Statistical analysis of biomedical data [Elektronische Ressource] / vorgelegt von Andreas Jung

Sujets

• Mathematics
• Physik
• Gesundheit

Informations

• Publié par: universitat_regensburg
• Publié le: 01 janvier 2004
• Nombre de lectures: 13
• Langue: English
• Poids de l'ouvrage: 4 Mo

Extrait

Statistical analysis of biomedical data
Dissertation
zur Erlangung des Doktorgrades
der Naturwissenschaften (Dr. rer. nat.)
derhaftlichen Fakult at II { Physik
der Universit at Regensburg
vorgelegt von
Andreas Jung
aus Munc hen
Dezember 2003Das Promotionsgesuch wurde am 17. Dezember 2003 eingereicht.
Das Promotionskolloquium fand am 29. Januar 2004 statt.
Prufungsaussc huss:
Vorsitzender: Prof. Dr. Werner Wegscheider
1. Gutachter: Prof. Dr. Klaus Richter
2.hter: Prof. Dr. Gustav Obermair
Weiterer Prufer: Prof. Dr. Matthias BrackMeinen ElternContents
Glossary iii
Introduction 1
1 Survey of the biomedical data sets 5
1.1 Anatomy and physiology of the human brain . . . . . . . . . . . . 5
1.2 Neuromonitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Electro-Encephalography (EEG) . . . . . . . . . . . . . . . . . . . 16
2 Time series analysis 19
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Application: Correlation between ... . . . . . . . . . . . . . . . . . 22
2.3.1 Invos on left and right hemisphere . . . . . . . . . . . . . . 22
2.3.2 Licox and Invos . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.3 Arterial blood pressure and oxygen supply . . . . . . . . . 29
2.3.4 blood and intracranial pressure . . . . . . 30
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Model of the haemodynamic and metabolic processes in the
brain 35
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Fluid dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.1 Assumption for the hydrodynamical model . . . . . . . . . 38
3.2.2 The compartments . . . . . . . . . . . . . . . . . . . . . . 39
3.2.3 Final set of equations . . . . . . . . . . . . . . . . . . . . . 42
3.2.4 Standard values . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2.5 Validation of the model . . . . . . . . . . . . . . . . . . . . 46
3.3 Oxygen transport . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.1 The blood . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.2 The Krogh cylinder . . . . . . . . . . . . . . . . . . . . . . 54
3.3.3 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4 Validation: Theory$ Experimental data . . . . . . . . . . . . . . 62
iii Contents
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4 Independent component analysis 69
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Basic theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2.1 Probability theory . . . . . . . . . . . . . . . . . . . . . . 73
4.2.2 Information . . . . . . . . . . . . . . . . . . . . . . 77
4.3 A geometric approach . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3.1 Geometric considerations . . . . . . . . . . . . . . . . . . . 81
4.3.2 The (neural) geometric learning algorithm . . . . . . . . . 83
4.3.3 Theoretical framework for the geometric ICA algorithm . . 85
4.3.4 Limit points of the geometric algorithm . . . . . . . . . . . 87
4.3.5 FastGeo: A histogram based . . . . . . . . . . . 90
4.3.6 Accuracy and performance of FastGeo . . . . . . . . . . . 92
4.3.7 Higher dimensions . . . . . . . . . . . . . . . . . . . . . . 98
4.3.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4 An information theoretical approach including time structures . . 101
4.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4.2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.4.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.4.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.4.5 A new concept for ICA { independent increments . . . . . 111
4.4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.5 Application to biomedical data . . . . . . . . . . . . . . . . . . . 113
4.5.1 Neuromonitoring data . . . . . . . . . . . . . . . . . . . . 113
4.5.2 Electro-Encephalography (EEG) data . . . . . . . . . . . . 113
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Conclusions and Outlook 123
A Mathematical tools and proofs 127
A.1 Correlation in the frequency domain . . . . . . . . . . . . . . . . 127
A.2 Proof: Uniqueness of geometric ICA . . . . . . . . . . . . . . . . . 129
A.3 Proof: Existence of only two xed points in geometric ICA . . . . 130
Bibliography 133
Dank 141Glossary
ABP Arterial Blood Pressure
BSS Blind Source Separation
CbO Oxygen content in the blood2
C(ajv)O content in the (arterialjvenous) blood2
cdf cumulative distribution function
CSF Cerebrospinal Fluid
CT Computer Tomography
DFT Discrete Fourier Transform
ECG Electro-Cardiography
EEG Electro-Encephalography
Hb Deoxyhaemoglobin
HbO Oxyhaemoglobin2
IC Independent Component
ICA Indeptonent Analysis
ICP Intracranial Pressure
Invos In-Vivo Optical Spectroscopy
Licox Liquor Oxygenation
MABP Mean Arterial Blood Pressure
MTM Multi Taper Method
NMR Nuclear Magnetic Resonance
PCA Principal Component Analysis
p mean partial oxygen pressure in the tissueti
pbO partial oxygen pressure in the blood2
p(ajv)O o in the (arterialjvenous) blood2
pdf probability density function
RMT Random Matrix Theory
SbO Saturation of the blood with oxygen2
S(ajv)O of the (arterialjvenous) blood with oxygen2
SNR Signal to Noise Ratio
STFT Short Time Fourier Transform
TSA Time Series Analysis
iiiIntroduction
Recently, the development of computer applications in the eld of life sciences,
in particular for the clinical and biomedical environments, has gained increasing
attention due to the promising results in the treatment of patients.
Today, the recording of the patient’s status in multivariate data sets is a
standard procedure in everydays clinical life. Apart from the immediate evalua-
tion of the data by the physicians, the (o -line) analysis by dedicated computer
processing tools can be a valuable information source.
In the context of data analysis, the methods derived in di eren t elds of
physics raise the question of applying these well known methods to life sciences,
in particular to biomedical data analysis: Is it possible to obtain a deeper un-
derstanding of the data with new (non)linear methods and physical models de-
scribing the biological system? Can we further improve the analysis to reveal the
hidden information in biomedical data by developing new algorithms overcoming
limitations of the existing methods?
At the university hospital in Regensburg, the department of neurosurgery
records di eren t biomedical data sets in the clinical environment. These record-
ings have motivated this work and we will focus on two of these data sets. On
one hand, neuromonitoring data is recorded on the intensive care unit from pa-
tients with severe head injuries. These data sets re ect mainly the following
brain status parameters: oxygen content in the blood and tissue of the brain, the
arterial blood pressure and the internal brain pressure. In addition, the patient’s
brain temperature is measured. On the other hand, the neural brain activity
of patients is recorded in the electro-encephalography (EEG), in particular in the
post-operative treatment, to monitor neurological diseases. These recordings rep-
resent { in contrast to the neuromonitoring data { highly multivariate data sets
with 21 or more signals.
The questions arising in the analysis of the data are very diverse. They all
focus on a deeper understanding of the mechanisms of the investigated system,
namely the human brain. In neuromonitoring data we may ask: Do the signals
in uence each other, are they correlated in some sense? Which processes trigger
the system? What is the underlying biological system generating the signals? In
the analysis of the EEG data we are mainly interested in whether new methods
can reveal more information or enhance the highly multivariate data.
12 Introduction
The techniques typically used in data analysis can be divided into three dif-
ferent categories. This depends on the amount of knowledge available about the
investigated system:
Time Series Analysis:
No knowledge is available about the system, only the time series originat-
ing from the outputs of the system can be analysed. Two main techniques
are used: either unimodal (fourier transforms, wavelets, (non)linear anal-
ysis using time embedding) or bimodal analysis methods (correlations and
couplings using nonlinear statistics). Symptoms or phenomena detectable
in the time series can be quanti ed in such a way that statistical tests can
be applied.
Independent Component Analysis (ICA):
Hidden sources are underlying the measured signals. ICA is in particular
useful for the analysis of highly multivariate data sets ( 2 signals). Using
the concept of ICA, the recorded signals can be described by a (non)linear
mixing of unknown statistically independent sources.
Model:
The basic mechanisms describing the system are known. Based on this
knowledge, the behaviour of the system can be modeled, however so