Download Open the publication - UEF Electronic Publications

Jukka Lipponen
Estimation Methods for
Cardiovascular Signal
Analysis
Publications of the University of Eastern Finland
Dissertations in Forestry and Natural Sciences
JUKKA LIPPONEN
Estimation methods for
cardiovascular signal
analysis
Publications of the University of Eastern Finland
Dissertations in Forestry and Natural Sciences
No 146
Academic Dissertation
To be presented by permission of the Faculty of Science and Forestry for public
examination in the Auditorium TTA in Tietoteknia Building at the University of
Eastern Finland, Kuopio, on September, 5, 2014,
at 12 o’clock noon.
Department of Applied Physics
Kopijyvä Oy
Kuopio, 2014
Editors: Prof. Pertti Pasanen, Prof. Kai Peiponen, Prof. Pekka
Kilpeläinen and Prof. Matti Vornanen
Distribution:
University of Eastern Finland Library / Sales of publications
http://www.uef.fi/kirjasto
ISBN: 978-952-61-1532-0 (printed)
ISSNL: 1798-5668
ISSN: 1798-5668
ISBN: 978-952-61-1533-7 (pdf)
ISSN: 1798-5676
Author’s address:
University of Eastern Finland
Department of Applied Physics
P.O.Box 1627
70211 KUOPIO
FINLAND
email: Jukka.Lipponen@uef.fi
Supervisors:
Docent Mika Tarvainen, Ph.D.
University of Eastern Finland
Department of Applied Physic
P.O.Box 1627
70211 KUOPIO
FINLAND
email: Mika.Tarvainen@uef.fi
Professor Pasi Karjalainen, Ph.D.
University of Eastern Finland
Department of Applied Physic
P.O.Box 1627
70211 KUOPIO
FINLAND
email: Pasi.Karjalainen@uef.fi
Reviewers:
Professor Jari Viik, Ph.D.
Tampere University of Technology
Department of Electronics and
Communications Engineering
P.O.Box 527
33101 TAMPERE
FINLAND
email: Jari.Viik@tut.fi
Professor Pablo Laguna, Ph.D.
University of Zaragoza
Department of Electrical Engineering
Mara de Luna
50015 ZARAGOZA
SPAIN
email: [email protected]
Opponent:
Professor Tapio Seppänen, Ph.D.
University of Oulu
Department of Computer Science and Engineering
P.O.Box 8000
90014 OULU
FINLAND
email: Tapio.Seppanen@oulu.fi
ABSTRACT
Cardiovascular measurements are nowadays important tools for
clinicians in the diagnosis of different cardiovascular diseases or
disease complications. Novel approaches for modeling of the cardiovascular system are intended to provide more accurate methods
for estimating the functioning or condition of the cardiovascular
system, and thus to improve can enhance the diagnostics of cardiovascular diseases. Previous studies have demonstrated that cardiovascular measurements could be beneficial in some situations
e.g. monitoring humans during their every day life, for example,
in order to prevent hypoglycemic events in diabetic subjects or to
monitor recovery. However, monitoring humans during their daily
life requires analyzing methods which can provide accurate parameters from even noisy recordings.
In this thesis, a principal component regression (PCR) based
method for analyzing noisy ECGs was developed. The use of datadriven prior information in the PCR model makes it very robust for
noise and thus the developed method can yield accurate parameter
estimates from ECG measured during a subject’s everyday life or
even from exercise ECG measurements.
Currently exercise performance estimation is done mainly using
blood lactate concentration analysis or using spirometric measurements. While these methods achieve accurate results, their suitability for normal fitness enthusiasts is limited because the analysis
requires access to expensive and complicated equipments. On the
other hand ECG is rather simple and inexpensive measurement and
thus heart rate monitors are commonly used for monitoring training intensity. Recent studies have shown that ECG parameters are
not entirely related to heart rate and parameter values differ as compared to exercise and recovery periods. Thus it was hypothesized
that these ECG parameters could provide additional information
beyond that which can be obtained wih heart rate to monitor exercise intensity and recovery. Results showed strong correlations
between blood lactate concentrations and ECG parameters such as
T-kurtosis and TCRT. Linear regression results also revealed that
additional information beyond the heart rate could be achieved using these ECG parameters.
The PCR method was also used to analyze changes of repolarization characteristics during a hypoglycemic event. The aim was
to study the possibility of detecting hypoglycemic events based on
ECG signal features. It was found that repolarization characteristics, specifically QT-time and the RT-amplitude ratio increase during hypoglycemia and this increase continued as long as the hypoglycemic event lasted. In addition, it was observed that changes in
repolarization characteristics caused by hypoglycemia became less
pronounced as duration of the diabetes increased. The results indicated that hypoglycemia detection could be possible, at least for
diabetics with a short history of diabetes.
In this thesis, a model for blood pressure regulation system was
also developed. More specifically, the associations between baroreflex sensitivity, autonomic nervous system (ANS) and artery elasticity were studied. An advanced causal multivariate autoregressive (MAR) model was used to model the dependencies between
different cardiovascular signals. The causal MAR-model can distinguish between direct effects, i.e. heart rate variation causing
blood pressure change, and the variation originating from baroreflex, i.e. blood pressure fluctuations which evoke a reflectory heart
rate change, from cardiovascular time series. The results indicated
that arterial elasticity can reduce the sensitivity of the blood pressure regulation system because stretch sensitive baroreceptors are
not capable to detect blood pressure fluctuations. Therefore, if
baroreflex sensitivity is used as a measure of ANS functioning, then
it is recommended that the assessment should be performed via arterial diameter variation rather than from blood pressure.
.
National Library of Medicine Classification: QT 36, WG 25, WG 140
Medical Subject Headings: Electrocardiography/methods; Cardiovascular System; Models, Cardiovascular; Blood Pressure; Baroreflex; Diabetes
Mellitus; Hypoglycemia/diagnosis; Lactic Acid/blood; Signal Processing,
Computer-Assisted; Principal Component Analysis; Regression Analysis;
Multivariate Analysis
Yleinen suomalainen asiasanasto: EKG; verenkiertoelimet; sydn; mallintaminen; estimointi; verenpaine; diabetes; hypoglykemia; maitohappo; signaalianalyysi; regressioanalyysi; monimuuttujamenetelmt
Acknowledgements
This research was carried out during the years 2008-2014 in the Department of Applied Physics, University of Eastern Finland. First
I would like to thank both of my supervisors for supporting my
work during the thesis project. I am grateful to my principal supervisor docent Mika Tarvainen for sharing his expertise of biosignal
analysis and guidance in the world of science. I also want thank to
my second supervisor Professor Pasi Karjalainen for giving me the
opportunity to work in his Biosignal Analysis and Medical Imaging
research group. I also want thank the entire Biosignal Analysis and
Medical Imaging group for the interesting discussions and enthusiastic atmosphere which has helped me during this project. I am
also grateful to all my co-authors for their significant contribution,
special thanks for Tiina and Tomi Laitinen for sharing their medical
expertise in cardiovascular research.
I am grateful to the preliminary examiners Professor Jari Viik
and Professor Pablo Laguna for valuable comments and suggestions on the thesis.
I want to thank to my parents Anja and Seppo and my brother
Jarkko for believing in me and supporting me during my whole
life. I am also grateful to Veli-Matti, Jouni, Mikael, Mika and Jani
for the support and friendship you have given me.
Finally I want to thank my dearest one Minna for her endless
love and understanding.
Kuopio 12 August, 2014
Jukka Lipponen
LIST OF PUBLICATIONS
This thesis is based on the following articles:
I Lipponen JA, Tarvainen MP, Laitinen T, Lyyra-Laitinen T and
Karjalainen PA, A Principal component regression approach
for estimation of ventricular repolarization characteristics, IEEE
Transactions on Biomedical Engineering, 57(5): 1062-1069 (2010)
II Lipponen JA, Kemppainen J, Karjalainen PA, Laitinen T, Mikola
H, Kärki T and Tarvainen MP, Dynamic estimation of cardiac
repolarization characteristics during hypoglycemia in healthy
and diabetic subjects, Physiological measurement, 32(6): 649-660
(2011)
III Lipponen JA, Gladwell VF, Kinnunen H, Karjalainen PA and
Tarvainen MP, The correlation of vectorcardiographic changes
to blood lactate concentration during an exercise test, Biomedical Signal Processing and Control , 8(6): 491-499 (2013)
IV Lipponen JA, Tarvainen MP, Laitinen T, Karjalainen PA, Vanninen J, Koponen T and Laitinen TM, Causal estimation of
neural and overall baroreflex sensitivity in relation to carotid
artery stiffness, Physiological measurement, 34(12) 1633-1644 (2013).
Throughout the overview, these papers will be referred by their
Roman numerals. The original publications have been reproduced
with permission of the copyright holders.
AUTHOR’S CONTRIBUTION
The publications of this thesis are original research articles and the
author has been the main author of each paper I-IV; in all papers
the co-operation provided by the co-authors has been significant.
The author participated in gathering of the measurement data for
study III, for measuring exercise ECG from healthy subjects. Measurements for study II were made in Turku University Hospital and
the measurements conducted in study IV were made in Kuopio
University Hospital. The author has developed all the mathematical methods and algorithms for signal analysis and cardiovascular
modeling whith the exception of the, QRS-detection algorithm and
arterial diameter detection algorithm. The author has analyzed all
the measurements and formulated all of the results presented in
studies I-IV
ABBREVATIONS
AIC
ANS
AR-model
AT
BGC
BP
BLC
BRS
BRSSBP,RR
BRSds ,RR
BRSSBP→RR
BRSds →RR
CAN
CO2/O2
cosRT
DBP
ECG
HR
HRV
MAR-model
PCR
PC
QT
QTc
RR
RT-ratio
Akaike information criterion
Autonomic nervous system
Autoregressive model
Anaerobic threshold
Blood glucose concentration
Blood pressure
Blood lactate concentration
Baroreflex sensitivity
Overall baroreflex sensitivity estimated
using noncausal analysis
Neural baroreflex sensitivity estimated
using noncausal analysis
Overall baroreflex sensitivity estimated
using causal analysis
Neural baroreflex sensitivity estimated
using causal analysis
Cardiac autonomic neuropathy
Carbon dioxide oxide ratio
see TCRT
Diastolic blood pressure
Electrocardiogram
Heart rate
Heart rate variability
Multivariate autoregressive model
Principal component regression
Pricipal component
QT-interval
heart rate corrected QT-interval
RR-interval
RT amplitude ratio
SBP
SD
SG-filter
SNR
SDNN
T1DM
T2DM
TCRT
VE
VECG
Systolic blood pressure
Standard deviation
Savitzky-Golay filter
Signal to noise ratio
Standard deviation of RR intervals
Type 1 diabetes mellitus
Type 2 diabetes mellitus
Total cosine of the spatial angle
between QRS complex and T wave
Minute ventilation
Vectorcardiogram
NOTATIONS
(·)∗
(·)T
|| · ||
λk
θ
θT
θR
φT
φR
aj
C
Coh
f
fs
P( f )
p
R
T
vk
w
Complex conjugate operator
Transpose operator
Euclidean norm
Prediction error
k:th eigenvalue
Model coefficients
Angle of T-wave vector loop
Angle of R-wave vector loop
Angle of T-wave vector loop
Angle of R-wave vector loop
j:th AR coefficient
Covariance matrix
Coherence
Frequency
Sampling frequency
Power spectral density matrix
Model order
Correlation matrix
System transfer function
k:th eigenvector
White noise
Contents
1 INTRODUCTION
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Scope and structure . . . . . . . . . . . . . . . . . . . .
2 CARDIOVASCULAR SYSTEM
2.1 Physiology of cardiovascular system . . . .
2.2 Electrophysiology of the heart . . . . . . .
2.3 Cardiovascular measurements . . . . . . . .
2.3.1 Cardiography . . . . . . . . . . . . .
2.3.2 Vectorcardiography . . . . . . . . .
2.3.3 Baroreflex sensitivity estimation . .
2.4 Arterial elasticity . . . . . . . . . . . . . . .
2.5 Cardiovascular signals in diabetes . . . . .
2.5.1 ECG changes during hypoglycemia
2.5.2 Cardiovascular changes in diabetes
2.6 Exercise physiology . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
3
7
7
8
9
9
13
15
17
18
18
19
21
3 AIMS
25
4 METHODS
4.1 Principal component regression approaches for ECG
denoising . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Principal component analysis . . . . . . . . . .
4.1.2 PCR model for electrocardiography . . . . . .
4.1.3 Augmented PCR model for vectorcardiography
4.2 Traditional ECG noise reduction methods . . . . . . .
4.3 Cardiovascular time series modeling . . . . . . . . . .
4.3.1 AR model . . . . . . . . . . . . . . . . . . . . .
4.3.2 Multivariate AR model . . . . . . . . . . . . .
4.3.3 Baroreflex sensitivity estimation using MARmodel . . . . . . . . . . . . . . . . . . . . . . .
27
27
27
28
31
32
32
32
35
36
4.3.4
5
6
7
Model validation . . . . . . . . . . . . . . . . .
MATERIALS
5.1 ECG signals for noise sensitivity analysis
5.2 Hypoglycemia measurements . . . . . . .
5.3 Exercise ECG measurements . . . . . . .
5.4 Cardiovascular measurements . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
43
43
43
45
46
RESULTS
6.1 Noise sensitivity analysis . . . . . . . . . . . . . . . .
6.2 Electrocardiographic changes during hypoglycemia .
6.3 ECG changes during exercise and performance estimation . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4 Modeling of arterial baroreflex . . . . . . . . . . . . .
49
49
52
DISCUSSION AND CONCLUSION
7.1 ECG preprocessing . . . . . . . . . . . . .
7.2 ECG changes during hypoglycemia . . .
7.3 VECG parameters during exercise . . . .
7.4 Overall and neural baroreflex sensitivity
69
69
71
74
77
REFERENCES
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
57
63
80
1 Introduction
1.1 BACKGROUND
The main function of the cardiovascular system is to deliver nutrients and oxygen to the tissues and organs. The blood delivery system consists of the heart, arteries, veins and capillaries which are
controlled by the autonomic nervous system (ANS) [1, 2]. ANS receives its input from afferent receptors, for example stretch sensitive
baroreceptors which detect fluctuations in blood pressure [3]. Integration of these signals in the medulla constructs then autonomic
outflow to the cardiovascular organs. ANS can regulate blood pressure in several ways e.g. by changing the cardiac output (blood
volume/minute), altering systemic vascular resistance and modulating local organ blood flow [2].
There are many possible ways to study the functioning of the
cardiovascular system, but probably the most common options are
electrocardiography and blood pressure (BP) measurements. In
electrocardiography, the electrical activity of the heart can be measured from the body surface using noninvasive electrodes. The electrocardiogram (ECG) displays electrical activation (depolarization)
and deactivation (repolarization) of atria and ventricles which can
be seen as repeated patterns called as P, QRS and T -waves [4, 5].
ECG can be used for diagnostic purposes since it is changed in several cardiac diseases but it can also provide information about the
balance of electrolytes or the level of glucose in the blood. Nowadays ECG is also important diagnostic tool for clinicians. Interpretation of the ECG requires often a trained clinician to analyze the
ECG signals, however there are now increasingly automated analyzing tools for developed specific purposes.
The fluctuations of the heart rate (HR) and BP have been shown
to be useful in measuring the function of cardiovascular autonomic
regulation. Measuring the HR responses to changing BP fluctuation is called an analysis of baroreflex sensitivity (BRS), and it has
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
been widely used as tool to analyze cardiovascular regulation system [6]. Many different methods have been proposed for BRS estimation, probably the most advanced option is the so called causal
BRS analysis which is based on a multivariate autoregressive model
(MAR) for cardiovascular system [7, 8]. A causal method can separate two causal directions in the cardiovascular signals; variation
where the heart rate change triggers blood pressure change and
variation which is of baroreflex origin i.e. changes in blood pressure are the reason for the heart rate changes.
Impairment of the baroreflex, i.e. reduced BRS, can lead to exaggerated blood pressure fluctuations [6]. There are two independent factors which can lead to reduced BRS. First an increase of
arterial stiffness can reduce the ability of stretch sensitive baroreceptors to detect blood pressure fluctuations. Secondly, if the ANS
is not functioning properly, baroreceptor neural impulses do not
reach the medulla [9]. Thus BRS can be subdivided into two different estimates; overall BRS and neural BRS, the first one describes
the general baroreflex functioning and second one reflects the sensitivity of the baroreflex neural branch, i.e. functioning of ANS. It is
important to divide BRS into these two estimates for example when
studying diabetic subjects, whose arterial stiffness may be increased
but furthermore the functioning of their ANS may also be reduced.
Diabetes is a disease were insulin production is either decreased
or the organs develop insulin resistance [10]. Insulin allows cells
to absorb glucose, and therefore in diabetics the lack of insulin
results elevated blood glucose concentrations (BGC) i.e. hyperglycemia. Hyperglycemia is problematic because it can damage
both the nerves and blood vessels, and over the long term it may
cause to a variety of complications such as cardiovascular autonomic neuropathy (CAN), myocardial ischemia or stroke. Diabetics
need to control their BGC by injecting insulin throughout the day,
because good BGC balance reduces the risk of long term complications. However, maintaining a suitable BGC level can be very difficult because unnecessary insulin injections can lead to decreased
BGC i.e. hypoglycemia which may result in seizures or cognitive
2
Dissertations in Forestry and Natural Sciences No 146
Introduction
problems, coma even death.
Hypoglycemia has been found to cause changes in the heart’s
electrical activity and these changes are reflected in the ECG waveforms, more specifically on the characteristics of cardiac repolarization [11–13]. It has been hypothesized that ECG changes could be
used for the detection of hypoglycemia. Online detection of these
ECG changes could enable development of a hypoglycemia alarm
system and seizures and other serious symptoms of hypoglycemia
could be avoided [14, 15].
Exercise has also been shown to induce changes in ECG waveforms in a similar manner as hypoglycemia. These changes can
be revealed when ECG characteristics are compared between exercise and recovery periods at same heart rate levels [16–18]. It has
been hypothesized that these changes are caused by alterations of
the ANS or increased secretion of hormones which affect the action
potentials in the heart.
1.2 SCOPE AND STRUCTURE
The present thesis focusses on the on development and validation
of cardiovascular signal analysis methods and time series models,
aiming to provide better tools for analyzing physiological changes
in the cardiovascular system. In studies I-III electrocardiographic
changes were the main focus whereas in IV study cardiovascular time series modeling was used to clarify the factors related to
baroreflex sensitivity.
In study I, an advanced ECG preprocessing method was developed, aiming to automated beat-to-beat analysis of ECG waveforms. This method was further developed for analyzing ECG
changes during hypoglycemic events (study II) and during exercise
(study III). These novel methods are based on principal component
(PC) modeling of ECG waveforms and by using this model characteristics of the ECG can be estimated. The developed method was
validated by comparing it to commonly used ECG preprocessing
methods. An automated ECG analysis was developed to enable the
Dissertations in Forestry and Natural Sciences No 146
3
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
evaluation of measurements made outside of the laboratory where
the signal quality can be low.
In study II, hypoglycemia related cardiac repolarization changes
were studied using continuous ECG analysis and BGC measurements. The specific aim was to estimate ECG characteristics in
beat-by-beat basis which made it possible to do a temporal comparison between BGC and ECG changes. In addition, repolarization characteristics were studied using three different groups of
subjects; healthy, recently diagnosed diabetics and diabetics with
long histories of the disease. In this way the impact of the duration
on changes in cardiac repolarization could be analyzed. These results are steps towards to creating a system which could accurately
detect hypoglycemic changes via ECG features.
The aim of study III was to analyze the relationships between
exercise intensity markers and ECG characteristics and to reveal if
there were any ECG parameters that could be used for performance
estimation. It was hypothesized that changes in ECG characteristics
between exercise and recovery periods could be related to some
specific analytes or hormones generated during the exercise, and
that these changes could be used to monitor recovery from intensive
exercise.
In study IV, multivariate AR models were created to estimate
neural and overall BRS. For the first time, a causal estimation method
was used for determining neural BRS values. It was hypothesized
that using causal analysis neural BRS estimates could provide an
accurate assessment of ANS functioning. The relationship between
artery stiffness and BRS was analyzed in diabetic subjects and the
difference between overall and neural BRS could be determined.
This thesis consists of seven chapters. Chapter 2 provides background information of cardiovascular system and a review of studies related to ECG changes during hypoglycemia and exercise, and
state-of-the-art baroreflex sensitivity estimation methods. Chapter
3 lists the aims of the thesis. In chapter 4, the ECG preprocessing methods developed in this thesis are introduced and the cardiovascular time series models used for BRS estimation are pre-
4
Dissertations in Forestry and Natural Sciences No 146
Introduction
sented. Chapter 5 details the cardiovascular measurements which
were used in this thesis. Chapter 6 presents the main results of
studies I-IV and Chapter 7 contains a discussion.
Dissertations in Forestry and Natural Sciences No 146
5
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
6
Dissertations in Forestry and Natural Sciences No 146
2 Cardiovascular system
This section briefly describes the main parts and functions of the
cardiovascular system. The main focus is on functioning of the
heart and electrocardiography.
2.1 PHYSIOLOGY OF CARDIOVASCULAR SYSTEM
The cardiovascular system consists of the heart, arteries, veins and
capillaries. Its main function is to deliver blood transporting nutrients and oxygen for tissues and organs and to transfer excreta
away from tissues [2]. The heart maintains BP by pumping blood
through the aorta into the systemic circulation. The maximum BP
achieved during the contraction of the heart is called systolic blood
pressure (SBP). SBP is mainly dependent on two factors, of the contraction strength of cardiac muscle and vascular resistance. After
the contraction, the aortic valves close and pressure difference between peripheral circulation and aortic arch causes blood to flow
towards the peripheral arteries. Because of this flow, BP decreases
until the heart beats again and the minimum BP just before next
contraction is called diastolic blood pressure (DBP) [2].
Cardiovascular system is controlled by the ANS which regulates cardiac output (blood volume/minute), systemic vascular resistance and local organ blood flow to ensure maintenance of optimal arterial pressure. Fluctuations in the HR are called as heart
rate variability (HRV) which reflects the dynamic response of ANS
to changing the internal and external conditions [1]. HR is controlled by the sinoatrial node which receives innervations from both
the sympathetic and parasympathetic branches of ANS [19]. Sympathetic activity increases HR and causes vasoconstriction which
evoke increased cardiac output and increased vascular resistance,
as a result, the arterial BP is increased. On the other hand, increased parasympathetic and decreased sympathetic activity tend
Dissertations in Forestry and Natural Sciences No 146
7
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
to decrease HR and decrease vascular resistance which leads to decrease in BP.
Long term regulation of the BP is mainly acquired by sympathetically activated renal regulation of blood volume, whereas the
arterial baroreflex is the mechanism provides beat to beat control of
HR and short-term control of BP. Baroreceptors are stretch-sensitive
sensory nerve endings and are mainly located at the carotid sinuses
and aortic arch [3, 6]. In simplified terms, increased BP causes vascular stretch and activates baroreceptors, which triggers ANS to
decrease HR and vascular resistance. Thus, an impairment of the
baroreflex leads to exaggerated BP fluctuations.
2.2
ELECTROPHYSIOLOGY OF THE HEART
Heart is an electrically activated muscle pump, which consists of
two atriums and two ventricles [2]. Blood entering the right atrium
is forced by atrial contraction into the right ventricle which pumps
the blood into the pulmonary artery and thence through the lungs.
From the lungs, oxygenated blood arrives into the left atrium and
finally the left ventricle pumps it into the aorta and through the
systemic circulation.
Electrical activation of the heart begins from the sinoatrial node
(SA-node) which is located in the right atrium. The SA-node consists of specialized muscle cells which generate action potentials
according to their natural rate, which has a normal resting state of
about 72 times per minute [2]. However, as mentioned earlier, ANS
can strongly influence on this rate. Action potentials originating
from SA-node spreads through the atria along the cardiac muscle fibers causing the atrial muscle to contract. From the atrium,
the impulse is propagated to the atrioventricular node (AV-node)
which delays the impulse to allow ventricles to fill [20]. From the
AV-node, activation spreads to ventricles along special conduction
fibers called Purkinje fibers. These Purkinje fibers propagate the action potential rapidly to all portions of the ventricles, which ensures
a simultaneous contraction of muscle fibers providing maximum
8
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
pumping power [2].
Electrical activation of the cardiac muscle cell is based on the
structure of the cell membrane [21, 22]. In its resting state, the cell
membrane is more permeable to potassium ions (K+ ) than it is to
sodium ions (Na+ ) and it has an active transport process which
maintains the concentration of K+ ions higher in the inside of membrane and the Na+ ion concentration higher on the outside. However when the voltage rises over a threshold due to action potential,
cell membrane becomes permeable to Na+ ions which flow inside
and impermeable to K+ ; this activation of the cell is called depolarization [21]. After depolarization the cell is more positive in the
inside, however there is gradual increase in K+ permeability and
this restores the intracellular potential to its resting value (repolarization phase).
2.3 CARDIOVASCULAR MEASUREMENTS
2.3.1 Cardiography
Measurements
Electrical potentials caused by heart depolarization and repolarization spread through our body. The measurement of those potentials
is called electrocardiography and the outcome of this procedure is
ECG [4, 5]. The ECG can be measured by placing electrodes on the
surface of the skin. The simplest ECG measures the potential difference between a reference point and the measurement point. For
example heart rate monitors use this kind of simply lead system.
The clinically most commonly used ECG lead system is the standard 12-lead system, which consist of six limb leads which are I,
II, III, aVL, aVR and aVF and six chest leads (or precordial leads)
V1 -V6 [4, 5]. Leads I, II and III are bipolar i.e. lead I is achieved
from the potential difference between the left arm and the right
arm. Similarly, lead II measures potential difference between left
leg and right arm and lead III between left leg and left arm. aVL
aVR and aVF are called augmented limb leads because the average
of the potential measured from two other limb electrodes is used as
Dissertations in Forestry and Natural Sciences No 146
9
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
I
[μV]
400
P
QRS
[μV]
V1
400
T
0
0
−400
−400
[μV]
II
[μV]
400
400
0
0
−400
V2
−400
[μV]
III
[μV]
400
400
0
0
−400
V3
−400
[μV]
aVR
[μV]
400
400
0
0
−400
V4
−400
[μV]
aVL
[μV]
400
400
0
0
−400
V5
−400
[μV]
aVF
[μV]
400
400
0
0
−400
V6
−400
0
0.5
time (s)
1
0
0.5
time (s)
1
Figure 2.1: An example of 12-lead ECG. Limb leads are presented in left column and chest
leads in right column.
a reference [4, 5]. Precordial leads measure the potential difference
between six chest electrodes and Wilson’s central terminal which is
the average potential of the three limb electrodes. Each lead views a
different projection of the heart’s electrical activity because of their
different viewpoints. The limb lead projections are in the frontal
plane and the chest leads are in a horizontal plane.
The ECG pattern which can be seen in figure 2.1 is the result
10
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
of the electrical activations and deactivations of cardiac cells, during repolarization and depolarization. Thus when one conducts an
ECG analysis then, the ECG is actually composed of waves, segments and intervals which reflect specific parts of the heart beat
[4, 5]. In figure 2.1, the waveforms present in a normal ECG are
shown, the first ECG wave is the P-wave which represents the discharge of the SA-node and atrial depolarization. After the P-wave,
depolarization slows down in the AV-node (ventricles fill), which
is revealed as a short pause (0.08-0.1s) in the ECG. Ventricular depolarization causes a rapid and angular waveform called the QRScomplex and 0.1s after the depolarization phase, the repolarization
of ventricles starts which is represented by the T-wave on ECG.
Parameters
Most commonly used ECG parameter is undoubtedly the HR which
is defined by the number of heartbeats per minute. RR interval
(RR) on the other hand is defined by the time difference between
two consecutive R-peaks i.e. heart beats. As described earlier, the
ANS regulates HR and thus RR-interval is changing from beat to
beat. Analysis of RR interval variation is called HRV analysis. HR
and HRV analysis have been used widely in sports (exercise performance) [23] and in clinical applications mainly as a measure of
ANS functioning [24].
Depolarization and repolarization phase are clearly visible in
the ECG thus several parameters have been defined for measuring
heterogeneity of these phases. Figure 2.2 shows a few commonly
used parameters for measuring depolarization and repolarization
characteristics. Probably the most widely used parameter to assess
repolarization abnormalities is the QT-interval which is a measure
of overall duration of the depolarization and repolarization times
of the ventricles. Normally QT-time is defined from Q-wave onset
to T-wave offset [25]. However, if the dynamics of the total repolarization and depolarization time are of interest, RT-time can be often
used instead of the QT-time because of its better accuracy. There is
two common ways to define RT-time (RTend and RTapex ) or QT-time
Dissertations in Forestry and Natural Sciences No 146
11
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
(QTend and QTapex ). The RTend -time is defined as the time elapsing from the R-wave apex to the intersection of the baseline and
steepest tangent fitted to T-wave downslope (see figure 2.2) [26]. In
estimation of RTapex -interval, T-wave apex is used as the endpoint
instead of T-wave steepest tangent.
It is widely known that QT (or RT) interval decreases at higher
HRs, and thus several methods for compensating this effect has
been developed. Two simple and commonly used methods for estimation of heart rate corrected QT-interval (QTc) are Bazett’s formula [27]
QT
(2.1)
QTc = √
2
RR
and Fridericia’s formula [28]
QT
QTc = √
3
RR
(2.2)
where HR level is corrected by dividing QT interval by the square
root or cubic root of the preceding RR interval. However it should
be noted that physiological factors such as gender and ANS alter
the relationship between the RR and QT intervals and thus this kind
of generalized formulas should be used with caution [29]. A Physiologically prolonged QTc-time or RTc-time signifies that it takes
longer than normal time for the myocardial cells to be ready for new
cycle. Prolongation of the QTc-time has been associated with the
risk of life-threatening arrhythmias and sudden cardiac death [30].
The depolarization phase consists of depolarization of the atria
and ventricles. Abnormalities in the depolarization of the ventricles
are shown in the PQ-duration which measures how the action potential is being transmitted through the AV-node from the atria to
the ventricles [4, 5]. On the other hand abnormalities in the shape
and duration of the QRS complex are related to intraventricular
conduction disturbances, increased QRS duration and elevated ST
segment are examples associated with bundle branch blocks [31].
Ventricular repolarization abnormalities are reflected in changes in
the T-wave. Parameters used to study ventricular repolarization
involve T-wave shape parameters: T-wave amplitude (T-amp) and
12
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
[μV]
QT
RT
700
R−amp
T−amp
0
T
PQ
QRS
−400
0
0.1
0.2
0.3
0.4
time (s)
0.5
0.6
0.7
Figure 2.2: One beat of ECG and parameters measuring depolarization and repolarization
characteristics. Presented parameters are R and T-wave amplitudes (R-amp and T-amp),
durations of PQ-interval, QRS-complex and T-wave. Parameters describing the total depolarization and repolarization time i.e RT and QT-intervals are also presented.
RT amplitude ratio (RT-ratio) and also T-wave duration has been
used as a measure of repolarization. For example T-waves abnormalities have been shown to be related to ischemia, i.e. reduced
blood supply in the myocardial tissue [32], and many other cardiac
dysfunctions [30, 32].
Morphological changes of ECG waveforms are also common in
certain diseases and thus parameters which reflects the morphology
of QRS-complex and T-wave have been developed. Here one such
parameter i.e. T-kurtosis is presented which descripes peakedness
of a T-wave [33]. T-kurtosis is defined as
1
T−kurtosis = N
2
1
N
∑kN=1 (zk − z)3
∑kN=1 (zk
− z )2
3
(2.3)
where zk is k’th data point of the T-wave epoch and z is mean value
of the same epoch.
2.3.2 Vectorcardiography
Vectorcardiography describes the electrical activity of the heart as
a 3 dimensional vector at each time instant. Each point of vector-
Dissertations in Forestry and Natural Sciences No 146
13
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
cardiogram (VECG) is a sum of the electrical vectors generated by
cardiac cells, i.e. it represents the average magnitude and direction of electrical forces of the cardiac cells. It is evident that the
direction and magnitude changes as a result of repolarization and
depolarization. The basic concept of the vectorcadiography is to
construct 3 orthogonal leads (X, Y and Z) which views the heart
in the three different orthogonal projections. The most commonly
used lead system for vectorcardiography is called Frank’s lead system [34]. Frank’s lead system is based on the assumption that the
heart can be modelled as a dipole which is located at the center of
the heart. Frank’s lead system was developed using a human torso
model taking into account inhomogeneities in the human body, although conductive properties are assumed to be homogenous and
linear.
The lead configuration of Frank’s lead system is shown in figure 2.3. It consists of 7 electrodes, the five chest leads construct a
horizontal plane (leads X and Z) and the leads on junction of the
neck and on left foot are used to construct vertical lead (Y). Orthogonal leads are then constructed by weighting lead voltages by the
weights acquired from torso model.
The right part of figure 2.3 shows vector loops of depolarization
(blue) and repolarization (red) waveforms and horizontal, frontal
and sagittal projections of the vector loops. One interesting VECG
parameter is the spatial angle between repolarization depolarization waveforms (TCRT). TCRT is defined as the cosine of the dominant vectors of the QRS vector loop and the main vector of the
T-wave loop [35]. Its value is limited in the range [-1,1] where -1 reflects the situation where vector loops are pointing in the opposite
directions and 1 reflects the situation where the loops are pointing
in the same direction. Recently rate dependence of VECG parameters has been studied and its parameters have been shown to be related with HR [36, 37]. Certain VECG parameters, especially TCRT,
have been shown to be potential tools for use in a risk stratification
of sudden cardiac death [22, 38].
Normal ECG parameters can be estimated from VECG measure-
14
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
−0.5
H
Z
y
0
M
A
I
Sagittal
0.5
1
X
1.5
E
C
−1
0
z
1
−0.5
Frontal
0
Y
x
y
z
0.5
1
1.5
y
−0.5
Horizontal
−1
0
x
1
F
z
0
0.5
1
1.5
−1
0
x
1
Figure 2.3: Electrode positions of Frank’s VECG are presented in red circles on the left
side of the figure. On the right side, 3D vector loops of depolarization (blue loops) and
repolarization (red loops) of ventricles during one heart beat are presented together with
sagittal, frontal and horizontal projections.
ment using a ”virtual” lead with a direction along the electrical axis
of the heart, i.e. QRS-loop orientation. Idea of this definition is to
reduce differences between electrode placements and heart orientations between participants [39].
2.3.3 Baroreflex sensitivity estimation
The ability of baroreceptors to control the heart rate and regulate
blood pressure is commonly referred to as baroreflex sensitivity
(BRS). BRS is normally analyzed as the ratio of SBP and RR fluctuations. Several methods for estimation of BRS have been presented,
Dissertations in Forestry and Natural Sciences No 146
15
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
perhaps the most widely used method is a spectral method where
BRS is estimated via the ratio of power spectral densities of SBP
and RR time series [40]. Another widely used method is the sequence method where BRS is estimated from the slope calculated
from sequences of three consecutive rising RR and SBP values [41].
There are several studies, where BRS has been used as a measure
of ANS function [42–44], and BRS has been found to be associated
with mortality after myocardial infarction [42, 43].
However cardiovascular signals are related to each other in a
closed loop system, which means that HR affects directly the BP
(feedforward) and the BP affects the HR through the baroreflex
(feedback). Traditional methods estimate BRS from both feedforward and feedback variations and thus these estimates can be biased [7, 8, 45]. Multivariate autoregressive modeling (MAR) of cardiovascular signals has been used for BRS estimation because MAR
model make it possible to separate of feedforward from feedback
variation in SBP and RR time series [7, 8]. In this method, RR
and BP time series are modeled using the MAR model and causal
gain and coherence are obtained by setting to zero the model coefficients representing feedforward variation from RR to SBP. The
causal approach has been tested with healthy subjects in the supine
position [46] and after head up tilt [45], and it was found that the
traditional methods had been overestimating BRS. Causal BRS estimation has been also utilized for important clinical problems. For
example, causal BRS was found to be a promising index for monitoring the safety of subjects during propofol anesthesia [47].
BRS is normally estimated from the variation in SBP and RR
time series. However, because baroreceptors are stretch sensitive
receptors, reduced artery elasticity causes impaired transduction
of pressure into the barosensory artery stretch and thereby limits
baroreceptors ability to detect BP fluctuations. Thus, in recent reports, BRS estimates have been categorized into overall BRS and
neural BRS estimates. Overall BRS refers to the BRS which is estimated from SBP and RR relationships. In the estimation of neural BRS use of arterial diameter instead of BP has been proposed
16
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
because arterial diameter variation provides direct information on
the stretching of the artery as detected by baroreceptors. Also in
this thesis, BRS estimated from variations between SBP and RR is
referred to as overall BRS and BRS estimated from variations between arterial diameter and RR is considered as neural BRS. The
neural BRS is thought to describe functioning of baroreflex neural
pathways and thus to be a better measure of ANS function because
it is not related to arterial stiffness [9, 48, 49]. Neural BRS has been
analyzed in healthy young individuals [9] and older subjects [50]
and it was found that there was a reduction in neural BRS as the individual aged. Exercise training has been found to improve neural
BRS in both young and older healthy subjects [51, 52].
2.4 ARTERIAL ELASTICITY
Arterial elasticity is an important feature of the cardiovascular system. Elastic arteries help to maintain appropriate BP level and
to buffer the pressure of the pulsation which protects the delicate
micro-circulation from pressure-induced damage and reduces heart
afterload [53, 54]. Arterial stiffness has been found to be related to
left ventricular hypertrophy [55], coronary ischemia [56] and to an
overall increased risk of cardiovascular events [57]. Several methods have been proposed for quantifying the mechanical properties
of arteries. Such as pulse pressure, stress-strain relation, pulse wave
velocity [58] and augmentation index [59] which is estimated from
the ratio between systolic pressure and the pressure of the waves
reflected from arteries.
In this thesis, the mechanical properties of carotid artery were
estimated from relation between pulse pressure (strain) and diameter variation (stretch) of the artery. One commonly used index for
defining arterial elasticity in this way is the β-index [60], which is
defined as
ln(SBP/DBP)
(2.4)
β=
(ds − dd )/dd
where SBP and DBP are systolic and diastolic blood pressures and
Dissertations in Forestry and Natural Sciences No 146
17
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
similarly ds and dd are systolic phase and diastolic phase artery
diameters.
2.5
CARDIOVASCULAR SIGNALS IN DIABETES
Diabetes is a group of metabolic diseases characterized by decreased
insulin production or by resistance of tissues to insulin [10]. Insulin
is the hormone which allows cells to absorb glucose. In diabetics, the lack of insulin results in increased BGC, i.e. hyperglycemia
(BGC >11 mmol/l). Long term hyperglycemia causes damage to
the nerves and blood vessels, thus it can evoke a variety of complications for example cardiovascular autonomic neuropathy (CAN),
myocardial ischemia or stroke. Diabetes is normally divided into
three different subtypes; type 1 diabetes (T1DM), type 2 diabetes
(T2DM) and gestational diabetes [10]. In T1DM, the beta cells which
produce insulin have been destroyed, thus T1DM subjects are dependent on exogenous insulin supplied by injections or they will
die from ketoacidosis. Subjects of T2DM have impaired insulin sensitivity and/or secretion. The progress of T2DM is slow and thus
it is not as lethal as T1DM in the short term. Diabetics maintaining
their BGC at the normal level (4-7mmol/l) through insulin injections, they should strive for good BGC balance since this has been
found to significantly decrease the risk of long term complications.
However this is often problematic because excessive insulin injections can lead to decreased BGC (hypoglycemia), which can result
in seizures or cognitive impairment rather rapidly.
2.5.1
ECG changes during hypoglycemia
Since the early 1990’s when many unexplained sudden deaths of
young people with T1DM (dead in bed syndrome) were reported
and cardiac arrhythmias triggered by hypoglycemia was considered
as the probable reason, hypoglycemia related changes to heart electrophysiology have been widely studied (see table 2.1). The most
common method for investigating hypoglycemia induced ECG chan-
18
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
ges is the glucose clamp technique, where a constant insulin infusion is used to clamp BGC to hypoglycemic range. In these studies
hypoglycemia has been demonstrated to affect cardiac repolarization. In the ECG this is reflected as prolonged QTc time and flattening of T-wave [11–13], and from VECG as a reduction of TCRT,
height and the width of T-wave loop [61]. Since the glucose clamp
technique is not fully comparable with spontaneous hypoglycemic
events, in recent studies continuous glucose monitoring during the
night has been used and repolarization characteristics between the
nights with and without hypoglycemic events have been compared.
QTc prolongation has been found to associate also with spontaneous hypoglycemic events [62–64]. Although the changes in repolarization during hypoglycemia have been widely studied, mechanisms underpinning these changes are not fully understood. There
are three putative mechanisms behind these changes: sympathoadrenal response to hypoglycemia causing increased adrenaline release [65], lowered potassium concentration which affects the activity of cardiac cells [13] or increased parasympathetic activity [66].
The ECG changes occurring during hypoglycemia have been
widely studied, however in many reports ECG changes are studied by manually annotating T-wave segments either from an averaged T-wave (T-waves averaged over a period of few minutes) or
from a few consecutive T-waves [62, 71, 72]. The normal variation
in ECG waveforms is however quite significant, and thus, the averaged waveforms can be more or less distorted.
2.5.2 Cardiovascular changes in diabetes
Hyperglycemia which is common in diabetic patients, causes damage to arteries. At the early stage, this can be seen by reduced
arterial elasticity [73]. In T1DM subjects, the reduced elasticity of
arteries leads to an increase in SBP because the same stroke volume
being ejected into a smaller artery volume generates a higher pressure. Reduced arterial elasticity has been widely studied in T1DM
subjects [74–77]. For example it has been shown to be more pro-
Dissertations in Forestry and Natural Sciences No 146
19
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
nounced in female than male T1DM subjects [74, 75]. Furthermore,
the reduction of elasticity has been reported to begin rather early, in
adolescents with T1DM as young as 15 years old [76, 77]. Reduced
elasticity has been thought to be an early sign of vascular disease
and it has been hypothesized that early detection of these cardiovascular abnormalities and a planned treatment could be one way
to slow down the natural progression of disease [78].
It is not sufficient that the hyperglycemia reduces artery elastic-
Table 2.1: Studies of the ECG changes during hypoglycemia.
Study
T. Linström
1992 [67]
J.L.B Marques
1997 [68]
B. Eckert
1998 [11]
L.
LandstedHallin 1999 [12]
Population
6, T2DM
Method
glucose clamp
7,
T1DM
8, T2DM
10, healthy
glucose clamp
glucose clamp
13, T2DM
glucose clamp
R.H Ireland.
2000 [69]
R. Robinson
2003 [70]
17, healthy,
glucose clamp
10, healthy,
glucose clamp
R. Robinson
2004 [62]
B. Suys
2006 [63]
T. Laitinen
2008 [13]
22, T1DM,
spontaneous
overnight
spontaneous
24h
glucose clamp
M.L. Koivikko
2008 [61]
G.V Gill
2009 [64]
20
9,
T1DM
children
18, healthy
16, T1DM
8, healthy
25, T1DM
glucose clamp
spontaneous,
24h
Main findings
ST-depression (p<0.05), Twave flattening
QTc prolongation (p<0.05)
QTc prolongation (p<0.001),
T-wave flattening (p<0.002)
QTc prolongation (p<0.001),
QT
dispersion
increase
(p<0.001)
QTc prolongation
QTc
prolongation
(p<0.0001),
QT dispersion increase (p<0.0001)
QTc prolongation (p<0.03)
QTc prolongation
QTc prolongation (p<0.001),
T-wave flattening (p<0.001),
ST depression (p<0.001)
and increase of R-wave
amplitude (p<0.001)
T-wave flattening (p<0.05),
TCRT reduction (p<0.05)
QTc prolongation (p<0.04)
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
ity, it also causes damage to the autonomic nerve fibers that innervate heart and blood vessels, resulting in abnormalities in HR and
vascular control. This serious complication is called cardiovascular
autonomic neuropathy (CAN) [79]. Based on these facts, it is not
difficult to believe that cardiovascular diseases are principal causes
of death in diabetic patients [80]. In its early stages, CAN may result in hypotension or exercise intolerance and in its later stages increased incidence of myocardial infarction and ischemia [79]. CAN
is clearly related to impaired HR and BP control, thus HR responses
to the Valsalva maneuver or from a lying to standing change are
used as indicators of CAN [81]. The reduction in HRV is one of the
earliest signs of CAN and thus also HRV during deep breathing
and resting conditions is used in CAN diagnosis [81]. Although
the impaired BP control is a clear sign of CAN, modern spontaneous BRS estimates in diabetic subjects have not been extensively
studied. However BRS has been reported to be able to separate
subjects with CAN from those without this disorder [82] and there
are a few reports that overall BRS could have prognostic value for
CAN detection in the early stages when it is not yet detectable with
conventional methods [44, 83].
2.6 EXERCISE PHYSIOLOGY
Exercise physiology is concerned with how the human body responds to the stress caused by exercise. Exercise physiology is
closely related to sport performance analysis, but on the other hand,
it is clinically relevant since one can evaluate how exercise can be
used to prevent cardiovascular diseases or how exercise testing can
be used directly as a diagnostic tool. In this section, changes in the
cardiac electrophysiology during exercise are shortly reviewed.
The measurement of blood lactate concentration (BLC) is an important tool for defining the anaerobic threshold (AT). BLC and AT
are important measures for athletes because they can be used to
prescribe training intensity. AT corresponds to the onset of blood
lactate accumulation, where the BLC threshold is normally about
Dissertations in Forestry and Natural Sciences No 146
21
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
2 mmol/l [84]. Another commonly used method for defining AT
is the ventilatory threshold which can be determined by on-line
gas analysis [85]. The ventilatory threshold is thought to be associated with the maximal steady state of BLC, normally about 4
mmol/l, however also conflicting reports have been published [86].
There are weaknesses as with these methods; i.e. ventilatory threshold measurements require expensive equipments and can be performed only under laboratory conditions and BLC measurements
are invasive and are not easy to acquire during dynamic exercise.
Therefore attempts have been made to estimate AT from easily measurable signals such as ECG, more specifically the relationship between HR/HRV parameters and exercise intensity markers have
been widely studied.
There are a number of previous studies which have evaluated
the relationship between BLC and heart rate or HRV [87–90]. There
has also been an attempt to determine AT using a heart rate deflection point [91], however conflicting reports have been published related to reliability of this method [92–94]. More recent reports have
claimed that HRV can be used determine the AT [95–97]. However
it is known that during exercise HRV is greatly reduced and this
complicates the accurate estimation of HRV parameters. In practical applications such as in athletic training, HR is actually used to
adjust the training intensity based on relationship between HR and
lactate accumulation. However BLC depends on many individual
factors such as training status and maximal muscle oxidative capacity and thus, the relationship between HR and BLC is not always
clear [98].
In healthy subjects exercise causes several detectable changes in
the ECG. The accelerating HR causes shortening of both the depolarization time (PQ-interval) and repolarization time (QT-interval).
The duration of QRS complex on the other hand does not display any significant changes during exercise, however the duration of R-wave shortens as the S-wave prolongs, similarly R-wave
amplitude decreases as the negative amplitude of S-wave increases
[16,17].There are several studies demonstrating that the QT-interval
22
Dissertations in Forestry and Natural Sciences No 146
Cardiovascular system
is significantly dependent on HR but this relationship differs between exercise and recovery periods i.e. the QT-time is shorter
during the recovery period compared to that obtained at the same
HR during the exercise period [16–18]. This phenomenon where
ECG parameter differs at the same HR depending on whether it is
measured during the exercise or recovery period is referred to as
electrocardiographic hysteresis [99]. Electrocardiographic hysteresis has been thought to be related to changes in the ANS function
or to the exercise induced increase in the level of hormones which
affect the durations of action potentials. Similar mechanisms have
been thought to induce ECG changes during hypoglycemia.
As mentioned previously, exercise ECG is an important tool for
clinical applications. The most important clinical application of exercise ECG is in the diagnostics of ischemic heart disease, but it can
be also used for diagnostics of some other diseases such as cardiovascular disease or testing vulnerability to arrhythmias [100, 101].
Dissertations in Forestry and Natural Sciences No 146
23
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
24
Dissertations in Forestry and Natural Sciences No 146
3 AIMS
This thesis focuses on the development and validation of methods
for cardiovascular signal analysis. In particular, work was done for
development of an ECG preprocessing method which can provide
an accurate parameter estimation even of signals with low signal to
noise ratios, such as the ECG measured outside of the laboratory or
during exercise. Secondly, cardiovascular models were developed
and validated for studying neural and overall BRS in relation to
arterial elasticity.
The specific aims of this thesis were:
1. To develop and validate the principal component regression
based ECG preprocessing method for analyzing ECG signals
with low SNR.
2. To investigate ECG changes during hypoglycemia and reveal
possibilities of development of ECG based alarm system for
hypoglycemic events.
3. To examine the potential of utilizing VECG changes during
exercise in the estimation of physiological performance.
4. To validate cardiovascular models for neural and overall BRS
estimation.
Dissertations in Forestry and Natural Sciences No 146
25
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
26
Dissertations in Forestry and Natural Sciences No 146
4 Methods
In this chapter, the methods which were used in this thesis are described. First, the principal component regression based ECG denoising method is presented. This method was developed in Study
I, applied in the hypoglycemia estimation in Study II, and modified
for VECG case in Study III. Secondly, the cardiovascular time series
models, used for neural and overall BRS estimation in Study IV, are
reviewed.
4.1 PRINCIPAL COMPONENT REGRESSION APPROACHES
FOR ECG DENOISING
4.1.1 Principal component analysis
This section introduces principal component regression (PCR) and
briefly reviews the theory behind principal components (PCs), for
more a detailed description see [102]. The basic concept of PCR is
to reduce the dimensionality in a data set. In practical terms this
is achieved by calculating a new set of variables, the PCs, so that
the first one contains as much as possible of the variation present in
original data. Principal components are calculated by projecting a
data set into a new coordinate system determined by eigenvectors
and the eigenvalues of a data covariance or correlation matrix, for
detailed proof see [102]. PCR has been commonly used for data
reduction, feature extraction, and visualization of multidimensional
data.
Assume that X is a data matrix of M observations from N random variables, i.e. random vectors xi (length N) form the columns
of X. Matrix X spans a vector space of at the most min{ M, N } dimensions. The number of K<min{ M, N } PCs which cover most of
the variance of the observations are computed as follows. First, the
Dissertations in Forestry and Natural Sciences No 146
27
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
data correlation matrix R is estimated as:
R=
1
XX T
M
(4.1)
and eigenvectors (v1 . . . vK ) and eigenvalues (λ1 ...λK ) with constrain
||vk || = 1 ∀ k, are estimated using eigendecomposition:
Rvk = λk vk
(4.2)
where vk is an eigenvector related to k:th largest eigenvalue. Since
vk are eigenvectors of symmetric matrix (correlation matrix R) and
constrain of ||vk || = 1 ∀ k was set, vk :s are orthonormal and thus
the projection of vector xi to the eigenspace Hs = [v1 , ..., vK ] can be
computed as
θiPC = HsT xi
(4.3)
where θiPC contains number of k principal components (PCs) related
to observation xi . After this procedure, the vector of the first PCs
contains as much variance of the observations as possible, using
orthonormal basis vectors.
4.1.2
PCR model for electrocardiography
This section describes how PCR can be used to model ECG waveforms. The PCR method is based on ECG waveform extraction and
the data driven model is then constructed using extracted wave
epochs. T waves and QRS complexes are modeled separately for
every heart beat within the whole measurement. Here brief details
will be provided to explain how a single wave epoch, i.e. T or QRS
epoch from a single ECG lead can be modeled and in the next section, an expansion of the PCR approach for multilead measurement
or x, y and z components of the VECG will be proposed.
Before the ECG wave extraction and modeling, R-wave fiducial points must be detected in order to obtain a reliable starting
point for wave extraction. Numerous QRS-detection algorithms
have been published for R-peak detection [103–108]. In general,
QRS detection consists of two parts, preprocessing and decision
28
Dissertations in Forestry and Natural Sciences No 146
Methods
parts. The simplest preprocessing part can consist of only bandbass filtering with 5-30 Hz pass band which includes most of the
frequencies of QRS complex, but power line noise, muscle noise
and baseline wander will be reduced. In the decision part, rules
are applied to determine if the QRS complex has occurred or not.
Typically decision rules include adaptive amplitude threshold and
adaptive average RR-interval. In this thesis, QRS detector similar as [105] is used for R peak detection for a detailed description
see [109].
After QRS detection, T wave and QRS complex epochs are extracted from ECG according to the detected R wave. Since the width
of QRS-complex remains relatively constant regardless of RR interval, extraction can be done using the simple window
Repoch = [−100, 100] ms
(4.4)
where R wave are used as a zero point (t=0). The width and position
of the T wave varies greatly depending on the RR interval [110], and
therefore the average RR interval (RRav ) is used for defining T wave
extraction window
[100, 500] ms,
if RRav > 700 ms
Tepoch =
(4.5)
[100, 0.7 RRav ] ms, if RRav < 700 ms.
After the wave epochs have been extracted, the PCR model can
be constructed. The i:th extracted wave epoch with N data points
is denoted as
⎛
⎞
xi,1
⎜
⎟
xi = ⎝ ... ⎠ .
(4.6)
xi,N
As an observation model, an additive noise model is used
x i = s i + ei
(4.7)
where si is the noiseless wave epoch and ei is the measurement
noise. Each epoch xi can be approximated as a linear combination
of basis vectors vk
xi = Hs θi + ei
(4.8)
Dissertations in Forestry and Natural Sciences No 146
29
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
where Hs = (v1 . . . vK ) is N × K matrix of basis vectors and θi is a
K × 1 column vector of weights related to i:th epoch.
The reason why this ECG model is called the PCR based model
is related to estimation of basis vectors vk . Basis vectors are selected
to be the eigenvectors of the correlation matrix R as described in
section 4.1.1. This basis is optimal in the sense that eigenvectors
describe as much as possible of the variance of the original data,
under the constraint of orthogonal basis vectors. Model basis vectors are estimated for ECG wave epochs as follows. First M wave
epochs xi are collected into a measurement matrix
X = ( x1 . . . xM )
(4.9)
and a sample correlation matrix is calculated using equation 4.1.
Eigenvectors (i.e. basis vectors vk ) of this correlation matrix can be
solved from the eigendecomposition.
After these optimal basis vectors are estimated, the dimensionality of the data is reduced by computing an approximation for a
waveform by using only few most significant eigenvectors Hs =
(v1 . . . vK ). As described in section 4.1.1, the eigenvectors of the
correlation matrix are orthonormal and therefore the least-squares
PC
solution for the parameters θˆi is
θ̂iPC = HsT xi
(4.10)
and the estimate for wave epoch can be computed as
PC
x̂iPC = Hs θˆi .
(4.11)
In normal situations, the most significant basis vectors contain information about the waveform and its normal variation. Random
noise, on the other hand, is mainly distributed into less significant
eigenvectors and thus most of the noise can be removed by using
only the few most significant basis vectors in the wave model.
If the ECG to be analyzed is thought to contain significant HR
changes or when analyzing long measurements, PCR basis vectors
can and should be updated dynamically to ensure that these vectors are capable of modeling the real variation in the wave epochs.
30
Dissertations in Forestry and Natural Sciences No 146
Methods
For example in exercise ECG measurements, model basis vectors
should be calculated using M/2 previous and M/2 following wave
epochs.
4.1.3 Augmented PCR model for vectorcardiography
In the case of multilead ECG or VECG measurement, the individual leads can be modeled separately. However, a more sophisticated way is to add waveforms from all of the measured leads to a
measurement matrix (augmented measurement matrix) when estimating the PCR basis vectors, this type of PCR modeling is called
here augmented PCR [111]. Therefore, the number of wave epochs
used for basis vector estimation is increased significantly and thus
stronger prior information of the wave shape can be attained. Here,
the augmented PCR model construction is presented for VECG
measurement with x,y and z leads, however a similar construction
can be used also for other lead systems. First, wave epochs are extracted from all leads and the measurement matrix is constructed
as
⎛
x1,1
⎜ ..
X=⎝ .
y1,1
..
.
z1,1
..
.
x1,N
y1,N
z1,N
···
..
.
···
x M,1
..
.
y M,1
..
.
⎞
z M,1
.. ⎟ .
. ⎠
x M,N
y M,N
z M,N
(4.12)
In other words, the x, y and z components of the wave for the M
consecutive beats are placed into the columns of X. The correlation
matrix is then calculated using equation (4.1) and the model basis
vectors are solved using eigendecomposition. The most significant
basis vectors are fitted to x, y and z components using equations
(4.10) and (4.11). This procedure is done for every beat within the
measurement. When all x, y and z components of T waves and
QRS complexes are modeled, these modeled epochs can be used
estimation of the desired parameters.
Dissertations in Forestry and Natural Sciences No 146
31
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
4.2
TRADITIONAL ECG NOISE REDUCTION METHODS
Two commonly used ECG denoising methods, time domain filtering and Savitzky-Golay (SG) filtering will be shortly described below. Performance of PCR denoising algorithm is later compared
to these approaches to the reveal strengths and weaknesses of PCR
denoising.
Time domain filtering is doubtlessly the most widely used ECG
denoising method. In this thesis, ECG has been preprocessed with
an order 8 Butterworth filter with a cutoff frequency 150 Hz for
R-wave and 40 Hz for T-wave denoising as recommended in [112].
SG filtering is also a well known and widely used method in many
biomedical applications, including ECG denoising [113–116]. The
SG filter is a polynomial smoothing filter [117], where polynomial
functions of a selected order are fitted piecewise to the signal. In
this thesis, segment length and polynomial order for R-wave and Twave denoising were optimized separately using the SNR increase
as a quality measure. A segment length of 0.011 s and second order
polynomial were used for R-wave denoising and for T-wave denoising segment a length of 0.061 s and a second order polynomial were
used.
4.3
CARDIOVASCULAR TIME SERIES MODELING
The focus of this section is on cardiovascular time series modeling.
First autoregressive (AR) and MAR models are presented and in
the third subsection, a cardiovascular model construction will be
presented.
4.3.1
AR model
Time series models are generally used for forecasting, spectrum
estimation and in the characterization of measured time series. The
most commonly used time series models are AR, MAR, moving
average (MA) and ARMA models [118–121]. Here basic principles
32
Dissertations in Forestry and Natural Sciences No 146
Methods
of AR model will be described as will spectral estimation with the
AR model.
In AR modeling, time series (xt ) is supposed to be an order p
AR process, i.e. it has to satisfy the equation
p
xt = − ∑ a j xt− j + wt
(4.13)
j =1
where (a1 ...a p ) are model parameters and wt is a white noise process. The model parameters a j are estimated so that the prediction
error (residual)
p
t = x t + ∑ a j x t − j
(4.14)
j =1
is minimized.
There are a few commonly used methods for AR model parameter estimation. These are Yule-Walker, Least-Squares (LS) and expanded covariance method, for detailed description see [119, 120].
Here only the LS method is presented which is based on minimization of squared norm of model residual ||||2 . The solution is based
on one step prediction equations which are constructed for every
time point in the time series
p
xt = − ∑ a j xt− j + t , t = p + 1, ..., N
(4.15)
j =1
in a matrix form
⎛
⎜
⎜
⎜
⎜
⎝
x p +1
x p +2
..
.
xN
x
⎞
⎟
⎟
⎟
⎟
⎠
⎛
=
⎜
⎜
⎝
xp
..
.
x N −1
=
x p −1
..
.
x N −2
···
..
.
···
x1
..
.
xN− p
⎞
⎟
⎟
⎠
⎛
⎜
⎜
⎜
⎜
⎝
H
a1
a2
..
.
ap
a
⎞
⎟
⎟
⎟
⎟
⎠
⎛
+
+
⎜
⎜
⎜
⎜
⎝
1
2
..
.
⎞
⎟
⎟
⎟.
⎟
⎠
N− p
(4.16)
It can be readily seen that the least squares solution for parameters
is
(4.17)
â LS = ( H T H )−1 H T x
where â LS contains estimates for AR parameters a = [ a1 , ..., a p ]
Dissertations in Forestry and Natural Sciences No 146
33
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
As described earlier, the AR model can be used for signal spectrum estimation. The model based spectrum estimation is commonly used in biomedical signal analysis. With short signals, the
AR spectrum provides better frequency resolution than Fourier based methods. When conducting a spectrum estimation, the ARmodel is written in the form
p
∑ a j x t − j = t ,
j =0
t = p + 1, ..., N
(4.18)
and taking z-transform from both sides
X (z) A(z) = E(z)
(4.19)
where X (z) is the z-transform of signal xt , A(z) is the z-transform
of AR polynomial and E(z) is the z-transform of the white noise
process. The AR model can be thought of as a system in which the
input is the white noise process and the output is signal xt . The
system function H (z) between input and output is then in the form
H (z) =
1
A(z)
(4.20)
and the power spectrum of the system function is achieved by transforming it into a frequency domain i.e. substituting z = e−i2π f / f s
2
1
P( f ) = H (e−i2π f / f s ) = A(e−i2π f / f s )2
(4.21)
where f s is the sampling frequency. Power spectrum density of
signal xt can be then estimated
1
σe2 / f s
Px ( f ) = 2 Pe ( f ) = 2
A(e−i2π f / f s )
p
−
i2πj
f
/
f
s
1 + ∑ j =1 a j e
(4.22)
where σe2 / f s is the spectrum of the white noise process.
34
Dissertations in Forestry and Natural Sciences No 146
Methods
4.3.2 Multivariate AR model
In analysis of biological signals it is often advantageous if one can
compare multiple recordings which are differently affected by the
same oscillation source. AR modeling permits only a quantification
of a single signal whereas multivariate parametric models provide
further information of cross-spectral patterns and causal interactions among the signals. Here multivariate AR model in the case of
two signals will be shortly described, for a more detailed description see [119, 120, 122]
Bivariate AR process for signals x1 and x2 is defined as follows
p
p
x1 (t) = ∑k=1 a11 (k) x1 (t − k) + ∑k=1 a12 (k) x2 (t − k) + w1 (t)
p
p
x2 (t) = ∑k=1 a21 (k) x1 (t − k) + ∑k=1 a22 (k) x2 (t − k) + w2 (t)
(4.23)
where p is model order and w1 (t), w2 (t) are uncorrelated zero mean
white noise processes with variances σw2 1 and σw2 2 . Now the parameters a11 are AR coefficients for signal x1 and a22 are AR coefficients
for signal x2 , and a12 and a21 are cross coefficients for signals x1 and
x2 respectively.
MAR model parameter estimation can be done similarly as presented in AR-model situation. One step prediction equations are
first constructed using equation (4.23). If we use a similar matrix
notation as in equation (4.16) the set of MAR prediction equations
become
x1 = H1 a11 + H2 a12 + 1
x2 = H2 a22 + H1 a21 + 2
(4.24)
where H1 is the observation matrix constructed using signal x1 and
similarly H2 is constructed using x2 values. a11 is a column vector
AR coefficients for signal x1 and a12 cross coefficients, a22 and a21
are similar coefficients for signal x2 . In a matrix form this can be
Dissertations in Forestry and Natural Sciences No 146
35
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
written as
=
H1
x1 x2
x
=
H2
H
a11 a21
a12 a22
a
+
+
1 2
(4.25)
Model parameters can be then estimated by minimizing the squared
residual norm, equation (4.17), similarly as for the AR model.
In the estimation of spectral density functions, equation (4.23) is
transformed into frequency domain
X1 ( f )
W1 ( f )
A11 ( f ) A12 ( f )
=
(4.26)
A21 ( f ) A22 ( f )
X2 ( f )
W2 ( f )
A( f )
X( f )
W( f )
where X( f ) and W( f ) are frequency transformations of the examined AR processes and white noise processes respectively. Matrix
A( f ) is defined as follows
p
p
1 − ∑k=1 a11 (k)e−i2πk f / f s − ∑k=1 a12 (k)e−i2πk f / f s
A( f ) =
p
p
− ∑k=1 a21 (k)e−i2πk f / f s 1 − ∑k=1 a22 (k)e−i2πk f / f s
(4.27)
The power spectral density matrix can be then defined as
P( f ) = A∗ ( f )−1 Cw (A( f )−1 ) T
(4.28)
where Cw is the covariance matrix of the bivariate noise process,
where variances σw2 1 and σw2 2 are in the diagonal and off-diagonal
terms are zero because w1 and w2 are uncorrelated. Auto spectrum
functions are obtained from diagonal elements of P( f ) and the off
diagonal elements form the cross-spectral density functions.
4.3.3
Baroreflex sensitivity estimation using MAR-model
Linear models are often used in the characterization of the cardiovascular system. In BRS estimation, the RR interval (output) depends on SBP (input) through a linear transformation T. A complex transfer function of this linear system can be presented in the
36
Dissertations in Forestry and Natural Sciences No 146
Methods
frequency domain as
T( f ) =
P12 ( f )
P1 ( f )
(4.29)
where P1 is power spectral density of input process and P12 is cross
spectral density between the input and output processes evaluated
at frequency f . The gain information of the system is achieved
from the modulus and the phase information from argument of the
transfer function T.
If a linear model is used for the characterization of the cardiovascular system, it is important to examine whether the linear coupling between input and output is strong enough to ensure linear
model feasibility. The strength of the linear correlation between
input and output processes is referred to as coherence and it is defined as
| P12 ( f )|2
.
(4.30)
Coh( f ) =
P1 ( f ) P2 ( f )
In a cardiovascular analysis, BP and RR interval interact with
each others in a closed loop system, see figure 4.1. RR interval
directly affects BP (feedforward) and BP affects RR interval through
the baroreflex (feedback). If a cardiovascular system is modeled
using spectrums achieved directly from measured time series, the
resulting coherence function reflects the coupling of both causal
directions and the transfer function is a global index which includes
all the variation from both signals.
MAR model can be used to model cardiovascular system and
to separate variations caused by feedforward and feedback effects
of the cardiovascular loop [7, 8, 46, 123]. Let us assume that signal x1 is the RR interval time series and x2 is the SBP time series.
Then the bivariate AR model, block A12 describes the effect from
SBP to RR and block A21 the effect from RR to SBP see figure 4.1.
To resolve the biasing effect of feedforward variation, it has been
proposed that one can use causal coherence and transfer function
analysis [7,8]. The causal relationship from SBP to RR is simply obtained by forcing coefficients in the block A21 to zero, and analyzing transfer function and coherence from resulting system. Power
Dissertations in Forestry and Natural Sciences No 146
37
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
w1
A21
SBP
Feedforward
direct
A22
Blood
pressure
regulation
A11
HR
control
A12
Feedback
Baroreflex
RR
w2
Figure 4.1: Cardiovascular closed loop system and corresponding MAR model coefficients.
w1 and w2 are white noise processes (input) and A11 , A12 , A21 and A22 are model coefficients. SBP and RR are system outputs.
spectral densities are estimated according to equation (4.28), with a
modification of the coefficients, the spectral matrix becomes
∗ ( f ) σ2
1
(1 − A22 ( f ))2 σ12 + | A12 ( f )|2 σ22 (1 − A11 ( f )) A12
1
P( f ) =
∗ ( f ) σ2
2 σ2
b
(1 − A11 ( f )) A12
(
1
−
A
(
f
))
11
2
2
(4.31)
where b = |(1 − A11 ( f )) (1 − A22 ( f ))|2 . Consequently, causal coherence (CohSBP→RR ) is defined by substituting these modified auto
and cross spectral density functions into equation 4.30 and causal
transfer function (TSBP→RR ) is obtained similarly from equation (4.29)
and it is reduced into the form
TSBP→RR =
A12 ( f )
1 − A11 ( f )
(4.32)
from which it can be easily seen that variance of the input or output
signals has no effect on baroreceptor function.
Absolute value (gain) of this causal transfer function is used as
a measure of BRS, i.e. the sensitivity of baroreflex modulation to
38
Dissertations in Forestry and Natural Sciences No 146
Methods
HR. Although, because the linear model is used to estimate BRS,
only those frequency points where the linear coupling is significant
can be used for BRS estimation, i.e. coherence between RR and SBP
time series must be significant. The significance of linear coupling
can be tested for example by using a surrogate data approach [124].
In this method, individual time series are first modeled separately
using the AR model and AR coefficients are then used to reproduce
an ensemble of surrogate time series from pairs of independent
white noise. The coherences of these surrogate pairs are then estimated and the threshold for significant coherence is set at the 95
percentile (significance p<0.05) of the coherences sampling distribution. The mean value of transfer gain of those frequency points
where coherence value is above this level is taken as the BRS estimate.
The overall BRS estimated using traditional non-causal analysis
is marked in this thesis as BRSSBP,RR and the overall BRS estimated
using causal analysis is designated as BRSSBP→RR where SBP →
RR indicates that only variation from SBP to RR has been taken
into account in the BRS estimation. Neural BRS can be estimated
similarly as overall BRS, only SBP time series are replaced by the ds
time series. Similarly, neural BRS as estimated with the non-causal
approach is designated as BRSds ,RR and if the causal approach has
been used then the BRSds →RR notation is used.
4.3.4 Model validation
The parameter estimation procedure finds the best model within
the chosen model structure, however then the question arises of
whether this model is good enough to BRS-estimation. Firstly, the
model order must be selected and for this purpose several selection
criteria have been proposed [119, 120]. These criteria are derived
from information theory and are based on the relationship between
the prediction error and the complexity of the model. Probably the
best known example is Akaike information criterion (AIC), which
Dissertations in Forestry and Natural Sciences No 146
39
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
is defined for MAR models by minimizing:
AIC[ p] = Nln(det(C )) + 2p
(4.33)
where N is length of the signal, p is the model order and C is the
covariance of the prediction errors.
Model order selection criteria can be used as guidance for order selection, but the basic assumptions of MAR model must be
fulfilled. Thus whiteness and uncorrelatedness of prediction errors
1 and 2 must be tested. Prediction errors must be also independent of past input values, however when the LS method is used, the
parameters are estimated such that prediction errors are orthogonal with the input, so this assumption needs no testing [120]. The
whiteness of prediction errors can be tested for example by using
an autocorrelation test; if is white noise then its correlation is zero
(r (τ ) = 0) except for zero lag (τ = 0). Sample autocorrelation can
be estimated as
r (τ ) =
1
N
N −τ
∑
t =1
( t + τ ) ( t ).
(4.34)
For a normalized sample correlation function, it must hold
r (τ )
(4.35)
→ N (0, 1/N )
r (0)
√
hyand thus probability P(|ρ (τ )| ≤ 1.96/ N) = 0.05. The null √
pothesis (r (τ ) = 0) can therefore be accepted if |ρ (τ )| ≤ 1.96/ N
for all τ = 0. Uncorrelatedness of the prediction errors can be
tested with similar procedure, although the null hypothesis must
hold also for τ = 0, for more details see [119].
In practical applications, model order selection criteria are first
used to select a sufficient model order. Secondly prediction errors
are used to test the validity of the basic assumptions within the
model and if some assumptions are rejected then the model order will need to be increased. Thirdly, after all assumptions are
accepted the model must be validated using common sense, plots
from acquired estimates, spectrums and transfer functions can be
ρ (τ ) =
40
Dissertations in Forestry and Natural Sciences No 146
Methods
used to validate the model abilities to capture the desired information from the modelled systems.
Dissertations in Forestry and Natural Sciences No 146
41
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
42
Dissertations in Forestry and Natural Sciences No 146
5 Materials
Studies I-IV consist data from five experimental setups. However,
in this thesis, only the most significant results of these studies will
be described and thus only three larger experimental setups are described here. Two smaller datasets which were used to investigate
possible applications of PCR based ECG denoising (study I) will
not be presented here. This chapter concentrates on recorded signals, measurement protocols and the characteristics of participated
subjects. All of the performed studies were approved by the local
ethical committees.
5.1 ECG SIGNALS FOR NOISE SENSITIVITY ANALYSIS
In study I, the PCR denoising method for one lead ECG, and in
study III an augmented PCR method for VECG denoising were
tested and compared with two commonly used noise reduction
methods. In this thesis, similar testing has been conducted using
high SNR ECGs (ten different recordings, from physionet [125,126])
which after preprocessing are assumed to be noise free signals and
correct values of ECG parameters could be acquired. To test noise
enhancement properties, ”real life” noise collected from the exercise ECG measurements was multiplied by a selected factor to the
attain desired SNR and this was added to the noise free ECG. Then
different preprocessing methods are applied to noisy ECG/VECG
and so that the acquired parameters could be compared to the true
parameter values which were estimated from the noise free ECG.
5.2 HYPOGLYCEMIA MEASUREMENTS
In study II hypoglycemia related ECG changes were analyzed. The
aim was decrease BGC into the hypoglycemic range under controlled laboratory conditions by using the glucose clamp technique.
Dissertations in Forestry and Natural Sciences No 146
43
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
BGC was monitored at 5 minute intervals accompanied by using
simultaneous ECG measurement so that the changes in repolarization characteristics as a function of BGC could be analyzed. Measurements were recorded in Turku University Hospital and 27 subjects participated in the tests. The study protocol was approved by
the Ethics Committee of the Hospital District of Southwest Finland,
and each subject provided written informed consent. Continuous
measurements of ECG signal which were recorded using a modified chest lead V5 with a sampling rate of 128 Hz, along with the
BGC measurements at 5 minute intervals.
The subjects were divided into three groups: 9 healthy control subjects (Healthy group), 6 otherwise healthy type 1 diabetics
(T1DM group) who did not have any disease related complications
i.e. control diabetics group (duration of diabetes 3.2±2.3 years), and
7 type 1 diabetics (duration of diabetes 22.7±12.7 years) who had
any disease related complications and problems with hypoglycemia
unawareness (T1DMc group). None of the participants had history
of heart or cardiovascular diseases. Five subjects were excluded
from the analyses because their ECGs showed extremely low amplitude T-waves, and this would have prohibited reliable T-wave
detection.
During the measurement, the BGC was changed using the glucose clamp technique [127]. A steady rate of insulin infusion (1.5
mU/kg/min) was given to subjects and the desired blood glucose
level was adjusted by changing the glucose infusion rate. Initially,
the blood glucose value was adjusted into a range of 5-7 mmol/l as
the normoglycemic period. The normoglycemic period lasted for
approximately 85 minutes and during that time, the subjects performed some additional tests, which are not described here. After
the normoglycemic period, the glucose infusion rate was decreased
and BGC started to decline. During this transition period, the BGC
was 5 - 3.5 mmol/l and this period lasted for approximately 40
minutes. After the transition period, BGC were kept in the hypoglycemic range (the target BGC was below 3.0 mmol/l). The hypoglycemic period lasted for approximately 55 minutes, after which
44
Dissertations in Forestry and Natural Sciences No 146
Materials
the glucose infusion rate was increased and BGC was restored into
the normoglycemic range.
5.3 EXERCISE ECG MEASUREMENTS
In study III, the aim was to analyze VECG changes during exercise and in particular to study the relationship of VECG parameters
with exercise performance and BLC. For this purpose, six (25-37
years old) healthy male participants performed two different incremental dynamic tests on a cycloergometer. During the test with
VECG, exhaled gas and BLC were monitored. The study was approved by ethical committee of Kainuu.
The measurement protocols are presented in figure 5.1. Both
protocols consisted of two minutes of baseline measurements, followed by a three minute warm-up period (cycling 75W), a 21 minute
exercise period and finally a 15 minutes recovery period. The total work required within the two protocols was equal and the used
crank velocity was 60 rpm. The idea behind the two different protocols was to obtain different BLC levels with different HR profiles.
In the first protocol (P1), HR and BLC were assumed to change similarly whereas the second protocol (P2) was designed to evoke BLC
changes that would be more independent of HR, and thus to help
to differentiate between those variables that depend mainly on BLC
and are not dependent exclusively only on HR.
VECG was measured using Frank’s vectorcardiographic lead
system with a Schiller CS-200 ECG unit (Schiller, Switzerland) at
a sampling rate of 1000 Hz. Ventilatory variables were measured
continuously using the power cube gas analysis device (Ganshorn
Medizin Electronic, Germany) which was integrated into the Schiller
CS- 200 cycloergometer system. Blood samples (25 μl) were collected from the left hand index finger, for measurement of BLC
concentrations. During both protocols, lactate was measured eleven
times, first before the exercise (normal concentration), six times during the exercise period and four times during the recovery period
(see figure 5.1). BLCs were measured using an enzymatic electro-
Dissertations in Forestry and Natural Sciences No 146
45
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
Load (W)
200
Warm
up
P1
Exercise
Recovery
Load
lactate
150
100
50
0
Load (W)
200
P2
150
100
50
0
0
10
20
time (min)
30
40
Figure 5.1: Measurement protocols of incremental dynamic tests. Used load is presented
as solid fill and lactate measurement time points are presented at vertical lines.
chemical lactate analyzer (Ebioplus, Eppendorf).
5.4
CARDIOVASCULAR MEASUREMENTS
In study IV, the aim was to analyze the relationship between arterial stiffness and overall and neural components of the baroreflex.
In the estimation of overall and neural BRS; ECG, BP and carotid
artery ultrasound images were recorded in Kuopio University Hospital from all participants. The duration of the measurements was
2 minutes while subjects were lying in a supine position. The ultrasound scanner in use was an Acuson Sequoia 512 (Siemens, USA).
The longitudinal direction of the carotid artery (10 mm proximal to
the bifurcation) was visualized with B-mode ultrasound with a 14
MHz linear transducer. BP was measured continuously using the
volume clamp method and a Finapres device (Ohmeda Englewood,
CO). ECG and BP signals were recorded with an A/D-converter
and the ultrasound images were stored using video card. The sam-
46
Dissertations in Forestry and Natural Sciences No 146
Materials
pling frequency of ECG and BP signals was 1000 Hz and the video
signal frame rate was 25 fps.
A total of 28 healthy subjects and 15 subjects with type 1 diabetes participated in this study. Group characteristics of the healthy
group were, as follows age 20±1 (years, mean±sd), 22 females and
6 males. In the diabetic subjects mean age 38±11 years and 10
females and 5 males. The duration of diabetes was 18±9 (years,
mean±sd). Two of the healthy subjects were excluded from the
final analysis, because their respiratory rates were below 0.17 Hz.
Dissertations in Forestry and Natural Sciences No 146
47
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
48
Dissertations in Forestry and Natural Sciences No 146
6 Results
6.1 NOISE SENSITIVITY ANALYSIS
The denoising properties of PCR preprocessing were studied using ECG recording containing one lead in study I and using VECG
recording in study III. Here the larger dataset (ten two minutes
recordings, Physionet PTB Diagnostic ECG Database [125,126]) with
real life noise collected from exercise ECG recordings was utilized
to test ECG enhancement properties of PCR method, details of PCR
method have been presented in sections 4.1.2 and 4.1.3. Time domain filtering and SG filtering were used to compare the performance of the PCR method, a description has been given in section
4.2
SNR enhancement properties of time domain filtering, SG filtering and PCR denoising were compared by applying them on simulated noisy ECGs. Simulated noisy ECGs were constructed by using
ten different VECG recordings with high SNR which after preprocessing were assumed to be noise free signals. Real life VECG noise
was then collected from 10 different exercise ECG recordings from
epochs between T-offset and P-onset. The noise was then multiplied
by an appropriate coefficient in order to obtain the desired SNR and
this was added to the noise free signal. Figure 6.1 presents examples of noiseless ECG (10s segment), gathered noise and the noisy
ECG (SNR 15dB).
Noisy ECGs were then preprocessed and the SNR of obtained
signal was estimated by subtracting the original ECG from the preprocessed one and calculating the variances of the error signal and
the original noise-free ECG. In figure 6.2, the top subfigure depicts
the increase of SNR after preprocessing methods had been applied.
The error bars represent mean ± standard deviation of SNR increase of all 10 simulated ECG recordings presented as a function
of noisy ECG SNR. The performance of time domain filtering is illustrated with red error bars, SG filtering is presented with green
Dissertations in Forestry and Natural Sciences No 146
49
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
and PCR denoising performance with blue error bars. If the measured ECG SNR was between 5-15dB, PCR denoising increased its
SNR 4 - 8dB as an average, SG 2-4dB and time domain filtering by
1-2dB.
The effects of preprocessing on the ECG parameters were also
tested by estimating true parameter values from the noiseless ECG
and comparing those to the values estimated from preprocessed
noisy ECGs. Figure 6.3 shows five typical ECG parameter time series from one measurement, parameters are TCRT, RT-apex, RT-end,
T-amp and R-amp. The true parameter values are depicted with
the blue line and parameters estimated from preprocessed ECG are
shown a black line. The first column represents parameters estimated from the time domain filtered ECG, the second column pa-
ECG (μV)
2000
1000
0
ECG + noise (μV) noise (μV)
−1000
500
0
−500
2000
1000
0
−1000
0
2
4
6
8
10
time (s)
Figure 6.1: Examples of noiseless ECG (upmost subfigure), real life noise gathered from
exercise ECG measurement (second subfigure) and noisy ECG with SNR 15dB (third
subfigure).
50
Dissertations in Forestry and Natural Sciences No 146
Results
SNR increase (dB)
rameters from the SG filtered ECG and the third column parameters
from PCR preprocessed ECG.
Figure 6.4 illustrates the error of individual parameter values
(mean ± SD) calculated from all ten measured recordings. The errors are shown as a function of SNR. Parameter errors which are
estimated from ECG preprocessed with time domain filtering are
presented as a red line and errors from SG filtered ECG are illustrated as a green line and PCR preprocessed as a blue line.
10
Time D. Filtering
SG Filtering
PCR Denoising
5
Noisy ECG
Noiseless ECG
0
5
15
SNR (dB)
20
3
2
1
25
4
SNR 15dB
ECG (mV)
ECG (mV)
4
10
SNR 10dB
3
2
1
0
0
0
0.2
time (s)
0.4
0
0.2
time (s)
0.4
Figure 6.2: Performance of ECG preproccessing methods. Top subfigure shows the improvement of SNR after time domain fitering (black), SG filtering (red) and PCR denoising
(blue) as a function of input ECG SNR. In the lower subfigures two examples of noiseless
ECG (green), ECG with added noise (gray) and ECGs after preprocessing with a time
domain filtering (black), SG (red) and PCR (blue) are presented.
Dissertations in Forestry and Natural Sciences No 146
51
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
6.2
ELECTROCARDIOGRAPHIC CHANGES DURING HYPOGLYCEMIA
In study II, the PCR method was used to analyze repolarization
characteristics during insulin induced hypoglycemia. PCR preprocessing allows accurate beat-to-beat parameter estimation and thus
repolarization characteristics can be compared continuously to the
R−amp (μV) T−amp (μV)
RT−end (s)
RT−apex (s)
TCRT
Time Domain Filtering
SG Filtering
PCR Denoising
0.7
0.6
0.5
0.28
0.26
0.24
0.35
0.33
0.31
1200
1050
900
4100
3850
3600
0
50
100 0
time (s)
50
100 0
time (s)
50
100
time (s)
Figure 6.3: Five beat to beat ECG parameter time series (TCRT, RT-apex, RT-end, Tamp and R-amp), estimated from noiseless ECG (blue thick line) and from noisy ECG
after preprocessing (black) time domain filtering (first column of subfigures), SG filtering
(second column of subfigures) and PCR denoising (third column of subfigures)
52
Dissertations in Forestry and Natural Sciences No 146
Results
TCRT
Time Domain Filtering
Error
0.2
SG Filtering
0
PCR Denoising
−0.2
RTapex
RTend
0.02
Error (s)
Error (s)
0.02
0
−0.02
0
−0.02
−0.04
T−amp
R−amp
Error (μV)
Error (μV)
400
200
0
−200
400
200
0
−400
5
15
SNR
25
5
15
SNR
25
Figure 6.4: Error of beat to beat parameters as a function of input ECG SNR. Errors
are calculated for TCRT, RTend -time interval, RTapex -time interval, R-amp and T-amp.
Parameter errors related to time domain filtering are presented as black line, parameter
errors from SG preprocessed ECG are presented as red line and errors related to PCR
denoising are presented as blue line.
BGC. As explained in the materials section, the measurements were
performed in 9 healthy control subjects (Healthy group), 6 otherwise healthy type 1 diabetics (T1DM group) and 7 diabetics who
had disease related complications and problems with hypoglycemia
unawareness (T1DMc group). The parameters to be estimated were
QT-interval, RR interval, RT-ratio and HR corrected QT-time (QTcinterval) which was calculated by using Fredericia’s method [28].
Dissertations in Forestry and Natural Sciences No 146
53
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
ECG (mV)
1
0
−1
0
0.2
0.4
0.6
time (s)
0.8
Figure 6.5: Example of ECG during normoglycaemia (blue line) and during hypoglycemia
(black line). Detected positions of Q-onset, T-wave apex and T-wave end are marked as red
stars.
Each of these time series was transformed into an evenly sampled
time series (sample rate 4 Hz) using cubic spline interpolation and
low frequency trends were estimated by using a smoothness priors
method [128] with a cutoff frequency of 0.003Hz.
Figure 6.5 presents examples of the ECG during normoglycemia
(blue line) and during hypoglycemia (black line). The Q-wave onset and T-wave apexes and offsets are marked with red stars and
the tangents which were used to T-wave offset detection are also
presented. A ECG segments were selected such that RR interval
is close to equal in both segments, which permits a time interval
comparison without RR correction. The typical hypoglycemia induced ECG changes are clearly visible, i.e. T-wave flattening and
QT-interval prolongation.
Figure 6.6, shows representative time series of a single subject
from each group, the first column depicts healthy, second column
T1DM and third column T1DMc subject time series. From top to
bottom, the time series are BGC, QT time interval, RR time interval,
QTc and R-T amplitude ratio. From the time series of the healthy
subject, it can be easily observed, that as the glucose decreases below 3.5 mmol/l, QTc and RT-ratio starts to increase slowly. In addition the QTc prolongation continues as long as BGC remained be-
54
Dissertations in Forestry and Natural Sciences No 146
Results
low 3.5 mmol/l in the healthy subject. The T1DM subject exhibited
a clear decrease in RR interval in the hypoglycemic period although
only a minor increase of QTc and RT-ratio could be observed. The
variation of RR interval was smaller in T1DMc subject as compared
to the other two subjects. The hypoglycemia related changes in the
ECG were also less apparent in T1DMc subject.
The group results for the changes in BGC and changes in the
QTc interval, RT-ratio and RR interval values during the test protocol are presented in figure 6.7. Since the duration of the transition period (transition from the normoglycemic to hypoglycemic
state) varied extensively between subjects, the individual time series cannot be compared in a continuous time frame. Therefore,
two time instants were fixed for each individual measurement, i.e.
the times when normoglycemic period ended (BGC declined below
5 mmol/l) and when hypoglycemic period started (BGC fell below
3.5 mmol/l). These two fixed time instants are shown as zero times
in figure 6.7. The thin lines in the figure show the individual BGC
values and trends of ΔQTc interval, ΔRT-ratio and ΔRR interval series, whereas the thick lines show the mean ± SD of individual
values. Note that the parameter values (ΔQTc interval, ΔRT-ratio
and ΔRR interval) are presented as changes from baseline where
the mean of the normoglycemic period was taken as the baseline
value. This removed the baseline differences between subjects and
the changes of the parameters became more apparent.
There were no major HR changes during the test protocol in diabetic groups. Although HR changed in the healthy subjects during
the hypoglycemic period, no similar pattern between subjects was
observed (see figure 6.7). In the comparison of the group results
of QTc times, QTc progressive prolongation could be detected during the hypoglycemic period althoughBGC remained more or less
at the same level (3 mmol/l). Similar progressive changes in the
RT-ratio were observed in the Healthy and T1DM groups. Overall,
it was noticeable that the repolarization changes were most clearly
apparent in the Healthy group and smallest in the T1DMc group.
In addition, even though these repolarization changes for the hypo-
Dissertations in Forestry and Natural Sciences No 146
55
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
QT (ms) BGC (mmol/l)
Healthy
8
>5 mmol/l
T1DM
<3.5 mmol/l
>5 mmol/l
T1DMc
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
6
4
2
400
360
RR (s)
320
1.1
0.9
RT ratio
QTc (ms)
0.7
400
360
320
3
2
1
0
50 100 150
time (min)
0
50 100 1500
time (min)
50 100 150
time (min)
Figure 6.6: Representative time series of one subject from each group: a healthy subject
(left column), subject from T1DM group (middle column) and subject from T1DMc group
(right column). Measured blood glucose values are shown in the uppermost axes and
underneath the glucose values (from top to bottom) are QT interval, RR interval, heart rate
corrected QT interval (QTc), and RT ratio time series. Original time series are depicted by
gray lines and trends of each time series as a black line.
glycemia (BGC <3.5mmol/l) varied between individuals, in most
subjects one could observe a lag time of a few minutes between the
start of hypoglycemia and appearance of repolarization changes.
56
Dissertations in Forestry and Natural Sciences No 146
Results
ΔQTc (ms)
40
ΔRT ratio
BGC (mmol/l)
Healthy
7
6
5
4
3
4
T1DMc
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
>5 mmol/l
<3.5 mmol/l
20
0
2
0
0.2
ΔRR (s)
T1DM
0
−0.2
−40 0 ... 0 40 80 −40 0 ... 0 40
time (min)
time (min)
−40
0 ... 0 40
time (min)
Figure 6.7: Group results of changes in blood glucose, ΔQTc interval, ΔRT ratio and ΔRR
interval during the measurement. Subject-specific estimates are presented by thin lines
and group-specific mean±SD values by thick lines.
6.3 ECG CHANGES DURING EXERCISE AND PERFORMANCE
ESTIMATION
In study III, VECG parameter changes during two different exercise
protocols were studied. As mentioned in the exercise physiology
section (2.6), earlier studies have shown that some ECG/VECG parameters are not completely related to HR and the parameter values
Dissertations in Forestry and Natural Sciences No 146
57
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
differ between exercise and recovery periods. It was hypothesized
that these ECG/VECG changes could be used for performance estimation in addition to HR. Thus, in study III, the VECG parameters were compared to traditional performance estimation markers
such as BLC, RR and carbon dioxide oxygen relation (CO2/O2).
The measurement protocols and measured signals have been described in section 5.3. Before parameter estimation, VECGs were
preprocessed using the PCR method. The following VECG parameters were estimated; TCRT (described in section 2.3.2), T-wave morphology parameter T-kurtosis which was estimated from a ”virtual”
lead with a direction along the electrical axis of the heart. Depolarization and repolarization loops were also be parameterized using
the cosines of the loop main vector angles as follows; θ R was defined as the angle between the horizontal projection of R-wave main
vector and positive x-axis, and φR as the angle between sagittal projection of R-wave main vector and positive z-axis. In addition, φT
and θ T are the main vector angles of the T-wave loop which were
defined similarly. The beat-to-beat variation of the parameters was
large and in order to achieve better comparability, low frequency
trend components for these parameters were estimated using the
smoothness priors method [128] with a cutoff frequency 0.01Hz.
In addition to VECG parameters, the minute ventilation (VE) and
CO2/O2 ratio were estimated from the spirometric measurements
and HRV parameters including mean RR interval, standard deviation of normal to normal RR intervals (SDNN), LF band power and
HF band power were estimated using a 3-minute moving window.
Figure 6.8 shows parameter trend changes occurring during the
protocols. A visual inspection of figure 6.8 reveals that RR interval and VE changed simultaneously with bicycle load and returned
rather rapidly towards baseline values after the exercise period had
ended, which was expected. Similarly BLC, CO2/O2-ratio and Tkurtosis followed the exercise load, but after exercise period these
parameters returned to baseline more slowly than the other parameters. Individual variation in the VECG loop angles was larger than
other parameters, however, similarities between participants could
58
Dissertations in Forestry and Natural Sciences No 146
Results
BLC (mmol/l)
RR (s)
0.8
0.6
0.4
100
VE (l/min)
1.1
CO2/O2
12
10
8
6
4
2
0.9
0.7
80
60
40
20
−3
0
8
TCRT
T−kurtosis
× 10
4
−0.1
−0.2
0
−0.3
0.2
0.1
φT
θT
0
−0.2
−0.4
−0.1
−0.6
0
20
35 0
time (min)
20
35
time (min)
0
20
35 0
time (min)
20
35
time (min)
Figure 6.8: Blood lactate concentration and trends of estimated parameters for both protocols, for VECG parameters resting value is set to zero for better comparability. Blue area
describes the used load (see figure 5.1 for more details). Presented parameters are RR interval, BLC, oxygen and carbon dioxide relationship (CO2 /O2 ), minute ventilation (VE),
T-kurtosis, TCRT, cosine of the T-wave angle to the xy plane (φT ) and cosine of the T-wave
angle to the zy plane (θ T ).
be observed. During exercise, the TCRT angle was smaller than at
baseline in all participants and it seemed to return rather slowly
towards to baseline. Similarly, a few minutes after the start of the
recovery period, θ T was smaller than at baseline in all participants
and it also reverted slowly towards its baseline value.
The linear correlations between the parameter trends presented
in figure 6.8 and BLC at different time lags (from the -240 to 240
Dissertations in Forestry and Natural Sciences No 146
59
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
seconds) were compared to reveal delays in the parameter changes.
The correlations at the different time lags are presented in the first
and second rows of figure 6.9. Individual correlations (for each
participant) as a function of time lag are shown as thin blue lines
and the averaged correlations are presented as thick black lines.
The time where maximum correlation was achieved is shown a as
black vertical line. In the third and fourth rows, parameters are
presented as a function of BLC, adopting the same lag for each
individual (the lag which produced the largest group correlation).
Individual correlation lines are designated with as thin gray lines
and the overall correlation (estimated from all measurements of all
participants) is illustrated as a thick dashed line. In addition, the
overall correlation coefficient is presented inside each subfigure.
Table 6.1 lists the correlations between the different parameters
and BLC, RR separately for protocols P1 and P2. For both BLC and
RR, the first column indicates the time lag giving the best group
correlation and the second column shows the observed group correlation (estimated directly from all BLC vs. parameter values; 11
BLC measurements 6 participants = 66 points). The third column
depicts the mean of the individual correlations, the fourth column
is the maximum with the fifth column being the minimum of the
individual correlations.
From figure 6.8 and table 6.1 one can be observe that BLC changes were delayed for 1-2 minutes when compared to RR interval,
CO2/ O2 relation, T-kurtosis and TCRT for all participants in both
protocols. VECG parameter inspection reveals that mean individual correlations were highest for T-kurtosis and TCRT, the highest correlation was achieved for negative time lag similarly as for
mean RR. TCRT and T-kurtosis also displayed high correlations
with mean RR. Significant correlations between φT , θ T and BLC
was also found, but the biggest correlation were achieved for positive time lags, which indicates that changes are delayed even when
compared to BLC. Even though these parameters did correlate with
BLC, the correlations related to RR changes were low.
Finally, a stepwise linear regression routine was applied to re-
60
Dissertations in Forestry and Natural Sciences No 146
0.8
0.6
0.4
0.2
0
r (BLC vs φT)
r (BLC vs TCRT)
0
−0.2
−0.4
−0.6
−0.8
0.4
0.2
0
0.5
0
−200
0
lag (s)
200
−200
CO2/O2
0.6
0.5
0
lag (s)
0.4
0.9
0.8
−0.2
r =−0.39
0.2
0.2
0
0
−0.2
θT
φT
T−kurtosis
0.1
5
10
BLC (mmol/l)
r =−0.60
−0.4
r = 0.66
0
−0.1
r = 0.53
−0.1
0
200
0
8
4
0
lag (s)
1
0.7
× 103
0
−200
200
1.1
r =−0.60
0.5
−0.5
−0.5
0.7
RR
0.8
0.6
TCRT
r (BLC vs T−kurtosis)
r (BLC vs RR)
0.2
0
−0.2
−0.4
−0.6
−0.8
r (BLC vs θT)
r (BLC vs CO2/O2)
Results
r = 0.56
0
5
10
BLC (mmol/l)
−0.6
0
5
10
BLC (mmol/l)
Figure 6.9: Parameter correlations with BLC estimated using lags between [-240 240]s
are presented in the first and second rows. Individual correlations are presented as blue
thin line and mean correlation ± standard deviation is a thick blue line. Correlation
plots from adopting the timelag which produced the best correlation are presented in the
third and fourth rows. Subjects’ individual correlations are depicted as thin blue lines and
correlation estimated from all the subjects’ BLC vs parameter values are presented as a
thick dashed line. Gray asterisks represent all the individual parameter values (22 values
per subject from two protocols).
veal if VECG parameters could provide information of physiological performance in addition of HR and HRV parameters. Linear
regression analysis for the associations between the different VECG
Dissertations in Forestry and Natural Sciences No 146
61
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
Table 6.1: Parameter correlations with BLC and mean RR-interval for protocols P1 and P2.
Lag defines the time where the best correlation is found and group, individual, maximum
and minimum correlations are estimated using this lag. Group correlation is estimated
directly from all BLC vs. parameter measurements (11 measurements × 6 subjects =
66 points, significance of correlation: ∗ p<0.01, ∗∗ p<0.001) and individual correlation is
mean of each subjects’ individual lactate vs. parameter correlation.
parameter
and protocol
RR (s)
CO2 /O2
TCRT
T-kurtosis
φT
θT
parameter
and protocol
RR (s)
CO2 /O2
TCRT
T-kurtosis
φT
θT
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
P1
P2
Lactate
lag (s)
-97
-91
-126
-176
-81
-37
-102
-51
189
203
194
190
RR
lag (s)
0
0
-112
-17
-52
-195
-81
-47
235
190
-68
-188
r(goup)
-0.65∗∗
-0.59∗∗
0.69∗∗
0.61∗∗
-0.44∗∗
-0.38∗
0.68∗∗
0.76∗∗
0.54∗∗
0.60∗∗
-0.54∗∗
-0.76∗∗
r(individ.)
-0.87
-0.76
0.75
0.76
-0.70
-0.62
0.86
0.80
0.39
0.54
-0.48
-0.53
max
-0.79
-0.64
0.88
0.83
-0.52
-0.23
0.92
0.88
0.70
0.85
0.05
0.34
min
-0.93
-0.84
0.59
0.67
-0.97
-0.88
0.77
0.68
-0.14
-0.14
-0.82
-0.91
r(group)
1.00
1.00
-0.82∗∗
-0.75∗∗
0.57∗∗
0.61∗∗
-0.81∗∗
-0.78∗∗
-0.02
-0.11
0.13
0.20
r(indivi.)
1.00
1.00
-0.87
-0.74
0.68
0.73
-0.88
-0.86
0.17
0.23
0.14
0.16
max
1.00
1.00
-0.81
-0.5
0.84
0.88
-0.70
-0.69
0.91
0.89
0.92
0.81
min
1.00
1.00
-0.96
-0.89
0.32
0.31
-0.96
-0.95
-0.92
-0.28
-0.72
-0.82
parameters and BLC, were applied as follows. BLC was used as a
dependent variable and the VECG parameters as explanatory variables. In each step of stepwise linear regression procedure, the explanatory variable which improved the model most was added to
the model until the improvement was statistically significant. Lin-
62
Dissertations in Forestry and Natural Sciences No 146
Results
ear regression results were calculated from the 132 data points (11
measurement values per protocol, 2 protocols per subject and 6 subjects).
According to the Stepwise linear regression procedure, the best
predictor for BLC was T-wave kurtosis and the best linear regression model for BLC concentration was attained by using T-kurtosis,
θ T , φT and SDNN as descriptive variables. Secondly, the mean RR
interval was taken as the first descriptive variable to reveal which
of the parameters could provide additional information to complement heart rate. When RR interval was selected into the model,
the predictive accuracy of the model was improved by adding φT ,
T-kurtosis and θ T parameters into the model.
6.4 MODELING OF ARTERIAL BAROREFLEX
In study IV overall and neural BRS were studied in relation to arterial elasticity. BRS estimation were made using the MAR-model
presented in methods section 4.3.3. The MAR-model enables separation of feedforward and feedback effects between time series measured from cardiovascular system, i.e. RR interval, BP and carotid
artery diameter. As mentioned in section 5.4, carotid artery ultrasound video, BP and ECG were measured simultaneously from
28 healthy subjects with a high probability to have normal BRS
(healthy group) and 15 old subjects with type 1 diabetes who were
very likely to have reduced arterial stiffness and reduced overall
BRS. The advanced BRS estimation methods combined with versatile subject group provided a good opportunity to study the relationship between arterial stiffness and BRS without any biasing
factors.
R-wave fiducial points were first detected from the ECG using
a QRS detection algorithm and secondly beat-to-beat SBP and DBP
values were detected from BP signal. From the ultrasound images
of carotid artery, arterial diameter was detected frame-by-frame using neural network edge detection algorithm similar in a previous
report [129]. Maximum diameter time series (ds ) and minimum
Dissertations in Forestry and Natural Sciences No 146
63
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
diameter time series (dd ) were constructed by detecting maximum
and minimum diameter values from each beat segment, similarly
as for blood pressure. Arterial elasticity was estimated as a mean
value of beat-to-beat β indexes (see equation (2.4)).
In the overall BRS estimation, a two dimensional MAR-model
was constructed using RR and SBP time series as described in section 4.3.2. Similarly for estimation of neural BRS, RR and ds time
series were used to construct the MAR-model. Figure 6.10 shows
RR, SBP and ds time series and spectrums of one representative
cases from diabetic and healthy groups. In addition, coherences
and transfer function gains are also presented at the bottom of figure 6.10. The constructed time series and MAR-spectrums reveal
that even though mean RR interval was similar in both subjects, RR
variation was slightly higher in the healthy subject. In contrast, the
BP changes observed in diabetic subject were higher, especially at
low-frequency band (0.04-0.1 Hz). Transfer function gains and coherences were estimated using the non-causal and causal approach
to reveal any bias caused by feedforward variation to these estimates, see figure 6.10. The coherence value estimated using causal
method (dashed line)was reduced as compared to non-causal coherence (solid line). The transfer function gain estimates on the
other hand did not reveal as clear changes between causal and noncausal methods.
In figure 6.11, shows a comparison between the traditional BRS
estimation method and the causal estimation with the Bland-Altman
plot. The estimation methods for overall BRS estimation are compared in the upper subfigure and methods for neural BRS estimation in the lower subfigure. With low BRS values, both methods
gave comparable results, however, when there were high overall
BRS values, the traditional method seemed to produce lower values than the causal method. When comparing the neural BRS estimates, in most cases the traditional method overestimated the BRS
gain, however there were a few subjects in whom the traditional
method gave smaller values than the causal method.
In figure 6.12, neural and overall BRS, estimated using causal
64
Dissertations in Forestry and Natural Sciences No 146
Results
Healthy
Diabetic
10
PSBP (mmHg2/Hz) PRR (s2/Hz)
1.2
1
0.8
150
(mm /Hz)
6.3
6.1
5.9
5.7
0
0
× 10−3
Pd
s
2
6
100
CohRR,d
s
1
50
time (s)
0.5
0
3
0
0
0.5
f (Hz)
1
0.5
0
× 10−3
2
s
40
TRR,d
CohRR,SBP
0
TRR,SBP
5
10
130
20
0
× 10−3
20
140
ds (mm)
SBP (mmHg)
RR (s)
1.4
0
f (Hz)
0.5
1
0
0
f (Hz)
0.5
Figure 6.10: Measured time series and estimated spectra, coherences and transfer function
gains, for one healthy (blue line) and one diabetic (black line) subject. The top three rows
of the subfigures represent RR-interval (RR), systolic blood pressure (SBP) time series and
maximum diameter (ds ) timeseries and their spectral density functions (PRR , PSBP and Pds
). Spectras are estimated using bivariate AR model. The two bottom rows of subfigures
represents coherences (Coh RR,SBP ) and transfer function gains (HRR,SBP ) between SBP
and RR and similarly between ds and RR (Coh RR,ds and HRR,ds ). Traditional coherences
and transfer function gains are presented as solid lines and their causal counterparts as
dashed lines.
and non-causal methods are presented as a function of arterial elasticity (β). As expected, overall BRS and arterial elasticity are reduced in diabetic subjects. Neural BRS was also reduced in dia-
Dissertations in Forestry and Natural Sciences No 146
65
20
(ms/mmHg)
BRSSBP,RR − BRSSBP → RR
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
0
−20
−40
0
10
→ RR
1
(ms/mm)
s
s
BRSd ,RR − BRSd
Healthy
Diabetic
20
30
Average BRSSBP (ms/mmHg)
40
0.5
0
−0.5
−1
0.2
0.4
0.6
0.8
1
1.2
Average BRSd (ms/mm)
1.4
1.6
Figure 6.11: Bland-Altman plot for differences between traditional and causal BRS estimates. Upper subfigure displays differences between overall BRS estimates and in the
lower subfigure the neural BRS estimates are compared. In both subfigures, BRS values of
diabetic subjects are marked with circles and BRS of healthy subjects with stars.
betic subjects however not as severely as overall BRS. It is interesting however that for three diabetic subjects, the overall BRS was
reduced but the neural BRS was close to the mean value of the
healthy group.
66
Dissertations in Forestry and Natural Sciences No 146
Results
Diabetic
2
BRSd ,RR (ms/mm)
40
0
1
s
20
2
6 β
9
0
12
2
6 β
9
12
2
6 β
9
12
(ms/mm)
2
→ RR
40
0
1
s
20
BRSd
BRSSBP → RR (ms/mmHg)
BRSSBP,RR(ms/mmHg)
Healthy
2
6 β
9
12
0
Figure 6.12: BRS estimates of healthy (stars) and diabetic subjects (circles) as a function
of artery elasticity. Upper subfigures display overall BRS (left) and neural BRS (right)
estimated using the traditional approach and the lower subfigures show the overall BRS
and neural BRS estimated with the causal method.
Dissertations in Forestry and Natural Sciences No 146
67
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
68
Dissertations in Forestry and Natural Sciences No 146
7 Discussion and conclusion
Accurate methods for cardiovascular signal analysis, parameter estimation and modeling can reveal diseases, disease complications or
other unwanted events in the human body. In this thesis, methods
for cardiovascular signal analysis, in particular for accurate ECG
parameter estimation and BRS, analysis were developed and validated. Developed ECG analysis method were used to study cardiac
repolarization changes during insulin induced hypoglycemia (studies I and II) and VECG changes during exercise (study III). It was
hypothesized that hypoglycemia and increased blood lactate concentration could induce detectable changes to cardiac repolarization parameters and thus that some of those parameters could be
used for detection of decreased blood glucose or increased lactate
concentrations. In study IV multivariate AR modeling was used
to study the relations between neural BRS, overall BRS and arterial elasticity. It was hypothesized that neural BRS estimated using
causal estimation method would describe better the neural segment
of BRS function than the traditionally used BRS estimates.
7.1 ECG PREPROCESSING
ECG and VECG parameters are important tools for clinicians for
diagnostic purposes, but, in the future, similar parameters could be
used for monitoring human beings during their daily lives. For example these parameters could be used for preventing hypoglycemic
events or for monitoring performance during exercise. However,
the accurate estimation of different ECG parameters is problematic
during the conditions when an individual is exercising or under
non-laboratory measurements, where SNR is low. In this thesis, a
PCR based method for single lead ECG preprocessing was developed (study I) and the method was expanded to multilead ECG
purposes in study III. The PCR method is based on a model for
Dissertations in Forestry and Natural Sciences No 146
69
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
the QRS-complex and the T-wave which is constructed using prior
information of either earlier T-waves or T-waves in the whole measurement. Because of the prior information, the method is robust
for noise, and thus, suitable for analyzing low SNR ECG recordings
acquired in different nonlaboratory environments.
The PCR method has been successfully used in many kind of
signal processing applications, for separating signals from different sources, i.e. differentiating maternal and fetal ECG [130], for
cardiovascular diagnostics [131] and for diagnostics of Parkinson
disease [132]. However, this is the first time when PCR has been
used for modeling ECG wave forms and improve the accuracy of
estimated ECG parameters. In study I, the performance of the PCR
method to denoise normally distributed random noise from a single
lead ECG was tested, and in study III, VECG denoising properties
were tested using one VECG measurement with real life noise. Section 6.1 of this thesis provides a more general validation of PCR
method because testing was conducted using ten individual measurements and real life noise collected from exercise ECG measurement.
More specifically, the performance of the PCR method was tested
using ten high SNR VECG recordings, the SNR of which were decreased by adding real life noise collected from exercise VECG measurements. It was observed that PCR achieved comparable results
with SG filtering and time domain filtering in high SNRs, but the
performance was better with low SNR situations, as shown in figures 6.2 and 6.4. PCR increased SNR approximately 8dB in low SNR
situations which was better than the performance of SG-filtering
(4dB) or time domain filtering (2dB). Parameters estimated from
ECG preprocessed with PCR method were also more accurate than
parameters estimated from SG or time domain filtered ECGs. For
example, from figure 6.4 can be seen that errors of RT-apex estimated from ECG with 15dB SNR were after PCR preprocessing:
0.5±6.2 ms (mean ± 2 × SD), SG-filtering 1.0±10.9 ms and after
time domain filtering 0.5 ± 13.6 ms.
The limitations of the PCR method are related to large and sud-
70
Dissertations in Forestry and Natural Sciences No 146
Discussion and conclusion
den ECG morphological changes, if morphology of QRS-complex
or T-wave changes abruptly then the PCR basis vectors are not able
to model these changed waveforms. This problem can be mitigated
by updating PCR base vectors dynamically during the measurement. However even the dynamic updating does not invariably
solve this problem, for example in subjects who have cardiogenic
disease, the morphology of the ECG waves can change quite suddenly.
In conclusion the noise simulations conducted in this thesis reveal that performance of PCR noise reduction method is better than
traditionally used methods. Thereby, the PCR approach can be recommended for use when estimating ECG or VECG parameters in
low SNR situations, for example during exercise or measurements
made outside the laboratory.
7.2 ECG CHANGES DURING HYPOGLYCEMIA
As explained in section 2.5, hypoglycemia may result in seizures or
cognitive problems which can be dangerous in specific conditions,
for example when driving a car [133]. It has also been hypothesized
that the abnormal cardiac repolarization characteristics observed in
hypoglycemia would be connected to the so-called dead-in-bed syndrome [134–136]. Therefore, changes in repolarization affected by
hypoglycemic event have been widely studied and some attempts
have even been made to predict hypoglycemic events based on ECG
measurement [14, 15].
Study II was a dynamic analysis of ECG repolarization characteristics during hypoglycemia. In many of the previous studies, only diabetic and/or healthy test subjects have been considered
without taking any account of the duration of diabetes. Although
it is known that the ANS response to hypoglycemic events is lower
in diabetics who have suffered the disease for a longer period as
compared to patients with shorter disease histories or to healthy
subjects. In study II, healthy and two groups of diabetic subjects
(divided by disease duration) were studied.
Dissertations in Forestry and Natural Sciences No 146
71
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
Many of the earlier studies have also utilized normoglycemic
and hypoglycemic clamps but studied them separately, however
that kind of set-up does not reveal the dynamics of cardiac repolarization changes during changes of BGC. In study II, ECG repolarization characteristics were estimated using the PCR method
which conferred advantage in comparison to traditional methods,
because of its better noise reduction. Thereby, all parameters could
be defined in a beat-to-beat basis and parameter changes during
the whole measurement could be tracked. The prolongation of QTc
interval occured during hypoglycemia in all groups (see figure 6.7),
which was in agreement with previous studies [13, 62, 70]. However, it was also noticed that there was progressive prolongation of
the QTc interval when BGC remained at a low level, this could be
observed in Healthy and T1DM groups as shown in figure 6.7. In
addition, the ECG changes related to hypoglycemia that were detected in the T1DMc group were smaller than the corresponding
changes in Healthy or T1DM groups. For example, the RT-ratio increased slightly in Healthy and T1DM groups, whereas in T1DMc
group no such change was observed. The findings of study II support the hypothesis that diabetics with a long history of the disease
might have become more habituated to hypoglycemia and their autonomic and/or hormonal responses to hypoglycemia would be impaired.
Recent studies have shown that there are three probable reasons
to explain how hypoglycemia changes heart electrical activity. The
first putative cause is hyperinsulinemia induced by hypokalemia.
As described in section 2.2, potassium ions have significant role
on the activation of cardiac cells and thus hypokalemia could affect cardiac repolarization. Hypokalemia also develops together
with hypoglycemia, and therefore, the ECG changes evoked by hypokalemia should be coincident with hypoglycemia [13]. The second possible source is that hypoglycemia can influence the neural
regulation which would then have a clear effect on cardiac activity [66]. For example, there was an increase in parasympathetic
activation observed during hypoglycemia. Neural regulation of car-
72
Dissertations in Forestry and Natural Sciences No 146
Discussion and conclusion
diac activity is rather rapid, and thus, the changes in ECG should
tend to be coincident with hypoglycemia. The third probable cause
is that hypoglycemia affects hormonal secretion (increase in the release of glucagon, growth hormone, adrenaline and noradrenaline).
For example adrenaline is known to exert significant effects on
cardiac cell activity [36]. The ECG changes caused by hormonal
effects would occur with a delay which could explain the time
lag observed between hypoglycemia and the cardiac repolarization
changes found in study II. It is also known diabetics with a long
disease duration experience reduced secretion of adrenaline and
glucagon during hypoglycemic events as compared to healthy or
T1DM subjects [137]. This could explain why the cardiac repolarization changes observed in study II were also less marked in
diabetics with a long history of diabetes.
One interesting reason for analyzing ECG changes during hypoglycemia is the possibility to detect hypoglycemic events directly
from ECG recordings. An ECG based hypoglycemia detector which
could display an alarm should the patient’s BGC fall below a certain
level and this could prevent the patient from suffering the symptoms of hypoglycemia. However, study II revealed a few major
challenges that would need to be overcome before such a system
could be developed. The first challenge is that cardiac repolarization changes seem to appear with a delay as compared to hypoglycemia itself. Secondly diabetics seem to habituate to hypoglycemic events and thus changes become less marked as a function of disease progression. The extensive individual differences
related to ECG changes occurring during hypoglycemia could be
also problematic. Furthermore, it should be noted that any anomalies in ECG waveforms such as significant changes in T-wave morphology (e.g. partly inverted or fully inverted T-wave) cause also
significant problems for such a detection algorithm, and thus, these
anomalies would need to be detected and handled appropriately.
However, the results of study II show that if the BGC is extremely
low (<2 mmol/l), the resulting ECG changes are so significant that
a hypoglycemia could be detected at least for diabetics with sort
Dissertations in Forestry and Natural Sciences No 146
73
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
disease history.
The biggest limitations of study II were the relatively small
group sizes (9 Healthy, 6 T1DM and 7 T1DMc subjects). However,
the inclusion of both diabetics with disease related complications
and otherwise healthy control diabetics was a clear strength of the
study. It would also be interesting to examine how repolarization
characteristics change when BGC started to be restored after hypoglycemia. Unfortunately study II this part could not be analyzed
because the measurements were terminated after the hypoglycemic
period. Another limitation of the study is that the blood potassium
concentration was not measured during the tests. Potassium can affect cardiac repolarization, and thus, one cannot fully evaluate the
reasons behind the changes noted in the repolarization characteristics.
In conclusion results of the study II revealed that ECG based
hypoglycemia detection is possible at least for diabetics with short
history of diabetes. Results revealed that diabetics with long history of disease are more habituated to hypoglycemic changes than
diabetics with short disease history. In addition results indicated
also that ECG changes seem to appear with a delay as compared
BGC changes.
7.3
VECG PARAMETERS DURING EXERCISE
Exercise performance estimation is an important tool for athletes
and sports since it allows optimization of training intensities. Nowadays there are practical applications of HR for monitoring exercise performance. Previous studies have shown that some VECG
parameters are not fully related to HR and the values differ between the exercise and recovery periods [16–18]. ANS or an exercise induced increase in the level of analytes has been thought to
change the action potential durations and to be reason for these
changes [16, 36]. In study III, it was hypothesized that more information from physiological performance could be attained by monitoring the depolarization and repolarization phase of heart muscle
74
Dissertations in Forestry and Natural Sciences No 146
Discussion and conclusion
via VECG measurements, instead of focusing simply on HR analysis.
In study III, repolarization and depolarization waveforms were
parameterized using φT , θ T , φR and θ R which present orientations
of VECG loops and TCRT which describes spatial deviation between repolarization and depolarization wave-fronts. These VECG
parameters were compared to different exercise intensity markers
including RR, ventilatory parameters (minute ventilation and carbon dioxide oxide ratio) and in particular to BLC. The aim of the
study was to evaluate if different VECG parameters could provide
more information, above what could be obtained via HRV parameters, on physiological performance during the exercise. Therefore,
correlations were calculated between VECG parameters, BLC and
ventilatory parameters during exercise and recovery periods. Two
cycloergometer protocols were designed to trigger changes in the
BLC levels but with different HR profiles in order to be able to identify those parameters actually correlating with BLC irrespective of
HR.
The VECG parameter correlations with BLC were examined using different time lags as shown in figure 6.9. It was observed that
BLC changes were delayed for 1-2 minutes when compared to RR
interval, T-kurtosis and TCRT in both protocols. During bicycle exercise, lactate is produced in the active muscles (mainly in the large
leg muscles). When one considers the blood volume of the vascular
system, it is clear that it takes some time before any change of BLC
can be measured in blood taken from the fingertip. During the exercise and recovery period kidney, liver, heart and inactive muscles
are organs which slowly metabolize and thus remove lactate from
the blood [138, 139].
The results of study III (see figure 6.9 and table 6.1) revealed
that the with respect of the measured parameters T-kurtosis correlation was highest for BLC with protocol P1, whereas in protocol
P2 the highest correlation with BLC was found from the VECG parameter θ T . The mean individual correlations to BLC were also high
for mean RR interval, TCRT and T-kurtosis and the highest corre-
Dissertations in Forestry and Natural Sciences No 146
75
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
lation was achieved using a negative time lag for all parameters.
TCRT and T-kurtosis also displayed high correlations with mean
RR interval; the similar time lag and the high correlation values are
evidence of the clear relationship between these parameters and the
mean RR interval. When changes of θ T and φT were examined there
was one subject whose BLC were very low and parameter changes
were opposite than those encountered in the other subjects and this
reduced the mean individual correlation value of these parameters.
Nonetheless the group correlations of θ T and φT with respect to BLC
were high and with respect to RR very low. With respect to the time
lag estimation, it was observed that θ T and φT were delayed even
when compared to BLC (which changed with delay as compared
to RR). Therefore it is hypothesized that the changes in these loop
orientation parameters were not directly related to changes in RR
interval.
Stepwise linear regression results revealed that VECG parameters could be related to exercise intensity. Furthermore, some of the
VECG parameters (namely T-kurtosis, θ T and φT ) seemed to provide information on exercise intensity beyond that provided by RR
and HRV. Although it should be noted, that the errors of the linear
stepwise regression are not independent due to the individual differences in the observed time series. This violates the assumptions
of the stepwise regression procedure and thus the actual predictive
values of different parameters could not be evaluated. Therefore,
the findings of this study will need to be verified with a larger sample size.
The biggest limitation of this study is the small number of participants (only 6). Although, two protocols were performed for each
participant, this was not sufficient to allow the creation of a properly formulated statistical model between lactate and VECG parameters. In addition, the length of time required for recovery was
underestimated because all parameters had not returned their baseline values at the end of the recovery period. Finally, a minor disadvantage is that BLC was defined from venous blood (taken from the
fingertip) which could had a higher BLC than arterial blood present
76
Dissertations in Forestry and Natural Sciences No 146
Discussion and conclusion
in the heart.
In conclusion could be said that BLC and spirometric measures
themselves contain several drawbacks for monitoring training intensity and there is a clear need for method which would adjust
training intensities and recovery times. There are several benefits
associated with the VECG measurement; noninvasive procedure,
possibility for continuous monitoring and real time feedback of exercise intensity. The findings of the study III revealed that there
were relations between VECG parameters and exercise intensity
markers and thus VECG parameters could be useful in determining the current physiological condition for fitness enthusiasts and
athletes during exercise.
7.4 OVERALL AND NEURAL BAROREFLEX SENSITIVITY
In study IV simultaneously acquired continuous ECG and BP signals and carotid artery ultrasound video were analyzed from 15
subjects with type I diabetes and 28 healthy subjects. These measurements made it possible to estimate arterial elasticity, overall
and neural BRS. The inclusion of diabetic subjects in the study population also made it possible to analyze BRS estimates over a wide
range of values of arterial elasticity. Both BRS parameters (neural and overall BRS) were estimated using traditional and causal
transfer function analysis. Study IV was the first time that causal
transfer function analysis used in the neural BRS estimation. The
advanced BRS estimation method combined with the diverse subject group represent a good way to study the relationship between
arterial stiffness and BRS without biasing factors.
In earlier studies, the non-causal method has been shown in
general to overestimate the BRS gain [7, 45, 46]. The results of study
IV support this claim only partly as can be seen from Bland-Altman
data presented in figure 6.11, which reveals that for high overall
BRS values, the causal method gave higher values than those obtained with traditional BRS estimates. On the other hand, the noncausal method seemed to overestimate slightly the neural BRS val-
Dissertations in Forestry and Natural Sciences No 146
77
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
ues (BRSds ,RR ). This was especially apparent for low neural BRS
values, although there were a few exceptions. The results of study
IV indicate that the causal analysis method would be advantageous
especially if the analyzed dataset is expected to contain low or very
low BRS values. In such cases, the effect of feedforward variation
can lead to an overestimation of BRS and misclassification of reduced BRS i.e. it would be classified as normal.
From figure 6.12 it was observed that neural BRS showed normal values even in a few old diabetic subjects. In these subjects,
overall BRS was reduced probably because of the low arterial elasticity, but ANS function seemed to be preserved as reflected in the
neural BRS. The results support the claim that in the BRS estimation
in is recommended to take account of diameter variation, if BRS is
being used as a specific measure of ANS. For example, it could be
argued that reduced BRS values in patients with myocardial infarctions [43] could have resulted from reduced elasticity in conjunction
with autonomic dysfunction. Overall BRS has been also used as a
marker of CAN in diabetic subjects [44], because the reduction in
the arterial elasticity, autonomic nervous system could be functional
even though the overall BRS is reduced (see figure 6.12). Thereby it
would be very interesting to compare commonly used CAN markers from diabetic subjects with neural BRS rather than the overall
BRS.
The greatest technical limitation of study IV was that blood
pressure was measured from the finger and not from the carotid
artery where ultrasound measurement was made which is also the
site where the baroreceptors are located. However only a noninvasive measurement was feasible from an ethical standpoint. Another
limitation relates to respiration which has been shown to be a latent variable when conducting a cardiovascular causality analysis
and therfore it has been recommended to include respiration into a
MAR-model [123]. Respiration was not measured in this study, but
the error attributable to this factor was minimized by limiting the
analysis to those subjects whose respiration rate (estimated from
the ECG) was above 0.17 Hz (2 subjects had to be excluded) and
78
Dissertations in Forestry and Natural Sciences No 146
Discussion and conclusion
BRS estimates were calculated only using the LF-band.
Neural BRS could provide valuable information on the functioning of neural pathways and cardiovascular system, although
there are a few technical challenges which complicate neural BRS
measurements. An appropriate measurement duration is one factor, in study IV 2-min measurements were used, which is a rather
short period with which to obtain reliable BRS estimation. However, longer ultrasound imaging of carotid artery is very challenging because of the movements caused by respiration and swallowing. Furthermore, questions have been raised about the accuracy
of BRS estimation from spontaneous BP and RR fluctuations in the
supine position [45], because of the minor variation in BP and RRinterval. In study IV it was observed that these technical limitations
led to a narrow frequency range where coherence was significant
between RR and SBP, and RR and ds , therefore the accuracy of BRS
estimate was reduced. The bias caused by these sources may partly
explain the increased causal BRS values when compared to traditional BRS estimates.
In conclusion, the results of study IV indicate that BRS estimates
which provide specific information from ANS function should be
computed from variations in arterial diameter using the causal estimation method. BRS values estimated from blood pressure variation or by using non-causal method may be biased thus providing
an unreliable indication of autonomic neural function, especially in
those subjects whose arterial elasticity has become reduced due to
old age, diabetes or vascular disease.
Dissertations in Forestry and Natural Sciences No 146
79
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
80
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[1] S. J. Crick, M. N. Sheppard, and R. H. Anderson, Neural
supply of the heart. The nervous system and the heart (Human
Press Inc. Totowa, New Jersey, 2000).
[2] A. C. Guyton, Physiology of human body (Saunders college
Publishing, Philadelphia, 1979).
[3] H. R. Kirchheim, “Systemic arterial baroreceptor reflexes,”
Physiol. Rev. 56 (1976).
[4] D. Dubin, Rapid interpretation of EKG’s: an interactive course
(Cover Publishing Company, 2000).
[5] J. Heikkilä and M. Mäkijärvi, EKG (Duodecim, 2003).
[6] A. M. Brown, “Receptors under pressure. An update on
baroreceptors,” Circ. Res. 46 (1980).
[7] R. Barbieri, G. Parati, and J. P. Saul, “Closed- versus OpenLoop Assessment of Heart Rate Baroreflex Implications for
Normal and Abnormal Blood Pressure Regulation and Noninvasive Assessment,” IEEE Eng. Med. Biol. Mag. 20 (2001).
[8] A. Porta, R. Furlan, O. Rimoldi, M. Pagani, A. Malliani, and
P. V. D. Borne, “Quantifying the strength of the linear causal
coupling in closed loop interacting cardiovascular variability
signals,” Biol. Cybern. 86 (2002).
[9] L. Kornet, A. P. Hoeks, B. J. Janssen, J. M. Willigers, and R. S.
Reneman, “Carotid diameter variations as a non-invasive tool
to examine cardiac baroreceptor sensitivity,” J Hypertens 20
(2002).
[10] L. Rydn, E. Standl, M. Bartnik, G. V. den Berghe, J. Betteridge,
M. J. D. Boer, and F. C. et. al., “Guidelines on diabetes, prediabetes, and cardiovascular diseases: executive summary
Dissertations in Forestry and Natural Sciences No 146
81
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
The Task Force on Diabetes and Cardiovascular Diseases of
the European Society of Cardiology (ESC) and of the European Association for the Study of Diabetes (EASD),” Eur.
Heart J. 28 (2007).
[11] B. Eckert and C. D. Agardh, “Hypoglycaemia leads to an
increased QT interval in normal men,” Clinical Physiology 18
(1998).
[12] L. Landstedt-Hallin, A. Englund, U. Adamson, and P. E. Lins,
“Increased QT dispersion during hypoglycaemia in patients
with type 2 diabetes mellitus,” J. Intern. Med. 246 (1999).
[13] T. Laitinen, T. Lyyra-Laitinen, H. Huopio, I. Vauhkonen,
T. Halonen, J. Hartikainen, L. Niskanen, and M. Laakso,
“Electrocardiographic alterations during hyperinsulinemic
hypoglycemia in healthy subjects,” Ann. Noninvas. Electro. 13
(2008).
[14] J. Meinhold, T. Heise, K. Rave, and L. Heinemann, “Electrocardiographic changes during insulin-induced hypoglycemia
in healthy subjects,” Horm. Metab. Res. 30 (1998).
[15] H. Nguyen, N. Ghevondian, S. Nguyen, and T. W. Jones, “Detection of hypoglycemic episodes in children with type 1 diabetes using an optimal Bayesian neural network algorithm,”
in 29th Annual International Conference of the IEEE (2007).
[16] M. L. Simoons and P. G. Hugenholtz, “Gradual changes of
ECG waveform during and after exercise in normal subjects,”
Circulation 52 (1975).
[17] J. W. Deckers, R. V. Vinke, J. R. Vos, and M. L. Simoons,
“Changes in the electrocardiographic response to exercise in
healthy women,” ’Brit. Heart J. 64 (1990).
[18] R. Lehtinen, “ST/HR hysteresis: exercise and recovery phase
ST depression/heart rate analysis of the exercise ECG,” J.
Electrocardiol. 32 (1999).
82
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[19] G. D. Berntson, J. T. Bigger, D. L. Eckberg, P. Grossman, P. G.
Kaufmann, M. Malik, H. N. Nagaraja, S. W. Porges, J. P. Saul,
P. H. Stone, and M. W. V. D. Molen, “Heart rate variability:
Origins, methods and interpretive caveats,” Psychophysiology
34 (1997).
[20] H. A. Fozzard, E. J. Haber, R. B. Jennings, A. M. Katz, and
H. Morgan, The heart and cardiovascular system (Raven Press,
New York, 1991).
[21] J. Malmivuo and R. Plonsey, Bioelectromagnetism: principles
and applications of bioelectric and biomagnetic fields (Oxford University Press, 1995).
[22] T. Kenttä, Dynamics of cardiac repolarization during exercise
(Doctoral dissertation, University of Oulu, 2012).
[23] A. E. Aubert, B. Seps, and F. Beckers, “Heart Rate Variability
in Athletes,” Sports Med. 33 (2003).
[24] F. Lombardi and A. Malliani, “Task force of the European
society of cardiology and the North American society of pacing and electrophysiology, Heart rate variability standards of
measurement, physiological interpretation, and clinical use,”
Circulation 93 (1996).
[25] P. M. Rautaharju, B. Surawicz, and L. S. Gettes, “AHA ACCF
HRS Recommendations for the Standardization and Interpretation of the ElectrocardiogramPart IV: The ST Segment, T
and U Waves, and the QT Interval A Scientific Statement
From the American Heart Association Electrocardiography
and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the
Heart Rhythm Society Endorsed by the International Society
for Computerized Electrocardiology,” J. Am. Coll. Cardiol. 53
(2009).
Dissertations in Forestry and Natural Sciences No 146
83
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
[26] G. X. Yan and C. Antzelevitch, “Cellular basis for the normal T wave and the electrocardiographic manifestations of
the long-QT syndrome,” Circulation 98 (1998).
[27] H. C. Bazett, “An analysis of the time-relations of electrocardiograms,” Heart 7 (1920).
[28] L. S. Fridericia, “The duration of systole in an electrocardiogram in normal humans and in patients with heart disease,”
Ann. Non. Elect. 8 (2003).
[29] M. Malik, P. Färbom, V. Batchvarov, and K. H. A. J. Camm,
“Relation between QT and RR intervals is highly individual
among healthy subjects: implications for heart rate correction
of the QT interval,” Heart 87 (2002).
[30] J. Karjalainen, A. Reunanen, P. Ristola, and M. Viitasalo, “QT
interval as a cardiac risk factor in a middle aged population.,”
Heart 77 (1997).
[31] P. Brugada and J. Brugada, “Right bundle branch block, persistent ST segment elevation and sudden cardiac death: A
distinct clinical and electrocardiographic syndrome: A multicenter report,” J. Am. Coll. Cardiol. 20 (1992).
[32] G. E. Hayden, W. J. Brady, A. D. Perron, M. P. Somers, and
A. Mattu, “Electrocardiographic T-wave inversion: Differential diagnosis in the chest pain patient,” Am. J. Emerg. Med. 20
(2002).
[33] T. Baas, K. Gräfe, A. Khawaja, and O. Dösselr, “Investigation
of parameters highlighting drug induced small changes of the
T-wave’s morphology for drug safety studies,” in 33rd Annual
international conference of IEEE EMBS (2011).
[34] E. Frank, “An accurate, clinically practical system for spatial
vectorcardiography,” Circulation. 13 (1956).
84
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[35] B. Acar, G. Yi, K. Hnatkova, and M. Malik, “Spatial, temporal and wavefront direction characteristics of 12-lead T-wave
morphology,” Med. Biol. Eng. Comput. 37 (1999).
[36] P. Smetana, V. N. Batchvarov, K. Hnatkova, A. J. Camm, and
M. Malik, “Ventricular gradient and nondipolar repolarization components increase at higher heart rate,” Am. J. PhysiolHeart Circ. 286 (2004).
[37] T. Kenttä, M. Karsikas, A. Kiviniemi, M. Tulppo, T. Seppänen,
and H. V. Huikuri, “Dynamics and Rate-Dependence the spatial angle between ventricular depolarization and repolarization wave fronts during exercise ECG,” Ann. Noninivas. Electro. 15 (2010).
[38] M. Zabel, B. Acar, T. Klingenheben, M. R. Franz, S. H. Hohnloser, and M. Malik, “Analysis of 12-lead T-wave morphology
for risk stratification after myocardial infarction,” Circulation
102 (2000).
[39] A. V. Oosterom, R. Hoekema, and G. J. H. Uijen, “Geometrical factors affecting the interindividual variability of the ECG
and the VCG,” J. Electrocardiol. 33 (2000).
[40] H. W. Robbe, L. J. Mulder, H. Ruddel, W. A. Langewitz, J. B.
Veldman, and G. Mulder, “Assessment of baroreceptor reflex sensitivity by means of spectral analysis,” Hypertension
10 (1987).
[41] G. Bertinieri, M. D. Rienzo, A. Cavallazzi, A. Ferrari, A. Pedotti, and G. Mancia, “Evaluation of baroreceptor reflex by
blood pressure monitoring in unanesthetized cats,” Am. J.
Physiol. 254 (1988).
[42] M. L. Rovere, J. Bigger, F. Marcus, A. Mortara, and
P. Schwartz, “Baroreflex sensitivity and heart-rate variability in prediction of total cardiac mortality after myocardial
infarction. ATRAMI (Autonomic Tone and Reflexes After Myocardial Infarction) investigators,” Lancet 351 (1998).
Dissertations in Forestry and Natural Sciences No 146
85
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
[43] N. Honzikova, B. Semrad, B. Fiser, and R. Labrova, “Baroreflex sensitivity determined by spectral method, and heart rate
variability, and two years mortality in patients after myocardial infarction,” Physiol. Res. 49 (2000).
[44] A. Frattola, G. Parati, P. Gamba, F. Palear, G. Mauri, M. D.
Rienzo, P. Castiglioni, and G. Mancia, “Time and frequency
domain estimates of spontaneous baroreflex sensitivity provide early detection of autonomic dysfunction in diabetes
mellitus,” Diabetologia 40 (1997).
[45] G. Nollo, L. Faes, A. Porta, R. Antolini, and F. Ravelli, “Exploring directionality in spontaneous heart period and systolic pressure variability interactions in humans: implications
in the evaluation of baroreflex gain,” Am. J. Physiol-Heart C.
288 (2005).
[46] L. Faes, A. Porta, R. Cucino, S. Cerutti, R. Antolini, and
G. Nollo, “Causal transfer function analysis to describe closed
loop interactions between cardiovascular and cardiorespiratory variability signals,” Biol. Cybern. 90 (2004).
[47] Z. Chen, P. L. Purdon, G. Harrell, E. T. Pierce, J. Walsh, E. N.
Brown, and R. Barbieri, “Dynamic assessment of baroreflex
control of heart rate during induction of propofol anesthesia
using a point process method,” Ann. Biomed. Eng. 39 (2011).
[48] B. E. Hunt, L. Fahy, W. B. Farquhar, and J. A. Taylor, “Quantification of mechanical and neural components of vagal
baroreflex in humans,” Hypertension 37 (2001).
[49] N. P. Saeed, S. Reneman, and A. P. Hoeks, “Contribution
of vascular and neural segments to baroreflex sensitivity in
response to postural stress,” J. Vasc. Res. 46 (2009).
[50] C. C. Callegaro and J. A. Taylor, “Age-related effects of vagotonic atropine on cardiovagal baroreflex gain,” Neurobiol. Aging. 33 (2012).
86
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[51] G. Deley, G. Picard, and J. A. Taylor, “Arterial Baroreflex
Control of Cardiac Vagal Outflow in Older Individuals Can
Be Enhanced by Aerobic Exercise Training,” Hypertension 53
(2009).
[52] H. Komine, J. Sugawara, K. Hayashi, M. Yoshizawa, and
T. Yokoi, “Regular endurance exercise in young men increases arterial baroreflex sensitivity through neural alteration of baroreflex arc.,” J. Appl. Physiol. 106 (2009).
[53] M. E. Safar, “Peripheral pulse pressure, large arteries, and
microvessels,” Hypertension 44 (2004).
[54] P. Hamilton, C. Lockhart, C. Quinn, and G. Mcveigh, “Arterial stiffness: clinical relevance, measurement and treatment,”
Clin. Sci. 113 (2007).
[55] X. Girerd, S. Laurent, B. Pannier, R. Asmar, and M. Safar,
“Arterial distensibility and left ventricular hypertrophy in patients with sustained essential hypertension,” Am. Heart J. 122
(1991).
[56] D. Fukuda, M. Yoshiyama, K. Shimada, H. Yamashita,
S. Ehara, Y. Nakamura, K. Kamimori, A. Tanaka,
T. Kawarabayashi, and J. Yoshikawa, “Relation between aortic
stiffness and coronary flow reserve in patients with coronary
artery disease,” Heart 92 (2006).
[57] C. M. McEniery, “Novel therapeutic strategies for reducing
arterial stiffness,” Brit. J. Pharmacol. 148 (2006).
[58] S. Laurent, J. Cockcroft, L. V. Bortel, P. Boutouyrie, C. Giannattasio, D. Hayoz, B. P. C. Vlachopoulos, I. Wilkinson,
and H. Struijker-Boudier, “Expert consensus document on
arterial stiffness: methodological issues and clinical applications,” Eur. Heart J. 27 (2006).
Dissertations in Forestry and Natural Sciences No 146
87
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
[59] R. Kelly, C. Hayward, A. Avolio, and M. O’rourke, “Noninvasive determination of age-related changes in the human
arterial pulse,” Circulation 80 (1989).
[60] T. Hirai, S. Sasayama, T. Kawasaki, and s. Yagi, “Stiffness of systemic arteries in patients with myocardial infarction. A noninvasive method to predict severity of coronary
atherosclerosis,” Circulation 80 (1989).
[61] M. L. Koivikko, M. Karsikas, P. I. Salmela, J. S. Tapanainen,
A. Ruokonen, T. Seppänen, H. Huikuri, and J. S. Perkiömäki,
“Effects of controlled hypoglycaemia on cardiac repolarisation in patients with type 1 diabetes,” Diabetologia 51 (2008).
[62] R. T. C. E. Robinson, N. D. Harris, R. H. Ireland, I. A. Macdonald, and S. R. Heller, “Changes in cardiac repolarization
during clinical episodes of nocturnal hypoglycaemia in adults
with Type 1 diabetes,” Diabetologia 47 (2004).
[63] B. Suys, S. Heuten, D. D. Wolf, M. Verherstraeten, L. O.
de Beeck, D. Matthys, C. Vrints, and R. Rooman, “Glycemia
and corrected QT interval prolongation in young type 1 diabetic patients what is the relation?,” Diabetes Care 29 (2006).
[64] G. V. Gill, A. Woodward, I. F. Casson, and P. J. Weston, “Cardiac arrhythmia and nocturnal hypoglycaemia in type 1 diabetesthe dead in bedsyndrome revisited,” Diabetologia (2009).
[65] S. Lee, N. D. Harris, R. T. Robinson, L. Yeoh, I. A. Macdonald,
and S. R. Heller, “Effects of adrenaline and potassium on QTc
interval and QT dispersion in man,” Eur. J. Clin. Invest. (2003).
[66] H. Schchinger, J. Port, S. Brody, L. Linder, F. Wilhelm,
P. Huber, D. Cox, and U. Keller, “Increased high-frequency
heart rate variability during insulin-induced hypoglycaemia
in healthy humans,” Clin Sci 106 (2004).
88
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[67] T. Lindström, L. Jorfeldt, L. Tegler, and H. J. Arnqvist, “Hypoglycaemia and cardiac arrhythmias in patients with type 2
diabetes mellitus,” Diabetic Med. 9 (1992).
[68] J. L. B. Marques, E. George, S. R. Peacey, N. D. Harris, I. A.
Macdonald, T. Cochrane, and S. R. Heller, “Altered ventricular repolarization during hypoglycaemia in patients with diabetes,” Diabetic Med. 14 (1997).
[69] R. H. Ireland, R. T. C. E. Robinson, S. R. Heller, J. L. B. Marques, and N. D. Harris, “Measurement of high resolution
ECG QT interval during controlled euglycaemia and hypoglycaemia,” Physiol. Meas. 21 (2000).
[70] R. T. C. E. Robinson, N. D. Harris, R. H. Ireland, S. Lee,
C. Newman, and S. R. Heller, “Mechanisms of abnormal cardiac repolarization during insulin-induced hypoglycemia,”
Diabetes 52 (2003).
[71] N. P. Murphy, M. E. Ford-Adams, K. K. Ong, N. D. Harris,
S. M. Keane, C. Davies, R. H. Ireland, I. A. MacDonald, E. J.
Knight, J. A. Edge, S. R. Heller, and D. B. Dunger, “Prolonged cardiac repolarisation during spontaneous nocturnal
hypoglycaemia in children and adolescents with type 1 diabetes,” Diabetologia 47 (2004).
[72] S. Giunti, G. Bruno, E. Lillaz, G. Gruden, V. Lolli,
N. Chaturvedi, J. H. Fuller, M. Veglio, and P. Cavallo-Perin,
“Incidence and risk factors of prolonged QTc interval in type
1 diabetes,” Diab. Care. 30 (2007).
[73] J. S. Romney and R. Z. Lewanczuk, “Vascular compliance is
reduced in the early stages of type 1 diabetes,” Diabetes Care
24 (2001).
[74] A. R. Ahlgren, G. Sundkvist, T. Sandgren, and T. Länne, “Female gender increases stiffness of elastic but not of muscular
arteries in type I diabetic patients,” Clin. Physiol. Funct. I. 22
(2002).
Dissertations in Forestry and Natural Sciences No 146
89
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
[75] D. Tryfonopoulos, E. Anastasiou, A. Protogerou, T. Papaioannou, K. Lily, A. Dagre, E. Souvatzoglou, C. Papamichael,
M. Alevizaki, and J. Lekakis, “Arterial stiffness in type 1 diabetes mellitus is aggravated by autoimmune thyroid disease,”
J. Endocrinol. Inves. 28 (2005).
[76] D. A. Stakos, D. P. Schuster, E. A. Sparks, C. F. Wooley,
K. Osei, and H. Boudoulas, “Cardiovascular effects of type
1 diabetes mellitus in children,” Angiology 56 (2005).
[77] A. Parikh, E. B. Sochett, B. W. McCrindle, A. Dipchand,
A. Daneman, and D. Daneman, “Carotid artery distensibility
and cardiac function in adolescents with type 1 diabetes,” J.
Pediatr. 137 (2000).
[78] C. D. A. Stehouwer, R. M. A. Henry, and I. Ferreira, “Arterial
stiffness in diabetes and the metabolic syndrome: a pathway
to cardiovascular disease,” Diabetologia 51 (2008).
[79] A. I. Vinik and D. Ziegler, “Diabetic cardiovascular autonomic
neuropathy,” Circulation 115 (2007).
[80] P. Deedwania and S. Srikanth, “Diabetes and vascular disease,” Expert review of cardiovascular therapy 6 (2008).
[81] D. Ziegler, “Diabetic cardiovascular autonomic neuropathy:
prognosis, diagnosis and treatment,” Diabetes Metab. Res. 10
(1994).
[82] D. Lucinia, G. Zuccotti, M. Malacarne, A. Scaramuzza, S. Riboni, C. Palombo, and M. Pagani, “Early progression of the
autonomic dysfunction observed in pediatric type 1 diabetes
mellitus,” Hypertension 54 (2009).
[83] D. Ziegler, D. Laude, F. Akila, and J. L. Elghozi, “Time-and
frequency-domain estimation of early diabetic cardiovascular
autonomic neuropathy,” Clin. Auton. Res. 11 (2001).
90
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[84] K. Tanaka, Y. Matsuura, S. Kumagai, A. Matsuzaka, K. Hirakoba, and K. Asano, “Relationships of anaerobic threshold
and onset of blood lactate accumulation with endurance performance,” Eur J Appl Physiol Occup Physiol 52 (1983).
[85] T. McLellan, “Ventilatory and plasma lactate response with
different exercise protocols: a comparison of methods,” Int J
Sports Med 6 (1985).
[86] R. Mocellin, M. Heusgen, and H. P. Gildein, “Anaerobic
threshold and maximal steady-state blood lactate in prepubertal boys,” Eur. J. Appl. Physiol. O.l 62 (1991).
[87] G. Borg, G. Ljunggren, and R. Ceci, “The increase of perceived exertion, aches and pain in the legs, heart rate and
blood lactate during exercise on a bicycle ergometer,” Eur. J.
Appl. Physiol. O. 54 (1985).
[88] P. Hofmann, R. Pokan, K. Preidle, H. Leitner, D. Szolar,
B. Eber, and G. Schwaberger, “Relationship between heart
rate threshold, lactate turn point and myocardial function,”
Int. J. Sports Med. 15 (1994).
[89] A. J. Coutts, E. Rampinini, S. M. Marcora, C. Castagna, and
F. M. Impellizzeri, “Heart rate and blood lactate correlates of
perceived exertion during small-sided soccer games,” J. Sci.
Med. Sport 12 (2009).
[90] V. F. Gladwell, G. R. H. Sandercock, and S. L. Birch, “Cardiac vagal activity following three intensities of exercise in
humans,” Clin Physiol. Funct Imaging 30 (2009).
[91] F. Conconi, M. Ferrari, P. Ziglio, P. Droghetti, and L. Codeca,
“Determination of the anaerobic threshold by a non-invasive
field test in runners,” J Appl Physiol 52 (1982).
[92] A. Jones and J. Doust, “The Conconi test is not valid for
estimation of the lactate turnpoint in runners,” J Sports Sci 15
(1997).
Dissertations in Forestry and Natural Sciences No 146
91
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
[93] A. Jones and J. Doust, “Lack of reliability in Conconi’s heart
rate deflection point,” Int J Sports Med 16 (1995).
[94] J. Bourgois, P. Coorevits, L. Danneels, E. Witvouw, D. Cambier, and J. Vrijens, “Validity of the heart rate deflection point
as a predictor of lactate threshold concepts during cycling,” J.
Strength Cond. Res. 18 (2004).
[95] G. Blain, O. Meste, T. Bouchard, and S. Bermon, “Assessment of ventilatory thresholds during graded and maximal
exercise test using time varying analysis of respiratory sinus
arrhythmia,” Brit. J. Sports Med. 39 (2005).
[96] F. Cottin, P. M. Lepretre, P. Lopes, Y. Papelier, C. Medigue,
and V. Billat, “Assessment of Ventilatory Thresholds from
Heart Rate Variability in Well-Trained Subjects during Cycling,” Int. J. Sports Med. 27 ((2006).
[97] F. Cottin, C. Medigue, P. Lopes, P. Lepretre, T. Heuber, and
V. Billat, “Ventilatory Thresholds Assessment from Heart
Rate Variability during an Incremental Exhaustive Running
Test,” Int. J. Sports. Med. 28 (2007).
[98] C. Thomas, P. Sirvent, S. Perrey, E. Raynaud, and J. Mercier,
“Relationships between maximal muscle oxidative capacity
and blood lactate removal after supramaximal exercise and
fatigue indexes in humans,” J. Appl. Physiol.. 97 (2004).
[99] J. S. Sarma, K. Venkataraman, D. R. Samant, and U. Gadgil,
“Hysteresis in the Human RR-QT Relationship During Exercise and Recovery,” Pace 10 (1987).
[100] R. J. Gibbons, G. J. Balady, J. W. Beasley, J. T. Bricker, W. F.
Duvernoy, V. F. Froelicher, et.al., and T. Ryan, “ACC/AHA
Guidelines for Exercise Testing. A report of the American College of Cardiology/American Heart Association Task Force
on Practice Guidelines (Committee on Exercise Testing),” J.
Am. Coll.Cardiol. 30 (1997).
92
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[101] R. J. Gibbons, G. J. Balady, J. T. Bricker, B. R. Chaitman, G. F.
Fletcher, V. F. Froelicher, et. al., and S. C. Smith, “ACC/AHA
2002 guideline update for exercise testing: summary articleA report of the American college of cardiology/American
heart association task force on practice guidelines (committee to update the 1997 exercise testing guidelines),” J. Am.
Coll.Cardiol. 40 (2002).
[102] I. Jolliffe, Principal Component Analysis (Springer- Verlag, New
York, 1986).
[103] N. V. Thakor, J. G. Webster., and W. J. Tompkins, “Optimal
QRS detector,” Med. Biol. Eng. Comput. 21 (1983).
[104] O. Pahlm and L. Sörnmo, “Software QRS detection in ambulatory monitoringa review,” Med. Biol. Eng. Comput. 22 (1984).
[105] J. Pan and W. J. Tompkins, “A real-time QRS detection algorithm,” IEEE T. Bio-Med. Eng 3 (1985).
[106] G. M. Friesen, T. C. Jannett., M. A. Jadallah, S. L. Yates, S. R.
Quint, and H. T. Nagle, “A comparison of the noise sensitivity of nine QRS detection algorithms,” IEEE T. Bio-Med. Eng.
37 (1990).
[107] Q. Xue, Y. H. Hu, and W. J. Tompkins, “Neural-networkbased adaptive matched filtering for QRS detection,” IEEE T.
Bio-Med. Eng 39 (1992).
[108] S. Kadambe, R. Murray, and G. F. Boudreaux-Bartels,
“Wavelet transform-based QRS complex detector,” IEEE T.
Bio-Med. Eng 46 (1999).
[109] M. Tarvainen, J. Niskanen, J. Lipponen, P. Ranta-aho, and
P. Karjalainen, “Kubios HRV Heart rate variability analysis
software,” Computer methods in biomedicine 113 (2014).
[110] P. Laguna, N. V. Thakor, P. Caminal, R. Jane, A. H. R. Yoon,
B. de Luna, V. Marti, and J. Guindo, “New algorithm for
Dissertations in Forestry and Natural Sciences No 146
93
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
QT interval analysis in 24-hour Holter ECG: performance and
applications,” Med. Biol. Eng. Comput. 28 (1990).
[111] E. Niskanen, M. P. Tarvainen, J. P. Niskanen, P. Ranta-aho,
M. Könönen, H. Soininen, and P. A. Karjalainen, “Principal
component approach for mapping functional connectivity in
event-related fMRI,” in World congress on medical physics and
biomedical engineering (2009).
[112] P. Kligfield and et al., “Recommendations for the Standardization and Interpretation of the ElectrocardiogramPart I: The
Electrocardiogram and Its Technology A Scientific Statement
From the American Heart Association Electrocardiography
and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the
Heart Rhythm Society Endorsed by the International Society
for Computerized Electrocardiology,” J. Am. Coll. Cardiol. 49
(2007).
[113] C. M. Fira and L. Goras, “An ECG Signals Compression
Method and Its Validation Using NNs,” IEEE T. Bio-Med. Eng.
55 (2008).
[114] L. Lehtola, M. Karsikas, M. Koskinen, H. Huikuri, and
T. Seppänen, “Effects of noise and filtering on SVD-based
morphological parameters of the T wave in the ECG,” J. Med.
Eng. Technol. 32 (2008).
[115] A. Karagiannis and P. Constantinou, “Noise-assisted data
processing with empirical mode decomposition in biomedical signals,” IEEE T. Inf. Technol. B. 15 (2011).
[116] E. Castillo, D. Morales, A. Garca, F. Martnez-Mart, L. Parrilla,
and A. Palma, “Noise Suppression in ECG Signals through
Efficient One-Step Wavelet Processing Techniques,” Journal of
Applied Mathematics (2013).
94
Dissertations in Forestry and Natural Sciences No 146
Bibliography
[117] A. Savitzky and M. J. Golay, “Smoothing and differentiation
of data by simplified least squares procedures,” Anal. Chem.
36 (1964).
[118] M. Marple, Digital spectral analysis with applications (PrenticeHalll, 1987).
[119] T. Söderström and P. Stoica, System identification (PrenticeHall, Inc., 1988).
[120] L. Lennart, System identification: theory for the user (PTR Prentice Hall, Upper Saddle River, 1999).
[121] S. Kay, Modern Spectral Estimation: Theory and Application (
Englewood Cliffs, 1988).
[122] G. Baselli, A. Porta, O. Rimoldi, M. Pagani, and S. Cerutti,
“Spectral decomposition in multichannel recordings based on
multivariate parametric identification,” IEEE T. Bio-Med. Eng.
44 (1997).
[123] A. Porta, T. Bassani, V. Bari, G. D. Pinna, R. Maestri, and
S. Guzzetti, “Accounting for Respiratory is necessary to reliably infer granger causality from cardiovascular variability
series,” IEEE T. Bio-Med. Eng. 59 (2012).
[124] L. Faes, G. D. Pinna, A. Porta, R. Maestri, and G. Nollo, “Surrogate data analysis for assessing the significance of the coherence function,” IEEE T. Bio-Med. Eng. 51 (2004).
[125] R. Bousseljot, D. Kreiseler, Schnabe, and A., “Nutzung der
EKG-Signaldatenbank CARDIODAT der PTB ber das Internet,” Biomedizinische Technik 40 (1995).
[126] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff,
P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. K.
Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and
PhysioNet: components of a new research resource for complex physiologic signals,” Circulation 101 (2000).
Dissertations in Forestry and Natural Sciences No 146
95
Jukka Lipponen: Estimation methods for cardiovascular signal analysis
[127] R. A. DeFronzo, J. D. Tobin, and R. Andres, “Glucose clamp
technique: a method for quantifying insulin secretion and
resistance.,” Am. J. Physiol-Endoc. M. 237 (1979).
[128] M. P. Tarvainen, P. O. Ranta-aho, and P. A. Karjalainen, “An
advanced detrending method with application to HRV analysis,” IEEE T. Bio-Med. Eng. 49 (2002).
[129] V. R. Newey and D. K. Nassiri, “Online artery diameter
measurement in ultrasound images using artificial neural networks,” Ultrasound Med. Biol. 28 (2002).
[130] P. P. Kanjilal, S. Palit, and G. Saha, “Fetal ECG extraction from
single-channel maternal ECG using singular value decomposition,” IEEE T. Bio-Med. Eng. 44 (1997).
[131] P. Okin, R. Devereux, R. Fabsitz, E. Lee, J. Galloway, and
B. Howard, “Principal component analysis of the T wave and
prediction of cardiovascular mortality in American Indians:
the Strong Heart Study,” Circulation 105 (2002).
[132] S. M. Rissanen, M. Kankaanpää, A. Meigal, M. P. Tarvainen,
J. Nuutinen, I. M. Tarkka, O. Airaksinen, and P. A. Karjalainen, “Surface EMG and acceleration signals in Parkinsons disease: feature extraction and cluster analysis,” Med.
Biol. Eng. Comput. 46 (2008).
[133] D. J. Cox, L. Gonder-Frederick, and W. Clarke, “Driving
decrements in type I diabetes during moderate hypoglycemia
Diabetes,” Diabetes 42 (1993).
[134] P. J. Weston, J. M. Glancy, P. G. McNally, H. Thurston, and
D. P. Bono, “Can abnormalities of ventricular repolarization
identify insulin debend diabetic patients at risk of sudden
cardiac death,” Heart 78 (1997).
[135] N. Harris and S. Heller, “Sudden death in young patients
with type 1 diabetes: a consequence of disease treatment or
both?,” Diabetic Med. 16 (1999).
96
Dissertations in Forestry and Natural Sciences No 146
[136] S. R. Heller, “Abnormalities of the electrocardiogram during
hypoglycemia: the cause of the dead in bed syndrome?,” Int.
J. Clin. Pract. Suppl. 129 (2002).
[137] P. Bottini, E. Boschetti, S. Pampanelli, M. Ciofetta, P. D. Sindaco, L. Scionti, P. Brunetti, and G. B. Bolli, “Contribution
of autonomic neuropathy to reduced plasma adrenaline responses to hypoglycemia in IDDM: evidence for a nonselective defect,” Diabetes 46 (1997).
[138] H. B. Nielsen, J. O. Clemmesen, C. Skak, P. Ott, and N. H.
Secher, “Attenuated hepatosplanchnic uptake of lactate during intense exercise in humans,” J. Appl. Physiol.. 92 (2002).
[139] L. B. Gladden, “Lactate metabolism: a new paradigm for the
third millennium,” J. Physiol. 558 (2004).
Jukka Lipponen
Estimation Methods
for Cardiovascular
Signal Analysis
Cardiovascular measurements
are important diagnostic tools
for clinicians and measurements
could be also used for monitoring
humans during their everyday
life, for example, in order to
detecting hypoglycemic events in
diabetic subjects. In this thesis
novel approaches for estimating
the functioning and condition of
the cardiovascular system were
developed. Findings demonstrate the
usability of these new methods in
cardiovascular monitoring.
Publications of the University of Eastern Finland
Dissertations in Forestry and Natural Sciences
isbn 978-952-61-1532-0