Analysis of Physiological Signals in Response to

Stress using ECG and Respiratory Signals of

Automobile Drivers.
Karthik Soman, Varghese Alex, Chaithanya Srinivas
#
M. Tech Bio Medical Dept.
#Asst. Prof. Bio Medical Dept.
VIT University, Vellore, Tamil Nadu, India.
karthi.soman@gmail.com

Abstract This paper gives an analysis of variation of the
physiological signals of a person with respect to the stress
developed within him/her. The analysis was done using ECG and
respiratory signals acquired from the automobile drivers who
were made to drive on different road conditions to get different
stress levels. As a part of analysis, we extracted two feature
signals from the above said physiological signals. QRS power
spectrum and the breathing rate were the two feature signals that
were extracted from the mentioned physiological signals. Heart
rate was used as the marker signal for analyzing the variations in
the extracted physiological feature signals. The variations in the
feature signals with respect to the stress were expressed in terms
of correlation coefficients and were tabulated. The analysis
clearly showed the changes in the feature signals with respect to
the stress of the driver. It showed a direct proportionate relation
in the QRS power and the breathing rate with respect to the
stress of the driver. The analysis also showed that QRS power
signal is a better feature signal for analyzing the stress since it
showed more correlation with the heart rate marker signal. The
analysis points out the fact that the physiological signals can be
used as a metric for monitoring the stress of a person.
KeywordsStress, QRS power, ECG, breathing rate,
respiratory signal, correlation coefficient.
I. INTRODUCTION
American Psychological Association defines stress as any
uncomfortable emotional experience accompanied by
predictable biochemical, physiological and behavioral
changes. Reaction of a person from a normal state to an exited
state in order to preserve his/her integrity can also be defined as
stress. An extreme amount of stress can produce extreme health
issues and it can affect the cardiovascular and central nervous
system. Analysis of stress of a person has got many
applications. Modern automobiles incorporate an intelligent
system to monitor the stress level of the driver inside it to
ensure the safety. According to the US census bureau, 10% of
the total road accidents in 2005 were caused due to the
distraction of the drivers during driving. The number rose to
16% in 2009. Another statistics given by the Network of
Employers for Traffic Safety (NETS) say that 20-50% of the
accidents are caused due drivers distraction. These statistical
results show the importance of analysis of stress in a person.
Stressful mind does not produce a constructive result.
In this paper, analysis of the stress is done using the ECG
and the respiratory signals acquired from the automobile
drivers. Stress of an automobile driver is highly correlated to
heart rate and Galvanic Skin Response (GSR) [1]. We intend to
analyze the stress in a different way using different feature
signals extracted from the ECG and the respiratory signal. This
study can pave the way for more options of physiological
signal features that reflect the stress of a person.
It is proved that the heart rate and the galvanic skin
resistance are the two features that are highly correlated with
the stress of a person. Using this result, the analysis was done
in this paper by taking heart rate as the marker signal.
In this paper QRS power was the feature signal extracted
from the ECG signal of the driver. The methodology of the
extraction of the feature signals is explained below. This
feature signal was further correlated with the marker signal. It
was able to see a clear variation in the correlation when the
stress varies. Breathing rate was the feature signal extracted
from the respiratory signals. This was also correlated with the
stress marker signal. Both feature signals showed a direct
relation with the stress marker signal. The analysis is finally
tabulated to check the change in the correlation coefficients of
the feature signals with respect to the stress marker signal.
The authors like to acknowledge Dr. Jennifer. A. Healy
and Dr. Rosalind. W. Picard for providing the signals required
for the analysis through Physionet database. In this paper, we
have discussed the method by which they have acquired the
signals.
Dr. Jennifer A. Healy and Dr. Rosalind W. Picard
conducted the experiment in such a way that the drivers were
made to drive the car in 3 different situations such as highway
road with less traffic, busy road like a city and road having
toll [1]. These 3 different types of roads gave less, high and
medium stresses to the driver respectively. The less traffic road
was completed within 1000 seconds and busy road was
completed in 500 seconds and toll was completed in around
150 seconds. The driving was continued and the total drive
took around 50 min to 1.5 hrs. In this paper, analysis was done
for the first 30 min since the driver has passed through all the
stress conditions within this 30 min. For acquiring the signals,
4 types of sensors were used such as ECG, EMG, EDR and
respiratory sensors. ECG was used in lead 2 configuration for
reducing the motion artifacts and getting a high R peak
amplitude. EMG was acquired from the trapezius muscle on
the shoulder. GSR/EDR was obtained from the foot and the
978-1-4673-5090-7/13/$31.00 2013 IEEE
574


hand. Hand GSR was acquired from the first and middle finger
of the left hand and the foot GSR was obtained from the arch
of the sole of the left foot. The respiratory signals were
acquired using an elastic Hall Effect sensor strapped around the
drivers diaphragm. The signals were acquired as per the
expansion of the chest cavity of the driver. All these analog
signals were converted digitally by a Flex Comp A/D
converter. This also provided isolation of the driver from the
power supply.
II. METHODOLOGY
A. ECG Feature extraction
Electro Cardio Gram is the graphical plot of electrical
activity of the heart. The important parts of an ECG are P
wave, QRS complex and T wave. Each wave has its own
biological importance. P wave corresponds to the atrial
depolarization, QRS complex corresponds to ventricular
depolarization along with atrial re polarization and finally T
wave represents ventricular re polarization. Since pumping of
the heart to all parts of the body via aorta takes place during
ventricular depolarization, QRS complex has got the maximum
amplitude. Hence QRS complex is a good feature that can be
extracted from the ECG for the studies of pumping of heart.
In this paper, QRS power from each cycle of ECG is used
as the feature signal. This analysis will help to know about the
pumping variation of the heart with respect to the stress of the
person.
For the feature extraction, the steps followed are:
QRS detection QRS extraction Power estimation
Feature Signal formation.
a. QRS detection
It is done by using Pan Tompkins Algorithm [2,3]. The
algorithm was implemented in the following way:
Input Bandpass filter Derivative filter Squaring
Integrator Output.
Bandpass filter was implemented by cascading one low
pass filer with cut off frequency of 11 Hz and a high pass filter
with cut off frequency of 5 Hz.
The recursive filter transfer function for the low pass filter
is
B (z) = 1S2
(1-z
-6
)
2
(1-z
-1
)
2
(1)
The High pass filter was implemented using an all pass
filter minus low pass filter.
Band pass filter = All Pass filter Low Pass filter.
The transfer function for the High pass filter is
Bhp = z
-16
- 1S2 Blp(z) (2)
Blp (z) =
(1-z
-32
)
(1-z
-1
)
(3)
Low pass filter provides attenuation greater than 35 dB at
60 Hz and hence it suppresses the power line interference. It
introduces a delay of 5 samples. High pass filter also introduces
a delay of 13 samples. This has to be compensated finally
while extracting the QRS complex from the ECG signal.
Derivative filter functions in such a way that it will
suppress the low frequency components such as P and T waves
and it will enhance the high frequency QRS complex. The filter
equation for the derivative operation is
y (n) = 18 |2 x(n) + x(n-1) - x(n-S) - 2x(n-4)] (4)
Squaring operation is done to enhance the QRS complex by
making the result positive and emphasizing large differences
from QRS complex. It also suppresses the P and T waves.
The output of the derivative operator will have multiple
peaks within the duration of a single QRS complex. In order to
smoothen this output a moving window integrator is used. The
window width has to be selected in such a way that it should
enclose the QRS complex only. If the window length is too
small it produces multiple peaks, if it is large width it will
enclose T wave also. In this paper a window width of 31 was
used and got a good output. Fig 2 shows the output of the Pan
Tompkins algorithm for the first 1000 samples.
After the QRS detection, the nest stage is to extract the
QRS complex from the ECG signal by taking the index of the
pulses in the output of the Pan Tomkins algorithm after giving
the time delay compensation. The extracted QRS complex
signal is further taken for the QRS power estimation.
The power estimation is done by using the Welch method
[4]. The reason for choosing Welch method for power
estimation is because it takes the average of the modified
periodogram of the portioned segments of the signal so that the
output of the Welch power spectrum will be a smoothened one.
Hanning window of length 6 is used for the PSD calculation.
Window equation is:
w (n) = u.S (1 - cos (2 (nN))), u < n <N (5)
The power equation using Welch method is:
P
QRS (fn) = 1K _ I
k
(
n
)
K
k=1
(6)
Ik (fn) = L0 Ak(n)
2
(7)
Ak (n) = 1L _ X
k
(])w(])c
-2k]nL L-1
]=0
(8)
0 = 1L _ w
2
(])
L-1
]=0
(9)
L is the length of the window.
K is the number of segments of the input.
A
k
is the Fourier transform of the k
th
segment.
I
k
is the modified periodogram of the k
th
segement.
P
QRS
is the average of the periodograms of all the K
segments.
Fig 1 shows the Power Spectral Density obtained using
Welch method for all the QRS complexes extracted.
The estimated power of the QRS complex from a single
cycle is averaged so that such averaged power of the QRS
575


complex formed the feature signal for the analysis. Fig 3 shows
the feature signal extracted from the ECG signal.
The feature signal thus formed is classified based on the
mean value of the QRS power. Those powers above the mean
value are taken as high power and the other as low power.
Heart rate reflects the stress of the person. Hence heart rate
was taken as the marker signal which reflected the stress of the
driver. Heart rate and the stress showed a direct relation i.e. as
the stress of the person increases the heart rate also increases.
Hence the heart rate signal was classified based on a threshold
value of 80 in such a way that those heart beat above the
threshold were classified as high stress marker signal and the
other as the low stress marker signal.
Comparison is made between the QRS power, both high
and low power, to high stress and low stress marker signals for
10 drivers. Comparison is made by calculating the correlation
coefficients between the corresponding QRS power and the
stress marker signals. The equation for the correlation
coefficient is:
i =
C (x,)
c
x
c
j


(11)
C (x, y) = E {(x (n) - x) (y (n + k) - y)] (12)
r = correlation coefficient
C (x, y) = Cross Covariance of x and y
= Standard deviation.
E { } = Expectation value
The correlation coefficients were tabulated. The
comparison between the correlation coefficients pointed out the
fact that the high stress condition of a person is more correlated
to high QRS power than low power and low stress condition is
more correlated to low QRS power. Table 1 shows the list of
correlation coefficients corresponding to the stress and power.
B. Respiratory signal Feature extraction
Respiratory signals are the reflections of breathing of a
person. Here, the respiratory signals were acquired in response
to the expansion of the chest using a Hall Effect transducer tied
on the diaphragm area of the person. The feature signal
extracted from the respiratory signal was the number of breaths
per minute or the breathing rate of the person. Fig4 shows the
respiratory signal for 5 min.
The feature extraction from the respiratory signal is
conducted by detecting the peaks in the signal per minute.
Since each peak corresponds to a chest expansion, the number
of peaks detected shows the breathing rate of the person. This
feature signal is divided in to high breathing rate and low
breathing rate by giving a threshold value. Here, mean of the
breathing rate in 30 min is taken as the threshold. Fig5 shows
the respiratory feature signal for 30 minutes.
The two respiratory feature signals are then compared with
the same stress marker signals used in ECG analysis by
calculating the correlation coefficients for 10 different drivers.
The result of comparison of the breathing rate with the
stress marker signal showed that high breathing rate was more
correlated to high stress condition while the low breathing rate
was more correlated to low stress condition.
III. RESULTS
The result of comparison is tabulated as shown in Table 1.
1. ECG results
It is clear from Table 1 that, high QRS power is more
correlated to high stress condition and low QRS power to low
stress condition.
Biological interpretation of the above result is that, during
high stress condition, the left ventricle of the heart pumps the
blood in high power, which in turn produces high blood
pressure. In case of low stress the left ventricle of the heart
pumps the blood in a low power so that the blood pressure also
will be normal. That is why it is said that stressful persons are
more prone to cardiovascular diseases [5]. Practices such as
meditation, yoga will help to reduce stress and will help the
person to be healthier.
2. Respiration results
From the table it is clear that, high breathing rate is more
correlated to high stress condition and low breathing rate is
more correlated to low stress condition. This leads to a fact that
during stressful conditions the breathing becomes rapid.
Biological interpretation of the above result is that at
stressful condition the pattern of breath becomes rapid and
shallow. It is proved that, during that time message will be sent
to the adrenal gland and releases stress hormones like cortisol
[5], whereas during lower stress the breathing pattern is slow
and deep thus the number breaths per minute is low.
The correlation value varies from person to person. But it is
clearly visible from Table 1 that even though the values change
from person to person the correlation between the stress and
the feature signals maintain the same gradation i.e. high stress
will be correlated more with high QRS power and high
breathing rate. The mean value of the correlation values for the
high stress and high QRS power taken for 10 drivers is
0.27156, for low stress and low QRS power it is 0.2574. In
case of high stress and high breathing rate the mean value is
0.2629 and for low stress and low breathing rate it is 0.252.
Hence, QRS power is a better feature signal compared to
breathing rate to analyze the stress of a person since it showed
a higher correlation value to heart rate.
IV. CONCLUSION
The analysis done in this paper clearly proved the variation
of physiological parameters with respect to the stress
developed. It is a scientific explanation for the cardiac and
respiratory problems occurring in the persons who are stressful.
Persons with stressful minds are prone to heart diseases since
their ventricles take an extra power to pump the blood and this
will increase the blood pressure which will cause the problems
of hypertension. In the present days most of the people are
undergoing stress and it is good to perform practices which can
reduce their stress which will make them healthy.
576


As a part of future research, it is possible to analyze other
physiological signals which can give a higher metric for stress
analysis. In this paper only a single category of subject was
taken i.e. automobile drivers. The comparison of stress
between different categories of subjects also will help to study
about the stress in a better way. That also can be taken as a
future aspect of the research.
From the light of the research conducted for this paper the
overall conclusion is that, how much stress we reduce, that
much healthier we become.



Fig 1. Power Spectral Density obtained using Welch method for all the QRS complexes extracted (Driver 1)
577



Fig 2. Output of the Pan Tompkins algorithm for the first 1000
samples (Driver 1)


Fig 3. ECG feature signal (Driver 1)




Fig 4. Respiratory signal for 5 minutes (Driver 1) Fig 5 Respiratory feature signal for 30 min (Driver1)





578


TABLE 1.
RESULT OF ANALYSIS OF ECG AND RESPIRATORY SIGNALS WITH RESPECT TO STRESS


Driver

ECG signal analysis correlation coefficients

Respiratory signal analysis correlation coefficients
High
Stress
High
Power
High
Stress
Low
Power
Low
Stress-
High
Power
Low stress
Low Power
High stress
High
breathing
rate
High stress
Low
breathing
rate
Low stress
High
breathing
rate
Low stress - Low
breathing rate
1 0.258 0.136 0.171 0.289 0.327 0.241 0.1076 0.228
2 0.260 0.1338 0.148 0.231 0.269 0.146 0.165 0.247
3 0.239 0.153 0.129 0.256 0.218 0.107 0.151 0.254
4 0.244 0.128 0.155 0.253 0.280 0.124 0.157 0.218
5 0.287 0.138 0.1303 0.242 0.239 0.131 0.122 0.214
6 0.398 0.201 0.150 0.257 0.257 0.180 0.116 0.283
7 0.262 0.110 0.103 0.245 0.266 0.149 0.185 0.263
8 0.2776 0.143 0.107 0.241 0.290 0.162 0.140 0.258
9 0.231 0.185 0.181 0.294 0.247 0.122 0.116 0.282
10 0.259 0.142 0.141 0.266 0.236 0.129 0.137 0.273
ACKNOWLEDGMENT
The authors like to acknowledge Dr. Jennifer A.
Healey and Dr. Rosalind W. Picard for acquiring the signals
which was used in this paper for the analysis.
REFERENCES
[1] Jennifer A. Healy and Rosalind W. Picard, Detecting Stress
During Real-World Driving Tasks Using Physiological Sensors,
IEEE Transactions on Intelligent Transportation systems, vol. 6, no.
2, June 2005.
[2] Jiapu Pan and Willis J. Tompkins, A Real Time QRS Detection
Algorithm, IEEE Transactions on Bio Medical Engienering, vol.
BME-32, no. 3, March 1985.
[3] Rangaraj M. Rangayyan, Bio Medical Signal Analysis A Case
Study Approach, 4
th
ed. IEEE press, 2009.
[4] P.D. Welch, A Direct Digital Method of Power Spectrum
Estimation,IBM Journal, April 1961.
[5] K.J.W Wilson, Anne Waugh, Janet S. Ross, Ross and Wilson
Anatomy and Physiology in Health and Illness, 8
th
ed. Churchill
Livingston, 1996.






















579