Review the Resources and select one current national healthcare issue/stressor to focus on. Reflect on the current national healthcare issue/stressor you selected and think ab

2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)

978-1-6654-0126-5/21/$31.00 ©2021 IEEE 2992

PERSONALIZED STRESS MONITORING AI SYSTEM FOR HEALTHCARE WORKERS

Raina Ghanshyam Bangani†,Vineetha Menon†,Emil Jovanov‡

† Department of Computer Science ‡ Department of Electrical and Computer Engineering

University of Alabama in Huntsville, USA

ABSTRACT

In the current COVID-19 pandemic scenario, healthcare workers, in particular nurses, face prolonged exposure to stress. This intense duress takes a toll on their health over- time, affects their quality of life, and in turn impacts the quality of care provided to the patients. Hence, real-time detection and monitoring of stress is extremely important for early detection of stress patterns, prevention of burnouts and chronic conditions in healthcare workers as well as facilitate improved patient-care outcomes. In this paper, we present a proof-of-concept case study using machine learning (ML) and artificial intelligence (AI)-based stress detection model that determines a personalized assessment of stress level us- ing heart rate, heart rate variability, and physical activity of the users. We used wearable electrocardiogram and iner- tial sensor to record heart activity and physical activity of nurses during their shifts. Our preliminary results indicate that the proposed stress tracking model can effectively pre- dict any stress occurrences. This study is a pivotal attempt to emphasize the significance of stress-detection and relief for healthcare workers and provide them a tool for an effective assessment of personalized stress levels.

Index Terms— Personalized stress monitoring, machine learning, CNN, AI, K-Means clustering, classification

1. INTRODUCTION

Stress represents our body’s response to physical or psycho- physiological conditions that threatens (physical or per- ceived) homeostasis. A ‘stressor’ is a stimulus that disrupts homeostasis. In the US, one-third of employees report their job as stressful. As a reaction to stressful events our body releases hormones such as cortisol and adrenaline to make the person more alert. After the event has transpired, other hormones are then released to relax a person’s body. This re- sponse is called as ‘fight, flight or freeze’ response. Long term

We thank the Gulf Research Program of the National Academies of Sci- ences, Engineering, and Medicine for supporting this work. DISCLAIMER : ”The content is solely the responsibility of the authors and does not necessar- ily represent the official views of the Gulf Research Program or the National Academies of Sciences, Engineering, and Medicine.”

exposure to stress leads to progressive increase in heart rate, elevated levels of stress hormones and blood pressure[1, 2].

Although the causes of stress may vary from one indi- vidual to another, some common causes include trauma, de- manding work schedules, juggling multiple responsibilities, etc. [1, 2]. Long-term exposure to stressful conditions can have adverse effects on health. It makes one more prone to conditions such as fatigue, high blood pressure, diabetes, stress-related heart conditions, obesity, mental disorders such as anxiety and depression [2]. Work environments are often one of the crucial contributing factors to observed stress lev- els in a person [3]. Especially healthcare workers like the nurses work in an intense stressful dynamic environment with utmost priority for patient-care which allows no margin of error. Such prolonged stressful work conditions can eventu- ally take a toll on the health and well-being of nursing and healthcare community at-large. Many studies have shown a definite correlation between the personal health of nurses and the quality of patient-care they provide [4–7]. Therefore, it is consequential to assist our frontline workers like the nurses and equip them with a stress monitoring tool that can detect early indications of stress. The proposed personalized stress monitoring system is an innovative life-saver biomedical tool can enable timely intervention and feedback in order to avert any long-term adverse health effects due to stress.

In this novel work, we introduce an automated personal- ized stress detection and monitoring system using ML and AI techniques to determine and track personalized stress levels based on biophysical indicators such as heart rate and heart rate variability (HRV) in real-time for continuous monitoring and stress assessment [8]. It is important to note that occupa- tional environments, exposure to stress, and various physical and psychological factors determine the perceived personal- ized stress levels which are innately different for every indi- vidual. For the proposed stress monitoring system, we first extract a sequence of RR intervals as the time between two consecutive R peaks in the QRS signal of the Electrocardio- gram (ECG) signal to find the immediate heart rate and cal- culate measures of HRV such as RMSSD, NN50 and pNN50. Detailed discussion on ECG signal preprocessing and pro- posed personalized stress monitoring system is presented in Section 2. Preliminary analysis and results are discussed in

20 21

IE EE

In te

rn at

io na

l C on

fe re

nc e

on B

io in

fo rm

at ic

s a nd

B io

m ed

ic in

e (B

IB M

) | 9

78 -1

-6 65

4- 01

26 -5

/2 1/

$3 1.

00 ©

20 21

IE EE

| D

O I:

10 .1

10 9/

B IB

M 52

61 5.

20 21

.9 66

93 21

Authorized licensed use limited to: Walden University. Downloaded on February 27,2024 at 04:02:05 UTC from IEEE Xplore. Restrictions apply.

2993

Section 3, and Section 4 summarizes our novel contributions in this work.

2. PERSONALIZED STRESS MONITORING AND DETECTION SYSTEM

Data Description: The case study presented in this paper uses an existing dataset from an experiment designed to moni- tor stress of nurses during a single work shift in 2011 [9]. This datasset consists of ECG and physical activity data of six sub- jects. The dataset is recorded using a wearable patch sensor BioStamp from mc10 [10]. The patch records person’s elec- trocardiogram (ECG) signal data and 3 axis of acceleration of the inertial sensor. Although a machine learning algorithms would benefit from a much larger number of subjects, the existing dataset presents an unique opportunity to evaluate a proof-of-concept for a personalized stress assessment system. Moreover, our preliminary analysis indicate that the number of epochs during a whole day recording provides a sufficient dataset for training and analysis using machine learning algo- rithms.

2.1. R-peak Detection

The first step is offline preprocessing of the ECG signal ac- quired from biosensors to determine the R peaks and RR- intervals present in the data. For this, we perform R-peak detection in the wavelet domain to exploit its denoising and simultaneous time-frequency resolution properties. Discrete Wavelet Transform (DWT) decomposes any given signal into various components such that each level describes the change in the signal for a given frequency band. The maximal overlap discrete wavelet transform (MODWT) with ’sym4’ wavelet was used to detect the wavelets till level 5. Matlab Signal Processing Toolbox is used to extract the R peak values and the corresponding RR-intervals are computed [11].

2.2. Proposed Stress Monitoring System

Since the ECG signal and RR-interval data available are un- structured, K-Means clustering method is used to determine the initial stress clusters. The goal is to determine whether the nurse is stressed at any given point in time based on the sub- ject’s RR-intervals and RMSSD data. The algorithm for the proposed stress monitoring techniques is detailed as follows:

• Step 1: Wavelet domain R-peak detection: R-peaks are detected using MODWT technique with ‘sym4’ wavelet[11]. Corresponding RR-intervals are com- puted.

• Step 2: Outlier detection: Outliers in the data were identified using matlab function ‘rmoutliers()’ and re- moved. This process removes datapoints that have value more than three median absolute deviation. This process was done in batches of 5000 RR-intervals.

• Step 3: RMSSD calculations for HRV: Root Mean Square of Squared Differences (RMSSD) is computed from the RR intervals as below:

RMSSD =

√√√√ 1

N − 1

N−1∑ i−1

((R−R)i+1 − (R−R)i)2

(1)

• Step 4: K-Means Clustering: K means clustering is used to categorize RR-interval vs RMSSD data into the three distinct stress categories: low, normal, and high stress clusters present in the data. The outliers xoutlier

in each stress cluster category as depicted in Figure 1 is estimated as follows:

xoutlier = √

(x(i, 1)− x(centroid, 1))2 (2)

Fig. 1. K-Means-based cluster preprocessing of stress regions for all subjects

The overlapping K-Means cluster centroids for all sub- jects are denoted by the black square marker in Figure 1. The distance from the centroids is used to measure the belongingness of all points to the respective stress clusters for a personalized stress cluster identification and processing for all subjects as described below:

– Low Stress cluster (Green cluster in Figure 1) – outliers are values that are greater than 1/4th of the average distance values in that cluster.

– Normal Stress cluster (Blue) –outliers are the val- ues that are greater than half of the average dis- tance values in that cluster.

– High Stress cluster (Red) –outliers are the values that are greater than 2/7th of the average distance values in that cluster.

Authorized licensed use limited to: Walden University. Downloaded on February 27,2024 at 04:02:05 UTC from IEEE Xplore. Restrictions apply.

2994

These thresholds were chosen empirically based on ex- pert input on density and generic spread of stress cluster data for all subjects.

Fig. 2. The three stress regions identified for subject 1

• The stress cluster information from Figure 1 is used to identify the desired stress categories to obtain Figure 2 and generate corresponding stress labels as follows:

– For ‘Low Stress’ cluster (green) in Figure 2:

∗ Select the dense green cluster data and green outliers from Figure 1 that have RR-intervals < xgreen−centroid and both RR-intervals and RMSSD values > xgreen−centroid.

∗ Select the blue outliers from Figure 1 that have RR-intervals > xblue−centroid.

∗ Select the red outliers from Figure 1 that have RR-intervals < xred−centroid and ≤ 0.6 and RMSSD values > xred−centroid.

– For ‘Normal Stress’ cluster (Blue) in Figure 2:

∗ Select green outliers from Figure 1 that have RR-intervals > xgreen−centroid and RMSSD values ≤ xgreen−centroid.

∗ Select blue cluster data from Figure 1. ∗ Select red outliers from Figure 1 that have

RR-intervals and RMSSD < xred−centroid

and in the range RR-intervals ∈≤ 0.6 and RR-intervals ∈ {0.6− 0.7}.

– For ‘High Stress’ cluster (Red) in Figure 2:

∗ Select blue outliers from Figure 1 that have RR-intervals ≤ xblue−centroid.

∗ Select red cluster data from Figure 1, blue outliers that have RR-intervals > xred−centroid

and data values that have RR-intervals <

xred−centroid, RR-intervals ∈ {0.6 − 0.7} and RMSSD values > xred−centroid. Also include blue outliers that have RR-intervals < xred−centroid and > 0.7.

These thresholds were empirically chosen based on the expert input and the outlier detection from step 2. The RR-intervals of 0.6 and 0.7 are chosen because the min- imum value of RR interval is given as 0.5 by the expert.

• Step 5: Supervised Classification: Five ML and AI supervised classifiers are then trained on the obtained stress categories for stress detection and monitoring:

Decision Tree –It represents the rules learning in a tree- like structure where every internal node denotes a feature. Starting from root node, at every decision node, it chooses the branch to go to next level based on the value of an attribute. The leaf nodes represent the predicted class label. The rules (branching conditions) are learnt from the training data.

Naı̈ve Bayes Classifier –It is a probabilistic classifier based on the Bayes Theorem. It assumes that the predictor variables are independent. The predicted class label is the class with maximum probability.

Logistic Regression –Logistic Regression is a binary classifier. It uses logistic function to map the relation be- tween the dependent and independent variables. In this work, we have used multinomial logistic regression classifier which predicts the probabilities for data points belonging to multi- ple classes. The predicted label is chosen as the class with maximum probability.

Support Vector Machine (SVM) –Principle of SVM is that it finds a hyperplane that maximizes margin to divide classes in the most optimal manner. Since SVM is a binary classifier, for this work, we have used the one vs. one en- semble SVM training model. In a one vs one approach, 3 binary classifiers are trained as follows: class 1 (Low Stress) vs. class 2 (Normal Stress), class 2 (Normal Stress) vs. class 3 (High Stress) and for class 1(Low Stress) vs. class 3 (High Stress). The class that receives the maximum votes is chosen as the predicted label.

Convolution Neural Network (CNN) –CNNs are deep learning techniques which are widely used for image and text classification. It has multiple convolution layers fol- lowed by fully connected neural network. The input is given in the form of (2×1) vectors (Average RR interval/minute, RMSSD/minute). Learning rate for the network is chosen as 0.01. A stochastic gradient descent with momentum (SGDM) optimizer is used in the network with number of epochs=50. The CNN architecture employed is illustrated in Figure 3.

To summarize, the proposed personalized stress detection and monitoring system uses K-means clustering to determine the underlying stress clusters present in the data. The clus- ter centroids obtained from K-means clustering algorithm are further used to refine the desired stress categories, i.e., low,

Authorized licensed use limited to: Walden University. Downloaded on February 27,2024 at 04:02:05 UTC from IEEE Xplore. Restrictions apply.

2995

Fig. 3. CNN Architecture used for stress detection

normal and high stress. The normal stress region of the indi- vidual is extracted from the data adaptively based on the clus- ter centroids creating a personalized stress monitoring system. This categorized RMSSD vs RR-interval stress data is further used to train our ML/AI classifiers.

3. EXPERIMENTAL RESULTS In this section, we demonstrate the efficacy and proof-of- concept of our proposed case study for personalized stress detection and assessment system using five ML/AI super- vised classification techniques, namely, decision tree, naive Bayes, logistic regression, SVM, and CNN. In this novel framework, we have employed a mix of supervised and un- supervised techniques for stress detection as discussed in Section 2 to generate class labels using K-Means clustering based on HRV bio-markers. We also determine two decision boundaries for stress levels: boundary 1: between low and normal stress, and boundary 2: between high and normal stress regions. This automated detection of stress regions based on personalized bio-markers is what makes this work novel and unique.

Fig. 4. The three stress regions identified for subject 2

Fig. 5. The three stress regions identified for subject 3

The proposed model was trained using the data of sub- ject 1 and tested on the rest of the 5 subjects. The goal is to classify the given RR intervals Vs. RMSSD for a one-minute window to determine whether or a person is stressed or op- erating under normal conditions. Therefore, from the given ECG signal data, average RR intervals and RMSSD values were calculated over one minute window and the proposed al- gorithm steps described in Section 2 were followed. The HRV measures considered for this work are RMSSD and pNN50. RMSSD is the root mean square of successive differences of RR intervals as defined in equation 1. pNN50 is defined as the NN50(number of consecutive RR intervals that differ by more than 50 ms) divided by the total number of RR intervals [12]. Experimentally, RMSSD had better statistical distribu- tion properties than pNN50, hence RMSSD was chosen for all trials [12].

The stress regions and decision boundaries for subject 1 which was used to train the supervised classification model is shown in Figure 2. The performance of the proposed stress detection model for subjects 2-6 is depicted in Figures 4-8. Although the proposed model was not trained using subject

Authorized licensed use limited to: Walden University. Downloaded on February 27,2024 at 04:02:05 UTC from IEEE Xplore. Restrictions apply.

2996

Fig. 6. The three stress regions identified for subject 4

2, it is seen that the proposed model provided good identifica- tion of stress regions for all the subjects under consideration. Thus, it was able to extract personalized stress regions for monitoring and assessment for all 6 subjects. The overall pre- diction accuracy for all the test subjects across all the methods using cross validation is as shown in Figure 9. It can be ob- served that decision tree has the highest overall classification accuracy and outperformed all methods for all subjects even at low training sizes as 10%. Overall decision tree and logis- tic regression classifiers perform better than others in com- parison. Therefore, we have experimentally substantiated the proof-of-concept of our proposed stress detection and moni- toring system. This work is an inspiring attempt to provide early stress detection and intervention support to healthcare workers and community at-large.

4. CONCLUSIONS In this paper, we have proposed a case study and proof-of- concept for a personalized stress detection and assessment system based on AI/ML techniques to provide a continu- ous stress monitoring and assessment. Overall decision tree classifier-based stress model gave superior stress detection accuracy over 95% for low training sizes in comparison to others. The motivation behind this work was that the health- care workers like nurses tend to have elevated stress levels that could significantly influence their personal health and quality of patient care they provide.

5. FUTURE WORK

For future work, we would like to extend the scope of our pro- posed AI-based stress monitoring and prediction system and integrate it with physical activity for a more comprehensive personalized stress analysis of healthcare workers. We hope

Fig. 7. The three stress regions identified for subject 5

that this research will sponsor more active efforts in under- standing the stress induced burnout in healthcare workers in this COVID-19 pandemic scenario.

6. ACKNOWLEDGMENT

This research work was supported by an Early Career Re- search Fellowship from the Gulf Research Program of the National Academies of Sciences, Engineering, and Medicine for supporting this work. DISCLAIMER: “The content is solely the responsibility of the authors and does not neces- sarily represent the official views of the Gulf Research Pro- gram of the National Academies of Sciences, Engineering, and Medicine.”

7. REFERENCES

[1] A. S. Jansen, X. Van Nguyen, V. Karpitskiy, T. C. Met- tenleiter, and A. D. Loewy, “Central command neurons of the sympathetic nervous system: basis of the fight-or- flight response,” Science, vol. 270, no. 5236, pp. 644– 646, 1995.

[2] L. R. Murphy, “Stress management in work settings: A critical review of the health effects,” American journal of health promotion, vol. 11, no. 2, pp. 112–135, 1996.

[3] J. Siegrist, M. Wahrendorf, and Siegrist, Work stress and health in a globalized economy. Springer, 2016.

[4] A. Najimi, A. M. Goudarzi, and G. Sharifirad, “Causes of job stress in nurses: A cross-sectional study,” Ira- nian journal of nursing and midwifery research, vol. 17, no. 4, pp. 301–305, 2012.

Authorized licensed use limited to: Walden University. Downloaded on February 27,2024 at 04:02:05 UTC from IEEE Xplore. Restrictions apply.

2997

Fig. 8. The three stress regions identified for subject 6

Fig. 9. Overall prediction accuracy of the proposed AI-based personalized stress monitoring system for all subjects

[5] H. K. Spence Laschinger and M. P. Leiter, “The impact of nursing work environments on patient safety outcomes: the mediating role of burnout/engagement,” The Journal of nursing administration, vol. 36, no. 5, p. 259—267, May 2006. [Online]. Available: https://doi.org/10.1097/00005110-200605000-00019

[6] N. Talaee, M. Varahram, H. Jamaati, A. Salimi, M. Attarchi, M. Kazempour Dizaji, M. Sadr, S. Hassani, B. Farzanegan, F. Monjazebi et al., “Stress and burnout in health care workers during covid-19 pandemic: vali- dation of a questionnaire,” Journal of Public Health, pp. 1–6, 2020.

[7] M. Milosevic, E. Jovanov, K. H. Frith, J. Vincent, and E. Zaluzec, “Preliminary analysis of physiological changes of nursing students during training,” in 2012 Annual International Conference of the IEEE Engineer- ing in Medicine and Biology Society. IEEE, 2012, pp. 3772–3775.

[8] W. D. Scherz, R. Seepold, N. M. Madrid, P. Crippa, and J. A. Ortega, “RR interval analysis for the distinction between stress, physical activity and no activity using a portable ecg,” in 2020 42nd Annual International Con- ference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 2020, pp. 4522–4526.

[9] E. Jovanov, K. Frith, F. Anderson, M. Milosevic, and M. T. Shrove, “Real-time monitoring of occupational stress of nurses,” in 2011 Annual International Confer- ence of the IEEE Engineering in Medicine and Biology Society, 2011, pp. 3640–3643.

[10] Biostamp wearable patch sensor. [Online]. Available: https://www.mc10inc.com/

[11] MATLAB. R wave detection in the ecg. [Online]. Available: https://www.mathworks.com/help/wavelet/ ug/r-wave-detection-in-the-ecg.html

[12] M. Malik, J. T. Bigger, A. J. Camm, R. E. Kleiger, A. Malliani, A. J. Moss, and P. J. Schwartz, “Heart rate variability: Standards of measurement, physiological in- terpretation, and clinical use,” European Heart Journal, vol. 17, no. 3, pp. 354–381, 03 1996.

Authorized licensed use limited to: Walden University. Downloaded on February 27,2024 at 04:02:05 UTC from IEEE Xplore. Restrictions apply.