[ieee 2011 ieee international conference on control system, computing and engineering (iccsce) -...

ECG Signals Based Mental Stress Assessment Using Wavelet Transform P Karthikeyan, M Murugappan, S Yaacob

School of Mechatronics Engineering Universiti Malaysia Perlis (UniMAP)

Perlis, Malaysia [email protected]

Abstract—This paper describes the mental stress assessment using Electrocardiography (ECG) signal. Stress reflects the changes in heart rates under stressful situation. In this work, Heart Rate Variability (HRV) from ECG signal is used to study the activity of Autonomic Nervous System (ANS) under stress states. The Stroop colour word test is used to induce stress and ECG signal was simultaneously acquired from the 10 female subjects in the age range of (20 - 25) years in non invasive manner. An acquired ECG signals are preprocessed using 4th order elliptic band pass filter. The High Frequency (HF) and Low Frequency (LF) bands of ECG signals were considered to extract the stress related features through Discrete Wavelet Transform (DWT) using "db4" wavelet function. The extracted features are mapped into two states such as stress and relax using a K Nearest Neighbour (KNN). The experimental results show the maximum average classification accuracy of 96.41% on classifying the stress and relax states from the ECG signals.

Keywords- stress; stroop colour word test; discrete wavelet transform; K Nearest Neighbor.

I. INTRODUCTION Recent years, stress is one of the serious factors for causing

many diseases, mental illness and disorders. There are several research investigations are gradually coming up to resolve the limitations on measuring, analyzing and indentifying the human stress levels. Physiological signals plays a significant role in assessing and estimating the stress compared to the other methods such as questionnaires and bio-chemical samples [1]. A detailed review on different stress inducement stimuli and various stress measures are detailed [2]. In [3], the stress level of the car driver is measured and classified by using three physiological signals such as ECG, Electromyography (EMG) and Galvanic Skin Response (GSR). On the other hand, Skin Temperature (ST), GSR, Pupil Diameter (PD) and Blood Volume Pulse (BVP) are used for detecting the stress level of the computer users through Support Vector Machine (SVM) and achieved the maximum classification accuracy of 90.10% [4]. The Stress Response Inventory (SRI) questionnaire were used to induce the stress and three physiological signals (ECG, ST, and BP) are used to estimate the stress level of the subject [5]. In addition, three other stress inducement methods namely: public speaking task, mental arithmetic task, and cold presser test are used to induce the stress [2]. In [6], a personalized stress detection

model has been developed to measure the stress by using ECG, Respiration Rate (RR), ST and GSR signals.

In general, physiological signals are severely contaminated during the data acquisition with different types of noises and other interferences. Digital filters have been mostly used by several researchers to remove the basic sources of noises [7]. The time frequency analysis is performed by using Smoothed Pseudo Wigner-Ville Distribution (SPWVD) to extract HRV based stress related features in [8]. In addition to the above methods, discrete wavelet transform is widely used for statistical feature extraction in ECG signals based stress assessment [9, 10].

In this work, we have used the stroop colour word test as stimuli to induce the stress and ECG signals acquired simultaneously during the experiment. The data acquisition protocol is designed efficiently to increase the stress levels gradually from relaxed state to stress state. The acquired ECG signal has been preprocessed using elliptic 4th order band bass filter to remove the effects of noises and other external interferences and HRV features have been derived using DWT. Several researchers have investigated the stress on analyzing the LF band (0.05-0.14) Hz and HF band (0.014-0.5) Hz of ECG signals [8]. Since, the HF indicates the activity of the parasympathetic division and LF indicates the sympathetic activity of the ANS. Hence, the ratio of LF/HF reflects the balance of sympathetic and parasympathetic divisions of the ANS during different stress levels. Then, the acquired features were normalized and classified into two states namely stress and relaxed using simple non-linear classifier (KNN).

II. METHODOLOGY

A. Data Acquisition and Protocol Design The ECG signal based stress assessment methodology is

shown in Fig: 1. initially, the ideal laboratory setup was designed to perform the stress inducing task and which allows reduce the environmental changes during the stress inducing task. The ECG signals were collected from the subject during the entire experiments and it sampled at a frequency of 1 kHz. The ECG electrodes (Ag/AgCl) are placed basis of einthoven triangle is shown in Fig: 2. AD instruments, Australia was used to acquire the data. In this work, the proposed protocol is designed to induce relax and stress. The stroop colour word

2011 IEEE International Conference on Control System, Computing and Engineering

978-1-4577-1642-3/11/$26.00 ©2011 IEEE 258

Fig: 1 Research methodology for measurusing ECG signal.

test is used as the laboratory stressor (stimstress on all the subjects. several researcheinduce the stress for stress level aphysiological signals [2, 11]. Initially, the rel(natural scenes) has been shown with duramake the subject into the relaxed state. Afperiod, the stroop colour word test has to besubjects. The stroop colour word test consistasks namely congruent and incongruent scongruent stroop, all the colours and its labwill be same and in the incongruent sessiolabel of the colours will be different. Durinverbally pronounces the colours of the wordwritten. The total time duration of stroop taTotally 2 trials were taken and in betweesubjects were asked to take complete relaxasubjects were asked to report the effectiveneafter finished the experimentation through thform. B. Subjects

A total of 10 female subjects are voluntarthe experiment in the age range of 20-25 yproper explanation about the purpose of expdesign and time duration of this study, aconcern was obtained from all the subjects. Anon-smokers, non-alcohol users and they previous history of medication. Initially, tasked to seat comfortably on the chair anplaced.

Fig:2 Einthoven triangle electrode place

Data Acquisition

ECG Signals

PreproUsing

IIR

KNN Classifier

FeExtr(W

Tran

Relaxed

Stressed

subject

ring mental stress

muli) to induce the ers have used to assessment using laxation video clip ation of 3 min to fter the relaxation e performed by the ts of two different troop. During the bel of the colours

on the colours and ng the test, subject d but not the label asks is 12.33 min. en each trails, the tion. Finally, the ess of the protocol he self assessment

rily participated in years. After giving periment, protocol an earlier written

All the subjects are did not have any the subjects were d electrodes were

ement [12]

C. Preprocessing The acquired signals are us

artefacts and noises and it suitable digital filtering techninterpretation. Researchers haremove the effects of noises [13-16]. However, the Ellipticthat has low transient responseElliptic 4th order band pass Infilter is chosen to remove thfrequency noises (f>100 Hz) efficiently holds the ECG signrange of (0.01 Hz - 100 Hz).

D. Wavelet trasform A wavelet is the small oscil

a wave and it has the ability tfrequency analysis. This featabout the time-frequency localnon-stationary nature of ECG sfunctions created by scaling afunction namely mother wave(Ψa,b) function is expressed by ψ , t √ ψ

where, ’t’ is the time parameterthe shifting factor, respectivelwavelet space. The mother wacondition (admissibility conditi

Cψ |ψ ω |ω

∞∞ dω

where, ψ (ω) is the Fourier function (ψa, b (t)). The timperformed by repeatedly by filfilters namely low pass and higinput signal frequency at the mThe results of low pass filteapproximation coefficients (coefficients are called as Furthermore, the CA is suapproximation and detailed coprocess is carried out until theachieved for the given analysis

In this work, "db4" wavelextraction. The wavelet scalingof "db4" wavelet is shown in Fwavelet family and it has mmoments. Here, the scaling orthogonal multi-resolution anasimilar to the one cycle of ECG

Fig: 3 Scaling function (φ) [17]

ocessing Elliptic Filter

eature raction avelet nsform)

ually contaminated with several should be removed by using

niques for efficient ECG signal ave used several methods to and artefacts from ECG signal c filter is a simple digital filter in pass band and stop band [7]. nfinite Impulse Response (IIR) he low (f<0.01 Hz) and high in the ECG signals. This filter

nal information in the frequency

llatory waveform that looks like to allow simultaneous time and ture will give the information lization of the input signal. The signals allow us to expand basis and shifting of single prototype let (Ψa,b). The mother wavelet .

a, b ∈ R, a>0, (1)

rs, 'a' and ‘b’ are the scaling and y, a, b ∈ R, a>0, and R is the velet must satisfy the following ion) in eq. (2).

ω ∞ (2)

transform of mother wavelet me-frequency representation is ltering the signal with a pair of gh pass filters and it cut-off the middle of its frequency domain. ered coefficients are called as (CA) and high pass filtered

detailed coefficients (CD). ubsequently divided into new oefficients. The decomposition

e required frequency response is . let function is used for feature g functions and wavelet function Fig 3 &4. This is an orthogonal

maximum number of vanishing function which generates an

alysis. These function moreover G morphology.

Fig: 4 Wavelet function (ψ) [17]


259

TABLE: 1 COMPUTATION OF STATISTICAL FEATURES USING DWT

F. Feature extraction

The preprocessed ECG signal is highly nonlinear and requires non linear feature extraction methods to extract the stress relevant features from the ECG signal. The Short Time Fourier Transform (SFT) and Fast Fourier Transform (FFT) had been used by several researchers for ECG signal processing. However, there are several limitations are still existing in FFT due to its lack of response of time information's while analyzing frequency domain and vice versa. In order to resolve this limitation, the time frequency domain (TFD) analysis required, which is either carried out by using the discrete wavelet transform, empirical mode decomposition (EMD), Smoothed Pseudo Wigner-Ville Distribution (SPWVD) and etc, to derive the time and frequency domain features. In this application, we have used DWT with "db4" wavelet function for extracting the statistical features. Because of “db4” wavelet characteristics functions looking similar to the ECG signal and it has been used by several researchers on ECG signal analysis and denoising [9, 18]. In order to extract stress related features, the LF (0.05-0.14) Hz and HF (0.14-0.5) Hz bands of ECG signals were chosen and the signal is decomposed in to 14 levels to get better resolution on these frequency bands.

G. Features In this work, there are six features were computed from LF

band, HF band, and ratio of LF and HF for classifying the

ECG signal into relaxed state or stressful state. The extracted features are mean, standard deviation, power, energy, covariance and entropy of wavelet coefficients. The mathematical expression of each feature is given in Table I. The feature vectors extracted from DWT have been normalized by using the simplest normalization method (Eq.3). In this method, on each subject feature vectors were divided to the maximum value of each subject features.

Norm= (xi/xmax ) (3)

where, xi is the input vectors and xmax is the maximum of value on the data set of the single subject Some Common Mistakes

H. Classfication The K-nearest neighbour algorithm (KNN) performs

classification based on closest training points on the feature space. It is the simplest machine learning algorithm compared to other algorithms like support vector machine (SVM), decision tree classifier etc [4]. The operation and classification is purely based on the majority of voting to the nearest neighbours. In this work, Euclidian distance is used as distance measure for KNN classifier with different K values from 1-10.

III. RESULTS AND DISCUSSION In this work, a sum of 6 six statistical features over two

different frequency bands (HF, LF, HF+LF, HF/LF, HF/(HF+LF), LF/(HF+LF)) on 10 subjects with two trails are used to form the feature vector for classifying the stress using KNN. The 70% data has been used for testing and 30% for training of each class. The K-value of the KNN classifier is varied from 1 to 10. For each K value, the algorithm computes the average classification rate over 10 trials that randomly taken features on each class to get the good accuracy. The results of various frequency bands are shown in the Table II to Table VI. The Table II shows the 2 class classification of HF features and we obtained the maximum accuracy in co-variance features over 79.38 %. Mean and standard deviation feature gives the maximum classification rate of above 70% on classifying relax and stress state, respectively over other features in KNN. Similarly, Table III shows the 2 state results of LF features in KNN. The covariance features shows the maximum classification accuracy of 91.96 % compared to all

TABLE II. CLASSIFICATION ACCURACY OF HF BAND

Features K Value

Class Average Accuracy Relax High

Mean 9 71.56 56 63.8

Std.dev 7 66.46 72.7 69.58

Energy 9 69.48 65.4 67.45

Power 9 66.15 67 66.56

Co-variance 8 79.38 79.2 79.27

Entropy 8 68.44 68.2 68.33

Features Expression Description

Power 1

ix is the wavelet

coefficient and n is the total no of wavelet coefficients

Energy

Mean 1

Standard Deviation

∑

Co- variance

Cov x E x1 µ … xn µ

where, E is the mathematical expectation and µi=E

ix

Entropy

where, x = (x1, x2. . . xn) is a set of random phenomena of wavelet co-efficient, and p is a probability of random phenomenon of wavelet co-efficient.


260

TABLE III. CLASSIFICATION ACCURACY OF LF BAND

Features K Value


Mean 10 66.3 64.7 65.47

Std.dev 9 68.8 58.9 63.8

Energy 10 69.5 66.9 68.18

Power 6 66.3 63.2 64.74

Co-variance

1 84.7 92 88.33

Entropy 9 64.8 59.8 62.29

TABLE IV. CLASSIFICATION ACCURACY OF LF/HF BAND

Features K Value


Mean 4 72.5 72.1 72.29

Std.dev 10 62.1 60 61.04

Energy 5 65.4 52.9 59.17

Power 5 59.8 53.2 56.51

Co-variance

6 71.3 75.7 73.49

Entropy 9 62.8 56.4 59.58

TABLE V. CLASSIFICATION ACCURACY OF (LF/HF+LF) RATIO

Features K Value


Mean 9 73 76.2 74.58

Std.dev 10 69.8 68 68.91

Energy 10 74.7 73.1 73.91

Power 10 77.5 71.4 74.43

Co-variance

2 72.1 78 75.05

Entropy 9 77.6 76.5 77.03

TABLE VI. CLASSIFICATION ACCURACY OF (HF/HF+LF) RATIO

Features K Value


Mean 5 92.8 100 96.41

Std.dev 6 57.8 56.8 57.29

Energy 10 62.9 65.2 64.06

Power 6 63.4 69.6 66.51

Co-Variance

7 75.6 77.4 76.51

Entropy 5 77.3 75.4 76.35

Fig: 5. Maximum accuracy of all frequency bands and ratio

analyzed the effect of frequency band ratio for assessing the stress (Table IV to Table VI). Among the different frequency band ratio, HF/(LF+HF) gives a maximum average classification accuracy of 96.41 % on mean feature (Table 6). The overall maximum average classification rate of stress and relaxed is shown in Fig: 5.

IV.CONCLUSION The stroop colour word test based stress inducement is

dependably induced the stress and that was confirmed by using questionnaire collected from the subject after the experiment. To recognize the subject response from the ECG signals, the DWT based stress recognition has done by using the KNN classifier over the 10 female subjects. In this work, we have analyzed six different statistical features derived from two different frequency bands for assessing the stress states from ECG signals. Among the different frequency band features, the features derived from the ratio of HF/LF+HF gives the maximum average accuracy of 96.41 % on mean feature on classifying the stress and relaxed states using KNN. However, the covariance features gives the maximum classification rate on LF, HF and LF/HF over other feature combinations. The accuracy (96.41 %) that we achieved by using KNN is higher than previous research (90.10 %) using SVM on similar stroop colour word test [4]. In future the classification will be done in between the class of stress namely low, medium and high by using further advanced classifier in large group of subjects.

ACKNOWLEDGMENT This project work is supported by Fundamental Research

Grant Scheme (FRGS), Malaysia. Grant Code: 9003-00245.

REFERENCES

[1] T. H. Holmes and R. H. Rahe, "The social readjustment rating scale," Journal of Psychosomatic Research, vol. 11, pp. 213-218, 1967.

[2] P. Karthikeyan, et al., "A review on stress inducement stimuli for assessing human stress using physiological signals," in Signal


261

Processing and its Applications (CSPA), 2011 IEEE 7th International Colloquium on, pp. 420-425.

[3] J. A. Healey and R. W. Picard, "Detecting stress during real-world driving tasks using physiological sensors," Intelligent Transportation Systems, IEEE Transactions on, vol. 6, pp. 156-166, 2005.

[4] J. Zhai and A. Barreto, "Stress Detection in Computer Users Based on Digital Signal Processing of Noninvasive Physiological Variables," in Engineering in Medicine and Biology Society, 2006. EMBS '06. 28th Annual International Conference of the IEEE, 2006, pp. 1355-1358.

[5] L. Salahuddin, et al., "Dependence of Heart Rate Variability on Stress Factors of Stress Response Inventory," in e-Health Networking, Application and Services, 2007 9th International Conference on, 2007, pp. 236-239.

[6] Y. Shi, et al., "Personalized Stress Detection from Physiological Measurements," in International Symposium on Quality of Life Technology, 2010

[7] S. C. Mahesh, et al., "Suppression of noise in the ECG signal using digital IIR filter," presented at the Proceedings of the 8th WSEAS International Conference on Multimedia systems and signal processing, Hangzhou, China, 2008.

[8] H. M. Seong, et al., "The Analysis of Mental Stress using Time-Frequency Distribution of Heart Rate Variability Signal," in Proceedings of the 26th Annual International Conference of the IEEE EMBS, San Francisco, CA, USA September 1-5, 2004, pp. 283-284

[9] G. Ranganathan, et al., "ECG Signal Processing Using Dyadic Wavelet for Mental Stress Assessment," in Bioinformatics and Biomedical Engineering (iCBBE), 2010 4th International Conference on, 2010, pp. 1-4.

[10] P. Rani, et al., "Online stress detection using psychophysiological

signals for implicit human-robot cooperation," vol. 20, ed: Cambridge JournalsOnline,2002,pp.673-685.

[11] B. Pehlivanoğlu, et al., "Computer Adapted Stroop Colour-Word Conflict Test as a Laboratory Stress Model," Erciyes Medical Journal, vol. 27, pp. 58-63, 2005.

[12] Knowledge Weavers Project -ECG. Available: http://library.med.utah.edu/kw/ecg/ecg_outline/Lesson1/lead_dia.html

[13] A. K. Ziarani and A. Konrad, "A nonlinear adaptive method of elimination of power line interference in ECG signals," IEEE Transactions on Biomedical Engineering, vol. 49, pp. 540-547, 2002.

[14] Z. Zhao and J. Liu, "Baseline Wander Removal of ECG Signals Using Empirical Mode Decomposition and Adaptive Filter," in Bioinformatics and Biomedical Engineering (iCBBE), 2010 4th International Conference on, pp. 1-3.

[15] T. Ya, et al., "ECG Signal Preprocessing Based on Change Step Iteration of the LMS Adaptive Filtering Algorithm," IEEE 2008.

[16] G. Tang and A. Qin, "ECG De-noising Based On Empirical Mode Decomposition," in The 9th International Conference for Young Computer Scientists 2008. ICYCS 2008, Hunan, 2008, pp. 903 - 906

[17] F. Wasilewski. (2011). Wavelet Daubechies 4 (db4). Available: http://wavelets.pybytes.com/wavelet/db4/

[18] E.-S. El-Dahshan, "Genetic algorithm and wavelet hybrid scheme for ECG signal denoising," Telecommunication Systems, vol. 46, pp.209-215,2010.


262

[ieee 2011 ieee international conference on control system, computing and engineering (iccsce) -...

Documents