization was specifically accomplished by the [14] Doulah, Abul, Tonmoy Ghosh, Delwar Hossain, Masudul H. Imtiaz,
SVM approach. The average F-score for classification issues and Edward Sazonov. "“Automatic ingestion monitor version 2”–a novel
wearable device for automatic food intake detection and passive capture of food
was 80.2%, and the average accuracy of the SVM model was images." IEEE journal of biomedical and health informatics 25, no. 2 (2020):
80.4%. In the study conducted by Qian Zhang et al., an 568-576.
accuracy of 88% was attained through data collection involving [15] Garcia-Ceja, Enrique, Venet Osmani, and Oscar Mayora. "Automatic
stress detection in working environments from smartphones’ accelerometer
23 participants. Despite our limited participant count, we data: a first step." IEEE journal of biomedical and health informatics 20, no. 4
achieved a notably higher accuracy of 80.4%. By expanding (2015): 1053-1060.
our participant pool under similar conditions, the potential for [16] Chen, Yange, Qinyu Mao, Baocang Wang, Pu Duan, Benyu Zhang,
and Zhiyong Hong. "Privacy-preserving multi-class support vector machine
achieving elevated accuracy levels becomes promising. The model on medical diagnosis." IEEE Journal of Biomedical and Health
study’s conclusions showed that it is possible to identify three Informatics 26, no. 7 (2022): 3342-3353.
different levels of stress by using information from the [17] Thakur, Dipanwita, Antonella Guzzo, and Giancarlo Fortino.
”Attention-based Multihead Deep Learning Framework for online activity
accelerometer sensors on nurses. Applications utilizing the monitoring with Smartwatch Sensors.” IEEE Internet of Things Journal (2023).
accelerometer sensor for automated stress detection stand to [18] Sakib, Mohammad, Syeda Shanaz Pervez, and A. B. M. S. U. Doulah.
gain significant advantages from future research in several ”Towards smart helmet for motorcyclists: automatic stress level detection using
wearable accelerometer sensor system.” 2023 International Conference on
areas: (i) Integrating physiological data, including heart rate, Communication, Circuits, and Systems (IC3S). IEEE, 2023.
electroencephalogram (EEG) signals, Photoplethysmography [19] Krstinic´, Damir, Ljiljana Sˇeric´, and Ivan Slapnicˇar. ”Comments
(PPG) sensor data, along with the inclusion of new datasets and on” MLCM: Multi-Label Confusion Matrix”.” IEEE Access (2023).
[20] Manimurugan, S., Saad Al-Mutairi, Majed Mohammed Aborokbah,
incorporating video or audio recordings. (ii) Exploring and Naveen Chilamkurti, Subramaniam Ganesan, and Rizwan Patan. "Effective
implementing more efficient feature extraction methods. attack detection in internet of medical things smart environment using a deep
belief neural network." IEEE Access 8 (2020): 77396-77404.
V. REFERENCES [21] Kulkarni, Ajay, Deri Chong, and Feras A. Batarseh. "Foundations of
[1] Islam, Md Irteja, Kh Md Watin Alam, Syed Afroz Keramat, Munzur E. data imbalance and solutions for a data democracy." In Data democracy, pp.
Murshid, Rezwanul Haque, Enamul Kabir, Rasheda Khanam, and Manjurul 83-106. Academic Press, 2020.
Haque Khan. "Working conditions and occupational stress among nurses in [22] Arpaia, Pasquale, Nicola Moccaldi, Roberto Prevete, Isabella Sannino,
Bangladesh: a cross-sectional pilot study." Journal of Public Health (2021): 1- and Annarita Tedesco. "A wearable EEG instrument for real-time frontal
9. asymmetry monitoring in worker stress analysis." IEEE Transactions on
[2] Zhu, Lili, Petros Spachos, Pai Chet Ng, Yuanhao Yu, Yang Wang, Instrumentation and Measurement 69, no. 10 (2020): 8335-8343.
Konstantinos Plataniotis, and Dimitrios Hatzinakos. "Stress Detection Through [23] Nath, Rajdeep Kumar, and Himanshu Thapliyal. "Smart wristband-
Wrist-Based Electrodermal Activity Monitoring and Machine Learning." IEEE based stress detection framework for older adults with cortisol as stress
Journal of Biomedical and Health Informatics (2023). biomarker." IEEE Transactions on Consumer Electronics 67, no. 1 (2021):
[3] Li, Yuhan, Ke Li, Jiaao Chen, Shaofan Wang, Haochang Lu, and [24] Han, Hyewon, Kyunggeun Byun, and Hong-Goo Kang.” A deep
Dongsheng Wen. "Pilot Stress Detection Through Physiological Signals Using learning-based stress detection algorithm with speech signal.” In proceedings
A Transformer-Based Deep Learning Model." IEEE Sensors Journal (2023). of the 2018 workshop on audio-visual scene understanding for immersive
[4] Rahman, Md Minhazur, Asm Mohaimenul Islam, Jonayet Miah, Sabbir multimedia, pp. 11-15. 2018..
Ahmad, and Muntasir Mamun. "sleepWell: Stress Level Prediction Through [25] Attaran, Nasrin, Abhilash Puranik, Justin Brooks, and Tinoosh Mohs-
Sleep Data. Are You Stressed?." In 2023 IEEE World AI IoT Congress enin.” Embedded low-power processor for personalized stress detection.” IEEE
(AIIoT), pp. 0229-0235. IEEE, 2023. Transactions on Circuits and Systems II: Express Briefs 65, no. 12 (2018):
[5] Zhang, Qian, Farzan Sasangohar, Pratima Saravanan, Nima Ahmadi, 2032-2036.
Tariq Nisar, Valerie Danesh, and Faisal Masud. "Real-time Stress Monitoring [26] Al-Zebari, Adel, and Abdulkadir Sengur. "Performance comparison of
for Intensive Care Unit (ICU) Nurses." In Proceedings of the Human Factors machine learning techniques on diabetes disease detection." In 2019 1st
and Ergonomics Society Annual Meeting, vol. 66, no. 1, pp. 779-782. Sage CA: international informatics and software engineering conference (UBMYK), pp.
Los Angeles, CA: SAGE Publications, 2022. 1-4. IEEE, 2019.

View publication stats