Quantitative analysis of EEG signals: Time-frequency methods and Chaos theory

Rodrigo Quian Quiroga

Institute of Physiology - Medical University Lubeck and Institute of Signal Processing - Medical University Lubeck

Aus dem Institut fur Physiologie vertreten in der Technisch-Naturwissenschaftlichen Fakultat der Medizinischen Universitat zu Lubeck durch das Institut fur Signalverarbeitung und Prozessrechentechnik Direktoren: Prof. Dr. med. Wolfgang Jelkmann (Institut fur Physiologie) Prof. Dr.-Ing. Til Aach (Institut fur Signalverarbeitung und Prozessrechentechnik)

Quantitative Analyse von EEG-Signalen: Zeit-Frequenz-Methoden und Chaos-Theorie

Inauguraldissertation zur Erlangung der Doktorwurde der Medizinischen Universitat zu Lubeck - Aus der Technisch-Naturwissenschaftlichen Fakultat Vorgelegt von

Rodrigo Quian Quiroga

aus Buenos Aires, Argentinien Lubeck, 1998

1. Berichterstatter: Prof. Dr.-Ing. Til Aach 2. Berichterstatter: Prof. Dr.-Ing Erol Basar Tag der mundlichen Prufung: 04.12.98 Zum Druck genehmigt, Lubeck, den 18.05.1999 gez. Prof. Dr.-Ing. Erik Maehle. - Dekan der Technisch-Naturwissenschaftlichen Fakultat -

ii

Zusammenfassung
Seitdem 1929 die ersten EEGs von Menschen abgeleitet wurden, hat sich das EEG zu einem der wichtigsten diagnostischen Hilfsmittel in der klinischen Neurophysiologie entwickelt. Bis jetzt beruht die EEG-Analyse jedoch weitgehend auf der visuellen Inspektion der EEG-Aufzeichnungen. Dieses Auswertungsverfahren ist sehr subjektiv und und erschwert statistische Auswertung und Standardisierung. Daher wurden mehrere Methoden vorgeschlagen, um die im EEG enthaltene Information quantitativ zu erfassen. Unter diesen Verfahren hat sich die Fourier-Transformation als ein sehr nutzliches Hilfsmittel erwiesen. Diese kann die Frequenzkomponenten des EEG-Signals charakterisieren und hat klinische Bedeutung erlangt. Die Fourier-Transformation hat jedoch einige Nachteile, die ihre Anwendung einschranken. Daher sind andere Methoden erforderlich, um verborgene Information aus dem EEG zu gewinnen. In dieser Arbeit habe ich Methoden zur Analyse verschiedener Arten von EEGSignalen beschrieben, erweitert und vergliechen, und zwar (1) Zeit-Frequenz-Methoden (Gabor- und Wavelet-Transformation) und (2) Methoden der Chaos-Analyse (AttraktorRekonstruktion, Korrelations-Dimension, Lyapunov-Exponent). Diese Methoden lieferten hinsichtlich der Quellen und der Dynamik von Grand-MalAnfallen neue Information, die mit konventionellen Methoden schwierig zu erhalten war. Wahrend Grand-Mal-Anfallen herrschten alpha (8 ; 15Hz) und theta (4 ; 7Hz) Rhythmen vor, die spater langsamer wurden, was in Beziehung zum Beginn der klonischen Phase stand. Die Dynamik der Gehirn-Oszillationen in dieser Phase ist von Interesse im Hinblick auf Prozesse neuronaler Ermudung, auf ein Ungleichgewicht der Neurotransmitter, auf Ahnlichkeit mit Tierversuchen und auf Computer-Simulationen. Analysen mithilfe der Chaos-Theorie zeigten, da Parameter wie die Korrelations-Dimension oder der Lyapunov-Exponent abnahmen (diese Parameter charakterisieren die Komplexitat und die Chaotizitat des Signals). Diese Ergebnisse zeigten einen Ubergang zu einem einfacheren System im Verlauf des epileptischen Anfalls. Um grundlegende Eigenschaften von Gehirn-Oszillationen zu untersuchen, habe ich ereigniskorrelierte Potentiale (also Veranderungen des EEG aufgrund externer oder interner Reize) analysiert, und zwar mit neueren Methoden der Zeit-Frequenz-Analyse. In diesem Zusammenhang zeigte die Untersuchung ereigniskorrelierter Alpha-Oszillationen (also der Alpha-Komponenten ereigniskorrelierter Potentiale) eine topographische Verteilung mit signi kanten Latenz-Unterschieden zwischen anterioren und posterioren Elektroden. Dies legte nahe, da diese ereigniskorrelierten Alpha-Oszillationen an multiplen Orten entstehen. Ferner wiesen (a) die Unabhangigkeit der Alpha-Antworten von der Bearbeitung einer kognitiven Aufgabe, (b) das deutlichste Auftreten dieser Antworten iii

an okzipitalen Positionen und (c) die kurze Latenz dieser Antworten auf eine Beziehung zwischen ereigniskorrelierten Alpha-Oszillationen und primar-sensorischer Verarbeitung hin. Die Untersuchung von Antworten auf bimodale Reize (simultane auditorische und visuelle Stimulation) zeigte eine signi kante Zunahme der Amplitude im Vergleich mit unimodalen Reizen. Demzufolge war es moglich, eine Beziehung zwischen Gamma (30 ; 60Hz) Oszillationen und einem Proze anzunehmen, der die Information tragt, da zwei sensorische Wahrnehmungen (im Rahmen der bimodalen Stimulation) zu ein und demselben Reiz gehoren. Insbesondere ist diese Arbeit die erste Untersuchung, in der die neue Methode Wavelet-Entropie fur die Analyse ereigniskorrelierter Potentiale angepa t und angewendet wurde. In ereigniskorrelierten Potentialen gehen signi kante Abnahmen der Wavelet-Entropie mit einer kognitiven P300-Antwort einher. Dies zeigte, da diese P300-Antwort mit einer Ordnung der spontanen EEG-Oszillationen assoziert ist.

iv

Quantitative analysis of EEG signals: Time-frequency methods and Chaos theory

Rodrigo Quian Quiroga

Institute of Physiology - Medical University Lubeck and Institute of Signal Processing - Medical University Lubeck 1998

vi

To my family: Mama, Consuelo, Huguito and Elisa and to my closest friends: Esteban and Samy.

vii

viii

Preface
In this work, I will describe and extend two new approaches that started to be applied to physiological signals: 1) the time-frequency methods, and 2) the methods based on Chaos theory. I will discuss their applicability and usefulness mainly in two types of brain signals: a) EEG recordings from \Grand Mal" epileptic seizures, and b) Eventrelated potentials. Moreover, I will compare all these new methods, comparison which was not performed so far, stressing their advantages over conventional approaches in the analysis of EEG signals. Furthermore, the results obtained will be closely linked with physiological interpretations. In particular, this thesis is the rst work where the novel method \Wavelet entropy" is adjusted and applied to the analysis of evoked responses. The structure is as follows: The rst part of the thesis consists in an introduction to basic concepts of electroencephalography and a review of previous approaches to its quantitative analysis. In particular, chapter x1 gives a brief description of the necessary background of neurophysiology focusing on the concepts needed for understanding the basics of brain signals, and chapter x2 describes the traditional Fourier analysis and its main applications to EEGs. Chapters x3 to x6 are the main part of the thesis, each chapter referring to one of the new quantitative methods. They all have the same internal structure: 1) they start with an introduction in which the goal of the method is described, 2) then, a theoretical background is given, 3) their application to di erent types of EEG is shown and nally, 4) a physiological interpretation of the results is given and advantages of the methods are discussed in comparison with other approaches. More speci cally, chapter x3 presents the Gabor Transform, a time-frequency method that solves some of the disadvantages of the Fourier Transform. Furthermore, since in many cases a more detailed information is required, as I will show with the study of Grand Mal seizures, I will introduce new de nitions that will allow a better quantitative analysis of the EEG. Chapter x4 describes the theoretical background of the Wavelet Transform. Studies where the Wavelet Transform is applied to Tonic-Clonic seizures and to eventrelated potentials will show the advantages of this new method in the analysis of EEG signals. Chapter x5 presents the approach based on the Non-linear Dynamics (Chaos) theory. I will show its application to di erent type of seizure recordings, correlating ix

these results with the ones obtained with the methods described in the previous chapters. I will remark several problems in the implementation of these methods in the analysis of physiological signals that in many cases lead to pitfalls and misinterpretations. Furthermore, I will establish some criteria for the analysis of EEG signals with Chaos methods. Chapter x6 introduces a new method based on the \information theory", the Wavelet-entropy, that gives quantitative information about the ordered/disordered nature of the EEG signals. I will show its application to event-related potentials. Furthermore, I will show how it avoids several disadvantages of Chaos methods allowing the study of similar concepts with a completely di erent approach. Finally, in chapter x7, I will compare the time-frequency and Chaos approaches, and I will discuss the main physiological results by joining the evidence obtained with the di erent methods. Acknowledgments This work was supported by the Bundesministerium fur Bildung und Forschung (BMBF), Germany and by the Medical University of Lubeck, Germany. I am very thankful to Prof. Erol Basar, director of the group of Neurophysiology of the Medical University of Lubeck, for giving me the opportunity to work under his direction and for his experienced advice in the development of this work. I am also very thankful to Prof. Til Aach for his criticisms and corrections to this thesis, especially in the mathematical formalisms, and to Prof. Rupert Lasser for his guiding in the mathematical background during the rst stage of my work. I would like to thanks Dr. Martin Schurmann for two years of invaluable scienti c discussions, non-scienti c activities and for his criticisms and comments after a careful reading of this thesis. I am also very thankful to Dr. Juliana Yordanova and Dr. Vasil Kolev for very helpful criticisms during the development of this work and for their warm friendship. During my staying in Lubeck I also appreciated very much the collaboration with Dr. Atsuko Schutt, Dr. Irina Maltseva, Oliver Sakowitz, Dr. Richard RascherFriesenhausen and Dr. Tamer Demiralp to whom I am also very thankful for software implementation. I am also very thankful to Prof. Wolfgang Jelkmann, director of the Institute of Physiology for giving me the opportunity to work at his institute. This thesis would have not been achieved without the help of the group of neurophysiology in Lubeck. I would also like to mention the very nice work atmosphere that they created. My special thanks to Dipl.-Ing. Martin Gehrmann, Dipl.-Ing. Ferdinand Greitschus, Gabriele Huck, Betina Stier and Gabriela Fletschinger. I am especially thankful to Beate Nurnberg for her constant support and personal help. x

I would certainly like to remember all my colleagues/friends from Argentina. A very special thanks to Dr. Osvaldo Rosso and Dr. Susana Blanco, from the Chaos and Biology group at the University of Buenos Aires, for giving me the rst push in my steps as a physicist and also for their friendship and constant scienti c and non-scienti c support. I also appreciated further collaboration with Alejandra Figliola of the same group. I am very thankful to Dr. Adrian Rabinowicz, director of the Epilepsy department of the Institute of Neurological Investigations (FLENI) for teaching me what I know about epilepsy. I can not forget all the support and constant good mood of the people of the Neurophysiology department at FLENI foundation, with whom I had the pleasure to work with during my research stage in Argentina. Many thanks to Isabel, Jorge, Claudia, Sonia, Cecilia, Sandra, Alexandra, Mary, Monica, Dr. Ribero, Dr. Estelles, Dr. Nogues, Dr. Camarotta, Dr. Navarro Correa and I hope I am not forgetting somebody. Finally I would like to thank the one who introduce me in this fascinating world of Neurophysiology, the one who encouraged and supported a young student of physics coming up with crazy ideas about Chaos and EEGs. My very special thanks to Dr. Horacio Garc a.

xi

xii

Contents
Zusammenfassung Preface
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iii ix
x

Summary 1 Outline of Neurophysiology: Brain signals

1.1 Electroencephalogram (EEG) . . 1.1.1 Brain oscillations . . . . . 1.1.2 EEG in Epilepsy . . . . . 1.2 Event related potentials (ERP) . 1.3 Relation between EEG and ERP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1
. 3 . 6 . 7 . 8 . 11 . . . . . . . . . . . . . . . . . . . .

2 Fourier Transform

2.1 Introduction . . . . . . . . . . . . . . . . . . . 2.2 Theoretical background . . . . . . . . . . . . . 2.3 Fourier Transform in EEG analysis . . . . . . 2.3.1 Frequency bands . . . . . . . . . . . . 2.3.2 Topographical mapping . . . . . . . . . 2.3.3 Frequency analysis of evoked responses 2.3.4 Coherence . . . . . . . . . . . . . . . . 2.4 Conclusion . . . . . . . . . . . . . . . . . . . .

13

13 13 15 16 16 18 18 20 21 22 25 25 25 26 30 30 30 31 31 32

3 Gabor Transform (Short Time Fourier Transform)

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Theoretical background . . . . . . . . . . . . . . . . . . . . Uncertainty Principle . . . . . . . . . . . . . . . . . . . 3.3 Application to intracranially recorded tonic-clonic seizures 3.3.1 Methods and Materials . . . . . . . . . . . . . . . . 3.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . 3.4 Application to scalp recorded tonic-clonic seizures . . . . . 3.4.1 Methods and Materials . . . . . . . . . . . . . . . . Statistical analysis: plateau criteria . . . . . . . . 3.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . xiii

21

3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4 Wavelet Transform

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Continuous Wavelet Transform . . . . . . . . . . . . . . . . . . . 4.2.2 Dyadic Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . 4.2.3 Multiresolution Analysis . . . . . . . . . . . . . . . . . . . . . . . 4.2.4 B-Splines wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.5 Wavelet Packets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Short review of wavelets applied to the study of EEG signals . . . . . . . 4.4 Application to scalp recorded tonic-clonic seizures . . . . . . . . . . . . . 4.4.1 Material and Methods . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5 Application to alpha responses of visual event-related potentials . . . . . 4.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.2 Material and Methods . . . . . . . . . . . . . . . . . . . . . . . . Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . Comparison between wavelets and conventional digital ltering . 4.5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6 Application to gamma responses of bisensory event-related potentials . . 4.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.2 Material and Methods . . . . . . . . . . . . . . . . . . . . . . . . Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1 Basic concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.2 Correlation Dimension . . . . . . . . . . . . . . . . . . . . . . 5.2.3 Calculation of the Correlation Dimension . . . . . . . . . . . . 5.2.4 Problems arising when calculating the Correlation Dimension 5.2.5 Lyapunov Exponents and Kolmogorov Entropy . . . . . . . . xiv . . . . . . . . . . . . . .

36

36 38 38 38 39 40 43 44 46 46 46 50 50 50 51 52 52 54 59 61 61 61 62 62 65 66 68 68 68 69 70 71 72

5 Deterministic Chaos

68

5.3 5.4 5.5 5.6 5.7

5.2.6 Calculating Lyapunov Exponents . . . . . . . . . . Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . Short review of Chaos analysis of EEG signals . . . . . . . 5.4.1 Correlation Dimension . . . . . . . . . . . . . . . . 5.4.2 Lyapunov Exponents . . . . . . . . . . . . . . . . . Application to scalp recorded EEGs . . . . . . . . . . . . . 5.5.1 Material and Methods . . . . . . . . . . . . . . . . 5.5.2 Results and Discussion . . . . . . . . . . . . . . . . Application to intracranially recorded tonic-clonic seizures 5.6.1 Material and Methods . . . . . . . . . . . . . . . . 5.6.2 Results and Discussion . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

73 74 75 75 77 79 79 79 82 82 82 84

6 Wavelet-entropy

6.1 Introduction . . . . . . . . . . . . . . . . . . . 6.2 Theoretical Background . . . . . . . . . . . . 6.3 Application to visual event-related potentials . 6.3.1 Methods and Materials . . . . . . . . . Statistical analysis . . . . . . . . . . 6.3.2 Results . . . . . . . . . . . . . . . . . . 6.3.3 Discussion . . . . . . . . . . . . . . . . 6.4 Conclusions . . . . . . . . . . . . . . . . . . .

86

86 88 89 89 90 90 94 99

7 General Discussion

7.1 Physiological considerations . . . . . . . . . . . . . . . . . . . . . . . 7.1.1 Dynamics of Grand Mal seizures . . . . . . . . . . . . . . . . . 7.1.2 Event-related responses . . . . . . . . . . . . . . . . . . . . . . 7.1.3 Are EEG signals chaos or noise? . . . . . . . . . . . . . . . . . 7.2 Comparison of the methods . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Fourier Transform vs. Gabor Transform . . . . . . . . . . . . 7.2.2 Gabor Transform vs. Wavelet Transform . . . . . . . . . . . . 7.2.3 Wavelet Transform vs. conventional digital ltering . . . . . . 7.2.4 Chaos analysis vs. time-frequency methods (Gabor, Wavelets) 7.2.5 Wavelet-entropy vs. frequency analysis . . . . . . . . . . . . . 7.2.6 Wavelet-Entropy vs. Chaos analysis . . . . . . . . . . . . . . . xv

101

101 101 102 103 104 104 105 107 108 108 109

Conclusion

110

A Time-frequency resolution and the Uncertainty Principle

A.1 Preliminary concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 A.2 Uncertainty Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 A.3 Time-frequency resolution of the Fourier, Gabor and Wavelet Transform 113

111

References Biographical sketch

117 129

xvi

Summary
Since the rsts recordings in humans performed in 1929, the EEG has become one of the most important diagnostic tools in clinical neurophysiology, but up to now, EEG analysis still relies mostly on its visual inspection. Due to the fact that visual inspection is very subjective and hardly allows any statistical analysis or standardization, several methods were proposed in order to quantify the information of the EEG. Among these, the Fourier Transform emerged as a very powerful tool capable of characterizing the frequency components of EEG signals, even reaching diagnostical importance. However, Fourier Transform has some disadvantages that limit its applicability and therefore, other methods for extracting \hidden" information from the EEG are necessary. In this work, I described, extended and compared methods of analysis of di erent types of EEG signals, namely time-frequency methods (Gabor Transform and Wavelet Transform) and Chaos methods (attractor reconstruction, Correlation dimension, Lyapunov exponents, etc.). Time-frequency methods provided new information about sources and dynamics of Grand Mal (Tonic-clonic) seizures, something very di cult to obtain with conventional methods. Grand Mal seizures were rst dominated by alpha (7:5 ; 12:5Hz) and theta (3:5;7:5Hz) rhythms, these rhythms later becoming slower in correlation with the starting of the clonic phase. The dynamics of the frequency patterns during these seizures was very interesting in relation to processes of neuronal fatigue, neurotransmitter disbalance, similarity with animal experiments and computer simulations. The analysis with Chaos theory showed a decrease in parameters as the Correlation Dimension or the maximum Lyapunov exponent, parameters that characterize the complexity and \chaoticity" of the signal. These results showed a transition to a more simple system during epileptic seizures. In order to study basic features of brain oscillations, I analyzed event-related responses (i.e. alterations of the ongoing EEG due to an external or internal stimuli) with recent methods of time-frequency analysis. In this context, the study of event-related alpha oscillations (i.e. event-related responses in the alpha range) showed that these responses were distributed along the scalp with signi cant di erences in their delays between anterior and posterior electrodes. This result implied that several sources were involved in the origin of the event-related alpha oscillations. Furthermore, their independence on the performance of a cognitive task, the best de nition in occipital locations and the short latency of the responses pointed towards a relation between event-related alpha oscillations and primary sensory processing. The study of the responses upon bimodal stimulation (simultaneous visual and audi1

tory stimulation) showed a signi cant increase of amplitude in comparison with the unimodal ones. Then, it was possible to conjecture a relation between gamma (30 ; 60Hz) oscillations and a process responsible of carrying the information that two sensory perceptions of bimodal stimulation correspond in fact to the same stimulus. In particular, this thesis is the rst work where the novel method \Wavelet entropy", a measure of the distribution of the signal in the frequency domain, was adjusted and applied to the analysis of event-related responses. In event-related potentials, signi cant decreases in the wavelet entropy correlated with the P300 cognitive response showed that this response was associated with an ordering of the spontaneous EEG oscillations.

1 Outline of Neurophysiology: Brain signals

This chapter presents some basic topics of neurophysiology necessary for understanding the experiments and results to be described in the following chapters. In this context, the concepts exposed and the detail of their treatment are not expected to provide a complete background on neurophysiology. On the contrary, this chapter is focused on describing the electroencephalogram and event-related potentials (ERPs), especially applied to the study of epilepsy and brain oscillations. Despite the wide application of these issues, some fundamental points are still controversial and due to the complex behavior of these signals, they are di cult to resolve with traditional approaches, thus being ideal candidates to be studied with new quantitative methods.

1.1 Electroencephalogram (EEG)

The EEG was originally developed as a method for investigating mental processes. Clinical applications soon became visible, most notably in epilepsy, and it was only with the introduction of ERP recordings that EEG correlates of sensory and cognitive processes nally became popular. The rst recordings of brain electrical activity were reported by Caton in 1875 in exposed brains of rabbits and monkeys, but it was not until 1929 that Hans Berger (Berger, 1929) reported the rst measurement of brain electrical activity in humans. EEG visual patterns were correlated with functions, dysfunctions and diseases of the central nervous system, then emerging as one of the most important diagnostical tools of neurophysiology. The electroencephalogram (EEG) can be roughly de ned as the mean electrical activity of the brain in di erent sites of the head. More speci cally, it is the sum of the extracellular current ows of a large group of neurons. Since the generation of the EEG from the action potentials of the neurons is beyond the scope of this thesis, for further details I suggest the comprehensive works of Steriade et al. (1990), Lopes da Silva (1991), Steriade (1993), Speckermann and Elger (1993), Pedley and Traub (1990) and Basar (1980). EEG recordings are achieved by placing electrodes of high conductivity (impedance < 5000 ) in di erent locations of the head. Measures of the electric potentials can be recorded between pairs of active electrodes (bipolar recordings) or with respect to a supposed passive electrode called reference (monopolar recordings). These measures are mainly performed on the surface of the head (scalp EEG) or by using special electrodes placed in the brain after a surgical operation (intracranial EEG).

Figure 1: 10-20 system of montage of electrodes. Notation: F means frontal, C means central, P means parietal, T means temporal, O means occipital and A means earlobe reference. Modi ed from Reilly, 1993. Scalp recordings In scalp, or surface, EEG recordings, the most widely used placement of electrodes is the so called 10;20 system, consisting in 20 electrodes (or sometimes less) uniformly distributed along the head, generally referenced to 2 electrodes in the earlobes (Reilly, 1993 see g. 1). Normal scalp EEG recordings are usually taken with the subject relaxed, in three modalities: 1) with eyes closed 2) with eyes open 3) under hyperventilation, photostimulation, etc. One important problem of scalp EEGs are the artifacts (Reilly, 1993). Artifacts are alterations of the EEG due to head movements, blinking, muscle activity, electrocardiogram, etc. (see an example in g. 2). Due to the very low amplitude of EEG signals (usually between 10 ; 200 V ), artifacts often contaminate the recordings restricting or making impossible any analysis or interpretation. Intracranial recordings With intracranially implanted electrodes a better resolution can be achieved. Moreover, possible alterations of the EEG signal in its way from the originating neural struc4

Figure 2: Normal 20 channels (10-20 system) eyes open EEG. Vertical lines show one second divisions. The arrow marks the occurrence of a blinking artifact (best visible in frontal electrodes). Note also the appearance of alpha waves ( 10cycles=sec:), best de ned in parietal and occipital locations

tures up to the scalp are overcome and most of the artifacts can be avoided. There are two main types of intracranial electrodes: 1) the subdural electrodes, which are arranged in grids or strips placed on the brain (Arroyo et al., 1993), and 2) the deep electrodes, which are arranged in a needle in order to study deep structures (Niedermeyer, 1993b). The application of intracranial electrodes involves a major surgery and their use is restricted to very particular cases, as for example the study of epileptic patients who are candidates for a surgical resection.

1.1.1 Brain oscillations

In his rst publication, Berger (1929) already mentioned the presence of what he called alpha and beta oscillations. The study of di erent types of rhythmicities of the brain and their relation with di erent pathologies and functions keep the attention of researchers since the beginnings of electroencephalography. Brain oscillations were divided in frequency bands that have been related with di erent brain states, functions or pathologies (see Niedermeyer, 1993a Steriade, 1993). during wakefulness, under relaxation and mental inactivity conditions. They are best seen with eyes closed and most pronounced in occipital locations (see g. 2 sources and functions of alpha band will be discussed in sec. x4.5).

Alpha rhythms (7:5 ; 12:5Hz): they appear spontaneously in normal adults

Beta rhythms (12:5 ; 30Hz): they are best de ned in central and frontal locations, they have less amplitude than alpha waves and they are enhanced upon expectancy states or tension. They are traditionally subdivided in 1 and 2 oscillations.

important role in infancy and childhood. In the awake adult, high theta activity is considered abnormal and it is related with di erent brain disorders.

Theta rhythms (3:5 ; 7:5Hz): they are enhanced during sleep and they play an

Delta rhythms (0:5 ; 3:5Hz): they are characteristic of deep sleep stages. Furthermore, delta oscillations with certain speci c morphologies, localizations and rhythmicities are correlated with di erent pathologies.

to the surface EEG, they gained popularity after the cellular level experiments of Gray and Singer (1989) and Gray et al. (1989) showing their relation with linking of stimulus features into a common perceptual information (binding theory). Gamma oscillations will be further studied in sec. x4.6. 6

Gamma rhythms (30 ; 60Hz): previously not of major interest with regard

1.1.2 EEG in Epilepsy

One important application of the EEG is to the study of epilepsy. The appearance of abrupt high amplitude peaks (spikes), abnormal rhythmicities, \slowing" of the recording and several other paroxysms are a general landmark of it, helping to identify, classify and localize the seizures. Epilepsy is a disorder of the normal brain function that a ects about 1% of the population, characterized by an excessive and uncontrolled activity of either a part or the whole central nervous system. Historically, the EEG has been the most useful tool for its evaluation, now complemented with the advances in imaging techniques, especially the ones achieved by the Magnetic Resonance Imaging (MRI). Given that ictal recordings (i.e. recordings during a seizure) were rarely obtained, EEG analysis of epileptic patients usually relied on interictal ndings. In those interictal EEGs, seizures are usually activated with photostimulation, hyperventilation and other methods. However, one disadvantage of these stimulation techniques is that provoked seizures do not necessarily have the same behavior as the spontaneous ones. The introduction of long term Video-EEG recordings has been an important milestone providing not only the possibility to analyze ictal events, but also contributing with valuable information in those candidates evaluated for epilepsy surgery (see Quian Quiroga et al., 1997a Porter and Sato, 1993 Kaplan and Leser, 1990 Gotman et al., 1985 Meierkord et al., 1991). In this setting and following strict protocols, seizures are elicited by gradually reducing antiepileptic drugs. Classi cation The classi cation of epileptic seizures is a very di cult task and leaded to several controversies. I will present a simpli ed classi cation (for further details see Niedermeyer, 1993c). Tonic-Clonic seizures will be described with more detail since these are the ones to be studied in the following chapters. Seizures can be classi ed in: 1. Partial seizures: they have a focal origin they can also evolve to a generalized seizure (secondarily generalized). Simple partial seizures: consciousness is not impaired. Depending on the zone of the brain involved, they are characterized by focal motor movements, sensory symptoms (simple hallucinations), autonomic symptoms (epigastric sensation, sweating, pupillary dilatation, etc.) or psychic symptoms. Complex partial seizures: involve a loose of consciousness. They are characterized by di erent psychical sensations and motor automatisms. Since there are many types of these seizures, the EEG is very variable but it can be gen7

erally characterized by low frequency discharges (3 ; 6 Hz) localized in some region of the brain. They can begin as simple partial seizures. 2. Generalized seizures: the whole brain is involved. Absence seizures (Petit Mal epilepsy): Consist of a sudden lapse of consciousness with impairment of mental functions. They last about 5-20 seconds and the EEG is characterized by generalized 3Hz spike-wave discharges (a spike followed by a slow half period oscillation). Tonic-clonic seizures (Grand Mal epilepsy): described below in more detail. Other type of generalized seizures are the myoclonic, clonic and atonic ones. Tonic-clonic seizures A tonic-clonic (Grand Mal) seizure normally lasts about 40 ; 90 sec and it is characterized by violent muscle contractions. An initial tonic phase with massive spasms (i.e. extreme muscular tension without movement) is supplanted some seconds later by the clonic phase with violent exor movements and characteristic rhythmic contractions of the entire body until the ending of the seizure (Niedermeyer, 1993c). During these seizures consciousness is impaired. Visual inspection of scalp recordings of these events is often troublesome because the signal is obscured by muscle artifacts (see g. 3). In most cases the analysis is con ned to the interpretation of electrical abnormalities that either precede or follow the tonic-clonic activity, thus neglecting the ictal phase (see sec. x3.4). Therefore, since these seizures are very di cult to study with traditional approaches, there will be one of the main type of signals to be studied with the methods described in the following chapters.

1.2 Event related potentials (ERP)

EEG activity is present in a spontaneous way or can be generated as a response to an external stimulation (e.g. tone or light ash) or internal stimulation (e.g. omission of an expected stimulus). The alteration of the ongoing EEG due to these stimuli is called event related potential (ERP), in the case of external stimulation also called evoked potential (EP). Owing to the low amplitude of the ERPs in comparison with the ongoing EEG, it is a common practice to average several responses in order to visualize the evoked activity. The reason for averaging is that ERPs have a de nite latent period determined from the stimulation time, and that they have a similar pattern of response which is more or less predictable under similar conditions (Basar, 1980). 8

Figure 3: Scalp EEG of 18 channels during a Tonic-Clonic seizure. Note how the artifacts obscure the recording completely. Vertical lines show one second divisions.

Modalities and paradigms There are mainly three modalities of stimulation (for more details see Regan, 1989): 1. Auditory: stimuli are single tones of a determined frequency, or clicks with a broadband frequency distribution. 2. Visual: stimuli are produced by a single light or by the reversal of a pattern as for example a checkerboard. 3. Somatosensory: stimuli are elicited by electrical stimulation of peripheral nerves. The rst two modalities can be combined in what is called bimodal stimulation. Bimodal stimulation will be studied in sec. x4.6. Sequence of stimuli are arranged in paradigms in order to study the responses to di erent tasks. The most widely used paradigms are: 1. No-task evoked potentials: subjects are relaxed and instructed to perceive the stimuli without performing any task. 2. Oddball paradigm: two di erent stimuli are presented in a pseudorandom order. A frequent NON-TARGET stimulus is presented in 75% of the trials, and a deviant stimuli called TARGET in the remaining 25% of the trials. Subjects are instructed to pay attention to the appearance of the TARGET stimuli. Transient responses: description of P100 and P300 peaks. Responses can be classi ed in: 1) transient (occurring after the delivery of a shortlasting stimulus, with interstimulus intervals greater than the responses), 2) driven (generated with rapidly repeating stimuli), or 3) sustained (occurring throughout the duration of a continuous stimulus). In the following we will only study transient responses. As an example, gure 4 shows the averaged transient responses to pattern visual ERPs generated with an oddball paradigm. One second pre- and one second post-stimulation EEG recorded from the left occipital electrode are plotted. A positive peak at about 100ms after stimulation (P100) can be observed upon NON-TARGET and TARGET stimulation. Since this peak does not depend on the task, it has a relatively short latency (100ms) and it is best de ned in the primary sensory visual cortex (occipital locations), it is mostly related with primary sensory processing (Regan, 1989). Another peak at about 450ms after stimulation (P300) appears only upon TARGET stimulation. Since this response is task dependent and has long latency, it is traditionally related to cognitive processes as signal matching, decision making, attention, memory updating, etc. (Polich and Kok, 1995 Regan, 1989 Basar-Eroglu et al., 1992). 10

Figure 4: Averaged responses upon NON TARGET and TARGET stimuli of a left occipital electrode in pattern visual evoked potentials

1.3 Relation between EEG and ERP

In contrast with the conventional view of ERPs as signals being added to the noisy EEG, Basar (1980) pointed out that ERPs might arise from the ongoing EEG by means of a resonance process. Resonance is a well known phenomenon of physics: if a driving is applied to a system, for example a pendulum, the system will show very large and increasing outputs if the driving frequency is similar to the natural frequency of the system (for a physical description see Feynman et al., 1964). Basar (1980) introduced the following concept for understanding the relation between EEG and ERP: In the EEG several oscillations are occurring at the same time in a non-synchronized way, and when a stimulus is applied, some of these frequencies can be enhanced by a resonance phenomenon. Furthermore, it was assumed that these di erent enhanced oscillations are related to transmitting information through the brain, having di erent \meanings" and \functions" (see sec. x2.3.3). According to a classi cation of Galambos (1992) event related oscillations can be evoked (originated by and time locked to the stimulus), induced (originated by, but not time locked to the stimulus) or emitted (originated by an internal process rather than by the stimulus). 11

I will summarize here some of the experimental evidence supporting this view (for more details see Basar, 1980, 1998, 1999 Basar et al., 1999): 1. In three year old children, neither spontaneous, nor evoked 10Hz activity is obtained (Kolev et al, 1994, Basar-Eroglu et al., 1994). Then, according to the resonance theory, in this case alpha rhythm does not belong to the natural brain frequencies. In other words, alpha oscillations can not be evoked because there is no spontaneous alpha activity. 2. High amplitude pre-stimulus 10Hz activity reduces the evoked potential amplitudes (Basar, 1980 Basar and Stampfer, 1985 Basar et al., 1992), thus showing an inverse relation between the pre-stimulus EEG and the evoked responses. An in uence of pre-stimulus theta EEG was also reported on visual evoked potentials (Basar et al., 1998). Then, since the background EEG in uences the evoked responses, it can not be merely considered as additive noise. 3. Converging results support the view that evoked or induced oscillations can be correlated, at least partially, with di erent functions, as it will be described in sec. x2.3.3, thus showing that the frequency analysis of the evoked responses is not arbitrary. In conclusion, following the resonance theory, the ERP can be seen as evoked synchronization, frequency stabilization, frequency selective enhancement and/or phase reordering of the ongoing EEG, and it should not be interpreted as an additive component to a noisy background EEG.

12

2 Fourier Transform
2.1 Introduction
Time signals can be represented in many di erent ways depending on the interest in visualizing certain given characteristics. Among these, the frequency representation is the most powerful and standard one. The main advantage over the time representation is that it allows a clear visualization of the periodicities of the signal, in many cases helping to understand underlying physical phenomena. Frequency analysis, developed by Jean Baptiste Fourier (1768-1830), reached innumerable applications in mathematics, physics and natural sciences. Furthermore, the Fourier Transform is computationally very attractive since it can be calculated by using an extremely e cient algorithm called the Fast Fourier Transform (FFT Cooley and Tukey, 1965). This chapter starts with a brief mathematical background of Fourier Transform and then, applications of Fourier analysis to EEG signals will be reviewed. Four main applications will be described: analysis of frequency bands, topographical mapping, analysis of evoked responses and coherence.

2.2 Theoretical background

Under mild conditions, the Fourier Transform describes a signal x(t) (I will consider only real signals) as a linear superposition of sines and cosines characterized by their frequency f .

x(t) =
where

+1
;1

X (f ) ei2 x(t) e;i2

ft ft

df dt

(1) (2)

X (f ) =

+1
;1

are complex valued coe cients that give the relative contribution of each frequency f . Equation 2 is the continuous Fourier Transform of the signal x(t). It can be seen as an inner product of the signal x(t) with the complex sinusoidal mother functions e;i2 ft , i.e.

X (f ) =< x(t) e;i2 ft > (3) Its inverse transform is given by eq. 1 and since the mother functions e;i2 ft are orthogonal, the Fourier Transform is nonredundant and unique. Let us consider in the following that the signal consists of N discrete values, sampled every time .
13

x(n) = fx0 x1 : : : xN ;1 g = fxj g where xj is the measurement taken at a time tj = t0 + j . The discrete Fourier Transform of this signal is de ned as: X (k) =
and its inverse as:
N ;1 X n=0

(4)

x(n) e;i2

kn=N

k = 0 ::: N ; 1

(5)

;1 X 1N x(n) = N X (k) ei2

k=0

kn=N

(6)

where, since we are considering real signals, the following relation holds: X (k) = X (N ; k), and the discrete frequencies are de ned as:

fk = Nk (7) Note that the discrete Fourier Transform gives N=2 independent complex coe cients, thus giving a total of N values as in the original signal and therefore being nonredundant. Clearly, the frequency resolution will be f = N1
(8)

Aliasing The frequency fN = 21 , corresponding to k = N 2 in eq. 7, is called the Nyquist frequency and it is the highest frequency that can be detected with a sampling period . If the signal x(n) has frequencies above the Nyquist frequency, these ones will be processed just as though they are in the range fk < fN , thus giving a spurious e ect called aliasing (Jenkins and Watts, 1968 Weaver, 1989). The greater this excess is, the greater the errors and the less adequate is the sampling rate for representing the signal. From the complex coe cients of eq. 5, the periodogram1 can be obtained as:
1 Due to the high complexity of the EEG signal, it will be convenient to consider it as a realization of a stochastic process (Lopes da Silva, 1993a Dummermuth and Molinari, 1987). However, it should be stressed that this is a mathematical and not a biophysical model. On the other hand, in section x5, the EEG will be considered as a realization of a non-linear deterministic \chaotic" system. The deterministic/stochastic nature of the EEG will be discussed in sections x5.7 and x7.1.3. Regarding the EEG as a random signal, the periodograms can be interpreted as estimators of the power spectrum.

14

Ixx(k) = jX (k)j2 = X (k) X (k)

(9)

where denotes complex conjugation. The sample cross spectrum can be similarly de ned by making the product of eq. 9 between two time series x(n) and y(n) as follows

Ixy (k) = X (k) Y (k)

(10)

where X (k) and Y (k) are the Fourier Transforms of x(t) and y(t) respectively. The sample cross spectrum gives a measure of the linear correlation between two signals for the di erent frequencies fk , de ned as in eq. 7. Since it is a complex measure, it can be described by an amplitude and a phase. A useful measure of the amplitude is obtained with its square value, normalized in the following way:

jIxy (k)j2 ;2 ( k ) = (11) xy Ixx(k) Iyy (k) Equation 11 is known as the squared coherence function, and it will give values tending to 1 for highly linearly correlated signals and to 0 for non correlated ones. The counterpart of the coherence is the phase function, de ned as
where = means imaginary part and < the real one. We should remark that the phase is meaningful only when the coherence is signi cant. If this is the case, gives a measure of the time delay between the two signals according to the following formula (Brazier, 1972): (13) t= 2 f k where is the phase di erence in degrees at a frequency fk (de ned as in eq. 7).

Ixy (k) = arctan ;= <I (k)

xy

(12)

2.3 Fourier Transform in EEG analysis

Up to the moment, after the introduction of digital recordings of EEGs, the spectral analysis based on the Fourier Transform is by far the most used quantitative method for the analysis of EEG signals. Moreover, it has reached importance as a diagnostical tool. Fourier Transform allows the separation and study of di erent EEG rhythms, a task di cult to perform visually when several rhythms occur simultaneously. 15

2.3.1 Frequency bands

Since EEG data is non-stationary and contains many artifacts, the rst step when making a Fourier analysis is to choose several representative data samples, generally of 1 or 2 seconds, free of artifacts (Nuwer et al., 1994). Before computing the Fourier Transforms, each epoch is multiplied by an appropriate windowing function (normally a Hanning window is used) in order to avoid border problems (leakage) (Jenkins and Watts, 1968 Dumermuth and Molinari, 1987). Then, the periodograms of the epochs are computed and averaged. As stated in sec. x1.1.1, it is very useful to group the frequencies in di erent bands. Several quantitative parameters were de ned from these frequency bands, as for example the relative power between them (being the most used the alpha/theta ratio), reactivity (ratio between eyes closed/eyes open alpha activity), asymmetry index ( (left power right power)/(left power + right power) ), etc. (Nuwer et al., 1994). Moreover, statistical techniques can be used in order to establish normal ranges and their deviation with several pathologies (John et al., 1987).

2.3.2 Topographical mapping

The information of the spectrograms of the di erent electrodes can be arranged in topographical maps (Gotman, 1990a Gevins, 1987 Lehmann, 1987 Lopes da Silva, 1993a). Usually, these algorithms use linear or quadratic interpolations between the 3 or 4 nearest recording sites. Since the use of single electrode references can distort the maps near the reference site (Lehmann 1971, 1987), one critical point in these plots is the election of the reference. Several suggestions were proposed in order to avoid this distortion, among them the use of averaged references and also the use of the average of derivatives of the EEG activity (Lehmann, 1987 Lopes da Silva, 1993b). One nal issue to be considered is how to project a three dimensional head into a two dimensional map. The use of topographic plots started more than 30 years ago by implementing different techniques (Walter et al., 1966 Lehman, 1971), but it was after the introduction of color topographic maps by Du y et al. (1979) that they became widely accepted, and started to be used in several medical centers with the appearance of commercial systems. With these plots is easy to visualize asymmetries and to localize frequency band activities. Furthermore, the topographic maps complement the quantitative parameters described in the previous sections in the characterization of normality and in the study and diagnosis of several pathologies (Du y, 1986 Maurer and Dierks, 1991 Pfurtscheller 16

17

Figure 5: Topographical mapping of the di erent EEG frequency bands from a normal EEG recording taken with eyes closed.

and Lopes da Silva, 1988 Garcia and Quian Quiroga, 1998). Figure 5 shows the topographic mapping of an EEG recording of a normal subject with eyes closed. Five di erent frequency bands are plotted. As expected, the power is homogeneously distributed in all frequency bands, except in alpha, where there is a simetrical increase in the posterior locations. This increase re ects the presence of the normal spontaneous alpha activity described in section 1.1.1.

2.3.3 Frequency analysis of evoked responses

Considering the proposed relation between EEG and ERP described in section x1.3, the physiological functions of the frequency bands can be analyzed by studying their response to di erent types of stimulation. Basar (1980) described a combined analysis procedure for achieving this goal. The procedure relies on the study of enhancements of di erent frequency bands as a response to stimulation. After a Fourier analysis of the pre- and post-stimulus EEG, frequency ranges are determined and bands are separated by using a selective ltering. I will summarize here some of these results (for more details see Basar, 1980): Delta enhancements were correlated with cognitive processes like signal matching or decision making by using auditory and visual oddball paradigms (Stampfer and Basar, 1985 Basar-Eroglu et al., 1992 Schurmann et al., 1995 Demiralp et al., 1999). Theta enhancements were correlated with cognitive tasks like focused attention and signal detection by using omitted stimulus paradigms in free moving cats (Basar-Eroglu et al., 1991) and in humans (Demiralp and Basar, 1992). Alpha enhancements were correlated to primary sensory processing in several experiments using cross-modality stimulation in cats and humans (Basar, 1980 Basar et al., 1992 Basar and Schurmann, 1996). Gamma enhancements were correlated with binding of features (Gray and Singer, 1989 Gray et al., 1989) and they were suggested to be involved in several other sensory and cognitive processes (Basar-Eroglu et al., 1996a,b Schurmann et al., 1997).

2.3.4 Coherence
The coherence function gives a measure of the linear correlation between two EEG electrodes. The main advantage over previous \cross-correlation" measures (see Gevins, 18

1987) is that coherence gives the correlation as a function of frequency, thus allowing the study of spatial correlations between di erent frequency bands. There is an extensive literature about the use of coherence in EEGs. This section will only show some examples and for a more complete review I suggest the work of Lopes da Silva (1993a). Lopes da Silva et al. (1973a) computed thalamo-cortical and cortico-cortical coherences for checking the validity of the assumption of a thalamic pacemaker of cortical alpha rhythms, as proposed by Andersen and Andersson (1968). They found that cortico-cortical coherences were higher than the thalamo-cortical ones, thus stressing the importance of cortical mechanisms in the spread of alpha rhythms, in disagreement with the thalamic pacemaker theory. Storm van Leeuwen et al. (1978) used coherence analysis for di erentiating alpha and mu rhythms. It is interesting to remark that since these frequencies are very similar, this is a very di cult task to achieve with normal spectrograms. Other interesting application of coherence is for the de nition of characteristic \lengths of synchrony" between neuronal populations. This is achieved by establishing lengths with signi cant coherence. Following this approach, after studying 189 children, Thatcher et al. (1986) proposed a \two-compartmental model" stating that two separate sources of EEG coherences are present: the rst one determined by short length axonal connections and the second one by long distance connections. Bullock and coworkers (1989, 1995a, 1995b) criticize the use of scalp electrodes for measuring spatial structure of synchrony and they studied the spatial structure of coherence with intracranial electrodes. In contradiction with previous reports of population synchrony of the order of several centimeters, they reported that coherence varied in the order of a few millimeters, thus showing a microstructure of the cell populations. Furthermore, due to this small range structure, they criticize the view of the EEG as a linear superposition of independent oscillators. Several attempts were performed in order to analyze epileptic seizures with coherence analysis. Gerch and Goddard (1970) used coherence for identifying the location of an epileptic focus in a cat brain. By seeing a signi cant decrease of coherence between two channels after the subtraction of the e ect of a third one (this method called partial coherence), they deduce that the third one was driving the other two, thus being the focus of the seizure. Brazier (1972), introduced the measurement of time delays between channels (see eq. 12) in order to follow routes of propagation and establish the origin of seizure activity. Gotman (1983, 1987) proposed a modi cation of this algorithm by calculating the time delay from the slope of a straight line tted to the phase spectrum when the coherence is signi cant. He showed that time delays most often indicate a lead from the side of 19

onset in di erent type of seizures in animals and humans. However, a straight line is not always possible to t in the phase spectrum and coherence values are often very low in order to extract any meaningful measure from the phase spectrum.

2.4 Conclusion
In this chapter the basic background and several applications of Fourier Transform to EEG analysis were described. One important application is the comparison between the power of di erent frequency bands and their topological distribution used to discriminate between normal subjects and pathological cases. This analysis is already adapted to many commercial systems and used in several medical centers. A word of caution should be mentioned at this point. Although quantitative parameters can be easily extracted from the EEG in an automated way, the visual inspection of the recordings should not be left aside. This helps to avoid misinterpretations due to nonstationarity, artifacts or others. At this respect, topographic mapping and quantitative values should be considered as a complement and not as a replacement of the visual inspection of the EEG. Fourier Transform can be used for studying the functions and sources of di erent brain oscillations by analyzing EEGs or ERPs. Some examples of how the measure of coherence can lead to interesting physiological interpretations were also presented. Measures of connectivity or synchrony between neural populations can be established and tested in several states and pathologies. Since an epileptic seizure is an abnormal synchronization of cell groups, coherence appears as an ideal tool for measuring it. Several works tried to use time delays calculated from phase di erences as a method for source localization. However, the measure of time delays is di cult since it depends on having high coherence values, this being not the usual case. Despite of the fact that Fourier Transform proved to be very useful in EEG analysis, I would like to mention here two main disadvantages. First, Fourier Transform requires stationarity of the signal, the EEGs being highly non-stationary. Second, with Fourier Transform the time evolution of the frequency patterns is lost. The methods to be presented in the following sections are developed in order to overcome these two limitations.

20

3 Gabor Transform (Short Time Fourier Transform)

3.1 Introduction
Since the Fourier Transform is based in comparing the signal with complex sinusoids that extend through the whole time domain, its main disadvantage is the lack of information about the time evolution of the frequencies. Then, if an alteration occurs at some time boundary, the whole Fourier spectrum will be a ected, thus also needing the requirement of stationarity. In many occasions, signals have time varying features that can not be resolved with the Fourier Transform. A traditional example is a chirp signal (i.e. a signal characterized by a well de ned but steadily rising frequency). In the case of a chirp, Fourier analysis can not de ne the instantaneous frequency because it integrates over the whole time, thus giving a broad frequency spectrum. This is partially resolved by using the Gabor Transform, also called Short-Time Fourier Transform. With this approach, Fourier Transform is applied to time-evolving windows of a few seconds of data smoothed with an appropriate function (Cohen, 1995 Chui, 1992 Qian and Chen, 1996). Then, the evolution of the frequencies can be followed and the stationarity requirement is partially satis ed by considering the signals to be stationary in the order of the window length (Lopes da Silva, 1993a). In other words, the procedure consists in breaking the signal in small time segments and then in applying a Fourier Transform to the successive segments. One application of Gabor Transform is to the analysis of Tonic-Clonic (Grand Mal) seizures (see sec. x1.1.2). As we will see in this chapter, these seizures have interesting time evolving frequency characteristics that can not be resolved with methods such as the visual inspection of the EEG or the Fourier Transform (Quian Quiroga et al., 1997b Blanco et al., 1998b). Moreover, since it is di cult to extract any information from the traditional time-frequency graphic representation of the Gabor Transform, called spectrogram, I will introduce three parameters, the relative intensity ratio and the mean and maximum band frequencies, de ned from the Gabor Transform, in order to obtain quantitative measures of the dynamics of epileptic seizures. After a brief theoretical outline, I will describe two di erent type of applications. First, the application to scalp recorded tonic-clonic seizures (Quian Quiroga et al., 1997b) and second, the application to intracranially recorded tonic-clonic seizures (Blanco et al., 1995b,1996a). 21

3.2 Theoretical background

The Gabor transform of a signal x(t) is de ned as follows:

GD (f t) =

1 ;1

x(t0 ) gD (t0 ; t) e;i2

ft0

dt0

(14)

where denotes complex conjugation. Note that GD (f t) is the same as the Fourier Transform (eq. 2) but with the introduction of the window gD (t0 ; t) of wide D and center in t. Furthermore, in analogy with the Fourier Transform, the Gabor Transform can be considered as an inner product between the signal and the complex sinusoidal functions e;i2 ft0 modulated by the window gD (compare with eq. 3), i.e. (15) The window function gD is introduced in order to localize the Fourier Transform of the signal at time t. For this reason, the window function must be peaked around t and falling o rapidly, thus emphasizing the signal in the rst case and suppressing it for distant times. Several window functions can be used for achieving this, among them, Gabor (1946) proposed the use of a Gaussian function:

GD (f t) =< x(t0 ) gD (t0 ; t) e;i2 ft0 >

e; t g (t) = (16) Since the Fourier Transform of a Gaussian is again a Gaussian, this function allows a simultaneous localization in time and frequency (see appendix xA). It should be noted that the Gaussian is described in function of a positive constant and the window function in eq. 14 was de ned in function of the window length D. This is because Gaussian functions do not have compact support (i.e. they are not 0 outside a certain boundary). However, they approach asymptotically to 0 with a rate determined by the parameter . In consequence, the Gaussian function can be truncated to have a length D, by assuming that in the borders their values are nearly 0 due to the use of a reasonable decay . Then, in the case of Gaussian windows the Gabor Transform can be explicitly de ned as a function of the parameter
2 2

1=4

(17) Let us return now to the general case of any arbitrary window function and consider the inverse transformation. Taking the inverse Fourier Transform of eq. 14 and after some algebraic manipulation we obtain:
Z 1 Z 1 1 x(t) = k g k GD (f t0 ) gD (t ; t0 ) ei2 D ;1 ;1

GD (f t) ! GD (f t)

ft

df dt0

(18)

22

1 jg (t0 )j2 dt0 . Eq. 18 is the inverse Gabor Transform and implies that where k gD k= ;1 D the original signal can be completely reconstructed from the coe cients GD (f t). Gabor Transform is highly redundant because it gives a time-frequency map from every time value of the original signal. In order to decrease redundancy, a sampled Gabor Transform can be de ned by taking discrete values of time and frequency, i.e.:
R

(19) where F and T represent the frequency and time sampling steps. Small F steps are obtained by using large windows, and small T steps are obtained by using high overlapping between successive windows. Depending on the resolution required, a proper choice of F and T will decrease the redundancy and save computational time, but the price to pay is that the reconstruction will be no longer straightforward as with eq. 18 (Qian and Chen, 1996). In the following, for convenience I will keep the notation of the continuous Gabor Transform de ned in eq. 14. The spectrum can be de ned as

GD (f t) ! GD (mF nT )

I (f t) = jGD (f t)j2 = GD (f t) GD (f t) (20) and a time-frequency representation of the signal can be obtained by using a recursive algorithm that slides the time window and plots the energy (I ) as a function of the frequency and time. These plots, called spectrograms (see g. 8), give an elegant visual description of the time evolution of the di erent frequencies, but it is di cult to extract from them any quantitative measure. In order to quantify this information, I will de ne the band power spectral intensity for each one of the EEG traditional frequency bands i (i = : : :) as: I (i) (t) =
Z

f (i) max f (i) min

I (f t) df i =

:::

(21)

where ( f (i) min f (i) max ) are the frequency limits for the band i. The division and grouping of the spectrum in frequency bands is not arbitrary, since as showed in section x1.1.1 and section x2.3.3, EEG bands are correlated with di erent sources and functions of the brain. Obviously the total power spectral intensity will be

IT (t) =

and we can de ne the band relative intensity ratio (RIR) for each frequency band i as 23

I (i) (t) i =

:::

(22)

Figure 6: Procedure for calculating the Gabor Transform and the RIR.
(i) (23) RIR(i) (t) = II ((tt)) 100 T Figure 6 illustrates the iterative procedure for calculating the RIR. First, a window centered in a time T1 is applied to the data (step A) and the Gabor Transform of this windowed data is calculated by using eq. 14 (step B). Then, the RIR is calculated and plotted (step C). Finally, the window is displaced a xed time and the procedure is repeated, obtaining a time representation of the RIR.

For further quantitative analysis, we de ne the band mean weight frequency as

< f (i) (t) > =

h Z

f (i) max

f (i) min

I (f t) f df = I (i) (t)

(24)

24

(i) (t)) of the i-band, as the frequency value for and the band maximum peak frequency (fM which I (f t) is maximum in the interval ( f (i) min f (i)max ) at a time t, i.e. (i) f (i) ) I (fM t) > I (f t) 8 f 6= fM (fmin max

(25)

Uncertainty Principle One critical limitation appears when windowing the data due to the Uncertainty Principle (Cohen, 1995 Chui, 1992 Qian and Chen, 1996 Kaiser, 1994). If the window is too narrow, the frequency resolution will be poor, and if the window is to wide, the time localization will be not so precise. Or in another words, sharp localizations in time and frequency are mutually exclusive because a frequency can not be calculated instantaneously. If we denote by t the time uncertainty (time duration) and by f the uncertainty in the frequencies (frequency bandwidth), in the case of normalized signals the Uncertainty Principle can be expressed as follows (see appendix xA for a proof): 1 (26) 4 This limitation becomes important when the signal has transient components localized in time as in the case of EEGs or ERPs. Gabor (1946) suggested a Gaussian function as the smoothing function, owing to its good localization in time and frequency. In fact, with a Gaussian function eq. 64 is an equality and furthermore, the equality also holds for the mother function of the Gabor Transform, gD ei2 ft (see appendix xA).
t f

3.3.1 Methods and Materials

3.3 Application to intracranially recorded tonic-clonic seizures

Seizure EEG recordings were obtained from a 21 years old male patient. A nine hours recording was performed with 12 depth electrodes ( each electrode having 5 to 15 contacts ) placed in the epileptogenic zone and propagating brain areas. Each signal was ampli ed and ltered using a 1 ; 40Hz band-pass lter. A 4 pole Butterworth lter was used as low-pass lter and as an anti-aliasing scheme. After 10 bits A/D conversion the EEG data was written continuously onto a hard drive with a sampling rate of 256Hz per channel. Selected data sets of ictal and interictal activity were stored for subsequent o -line analysis. We established the necessary frequency resolution in f = 0:25Hz in order to have enough frequency values for calculating the mean and maximum band frequencies and the time resolution was set to t = 0:25sec. This was done by using a Gaussian window of D = 4sec width2 with slide displacement steps of 0:25sec. The
1 1 2 from eq. 8, f = 1 = N 4 256Hz (1=256Hz) = 4 Hz = 0:25Hz

25

decay constant of the Gaussian was set to = 2 =D2 (see eq. 16), what makes their values at the borders less than 5% than the ones of the peak, thus avoiding leakage problems3. Finally, the frequency band limits were de ned in the following way: delta (0:5 ; 3:5Hz), theta (3:5 ; 7:5Hz), alpha (7:5 ; 12:5Hz ), beta-1 (12:5 ; 18Hz) and beta-2 (18 ; 30Hz).

3.3.2 Results
Figure 7 shows 64 sec. of the EEG signal corresponding to one depth electrode in the left hippocampus. The epileptic seizure starts about second 10, and nishes about second 54. The spectrogram corresponding to the previous EEG recording is shown in g. 8. We can observe from this plot that after second 10, when the seizure starts, there is an increase of the frequencies between 5 ; 20Hz, in agreement with what it can be seen by visual inspection of the EEG. Furthermore, there is an abrupt increase of the low frequencies (less than 3Hz) in the post-seizure stage, clearly correlated with the slow oscillations dominating the last part of the recording in g. 7. Although this plot gives a very elegant representation of the signal, it is di cult to extract more information from it. Figure 9 shows the relative intensity ratio (RIR), corresponding to the same EEG recording. Looking at the time evolution of the RIR a good agreement with the signal morphology can be also established, but in this case the evolution of the frequency bands can be followed with more detail. The main feature seen in this plot is a clear decrease of the delta band during the seizure, with a dominance of alpha and some contribution of theta. In this case, higher frequencies ( 1 and 2) do not give an important contribution during the seizure. In gure 10 we show the time evolution of the mean and the maximum frequency ( i (fM) (t) and < f (i) (t) > see eq. 24 and eq. 25) for the di erent bands considered in this work. The most interesting result is the similarity between the mean and the maximum frequency in the delta band between seconds 28 and 40. This re ects the presence of a well de ned peak in this band, although, as seen in g. 12 its intensity at this time is relatively low.
3 for simplicity the normalization constant ; 1=4 of eq. 16 was arbitrarily set to 1.

26

Figure 7: Intracraneal seizure recording. Seizure starts at about second 10 and ends at about second 54.

27

Figure 8: Spectrogram of the previous EEG recording.

Figure 9: RIR of the previous seizure recording. 28

Figure 10: Mean (soft line) and maximum (line with breaks) band frequencies corresponding to the same seizure recording. 29

3.3.3 Discussion
The RIR and the mean and maximum band frequencies give quantitative information di cult to obtain from the visual inspection of the EEG recording or from the spectrogram. Moreover, with the RIR the evolution of the di erent frequency bands can be followed accurately. The evolution of the frequency bands can be more interesting than the following of single peaks, since the rst ones can be related with processes and sources of the brain, as described in sec. x1.1.1. Moreover, the access to quantitative values allows the statistical study of interesting features with a large sample of seizures, as it will be shown in sec. x3.4. The main result was that during the seizure there was an abrupt decrease of delta activity until the ending of the seizure when it raised up again. The seizure was mostly dominated by alpha oscillations, result that will be statistically veri ed in sec. x3.4 with the analysis of several scalp seizure recordings. Although during the seizure delta activity did not reach high amplitudes in comparison with the ones of alpha and theta, it was the only one who showed a signi cant similarity between the mean and maximum frequencies, thus showing the presence of a well de ned peak. It can be conjectured that during the seizure delta oscillations act as a latent pacemaker obscured by the higher amplitude of the alpha and theta ones, until the ending of the seizure when it recruits more cell assemblies and completely dominates.

3.4 Application to scalp recorded tonic-clonic seizures

3.4.1 Methods and Materials
Subjects and data recording Twenty tonic-clonic seizures from eight epileptic patients were analyzed. Scalp electrodes with linked earlobes references were applied following the 10 ; 20 international system (see sec. x1.1). Each signal was digitized at 409:6Hz through a 12 bit A/D converter and ltered with an \antialiasing" eight pole lowpass Bessel lter, with a cuto frequency of 50 Hz. Then, the signal was digitally ltered with a 1 ; 50Hz bandwidth lter and stored, after decimation, at 102:4Hz in a PC hard drive. Gabor Transform and data processing Analysis for each event included one minute of EEG before the seizure onset and two minutes containing the ictal and post-ictal phases. All 3 minutes were analyzed from the right central (C4) electrode, choosing this electrode after visual inspection of the 30

EEG as the one with least amount of artifacts. The RIR was calculated as described in g. 6 by applying a Gaussian window with a width of D = 2:5sec: and slide displacement steps of 1:25sec: (half-overlapping windows). The selection of D = 2:5sec: allows a frequency resolution f = 0:4Hz, which was considered enough for obtaining a good measure of the RIR. The constant of the Gaussian function (see eq. 16) was set as in section x3.3.1. Traditional frequency bands were set to 1 ; 3:5, 3:5 ; 7:5, and 7:5 ; 12:5Hz respectively. Given that beta 1 and beta 2 are closely related to muscle artifact those bands were excluded from the analysis. Plateau criteria Clear decrements in delta activity were seen during the seizures. In order to quantify this observation, the mean relative intensity ratio (MRIR) for the delta band was calculated in the pre-ictal phase and compared with the value in the low intensity zones (plateaus) observed during the seizure. Pre-ictal MRIR was de ned as the mean value of the RIR in the minute before the seizure, discarding those areas contaminated by artifacts (de ned by visual inspection of the EEG). In the ictal phase, plateaus were de ned according to the following criteria: 1. Plateaus must last at least 10sec in order to avoid local variations 2. MRIR(ictal) < 0:3 MRIR(pre ; ictal) and 3. Standard error of the plateau MRIR (SEM ) must be less than 1 to con rm low variance. Although the choice of these criteria is arbitrary, no plateaus were identi ed in the pre-ictal phase, strongly indicating that the appearance of a plateau in the ictal phase re ects a dynamical change rather than a statistical phenomenon.

3.4.2 Results
As an example of the analysis performed I will describe the results of one of the seizures and then I will summarize the global ndings. Figure 11 discloses three minutes of EEG data of one typical seizure and the RIR of this signal is shown in gure 12. Pre-ictal phase is characterized by a signal of 50 V amplitude with a dominance of delta rhythms (pre-ictal delta MRIR 50%). Seizure starts at second 80 (marked with an S in both graphs) with a discharge of slow waves superposed by fast ones with lower amplitude. This discharge lasts approximately 8 seconds, has a mean amplitude of 100 V , and, as seen in gure 12, produces a marked rise in delta band, which reaches 90% of the RIR. Afterwards, seizure spreads making 31

the analysis of the EEG more complicated due to muscle artifacts however, it is possible to establish the beginning of the clonic phase at around second 123 and the end of the seizure at second 155 (marked with an E in both graphs) where there is an abrupt decay of the signal. Although it is di cult to extract any information from the EEG during the seizure, in gure 12 we can follow its frequency pattern. From second 90, delta activity decreases abruptly to values lower than 10% of the RIR, and theta and alpha bands alternatively dominate. We also observe that the starting of the clonic phase is correlated with a rise of theta frequencies and after second 140, when clonic discharges became more separated, delta activity rises up again until the end of the seizure, also maintaining this behavior in the post-ictal phase. We can conclude from this example that the seizure was dominated by alpha and theta rhythms with a corresponding abrupt decrease of delta activity. By applying the criteria explained in the previous section, a plateau of delta decrement was de ned between seconds 99 and 138, lasting 39 seconds, having very low variance (0.14) and also having a very low ictal to pre-ictal delta MRIR (2%). Considering the whole group, a stereotyped pattern was identi ed in 14=20 (70%) of the seizures. This pattern was characterized by a signi cant reduction in delta activity during the seizures, implying that they were dominated by theta and alpha rhythms until the ending of the seizure when delta activity started to rise again. One critical point in our results is the possible distortion due to spatial propagation of the seizure, due to the fact that data of C4 electrodes was analyzed, and the sources of the seizures were mostly in temporal locations. In order to overcome this, Gabor Transform was also applied to T3 and T4 electrodes, obtaining similar results to the ones reported with C4 electrodes.

3.4.3 Discussion
The most robust nding is that in 70% of the seizures a signi cant reduction in delta activity was observed by applying the plateau criteria. This was usually seen at the beginning of the seizure, thus emphasizing the dominance of alpha and theta bands during the initial phases of a tonic-clonic seizure. This result is in agreement with our previous observations using Gabor Transform to analyze ictal patterns recorded with depth electrodes (Blanco et al., 1995b, 1996a, 1996b, 1997). Although the selection of the plateau criteria might have in uenced these results, slight changes in the de nition of the plateau showed no signi cant variations. Also, results do not depend on the patient studied, and owing to the great variation in age, antiepileptic drugs, and source of the seizure between patients, we can conclude that the results are not dependent on 32

Figure 11: Scalp EEG of the right central channel during a Tonic-Clonic seizure

Figure 12: RIR corresponding to the previous seizure 33

these factors. A rise in delta activity was observed towards the end of the seizures. This pattern was also seen with depth recordings, thus making unlikely the possibility of being generated by artifacts. Neuronal fatigue, decrement in neuronal ring and preponderance of inhibitory mechanisms are critical factors in the mechanisms underlying this observation which prompts further research. These results were obtained by avoiding muscle artifacts with Gabor Transform, thus supporting their utility for the evaluation of tonic-clonic seizures, something di cult to assess in scalp recorded EEGs. As Niedermeyer (1993c) pointed out:
\Muscle activity rapidly obscures the recording the vertex deviation, however, may remain artifact-free due to the lack of underlying muscles. Informative Grand Mal recordings can be secured only from patients with muscle relaxation from curarization 4 and arti cial respiration".

Gotman et al. (1981) avoid this problem by the use of lters. However, he pointed out that ltering the signal has several disadvantages. On the one hand it is impossible to separate brain and muscle activity in the EEG, and further, it is well known that ltering high frequencies also a ects the morphology of the low ones. Gastaut and Broughton (1972) instead, described a frequency pattern during a tonic-clonic seizure of a curarizated patient. In the rst seconds after the seizure onset, they found an epileptic \recruiting rhythm" of 10Hz associated with the tonic phase that lasted approximately 10 seconds later, as the seizure ended, there was a progressive increase of the lower frequencies (5 ; 6Hz) associated with the clonic phase. Our ndings were similar to the ones highlighted by Gastaut and Broughton, but recruiting rhythms were observed to be also in the theta range or, as seen in most cases, uctuating between alpha and theta. Darcey and Williamson (1985) also obtained similar patterns by analyzing seizures recorded with depth electrodes. They reported an activity characterized by 10Hz at the onset of the seizure, frequency that declined as the seizure ended. The results presented here di ers from these previous attempts in that a quanti cation of the Gabor Transform was performed with the RIR, thus allowing the analysis of the frequency content of the seizure. One important point to note is that these results were obtained with scalp recordings and without the need for curare or any ltering method.
4 Curare is a drug that blocks neuromuscular transmission, thus inhibiting the muscular artifacts characteristic of the Grand Mal seizures

34

3.5 Conclusion
In this chapter I described the use of quantitative parameters de ned from the Gabor Transform applied to the analysis of EEG signals. These parameters give an accurate description of the time evolution of EEG rhythms, something di cult to perform from the spectrograms. Spectrograms are more suitable for the analysis of long recordings in which only large scale variations, of the order of minutes or hours, are relevant (see examples of spectrogram analysis in Lopes da Silva, 1993b Gotman, 1990a). In this context, spectrograms are for example very useful for discriminating between di erent sleep stages, due to the slow variation of the frequency patterns. However, as showed in sec. x3.3.2, spectrograms are not so suitable for the analysis epileptic seizures, in which frequency patterns change in the order of seconds. On the other hand, the quanti cation by means of the RIR allowed the performance of a statistical analysis of the frequency content of Grand Mal seizures, by studying 20 scalp seizure recordings. This analysis showed marked decreases of the delta band during the seizure due to the dominance of alpha and theta band at this stage. Since intracranial recordings are nearly free of artifacts, the fact that the same pattern was seen in both situations reinforces the idea that the results obtained with scalp electrodes were not an spurious e ect of muscle activity, or of the procedure achieved for their elimination. The introduction of the mean and maximum band frequencies allowed the study of the time evolution of frequency patterns within each frequency band. It is interesting to note that with this analysis it was possible to establish the presence of a well de ned delta oscillation during the seizure, in spite of the fact that its energy was very low and remained latent until the ending of the seizure, when it reached very high amplitudes and dominated the EEG. Other interesting point to note is that although the grouping in frequency bands implies a loose of frequency resolution, it can be more useful than the study of single frequencies or peaks, due to the relations between frequency bands and functions or sources of the brain. The use of the present time-frequency analysis, together with the clinical patient history and the visual assessment of the EEG, can contribute to the identi cation of the source of epileptic seizures and to a better understanding of its dynamic. Certainly, Gabor Transform is not intended to replace conventional EEG analysis, but rather to complement it and also to provide further insights into the underlying mechanisms of ictal patterns. In this context, RIR allows a fast interpretation of several minutes of frequency variations in a single display, something di cult to perform with traditional scalp EEG. 35

4 Wavelet Transform
4.1 Introduction
Fourier Transform consists in making a correlation between the signal to be analyzed and complex sinusoids of di erent frequencies (see upper part of g. 13). As stated in chapter x2, Fourier Transform gives no information about time and requires stationarity of the signal. By \windowing" the complex sinusoidal mother functions of the Fourier Transform, a time evolution of the frequencies can be obtained just sliding the windows throughout the signal. This procedure, called the Gabor Transform, consists in correlating the original signal with modulated complex sinusoids as showed in the middle part of g. 13 (for details see section x3.2). Gabor Transform gives an optimal time-frequency representation, but one critical limitation appears when windowing the data due to the Uncertainty Principle (see x3.2 and xA Cohen, 1995 Strang and Nguyen, 1996 Chui, 1992). If the window is too narrow, the frequency resolution will be poor, and if the window is too wide, the time localization will be not so precise. Data involving slow processes will require wide windows and on the other hand, for data with fast transients (high frequency components) a narrow window will be more suitable. Then, owing to its xed window size, Gabor Transform is not suitable for analyzing signals involving di erent range of frequencies. Grossmann and Morlet (1984) introduced the Wavelet Transform in order to overcome this problem. The main advantage of wavelets is that they have a varying window size, being wide for slow frequencies and narrow for the fast ones (see lower part of g. 13), thus leading to an optimal time-frequency resolution in all the frequency ranges (Chui, 1992 Strang and Nguyen, 1996 Mallat, 1989). Furthermore, owing to the fact that windows are adapted to the transients of each scale, wavelets lack of the requirement of stationarity. This chapter starts with a brief theoretical background of the Wavelet Transform. Details of a straightforward implementation called the multiresolution analysis and an alternative decomposition of time signals called the Wavelet Packets analysis will be given. In section x4.4, I will show the results obtained by applying trigonometric wavelet packets (Serrano, 1996) to the study of tonic-clonic seizures (Blanco et al., 1998a). In sections x4.5 and x4.6, I will analyze two di erent types of event related potentials. By using the multiresolution decomposition method, in the rst case I will study the alpha band responses (Quian Quiroga and Schurmann, 1998, 1999) and in the second case the ones corresponding to the gamma band (Sakowicz et al., 1999). Finally, I will discuss advantages of Wavelets in the study of brain signals. 36

Original signal

FOURIER TRANSFORM

GABOR TRANSFORM

WAVELET TRANSFORM

Figure 13: Frequency and time-frequency methods. Fourier Transform is obtained by correlating the original signal with complex sinusoids of di erent frequencies (upper part). In Gabor Transform, the signal is correlated with modulated sinusoidal functions that slides upon the time axis, thus giving a time-frequency representation (middle part). Wavelets give an alternative time-scale representation but due to their varying window size, a better resolution for each scale is achieved (bottom part). Furthermore, the function to be correlated with the original signal can be chosen depending on the application (e.g. in the graph quadratic B-Splines wavelets are shown).

37

4.2 Theoretical Background

4.2.1 Continuous Wavelet Transform
A wavelet family a b is a set of elemental functions generated by dilations and translations of a unique admissible mother wavelet (t):
a b (t)

= jaj;1=2

t;b a

(27)

where a b 2 R, a 6= 0 are the scale and translation parameters respectively. As a increases the wavelet becomes more narrow and by varying b, the mother wavelet is displaced in time. Thus, the wavelet family gives a unique pattern and its replicas at di erent scales and with variable localization in time. The continuous wavelet transform of a signal X (t) 2 L2 (R) ( nite energy signals) is de ned as the correlation between the signal and the wavelet functions a b , i.e. (W X )(a b) =

jaj;1=2

1 ;1

b ) dt = < X (t) X (t) ( t ; a

ab

>

(28)

where denotes complex conjugation. Then, the di erent correlations < X (t) a b > indicates how precisely the wavelet function locally ts the signal at every scale a. Since the correlation is made with di erent scales of a single function, instead of with complex sinusoids characterized by their frequencies, wavelets give a time-scale representation.

4.2.2 Dyadic Wavelet Transform

The continuous Wavelet Transform maps a signal of one independent variable t onto a function of two independent variables a b. This procedure is redundant and not e cient for algorithm implementations. In consequence, it is more practical to de ne the Wavelet Transform only at discrete scales a and discrete times b. One way to achieve this, is by choosing the discrete set of parameters faj = 2j bj k = 2j kg, with j k 2 Z . Replacing in eq. 27 we obtain the discrete wavelet family
j k (t)

= 2;j=2 ( 2;j t ; k )

j k 2 Z

(29)

that forms a basis of the Hilbert space L2(R), and whose correlation with the signal is called Dyadic Wavelet Transform. 38

4.2.3 Multiresolution Analysis

Contracted versions of the wavelet function will match high frequency components of the original signal and on the other hand, dilated versions will match low frequency oscillations. By correlating the original signal with wavelet functions of di erent sizes we can obtain the details of the signal in di erent scale levels. Then, the information given by the dyadic Wavelet Transform can be organized according to a hierarchical scheme called multiresolution analysis (Mallat, 1989 Chui, 1992). If we denote by Wj the subspaces of L2 generated by the wavelets j k for each level j , the space L2 can be decomposed as a direct sum of the subspaces Wj ,

L2 =
Let us de ne the closed subspaces

j 2Z

Wj

(30)

j2Z (31) The subspaces Vj are a multiresolution approximation of L2 and they are generated by scalings and translations of a single function j k called the scaling function. Then, for the subspaces Vj we have the orthogonal complementary subspaces Wj , namely: Vj = Wj+1 Wj+2 : : : Vj;1 = Vj Wj j 2 Z (32) Let us suppose we have a discrete signal X (n), which we will denote as x0 , with nite energy and without loss of generality, let us suppose that the sampling rate is t = 1. Then, we can successively decompose it with the following recursive scheme xj;1( n ) = xj ( n ) rj ( n )
(33) where the terms xj (n) 2 Vj give the coarser representation of the signal and rj (n) 2 Wj give the details for each scale j = 0 1 N . Then, for any resolution level N > 0, we can write the decomposition of the signal as

X (n) =

xN (k)

( 2;N

n ; k) +

N X X j =1 k

Cj (k)

j k (n)

(34)

where ( ) is the scaling function, Cj (k) are the wavelet coe cients, and the sequence fxN (k)g represents the coarser signal at the resolution level N . The second term is the wavelet expansion. The wavelet coe cients Cj (k) can be interpreted as the local residual errors between successive signal approximations at scales j ; 1 and j , and 39

rj (n) =

is the detail signal at scale j . Figure 14 shows as an example the multiresolution decomposition method applied to an event-related potential and the corresponding reconstruction achieved by using the inverse transform. The method starts by correlating the signal with shifted versions (i.e. thus giving the time evolution) of a contracted wavelet function, the coe cients obtained thus giving the detail of the high frequency components. The remaining part will be a coarser version of the original signal that can be obtained by correlating the signal with the scaling function, which is orthogonal to the wavelet function. Then, the wavelet function is dilated and from the coarser signal the procedure is repeated, thus giving a decomposition of the signal in di erent scale levels, or what it is analog, in di erent frequency bands. This method gives a decomposition of the signal that can be implemented with very e cient algorithms due to the fact that it can be achieved just by applying simple lters, as showed by Mallat (1989, see table 1 for an example). The lower levels give the details corresponding to the high frequency components and the higher levels the ones corresponding the the low frequencies. As pointed out in sec. x4.1, the levels related with higher frequencies have more coe cients that the lower ones, due to the varying window size of the Wavelet Transform.

Cj (k)

j k (n)

(35)

4.2.4 B-Splines wavelets

An important point to be discussed is how to choose the mother functions to be compared with the signal. In principle, the wavelet function should have a certain shape that we would like to localize in the original signal. However, due to mathematical restrictions, not every function can be used as a wavelet. Then, one criterion for choosing the wavelet function is that \it looks similar" to the patterns of the original signal. In this respect, B-Spline functions seem suitable for decomposing ERPs (see bottom part of gure 13). B-Splines are piecewise polynomial functions of a certain degree that form a base in L2 (R) (see e.g. Chui, 1992). Filter coe cients corresponding to quadratic B-Splines are shown in table 1. We can remark the following properties that makes them very suitable for the analysis of ERP (Unser et al., 1992, 1993 Unser, 1997 Chui, 1992 Strang and Nguyen, 1996): 1. Smoothness: the smooth behavior of B-Splines is very important in order to avoid border e ects when making the correlation between the original signal and a wavelet function with abrupt patterns. 40

-15

Original signal

0 15

-1sec

1sec Multiresolution Reconstruction

Multiresolution Decomposition

Level 1 Level 2 Level 3

Level 4 Level 5
-15 -15 0 15

Level 6

0 15

-1sec

1sec

-1sec

1sec

Figure 14: Multiresolution decomposition method. The signal is decomposed in scale levels each one representing the activity in di erent frequency bands. The wavelet coefcients show how closely the signal matches locally the di erent dilated versions of the wavelet mother function (in this case a quadratic B-Spline). Furthermore, by applying the inverse transform, the signal can be reconstructed from the wavelet coe cients for each scale level.

41

h(k)

g (k )

p2(k)

q2(k)

-10 0.00157 -0.00388 -9 0.01909 -0.03416 -8 -0.00503 0.03416 -7 -0.0444 0.07933 -6 0.01165 -0.02096 -5 0.10328 -0.18408 -4 -0.02593 0.04977 0.00208 -3 -0.24373 0.42390 -0.06040 -2 0.03398 -0.14034 0.25 0.30625 -1 0.65523 -0.90044 0.75 -0.63125 0 0.65523 0.90044 0.75 0.63125 1 0.03398 0.14034 0.25 -0.30625 2 -0.24373 -0.42390 -0.06040 3 -0.02593 -0.04977 -0.00208 4 0.10328 0.18408 5 0.01165 0.02096 6 -0.0444 -0.07933 7 -0.00503 -0.00901 8 0.01909 0.03416 9 0.00157 0.00388 Table 1: Filter coe cients for quadratic B-Splines. h and g are the decomposition lters and p2, q2 are the reconstruction ones (from Ademoglu et al, 1997 see also Strang and Nguyen, 1996). 2. Time-frequency resolution: it was shown that B-Spline functions have nearly optimal time-frequency resolution (i.e. nearly the maximum allowed by the uncertainty principle Unser et al., 1992). 3. Compact support: this means that the wavelet function does not extend to in nity. This fact is important in order not to include in a certain wavelet coe cient correlations with points far in the past or in the future. 4. Semi-orthogonality: the use of B-Splines as mother functions when applying the multiresolution decomposition guarantees orthogonality between the di erent scales. 42

4.2.5 Wavelet Packets

In the approach described before, only the successive signals xj (n) are decomposed, but in many cases, it is also interesting to decompose the details rj (n). If we decompose both xj (n) and rj (n), then the original signal can be represented in di erent ways as combinations of xj (n) and rj (n) of di erent levels j . In this work, a decomposition called trigonometric wavelet packets was used (Serrano, 1996). The main idea is to decompose the components rj (n) in portions. We de ne any portion or local signal as

rj(m l)( n ) =

m ;1 l+2 X

where the parameters m and l are chosen for rj(m l) (n) to cover the full time interval 2;j l n 2;j (l + 2m), which is a relative long interval of length 2m;j . Note that we de ned the local wavelet packet with 2m basic functions j k (n) for k = l l +2m ; 1. Now, we de ne the set of fundamental frequencies

k=l

Cj (k )

j k(

n)

(36)

!mh = + 2 h =2m
with 0 h 2m;1 and associated Fourier matrix M(m) given by
8 > > > <

(37)

sin (k + 1=2) ] if d = 1 1 = 2 (m) = 2;m=2 2 cos !mh (k + 1=2) ] if d is even Mdk (38) > 21=2 sin !mh (k + 1=2) ] if d is odd > > : cos 2 (k + 1=2) ] if d = 2m with 1 d 2m , 0 k < 2m and h = d=2]], where ] denotes the integer part. It can be demonstrated that M(m) is a 2m 2m dimensional orthogonal matrix (Serrano, 1996). Then, we can de ne the new set of elemental functions in order to expand rj(m l)(n) as a 2m dimensional vector obtained from
(m l ) jd ( n ) =
m ;1 l+2 X

for 1 d Clearly, these functions constitute a new local orthonormal basis covering the interval under analysis 2;j l n 2;j (l + 2m ). Therefore we can give a second description of the local signal as 43

2m .

k=l

(m) Mdk

j k(

n)

(39)

rj(m l)( n ) =

2m

and the corresponding coe cients are easily computed as

d=1

l) Dj(m l) ( d ) j(m d (n):

(40)

Dj(m l) ( d ) =
2m.

m ;1 l+2 X

where 1 d l) The trigonometric wavelet packets j(m d (n) have zero mean, oscillate on the interval 2;j l n 2;j (l + 2m ) and decay with exponential ratio. Moreover, their wave-forms resemble modulated sines or cosines. In fact, it can be demonstrated that each Fourier l) transform ^j(m d (! ) is centered at the fundamental frequency !mh , when d = 2h or l) d = 2h + 1. Moreover, ^j(m d (! ) = 0 on the other fundamental frequencies. In other words, the coe cients fDj(m l)(d)g can be considered as the discrete Fourier spectrum for the local signal rj(m l)(n). Summing up, we can resume in the double set of coe cients fCj (k) Dj(m l) (d)g the time-scale-frequency information of the local signal rj(m l) (n). Finally, to analyze the complete function rj (n), that is, the details at level j , we choose some partition in local components rj(mi li) (n), according the structure of the signal,

k=l

(m) C ( k ) Mdk j

(41)

rj ( n ) =

where the sequence of index li veri es li+1 = li + 2mi . Then, we implement the above refereed time-scale-frequency technique for each local signal.

mi

rj(mi li) ( n )

(42)

4.3 Short review of wavelets applied to the study of EEG signals

Several works applied the Wavelet Tranform to the study of EEGs and ERPs (see a review in Unser and Aldroubi, 1996 or in Samar et al., 1995). One rst line of applications is for pattern recognition in the EEG. This is achieved by correlating di erent transients of the EEG with wavelet coe cients of di erent scales. Schi et al. (1994a) used a multiresolution decomposition implemented with B-Splines mother functions for extracting features of EEG seizure recordings. They showed a better performance of wavelets in comparison with Gabor Transform, and a similar resolution of the multiresolution decomposition compared with the continuous Wavelet Transform but with a 44

marked decrease in computational time. Other works also used this approach for automatic detection of spike complexes characteristic of epilepsy, thus helping in the analysis of EEG recordings from epileptic patients (Schi et al., 1994b Senhadji et al., 1995 Clark et al., 1995). Demiralp et al. (1999) used coe cients in the delta frequency band for detecting P300 waves in single trials of an auditory oddball paradigm. Furthermore, they used this approach for making a selective averaging of the single trials, thus obtaining a better de nition of the P300. Basar et al. (1999) reported the utility of Wavelet Transform for classifying di erent type of single sweep responses to cross-modality stimulation. A digital ltering of ERPs based on the Wavelet Transform was proposed by Bertrand et al. (1994). They used the method as a noise reduction technique, reporting better results than the ones obtained with Fourier based methods, especially when applied to non-stationary signals. The main goal of this type of ltering is to extract the eventrelated responses from the single sweeps by eliminating the contribution of the ongoing EEG. This procedure would avoid the necessity of averaging the single sweeps. In this context, Bartnik et al. (1992) characterized the event-related responses from the wavelet coe cients, then using selected coe cients for isolating the event-related responses from the background EEG in the single trials. A similar approach has being later proposed by Zhang and Zheng (1997). Akay et al. (1994) used the Wavelet Transform for characterizing electrocortical activity of fetal lambs, reporting much better results than the ones obtained with the Gabor Transform. Thakor et al. (1993) analyzed somatosensory EPs of anesthetized cats with cerebral hypoxia by using the multiresolution decomposition. They report that selected coe cients are sensitive to neurological changes, but having comparable results than the ones obtained with Fourier based methods. Ademoglu et al. (1997) used wavelet analysis for discriminating between normal and demented subjects by studying the N70-P100-N130 complex response to pattern reversal visual evoked potentials (N70 and N130 are negative peaks of the ERP with a latency of 70 and 130ms respectively). Kolev et al. (1997) used the multiresolution decomposition for studying the presence of di erent functional components in the P300 latency range in an auditory oddball paradigm. Basar et al. (1999) used the wavelet decomposition for studying the alpha responses to cross-modality stimulation, reporting similar results than the ones obtained with digital ltering. 45

Figure 15: Scalp EEG seizure recording.

4.4 Application to scalp recorded tonic-clonic seizures

4.4.1 Material and Methods
An EEG time series corresponding to a tonic-clonic seizure of an epileptic patient was analyzed. Scalp electrodes were applied following the 10-20 international system. The signal was digitized at 409:6 Hz through a 12 bit A/D converter and ltered with an antialiasing eight pole lowpass Bessel lter with a cuto frequency of 50 Hz. Then, it was digitally ltered with a 1 ; 50 Hz bandwidth Butterworth lter and stored, after decimation, at 102:4 Hz in a PC hard drive. The recording included one minute of the EEG before the seizure onset and two minutes which included the ictal and post-ictal phases. All 3 minutes were analyzed at the right central (C4) electrode, choosing this electrode after visual inspection of the EEG as the one with the least amount of artifacts. Wavelet Transform was applied by using a cubic Spline function as mother wavelet. The multiresolution decomposition method (Mallat, 1989) was used for separating the signal in 7 frequency bands: B1 = 25:8 ; 51:2Hz B2 = 12:8 ; 25:2Hz B3 = 6:4 ; 12:8Hz B4 = 3:2 ; 6:4Hz B5 = 1:6 ; 3:2Hz B6 = 0:8 ; 1:6Hz B7 = 0:4 ; 0:8Hz).

4.4.2 Results
Figure 15 shows 90sec of the Tonic-Clonic seizure studied. The whole recording was already shown in g. 11. In this case seizure starts at second 10 and ends at second 85. Due to the fact that the aim of this work was to analyze middle and low frequencies 46

brain activity during an epileptic seizure, we eliminated B1 and B2 bands, both containing high frequency artifacts that obscure the EEG (see sec. x3.4). The relative band intensity ratio (RIR) (de ned as in sec. x3.2 but in this case from the wavelet scales) had a similar behavior as the one showed with Gabor Transform in gure 12. Frequency bands B3 and B4 were chosen for performing an analysis with Wavelets Packets, these bands being important in the development of the tonic-clonic seizures as showed in Chapter x3.4 (see also Quian Quiroga et al., 1997b). B3 band coe cients were segmented with sliding windows of l = 32 samples corresponding to time intervals of t = 2:5 sec. Discrete sets of frequencies between 6:4 and 12:8 Hz with intervals of 0:4 Hz were obtained as showed in g. 4.4.2 (squared values shown). From second 50, we can see an increase of the activity in nearly all the packets. Due to the good time-frequency resolution of the Wavelet Packets it is possible to follow the evolution of the frequency peaks. For example, the peak marked with an arrow in the wavelet packet corresponding to 8:4Hz at about second 50, is also visible in the packets corresponding to 8:0, 7:6 and 7:2Hz, appearing with higher amplitude and some delay. Then, this peak originated in the 8:4Hz packet, or probably in higher frequencies but with lower amplitude, is evolving with time to lower frequencies. Figure 17 shows the Wavelet Packets corresponding to the B 4 band (squared values shown). They were generated by using l = 16 samples corresponding to time intervals of t = 2:5 sec and discrete sets of frequencies between 3:2 and 6:4 Hz with intervals of 0:4 Hz. Note that the j = ;4 level has half dispersion in frequencies compared with the j = ;3 level and for this reason we used a window of 16 samples in order to obtain the same de nition. In the B 4 band there is a very clear peak, marked with an arrow, after second 60 in the frequencies around 3 ; 4Hz, this increase being correlated with the starting of the clonic phase of the seizure. This peak is also visible, but appearing earlier in time, in the higher frequency packets (also marked with an arrow). Although this behavior is predictable with a visual inspection of the EEG, it is very interesting to note that this peak is in fact the same peak described in the gure 4.4.2 but appearing more delayed. Analyzing both gures together, we can observe very clearly how this high amplitude, low frequency peak (3 ; 4Hz) at about 65sec was in fact rst observed in the higher frequencies (about 9Hz), then evolving with time to lower frequencies until reaching a very high amplitude when the clonic phase starts, this evolution being very di cult to identify from a visual inspection of the EEG or by using traditional methods as the spectrograms (see discussion of sec. x3.5). 47

7.2 Hz
2 2

7.6 Hz

0 0 3 3 3 20 40 60 80

0 0 3 20 40 60 80

8.0 Hz
2 2

8.4 Hz
2

8.8 Hz
2

9.2 Hz

0 0 3 20 40 60 80

0 0 3 20 40 60 80

0 0 3 20 40 60 80

0 0 3 20 40 60 80

0 0 3 20 40 60 80

0 0 3 20 40 60 80

0 0 3 20 40 60 80

0 0 3 20 40 60 80

11.2 Hz
2 2

11.6 Hz
2

12.0 Hz
2

12.4 Hz

0 0 20 40 60 80

0 0 20 40 60 80

0 0 20 40 60 80

0 0 20 40 60 80

Figure 16: Wavelet packets for the scale level B 3.

48

9.6 Hz

10 Hz

10.4 Hz

10.8 Hz

3.2 Hz
2 2

3.6 Hz
2

4.0 Hz
2

4.4 Hz

0 0 20 40 60 80

0 0 20 40 60 80

0 0 20 40 60 80

0 0 20 40 60 80

4.8 Hz
2 2

5.2 Hz
2

5.6 Hz
2

6.0 Hz

0 0 20 40 60 80

0 0 20 40 60 80

0 0 20 40 60 80

0 0 20 40 60 80

49
3 2 1 0

6.4 Hz

20

40

60

80

Figure 17: Wavelet packets for the scale level B 4.

4.4.3 Discussion
The high time-frequency resolution achieved with Wavelet Packets allowed a very detailed study of the time evolution of the frequency peaks during the seizure. In fact, it was possible to establish that the high amplitude peaks of about 3 ; 4Hz, characteristic of the clonic activity, were generated in higher frequencies. This result is in agreement with the dynamic described with the Gabor Transform. In this case, with the RIR (see sec. x3.4) I showed how during the seizure the relevant brain activity was dominated by alpha and theta bands, until the ending of the seizure, when delta activity dominated again. Then, it is reasonable to conjecture that the violent clonic contractions characteristic of the clonic phase of the Grand Mal seizures are the response to brain oscillations generated in higher frequencies, but owing to the fact that the muscles can not contract so fast, the muscle activity is limited to a tonic contraction until the oscillations become slower and the muscles are capable to oscillate in resonance with them. It is interesting to note that the frequency behavior described is in agreement with simulations of thalamic slices performed by Wang and coworkers (Wang et al., 1995 Golomb et al., 1996), who observed a slowing from 10Hz to 4Hz in their simulations after the suppression of GABAA inhibitors. This result is also in agreement with in vitro experimental results of Bal et al. (1995a,b). These experiments suggest that it would be very interesting to investigate if the \slowing" of the frequencies observed during the seizures can be related with a variation in the concentration of neurotransmitters, especially of the GABA inhibitors, possible due to a neuronal fatigue provoked by the abnormal ring rate of the neurons during the seizure.

4.5 Application to alpha responses of visual event-related potentials

4.5.1 Introduction
EEG alpha rhythms can be de ned as oscillations between 8 and 13 Hz, with an amplitude usually below 50 V and localized over posterior regions of the head. Alpha rhythms appear spontaneously during wakefulness, best seen with eyes closed, under relaxation and mental inactivity conditions (Niedermeyer, 1993a). Although alpha oscillations have been widely studied, their sources and functional correlates are still under discussion. Sources of alpha rhythms have been investigated leading to controversies whether they are thalamic, cortical or whether other structures are involved in their generation (Adrian, 1941 Andersen and Andersson, 1968 Lopes da 50

Silva et al.,1973a,1973b Lopes da Silva and Storn van Leewen, 1977 Basar et al., 1997). Moreover, many studies were performed in order to understand their functional meanings. These studies showed that alpha rhythms could be correlated even to sensory or cognitive processes depending on the task performed and generators involved, therefore not having an unique and speci c function (for a review, see Basar et al., 1997).

4.5.2 Material and Methods

In 10 voluntary healthy subjects (no neurological de cits, no medication known to a ect the EEG) two types of experiments were performed: 1. No-task visual evoked potential (VEP): subjects were watching a checkerboard pattern (sidelength of the checks: 50'), the stimulus being a checker reversal (N = 100 stimuli). 2. Oddball paradigm: subjects were watching the same pattern as above. Two different stimuli were presented in a pseudorandom order. NON-TARGET stimuli (75%) were pattern reversal, and TARGET stimuli (25%) consisted in a pattern reversal with horizontal and vertical displacement of one-half of the square side length. Subjects were instructed to pay attention to the appearance of the target stimuli (N = 200 stimuli). The inter-stimulus interval varied pseudo-randomly between 2.5 and 3.5 s. After each pattern reversal, the reverted pattern was shown for one second, then the pattern was re-reverted. Recordings were made following the international 10 ; 20 system in seven di erent electrodes (F3, F4, Cz, P3, P4, O1, O2) referenced to linked earlobes. Data were ampli ed with a time constant of 1:5sec: and a low-pass lter at 70Hz. With each stimulus, a single sweep of EEG data was recorded, i.e.: for each single sweep, 1sec: pre- and post-stimulus EEG were digitized with a sampling rate of 250Hz and stored in a hard disk. After visual inspection of the data, 30 sweeps free of artifacts were randomly selected for each type of stimuli (VEP, NON-TARGET and TARGET) for future analysis. A Wavelet Transform was applied to each single sweep using a quadratic B-Spline function as mother wavelet. The multiresolution decomposition method (Mallat, 1989) was used for separating the signal in frequency bands, de ned in agreement with the traditional frequency bands used in physiological EEG analysis. After a ve octave wavelet decomposition, components corresponding to the alpha band (8 ; 16Hz) were analyzed. For each subject the alpha components of the 30 single sweeps were averaged. Finally, results for each subject were averaged to obtain a \grand average". The temporal reso51

lution of the scale corresponding to the alpha band was of 64 ms 5 . However it should be remarked that the \real" resolution will depend on the characteristics of the signal and the mother function (i.e. how the mother function matches the signal see section 4.2.4). In this respect, the optimal resolution of B-Splines was shown with numerical computations (Unser et al., 1992). It is also interesting to note that non-redundancy is important for increasing the computational speed. Statistical analysis The wavelet coe cients processed were the ones obtained after averaging the responses of the 30 single trials for each electrode and subject. Then, wavelet coe cients were recti ed and the maximum coe cients and their time delay with respect to the stimulus occurrence were computed in the rst 500 ms post-stimulation. The analysis was limited to the rst 500 ms owing to the fact that no subject showed meaningful alpha responses after this time. Comparison between modalities and electrodes were done by using a multiple factor ANOVA test. Comparison between wavelets and conventional digital ltering Figure 18 gives some examples of single-trial evoked potentials, showing the comparison of results obtained with Wavelet Transform and with digital ltering. In addition, the gure shows the relation between the wavelet coe cients (used for all statistical analyses) and the waveforms reconstructed from the wavelet coe cients for a speci c level (which will be shown in the gures). I would like to remark that the sweeps selected do not necessary show a clear event-related response, but they are suitable for showing the better resolution achieved with the multiresolution decomposition based on the Wavelet Transform in comparison with conventional digital ltering. The digital lter used was an ideal lter (i.e. a digital lter based on band pass ltering in the Fourier domain as used in several earlier papers, Basar, 1980) with the lter limits set in agreement with the limits obtained with the multiresolution decomposition for the alpha band. As a general remark it can be stated that with the wavelet coe cients a better resolution and localization of the features of the signal is achieved. In between the vertical dashed lines in sweep #1 three oscillations in the alpha range are shown, having the last oscillation a larger amplitude as observed in the original sweep. This is well resolved with the wavelet coe cients and also in the reconstructed form. However, this ne structure of this train of alpha oscillations is not resolved by digital ltering i.e. reading a maximum from this curve is imprecise. In sweep #2, in between the dashed
5 From sec. 4.2.2 the time steps bj k are 2j data points, and since the alpha band corresponds to j = 4 and the sampling rate was of 250Hz we have t = 24 =250Hz = 64ms

52

sweep #1

sweep #2

sweep #3

sweep #4

sweep #5

-50

-50 0 50
-0.5 0.0 0.5 1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5 1.0

Original sweep

0 50
-1.0

-50 -50 -50 0 50

-0.5 0.0 0.5 1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5

-50 0 0 50
1.0 -1.0 -0.5 0.0 0.5 1.0

Digital filtering

0 50

50
-1.0 -0.5 0.0 0.5 1.0

53
Wavelet coeff. Wavelet reconstr.

-1.0

-150

-150

-150

-150

-150

150 -1.0

-0.5

0.0

0.5

150 1.0 -1.0

-0.5

0.0

0.5

150 -1.0 1.0

-0.5

0.0

0.5

150 1.0 -1.0

-0.5

0.0

0.5

1.0

150 -1.0

-0.5

0.0

0.5

1.0

-50 0 50
-1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5

-50 0 50
1.0 -1.0 -0.5 0.0 0.5 1.0

Figure 18: Examples of the better performance of the Wavelet Transform in comparison with an \ideal" digital ltering for ve single sweeps.

vertical lines, is showed a transient with a frequency clearly lower than the range of alpha band. The digital ltering does not resolve this transient and it spuriously \interpolates" alpha oscillations in continuity with the ones that precede or follow the transient. On the other hand, the wavelet coe cients show a decrease in this time segment, this phenomenon being also visible in the reconstructed form. Something similar occurs in sweep #3 with the transient marked with an arrow. In fact in this last case, the transient is due to the cognitive P300 wave typically obtained upon target stimuli. With wavelets it is visible that, as in the original signal, there is no important contribution of alpha oscillations in this time range, the digital ltering having not enough resolution for resolving this. The better time-frequency resolution of wavelets (in this case a better frequency localization for a certain time range) can be also seen in sweep #4. In the original signal, in between the vertical dashed lines there is a marked oscillation of about 4 ; 6Hz, corresponding to the theta band. The digital ltered signal shows an alpha oscillation not present in the original signal. On the other hand, due to its better resolution the wavelet coe cients and the reconstructed signal show a clear decrease for this time range. Finally, sweep #5 shows a ringing e ect (i.e. spurious oscillations appearing before the stimulation time point due to time resolution limitations). The oscillation before the stimulation time, marked with an arrow, appears in the digital ltered signal with more amplitude than in the original signal, this e ect being overcome with wavelets.

4.5.3 Results
The grand average wideband ltered (0:5 ; 100Hz) event-related potentials are shown in gure 19. The P100 response is clearly visible upon all stimuli types at about 100ms, best de ned in occipital locations, where it reaches amplitudes about 8 V . In the case of target stimulation, a marked positive peak appears between 400 ; 500ms, according to the expected cognitive (P300) response (see sec. x x1.2). Figure 20 shows the wavelet components in the alpha band (for brevity, \alpha responses") for the subject JA (for a better visualization of the responses, the signal reconstructed from the alpha band wavelet coe cients is shown). One second pre- and one second post-stimulation EEG are plotted. Alpha components corresponding to the pre-stimulus EEG have about 5 V and upon all stimulus types, amplitude enhancements are clearly marked in posterior locations reaching values up to 20 V . Furthermore, in posterior electrodes responses upon TARGET stimulation are prolonged compared with the other two stimulus types. Results for the grand average of the 10 subjects ( g. 21) are similar to the ones outlined for the rst subject. Amplitude increases were distributed over the whole scalp 54

VEP

non-target

target

F3

F4

F3

F4

F3

F4

V -15

V -15

V -15

Cz

Cz

Cz

55
P3

15
-1sec 0 1sec

15
-1sec 0 1sec

15
-1sec 0 1sec

P4

P3

P4

P3

P4

O1

O2

O1

O2

O1

O2

Figure 19: Grand average of the wide-band frequency responses.

Electrode F3 F4 Cz P3 P4 O1 O2 Mean 6.27 6.72 9.80 8.96 8.85 12.65 11.63 SEM 0.56 0.64 0.99 0.85 0.85 1.36 1.22 F3 XXX < 0.05 < 0.01 < 0.01 F4 XXX < 0.05 < 0.01 < 0.01 Cz XXX < 0.05 P3 XXX < 0.01 P4 XXX < 0.01 < 0.05 O1 XXX O2 XXX
Table 2: Multiple factor ANOVA comparison of the maximum alpha band wavelet coe cients for the factor electrode. Note: SEM means standard error of the mean, means no signi cance for the three stimulation types, best de ned in the occipital electrodes. The multiple factor ANOVA test showed no signi cant di erences between stimuli type. Electrode di erences, instead, were signi cant, con rming the predominant localization of the enhancements in the occipital locations, with a lower response in the anterior electrodes (see table 2). It is also interesting to note that in some of the subjects, responses in posterior electrodes upon TARGET stimulation are prolonged in comparison with the NON-TARGET and VEP ones. This coherent alpha activity extended up to a second post-stimulation. With the other stimulus types, enhancements have an abrupt decay after 200 ; 300ms post-stimulus. However, we should remark that this result was not consistent for the whole group. The delay of the maximum response in occipital electrodes was about 180ms after stimulation (see table 3). In parietal electrodes the maximum appeared about 30ms later, and in central and frontal electrodes between 50 ; 100ms after the occipital one. After applying a multiple factor ANOVA test, we veri ed statistically that frontal and central responses were signi cantly delayed in comparison to the occipital ones (p < 0:05).

56

VEP

non-target

target

F3

F4

F3

F4

F3

F4

-15

V -15

V -15

Cz

Cz

Cz

15
-1sec 0 1sec

15
-1sec 0 1sec

15
-1sec 0 1sec

57
P3

P4

P3

P4

P3

P4

O1

O2

O1

O2

O1

O2

Figure 20: Alpha responses of one typical subject.

VEP

non-target

target

F3

F4

F3

F4

F3

F4

-5

-5

-5

Cz

Cz

Cz

-1sec

1sec

-1sec

1sec

-1sec

1sec

P3

P4

P3

P4

P3

P4

O1

O2

O1

O2

O1

O2

Figure 21: Grand average (over N = 10 subjects) of the alpha responses.

58

Electrode Delay SEM F3 F4 Cz P3 P4 O1 O2

F3
246.03 21.09 XXX

F4
290.86 23.35 XXX

Cz
248.20 22.76 XXX

P3
209.20 20.51

P4
209.27 22.10

O1
172.37

O2
181.03

< 0.01
XXX

< 0.01
XXX

< 0.05 < 0.01 < 0.05

XXX

18.42

< 0.05 < 0.01 < 0.05

XXX

17.45

Table 3: Multiple factor ANOVA comparison of the time delays of the maximum alpha band wavelet components for the factor electrode. Note: SEM means standard error of the mean, - means no signi cance, delays in ms.

4.5.4 Discussion
Resonance theory In the present work, occipital maximum enhancements were located between 100 and 200ms, strongly stressing the importance of the alpha band in the generation of the P100 peak and the following N-200 negative rebound. This is in agreement with the results of Stampfer and Basar (1985), who, by digitally ltering auditory evoked potentials in humans, showed that sensory evoked responses were generated and shaped mostly by alpha and theta oscillations. This result is directly related with the resonance theory. In principle, if the alpha enhancements were found only at the time of appearance of the P100-N200 peaks, it can be argued that they were just a result of the morphology of these peaks. However, the prolonged alpha response observed in some of the subjects showed that this is not the case, and that it is more plausible to think about enhanced spontaneous oscillations as a response to the stimulus, according to the resonance theory presented in section x1.3. Functional correlates of event-related alpha oscillations Post-stimulus amplitude increases of alpha band were observed in all electrodes, being signi cantly higher in the occipital ones. Furthermore, anterior responses were delayed in comparison with the posterior ones. Frontal maximum responses appeared between 50 and 100ms after the occipital ones, this delay being statistically signi cant. Consequently, a primary response was reached at occipital locations, later propagating to parietal, central and frontal electrodes. Then, owing to the anatomy of the visual 59

pathway (Mason and Kandel, 1991 Shepherd, 1988), the occipital localization and the short latency of this response points towards a relation between event-related alpha oscillations and the rst stages of sensory processing. Moreover, the idea of sensory processing is reinforced by the independence of the response from the type of stimulus used. Since the 1940s, the event-related alpha response was interpreted as the reactiveness of the brain to sensory stimuli. A sensory alpha response was also observed in several intracortical structures of the cat brain upon auditory and visual stimulation (for a review of these works see Basar et al., 1997). However, this process should be not considered as the only one related with the alpha band. For example, in some subjects alpha responses were prolonged upon TARGET stimuli in posterior locations. Spontaneous alpha activity, low amplitude long-latency responses and low phase locking could be responsible of the lack of statistical signi cance when considering the whole group of subjects. Consequently, it could be conjectured that other processes related with a cognitive function, due to the long latency and the appearance only upon target stimuli, were present in some of the subjects. The cognitive function of alpha band was already pointed out in previous works (Stampfer and Basar, 1985 Kolev and Schurmann, 1992 Basar et al., 1992 Klimesch, 1997 Maltseva and Masloboev, 1997 and Petsche et al., 1997). Further experiments applying di erent types of tasks should be performed in order to learn more about cognitive related alpha responses. Sources of event-related alpha oscillations Another interesting question about alpha rhythms deals with its sources of generation. Alpha responses were best visible in occipital electrodes, but enhancements were also present in other locations with some time delay. Owing to the high di erences in the delays between occipital and anterior electrodes, the conjecture of a single generator and dispersion by volume conduction must be ruled out. A more plausible idea is to think about several generators activated at di erent times and with di erent strength depending on the stimuli and paradigm used. Several previous ndings support a multiple alpha generator theory. Alpha rhythms were described mainly in the thalamus and cortex (Adrian, 1941 Andersen and Andersson, 1968 Lopes da Silva and Storm van Leewen, 1977), but alpha activity was also found in other structures like the brain stem, cerebellum and limbic system, (for a review see Basar et al., 1997). Further evidence showing that alpha rhythms cannot be explained through generators only in the thalamus or cortex are the experiments performed on the cerebral ganglion of aplysia and with the isolated ganglia of Helix pomata (Schutt et al., 1992 Schutt and Basar, 1992). These studies provided evidence that 10Hz activity can be recorded in vitro in such small neural populations, each consisting only of approximately 2000 neurons. 60

In conclusion, our results point towards a distributed alpha system with functional relation to sensory processing and possibly to further processes.

4.6.1 Introduction

4.6 Application to gamma responses of bisensory event-related potentials

During the last ten years gamma-band phenomena gained increasing popularity since the reports of Gray and Singer (1989). Measuring the spike activity of cellular units and multi-units of the cat's striate cortex they found burst activities to be in congruence with oscillatory LFPs (local eld potentials) in orientation-speci c columns. Later ndings (Eckhorn et al., 1988, Gray et al. 1989) led to the proposal of a mechanism sufcient to explain the formation of distant neuronal units to assemblies. Such assembly formation has been suggested as the brain's solution to the problem of encoding. The so-called \binding-theorem" provided a solid explanatory base in congruence with the experimental data. Although being very elegant in solving the problem to encode a quasi-unlimited number of real object features and combinations in a limited number of neural elements interest has, nevertheless, not swayed from the experimental results, which were not directly related to theoretical questions. As, for example, the studies of Tiitinen et al. (1993) and Pulvermuller et al. (1995) could demonstrate, the 40Hz oscillatory activity is not restricted to sensory processing, but can be rather modulated or triggered by cognitive processes as well.

4.6.2 Material and Methods

Event-related potentials were measured in 15 healthy subjects who had neither any known neurological de cit nor reported intake of drugs known to a ect the EEG. Electrodes were placed according to the 10 ; 20 system in F3, F4, Cz, C3, C4, T3, T4, P3, P4, O1 and O2 locations, referenced to linked earlobes. Data were ampli ed with a time constant of 1:5sec: and a low-pass lter at 70Hz. For each single sweep, 1sec: pre- and post-stimulus EEG were digitized with a sampling rate of 250Hz and stored in a hard disk. Subjects were instructed to view and listen passively to the stimuli. A recording session consisted of 3 parts: 1. Registration of 120 sweeps covering 2s time-windows with application of unimodal stimulation presenting a 2 kHz sinusoidal tone-step binaurally. 61

2. Registration of 120 sweeps covering 2s time-windows with application of unimodal stimulation using a rectangular light-step centered in the visual eld at a distance of 1.5 m. 3. Registration of 120 sweeps covering 2s time-windows with application of multimodal stimulation. Stimuli of (1) and (2) were applied simultaneously. After visual inspection of the data, 30 sweeps free of artifacts were selected for each modality for future analysis. A Wavelet Transform was applied to each single sweep using a quadratic B-Spline function as mother wavelet. After a ve octave wavelet decomposition, the components of the gamma band 31 ; 62Hz were analyzed. For each subject the gamma components of the 30 single sweeps were averaged. Results for each subject were averaged obtaining a \grand average". Statistical analysis The wavelet coe cients processed were the ones obtained after averaging the responses of the 30 single trials for each electrode and subject. Then, wavelet coe cients were recti ed and the maximum coe cient was computed in a time window of 250ms post-stimulation. A multiple factor ANOVA (MANOVA) with repeated measures test was applied in order to compare di erences between modalities and electrodes. Furthermore, maximum wavelet components pre- and post-stimulation were compared with t-tests in order to analyze statistical signi cance of the amplitude enhancements for each electrode and modality.

4.6.3 Results
Gamma wavelet coe cients of one typical subject are shown in gure 22. In central locations gamma amplitude enhancements are more pronounced upon bimodal stimulation. In this case, pre-stimulus EEG amplitude is less than 5 (in arbitrary units) and post-stimulus enhancements reach values up to 12. Further increases are seen in right frontal and right occipital electrodes, in the latter case appearing with some delay. The grand average of all the subjects is shown in gure 23. With auditory stimulation enhancements are visible in the central electrode and more poorly de ned in the occipital ones. Visual stimulation shows enhancement only in the central electrode. On the other hand, bimodal stimulation evokes signi cantly higher enhancements than the other two modalities (p < 0:01 see also g. 24) and they are distributed all along the surface EEG, being more pronounced in the central location. Furthermore, bimodal responses are clearly not a linear superposition of the auditory and visual ones, this fact being very clear in central electrodes and specially in the frontal ones, where enhancements occur only upon bimodal stimulation. 62

63 Figure 22: Gamma responses for a typical subject.

Figure 23: Grand average (over N = 15 subjects) of the gamma responses.

64

Figure 24: T-test comparison of the pre- and post-stimulus gamma amplitudes. Finally, in gure 24 comparisons between maximum components pre- and poststimulation are shown for each electrode and modality. Enhancements are clearly higher upon bimodal stimulation in right frontal, posterior and central electrodes, where they reach high signi cance (p < 0:001).

4.6.4 Discussion
Gamma band enhancements were signi cantly higher upon bimodal stimulation, especially in central electrodes, this phenomenon being observed in the grand average as well as in the single subject plots. Furthermore, bimodal responses were not a superposition of the auditory and visual ones. Then, a new behavior solely related with the appearance of both stimuli together (and not with the mechanisms originated by each stimulus when applied separately) should be considered. One possible explanation of this bimodal response is to hypothesize a fast synchronized activity, responsible for transmitting the 65

information that two di erent stimuli are manifestations of the same process. However, the possibility that an intensity rise on joined-polymodal conditions leads to behavioral changes in terms of attention and arousal functions can not be discarded and deserves further studies. Less data is available on the role of 40 Hz-oscillatory processes in bisensory integration. Sheer and Schrock (1986) conducted their studies on focused attention by looking for modi cations in 40 Hz-SSRs (steady-state responses, see sec. x1.2) driven by means of simultaneously applied click-and- ash stimuli, from which one or the other modality had to be ignored. With the stimulus trains being not in phase, they found the 40 Hz-SSRs of the non-focused stimulus modality reduced. Other modes of intermodal perturbation of the 40 Hz-SSRs have been tested by Rohrbaugh et al. (1990). Both salient \foreground" visual and auditory stimuli implied a latency and amplitude decrease in the 40Hz auditory steady-state responses, stimuli of the same modality thereby dominating in e ect. Thus both exemplary studies have discriminative functions in common, which are modi ers of the 40Hz surface EEG.

4.7 Conclusion
Wavelet decomposition proved to be a very useful tool for analyzing brain signals. I would like to remark two advantages of Wavelet Transform over Fourier based methods. First, Wavelet Transform lacks of the requirement of stationarity, this being crucial for avoiding spurious results when analyzing brain signals, already known to be highly nonstationary. Second, owing to the varying window size of the Wavelet Transform, a better time-frequency resolution can be achieved when the signal has patterns involving di erent scales. This is particularly important in the case of event-related potentials, where the relevant response is limited to a fraction of a second, a good time resolution thus being crucial for making any physiological interpretation. In section 4.5.2, I showed with some selected sweeps how the multiresolution decomposition implemented with B-Splines functions leads to a better resolution of the event-related oscillations in comparison with a conventional ideal lter used in several previous works. Moreover, due to the fact that the multiresolution decomposition method is implemented as a ltering scheme, it can be seen as a way to construct lters with an optimal time-frequency resolution. This exempli es and complements the theoretical description of the advantages of wavelets introduced in section 4.2. Furthermore, the multiresolution decomposition is a way of data reduction, thus giving relevant (i.e. non-redundant) coe cients that allows a straightforward implementation of statistical tests. Alpha responses to pattern visual event-related potentials were related with primary sensory processing, having several generators. Furthermore, the analysis of discrete 66

wavelet coe cients allowed an easy design and implementation of statistical tests. In this context, the resolution achieved with Wavelet Transform was very important for obtaining signi cance of the results. The frequency dynamics during the seizures was already described with Gabor Transform (sec. x3.4). However, with wavelets packets it was possible to follow with better accuracy the time evolution of the frequency peaks. Further advantages can be obtained when comparing frequency patterns of di erent channels in order to obtain information about the sources of the seizures. In this case, the time resolution of wavelets can be crucial due to the fact that the seizure spread, from the focus to other locations, can take place in a few milliseconds. Wavelet Transform, due to its varying window size, is more suitable for analyzing signals involving di erent ranges of frequencies. In fact, as showed with alpha and gamma responses to ERPs, frequency behaviors can be resolved up to fractions of a second. On the other hand, with the election of a adequate window, Gabor Transform is more suitable for the analysis of signals with a more limited frequency content as shown in the previous chapter with Grand Mal seizures, in which the interesting activity was limited to the lower frequencies of the EEG (up to 12:5Hz).

67

5 Deterministic Chaos
5.1 Introduction
In the previous chapters I described the application of several time-frequency methods to the analysis of brain signals. Since these methods are linear, a completely di erent approach can be achieved by applying the concepts of non-linear dynamics, also known as Deterministic Chaos theory. Chaotic systems have an apparently noisy behavior but are in fact ruled by deterministic laws. They are characterized by their sensibility on initial conditions. That means, similar initial conditions give completely di erent outcomes after some time. Since chaotic signals look like noise and furthermore, since they also have a broadband frequency spectrum, linear approaches as the ones described in the previous sections are sometimes not suitable for their study. Several methods were developed in order to calculate the degree of determinism (or random nature), complexity, chaoticity, etc. of these signals. Among these, the Correlation Dimension, Lyapunov Exponents and Kolmogorov Entropy have been the most popular. In this chapter I will give a mathematical background of these methods and then I will describe their application to the study of EEG signals. The chapter is organized as follows. After a brief de nition of the basic concepts related with chaos, the mathematical background for the calculation of the Correlation Dimension and Lyapunov Exponents will be described. In section x5.2.4, I will remark some problems in the selection of computational parameters. One of these problems is the stationarity of the signal to be studied. In Section x5.3, I will describe a criterion based on weak stationarity (Blanco et al., 1995a). In section x5.4, I will give a brief review of previous results of chaos analysis of EEG signals. In sections x5.5 and x5.6, I will show the application of Chaos methods to scalp normal EEG recordings and to intracranially recorded Grand Mal seizures (Blanco et al., 1995a,b, 1996a). Finally in section x5.7, I will discuss about the applicability and results of chaos analysis to EEG signals.

5.2 Theoretical Background

5.2.1 Basic concepts
Phase space: It is a common practice in physics to represent the evolution of a system in phase spaces. For example, in mechanics, it is very useful to represent the movement of a system as a pendulum by plotting the position vs. the velocity or more generally, 68

the position vs. the linear momentum. In general, the phase space is identi ed with a topological manifold. An n-dimensional phase space is spanned by a set of n-dimensional \embedding vectors", each one de ning a point in the phase space, thus representing the instantaneous state of the system. The sequence of such states over the time scale de nes a curve in the phase space, called trajectory. Attractor: In some cases, trajectories of dissipative dynamical systems (systems with a volume contraction in the phase space) converge with increasing time to a bounded subset of the phase space. This bounded region to which all su ciently close trajectories (trajectories lying in the basin of attraction) converges asymptotically is called the attractor (Shuster, 1988). According to their topology, several types of attractors can be distinguished (see Abraham and Shaw 1983 Shuster, 1988 Ott et al, 1994 Basar and Quian Quiroga, 1998): 1. Fixed Point: Trajectories in phase space tend to a point at rest. A typical example is a damped pendulum that has come to rest after some time. 2. Limit cycle: The attractor is a closed (one dimensional) curve in the phase space representing a periodic motion. The standard example is a damped periodically driven oscillator. circumstances 3. Torus: The attractor is a two-dimensional toroidal surface. This type of attractor represent a quasiperiodic motion where two incommensurable frequencies correspond to the movement around and along the torus. 4. Strange attractor: The main property of strange attractors is their sensitive dependence on initial conditions. Points that are initially close in the phase space, become exponentially separated after some time. All known strange attractors have a non-integer dimension. Signals corresponding to strange attractors have a random appearance.

5.2.2 Correlation Dimension

The Correlation Dimension (D2) has become the most widely used quantitative parameter to describe attractors. It is a measure of complexity of the system related with its number of degrees of freedom, or in a more intuitive way with its topological dimension. It is already a common practice to calculate D2 and to investigate how it changes upon di erent . Furthermore, since in principle D2 converges to nites values for deterministic systems and do not converge in the case of a random signal, D2 is a good parameter for evaluating the deterministic or noisy inherent nature of a system. As we will see later 69

in this chapter, the question of a deterministic or noisy nature of EEG signals keep the attention of many research groups during the last years.

5.2.3 Calculation of the Correlation Dimension

I will illustrate brie y a proposal made by Grassberger and Procaccia (1983) for computing the correlation dimension D2. First of all, a phase space must be constructed, this space being spanned by a set of embedding vectors. In the case of univariate signals, following a proposal made by Takens (1981) called the \time shift method", n-dimensional vectors are constructed in the following way:

~ x = fx(t) x(t + ) : : : x(t + (n ; 1) )g (43) where is a xed time increment and n is the embedding dimension. Every instantaneous state of the system is therefore represented by the vector ~ x which de nes a point in the phase-space. Once the phase space is constructed, the correlation integral as a function of variable distances R is de ned as:
1 X (R ; j~ xi ; ~ xj j) C (R) = Nlim !1 N 2
i6=j

(44)

where N is the number of data points and is the Heavyside function. Then C (R) is a measure of the probability that two arbitrary points ~ xi , ~ xj will be separated by a distance less than R. Theiler (1986) made a correction to this method in order to avoid spurious temporal correlations. He proposed that the vectors to be compared when calculating the correlation integral, should be distanced at least W data points (ji ; j j > W ), where W is a measure of the temporal correlation of the signal (e.g. the rst zero of the autocorrelation function.) In the case the attractor is a simple curve in the phase space, the number of pair of vectors whose distance in less than a certain radius R will be proportional to R1 . In the case the attractor is a two dimensional surface, C (R) R2 and for a xed point, C (R) R0. Generalizing we can write the following relation:

C (R) RD (45) then, if the number of data and the embedding dimension are su ciently large we obtain
2

The main point is that C (R) behaves as a power of R for small R. By plotting log C (R) versus log R, D2 can be calculated from the slope of the curve. For an attractor with 70

log C (R) D2 = R lim !0 log R

(46)

unknown topology it is necessary to calculate C (R) for several embedding dimensions6. Then, for deterministic signals, the correlation dimension D2 should converge towards a saturation value, for some n > no. Singular Value Decomposition (SVD) In many occasions the amount of noise in a signal makes impossible the calculation of D2. One way of reducing the noise from the signal is by using the SVD method. The main idea is to form a matrix from the embedding vectors (eq. 43) and then, after calculating the eigenvectors and eigenvalues, to make a reconstruction of the signal by rotating the matrix to a reduced base composed by the eigenvectors considered signi cant. Since the eigenvectors with relative small eigenvalues are supposed to be dominated by noise, this phase space reduction is a way of eliminating noise from the signal.

5.2.4 Problems arising when calculating the Correlation Dimension

In order to make the phase space reconstruction described in eq. 43 an appropriate time lag and an appropriate embedding dimension n should be chosen. Since an unfortunate election could lead to wrong results, several criteria were proposed (Blanco et al., 1996a Abarbanel et al., 1993). Concerning the time lag, if it is to small, the components in eq. 43 will have nearly the same value and they will not properly span the phase space (redundancy). In the extreme case, the attractor will be constricted to a diagonal line and its topology will be lost. On the other hand, if the time lag is very high, the components of each embedding vector will become totally unrelated to each other, thus becoming meaningless (irrelevance). Several methods were developed for estimating the optimal time lag, among them, the rst zero of the autocorrelation function has been the most used. This criterion assures linear independence of the components. Some variations were also proposed as for example the value of a certain decay of the autocorrelation function. Furthermore, for taking into account non linear e ects, criteria based on non-linear correlations were de ned (average of mutual information) (Abarbanel et al., 1993 Pjin, 1990). Other approach was given by Rosenstein et al. (1994), who proposed a geometrical-based method. The main idea is to look for a time lag that gives an optimal expansion of the embedding vectors with respect to the diagonal in the phase space. The election of a proper n was also subtle to discussion. If n is too small, the attractor will be not completely unfolded and on the other hand, if n is too high, calculations
6 In fact, following Takens's proposal the dimension n of the embedding phase-space should be chosen at least twice the dimension of the attractor.

71

will be dominated by noise. One rst criteria is to take n > 2D2 (Takens, 1981). In this case, the attractor will be completely unfolded but this criterion assumes a previous estimation of D2. One solution is then to repeat the calculations for di erent sets of embedding dimensions and in case of nding convergence, then to cheek if this criterion is ful lled. However, this is sometimes di cult to implement because the calculations do not depend solely on or on n, on the contrary, they depend on the combination of both (Broomhead and King, 1986 Mees et al., 1987 Albano et al., 1988 Palus and Dvorak, 1992). Then, an estimation of a minimum n would be helpful. In this respect, Kennel et al. (1992) proposed the false nearest neighbors method. The main idea is to calculate if for a certain n nearest neighbors in the phase space still remain close for a dimension n + 1. If this is not the case, then the attractor was not completely unfolded and the embedding dimension must be higher. The procedure is repeated for increasing embedding dimensions until neighbors remain close. Another important problem arises when choosing the linear region in the plots of log C (R) vs. log R and also in how to calculate the slope. Up to the moment there is no unique solution to this question and di erent groups have di erent approaches for obtaining the values of D2 from the log C (R) vs. log R plots. Some researchers prefer to show directly the plots (Pjin et al., 1997) or to limit the calculation to the correlation integral (eq. 44).

5.2.5 Lyapunov Exponents and Kolmogorov Entropy

Another useful tool for characterizing the attractor are the Lyapunov exponents. Lyapunov exponents provide a quantitative indication of the level of chaos of a system. They measure the mean exponential divergence of initially close phase space trajectories with time. As more rapidly two trajectories diverges for a certain period of time, more chaotic is the system and more sensitive to initial conditions. Let us consider a small spherical hypervolume in the phase space. After a short time, as trajectories evolve, the sphere will have an ellipsoid shape with its axes deformed according the Lyapunov exponents. If the system is known to be dissipative, the volume in the phase space will tend to contract and the sum of the Lyapunov exponents will be negative. The longest axis of the ellipsoid will correspond to the most unstable direction, determined by the largest Lyapunov exponent. Usually only this exponent is computed. If it is positive, trajectories will diverge otherwise, they will get closer reaching a non chaotic attractor. Following this argument, a necessary condition for a system to be chaotic is that at least one of the exponents (the largest one) is positive. Lyapunov exponents also give an indication of the period of time in which predictions are possible and this is strongly related with the concept of information theory and entropy. In fact, 72

the sum of the positive exponents (i.e. the ones giving the rate of expansion of the volume), equals the Kolmogorov entropy (Pesin, 1977).

K2 =

>0

(47)

5.2.6 Calculating Lyapunov Exponents

Wolf Method Wolf et al. (1985) proposed an algorithm for calculating the largest Lyapunov exponent. First, the phase space reconstruction is made (eq.43) and the nearest neighbor is searched for one of the rst embedding vectors. A restriction must be made when searching for the neighbor: it must be su ciently separated in time in order not to compute as nearest neighbors successive vectors of the same trajectory. Without considering this correction, Lyapunov exponents could be spurious due to temporal correlation of the neighbors. Once the neighbor and the initial distance (L) is determined, the system is evolved forward some xed time (evolution time) and the new distance (L0 ) is calculated. This evolution is repeated, calculating the successive distances, until the separation is greater than a certain threshold. Then a new vector (replacement vector) is searched as close as possible to the rst one, having approximately the same orientation of the rst neighbor. Finally, Lyapunov exponents can be estimated using the following formula:

where k is the number of time propagation steps. Rosenstein Method Rosenstein et al. (1993) developed another algorithm for the calculation of the largest Lyapunov Exponent in short and noisy time series. As before, the rst step is to make a phase space reconstruction. Then the nearest neighbor for each embedding vector is found. After this, the system is evolved some xed time and the largest Lyapunov Exponent can be estimated as the mean rate of separation of the neighbors. Assuming that the separation is determined by the largest Lyapunov Exponent ( ), then at a time t the distance will be:

L0(ti ) L1 = (t ; t ) ln L (ti;1 ) k 0 i=1

1

k X

(48)

d(t) C e t
73

(49)

where C is the initial separation. Taking the natural logarithm of both sides we obtain:

ln d(t) ln C + t (50) This gives a set of parallel lines for the di erent embedding dimensions, and the largest Lyapunov Exponent can be calculated as the mean slope, averaging all the embedding vectors. Lyapunov exponents are very sensible to the election of the time lag, the embedding dimension and especially to the election of the evolution time. If the evolution time is too short, neighbor vectors will not evolve enough in order to obtain relevant information. If the evolution time is too large, vectors will \jump" to other trajectories thus giving unreliable results.

5.3 Stationarity
Due to the fact that Chaos methods require stationary signals and EEGs are known to be highly non stationary, I will brie y discuss this topic and I will propose a criterion of stationarity. In principle, non-stationarity means that characteristics of the time series, such as the mean, variance or power spectra, change with time. More technically, if we have a time series of discrete observed values fx1 x2 : : : xN g, stationarity means that the joint probability distribution function f12 (x1 x2) depends only on the time di erences jt1 ; t2 j and not on the absolute values t1 and t2 (Jenkins and Watts, 1968). Statistical tests of stationarity have revealed a variety of results in EEGs, and estimates of stationary epochs range from some seconds to several minutes (Lopes da Silva, 1993a). However, whether or not the same data segment is considered stationary, depends on the problem being studied and the type of analysis to be performed. In the case of EEGs, due to the large amount of data needed for the application of the non linear dynamic methods, strict stationarity is almost impossible to achieve. This problem has brought a great variety of results exposed by di erent authors (Basar, 1990 Basar and Bullock, 1989). A less restrictive requirement, called \weak stationarity" of order n, is that the moments up to some order n are fairly stable with time. If the probability distribution of a signal is Gaussian, it can be completely characterized by its mean ( m ), its variance ( 2 ) and its autocorrelation function. In this case, second order stationarity ( n = 2 ), plus an assumption of normality, is enough to assure complete stationarity (Jenkins and Watts, 1968). Consequently, in order to check for stationarity of the EEG segments to be used for analysis, the following procedure was used (Blanco et al., 1995a). 1. The total EEG time series was divided in bins with a xed number of data. The election of the bin length depends on the type of data and on the analysis to be 74

performed. On one hand, must be large enough in order to give a reliable statistic and on the other hand if it is too large, it will not be possible to observe fast changes. 2. The mean and variance for each bin were calculated, and zones in which their values do not have signi cant changes (e.g. less than 20%) were selected. 3. Finally, the corresponding histogram for this zone was constructed, and the normality of the obtained distribution was veri ed. I would like to remark that although the above criterion is arbitrary and does not prove stationarity7, it was very useful as a rst approximation in order to select EEG data segments to work with, or to reject others with clear non stationary behavior. However, I would like to stress that in general elaborated statistical criteria can be hardly ful lled by EEG signals. For a more complete discussion about stationarity related with the application of Chaos methods, see for example Schreiber (1997). Eckman et al. (1987) presented another approach to the problem of stationarity by using the so called recurrence plots. Recurrence plots are based on distance calculations in the phase space. After a phase space reconstruction, the Euclidean distance between the embedding vectors (lets say N in total) is calculated. Then, an N xN plot is performed in the following way: for each pair of embedding vectors (i j ) (i represented in the horizontal axis and j in the vertical axis) a point is plotted if its distance is less than a certain value r. Stationary signals will be characterized by homogeneous plots and non-stationary signals will show inhomogeneities due to varying features of the embedding vectors in the phase space. As an example, Fig. 25 shows the recurrence plot of an stationary one minute EEG signal (see Blanco et al., 1995a). Recurrence plots give an elegant representation of the stationarity of the signal, but in many occasions their interpretation is subjective and further analysis is required.

5.4 Short review of Chaos analysis of EEG signals

5.4.1 Correlation Dimension
Since the pioneering work of Babloyantz and coworkers (Babloyantz, 1985), several groups started to study the dynamics of the brain activity by reconstructing the trajectory of the system in the phase space and by calculating parameters as the Correlation Dimension (D2). Low D2 values correspond to simple systems and high D2 values to
7 Autocorrelation function is not checked, and the choice of what is considered as stable mean and variances is in fact arbitrary.

75

Figure 25: Recurrence Plot of a stationary EEG signal. Horizontal axis (i) and vertical axis (j ) represent the 3200 embedding vectors obtained from the signal after phase space reconstruction. A point is plotted in the position (i j ) if the distance between vectors i and j is less than a certain value r. more complex ones this value tending to in nity, in theory, in the case of noise. D2 proved to be very useful for characterizing the brain dynamics in di erent sleep states. It was found that D2 decreases in deep sleep stages, thus re ecting a synchronization of the EEG (Babloyantz, 1986, Roschke and Aldenho , 1991 Achermann et al., 1994a,b Pradhan et al., 1995). D2 was also used for characterizing the brain dynamics doing mental tasks (Lutzenberger et al., 1995 Molle et al., 1996). Furthermore, D2 was compared between normal subjects and patients with di erent pathologies, with the general result that decreases in D2 are related with abnormal synchronizations of the EEG, as demonstrated in epilepsy (Blanco et al., 1996a Babloyantz and Destexhe, 1986 Pijn et al., 1991 Iasemidis et al., 1990 Lehnertz and Elger, 1995, 1998) and other pathologies as Alzheimer, dementia, Parkinson, depression, schizophrenia or Creutzfeld-Jakob coma. (Pritchard et al., 1994, Besthorn et al., 1995 Stam et al., 1995, Roschke et al., 1994 Gallez and Babloyantz, 1991). For a more detailed review of di erent application of chaos analysis in brain signals I suggest the works of Basar (1989), Pritchard and 76

Duke (1992) and Elbert et al. (1994) and Basar and Quian Quiroga (1998). The Correlation Dimension was also used for characterizing the nature of EEG signals. In principle, converging D2 values point towards a non-linear deterministic nature and diverging D2 values would stress the interpretation of EEG signals as noise. First results showed converging values of D2 in several situations. However, Osborne and Provenzale (1989) showed that ltered noise can also give nite D2 values. Pijn et al. (1991) proposed the use of surrogate tests in order to validate the results obtained with D2. In brief, random (surrogate) signals are constructed from the original one with the same linear characteristics (frequency spectrum) and then, converging D2 values of the original signal should be considered valid, only if the ones of the surrogates diverge. By applying this procedure, they found that values of the original and surrogate data di er in the case of epileptic seizures in the case of the normal EEG, this analysis showing that it is indistinguishable from Gaussian noise. Achermann et al. (1994a) also reported no di erences between EEG in sleep stages and noise.

5.4.2 Lyapunov Exponents

Although the determination of the existence of a positive Lyapunov exponent could be a sign of chaos, publications about Lyapunov exponents are rare in comparison to the ones with Correlation Dimension. I will report here some important ndings with Lyapunov exponents in epilepsy, sleep stages and in di erent pathologies. Epilepsy One of the rst attempts to apply Lyapunov exponents to EEG data was done by Babloyantz and Destexhe (1986) using the Wolf method for the evaluation of a short epileptic \Petit Mal" seizure. They obtained a value of = 2:9 0:6, concluding that although the attractor has a global stability during an epileptic seizure (due to a very low Correlation Dimension), the presence of a positive Lyapunov exponent shows a great sensitivity to initial conditions, giving a rich response to external outputs. Frank et al. (1990) studied a longer \Grand Mal" epileptic seizure. They proposed a modi ed version of the Wolf method, choosing in a di erent way the replacement vectors and making multiple passes through the time series. They point out that Lyapunov exponents are sensible to the evolution time and to the embedding dimension, reporting a value of = 1 0:2 estimated across di erent embedding dimensions. Iasemidis and Sackellares (1991) also used a modi ed Wolf algorithm and analyzed seizures recorded with subdural electrodes. They observed a drop in the Lyapunov Exponents during seizures, with greater values (implying a more chaotic state) postictally than ictally or pre-ictally. Furthermore, they found a phase-locking of the focal 77

sites minutes before the starting of the seizure, with a progressive phase entrainment of the nonfocal ones. They propose this phase-locking as a method for assigning degrees of participation of each focal site and for classifying their importance in the developing of the seizure. Krystal and Weiner (1991), using the same algorithm, obtained similar results in electroconvulsive therapy seizures. Sleep Some groups also had success in evaluating Lyapunov exponents during di erent sleep stages. Babloyantz (1988) reported positive Lyapunov exponents during deep sleep, obtaining a value of = 0:4 ; 0:8 for stage II and a value of = 0:3 ; 0:6 for stage IV. Principe and Lo (1991) reported a greater value of = 2:1 for sleep stage II, but they remark that an accurate value is impossible to obtain because of the complexity of the signal, its time varying nature and the sensibility of the results with the election of the parameters for the calculations. Roschke et al. (1993), following the modi cation of the Wolf algorithm proposed by Frank et al. (1990) calculated the Lyapunov exponent of recordings from 15 healthy male subjects in sleep stages I, II, III, IV and REM. They found in all cases positive values, thus stating that EEG signals are neither quasiperiodic waves, nor simple noise. They also report a decrement in the Lyapunov exponents as sleep becomes slower. Roschke et al. (1994) studied di erences in Correlation Dimension and Lyapunov exponents in sleep recordings of depressive and schizophrenic patients compared with healthy controls. They mainly found alterations during slow sleep in depression, and during REM sleep in schizophrenia. Other studies Gallez and Babloyantz (1991) analyzed the complete Lyapunov spectrum in awake \eyes closed" state (alpha waves), deep sleep (stage IV) and Creutzfeld-Jakob coma. They found in all cases studied at least two positive Lyapunov exponents, increasing this number up to three in the case of alpha waves, implying that alpha waves correspond to a more complex system than the one present during sleep. Stam et al. (1995) studied 13 Parkinson and 9 demented patients against 9 healthy subjects by using the Correlation Dimension, the Lyapunov Exponents and the Kolmogorov entropy calculated from a spatial reconstruction of the embedding vectors (multichanneling). They report a value of = 6:17 for the control subjects, a similar value of = 6:12 (but with lower Correlation Dimension) for Parkinson patients and a signi cant lower value of = 4:84 for demented patients. Wallenstein and Nash (1991) implemented the Wolf algorithm with a varying prop78

agation time. They applied this procedure to study ERPs in central and parietal locations, nding lower values in non-target than in target stimulus, without signi cant di erences between electrodes. In the central location (Cz) they report a value of = 0:397 0:18 for non-target stimulus and of = 0:794 0:44 for the target ones. Although there is a great variance in the results of di erent groups, there is a general agreement that EEG signals have at least one positive Lyapunov exponent, implying that EEGs (in case of having a deterministic origin) re ect a chaotic activity.

5.5 Application to scalp recorded EEGs

5.5.1 Material and Methods
EEG recordings of 6 subjects were studied. Recordings were performed in a \no-task" awake state with eyes closed. Three of these subjects had normal EEG recordings ( N1, N2, N3 ) and other three had abnormal ones ( A1, A2, A3 ). The main abnormalities of these last ones were: slow waves, increases in theta rhythms, very low alpha reactivity and important decrease in the alpha/theta ratio. EEGs were digitized using a 8 bits analog-to-digital (A/D) converter. 20 electrodes were disposed according to the 10 ; 20 system with earlobe references. The data was sampled with a frequency of 256 Hz, and ltered with a high-pass lter at 0.5 Hz and a low-pass lter at 32 Hz. We chose for our analysis the central electrodes because they were the ones less contaminated by artifacts.

5.5.2 Results and Discussion

The objective of this section is to establish some criteria for the analysis of EEG signals previous to any calculation with Chaos methods. This is done in order to avoid spurious results due to unfortunate selections of the segments of data to be analyzed. The rst step is to choose data segments without artifacts. Then, after selecting proper zones free of artifacts or with very few interruptions, the stationarity of the data can be veri ed by applying the criteria described in sec. x5.3. In this case, the length of the data bins was 1000 data points (about 4 seconds of digitized EEG signal). In Fig. 26 and Fig. 27 the mean and variance for the time series denoted by N1 are showed. In this case, mean and variance were considered stable when their changes were in a range of 20% (see sec. 5.3). From these gures it can be concluded that the stationarity criterion previously proposed is satis ed between bins 9 and 26. Note that in this case, the variance does not give any restriction about selected bins to be used in the following analysis. 79

Figure 26: Mean values for bins of 1000 data points for the EEG data N1.

Figure 27: Standard deviation for bins of 1000 data points for the EEG data N1. Fig. 28 shows that the histogram for this selected segment of N1 has approximately a normal distribution (at least, no higher order moments seem to be relevant). The algorithm for the calculation of Correlation Dimension assumed stationarity and noise free time series. Since EEGs are usually contaminated by noise it is convenient prior to the calculation of the Correlation Dimension to apply noise reduction techniques as the Singular Value Decomposition (SVD) (Albano et al., 1988 Broomhead and King, 1986)(see sec. x5.2.3). The SVD has the advantage of reducing the embedding dimension of the system and of partially eliminating the noise of the signal. In Table 4 the total length of the series, the stationary portions and the results of the calculations of D2 and are presented. The time lag was chosen as the rst zero of the autocorrelation function and the embedding dimension by satisfying the Takens 80

Figure 28: Histogram for the EEG data N1 in the selected zone (see text).

N0 Nd De D2 1 N1 34000 15000 13 5:5 ; 6:0 0:26 N2 32000 13000 13 5:0 ; 6:0 0:23 N3 35000 14000 13 4:5 ; 5:5 0:25 A1 35000 15000 13 4:5 ; 5:5 0:28 A2 34000 14000 13 5:0 ; 5:5 0:26 A3 36000 15000 13 4:5 ; 5:0 0:27 Table 4: Correlation Dimension (D2 ) and maximum Lyapunov exponent ( 1) for the di erent EEG signals (N means normal recordings and A abnormal ones). N0 is the total length of the EEG recordings, Nd the length employed for the calculations and De the embedding dimension.

EEG

81

criterion (see sec. x5.2.4). D2 varied between 4:5 ; 6 without a clear di erence between normal and pathological EEG recordings. Maximum Lyapunov exponents were positive in all the cases giving an evidence of chaotic activity (in the case of signals having a deterministic origin).

5.6 Application to intracranially recorded tonic-clonic seizures

5.6.1 Material and Methods
The subject and data analyzed was described in sec. x3.3. The EEG corresponds to a 9-hour intracranial recording from a 21 years old patient with tonic-clonic seizures.

5.6.2 Results and Discussion

Figure 7 shows the EEG signal corresponding to a segment of 64 sec from one depth electrode in the left hippocampus. As seen in the gure, the epileptic seizure starts about second 10 and nishes about second 54. In order to study changes in the EEG behavior the signal was divided in intervals of 8 sec as follows: ( i ) pre-seizure (0 ; 8sec) ( ii ) starting of the seizure (10 ; 18sec) ( iii ) full development of the seizure (21 ; 29sec) and (29 ; 37sec) ( iv ) ending of the seizure (37 ; 44sec) and (44 ; 52sec). For these intervals, stationarity was tested following the criterion described in sec. x5.3. The length of the bins used was of 512 data. In the second interval (10 ; 18sec) the criterion of stationarity was not satis ed due to the fast changes in the EEG morphology. The time delay and the minimum embedding dimension were estimated with the geometrical method introduced by Rosestein and the False Nearest Neighbor method, respectively (see sect. x5.2.4). Fig. 29 displays the two dimensional projection of the phase portraits for the selected intervals. Each EEG segment was characterized by its Correlation Dimensions and maximum Lyapunov exponents. Table 5 shows the stationary zones, optimal time lags , minimum embedding dimensions, Correlation Dimension and maximum Lyapunov exponent for the di erent EEG intervals. These last two parameters were obtained as an average among the corresponding values obtained for all the considered embedding dimensions. These values are in good agreement with those given in the literature in background activity and in epileptic seizures (see sec. x5.4). The seizure is characterized by a drop in the value of the maximum Lyapunov exponent. A similar situation is observed for the Correlation Dimension. From the values 82

Figure 29: Attractors corresponding to the selected zones (see text) during a Grand Mal seizure. 83

EEG segment 0 ; 8 sec 21 ; 29 sec 29 ; 37 sec 37 ; 44 sec 44 ; 52 sec

0:0195 0:0156 0:0156 0:0156 0:0156

(min) De 8 13 10 11 13

D2 4:30 2:60 2:50 2:05 2:15

4:6 4:0 4:0 3:5 3:0

(min) , Table 5: Optimal value of time lag ( in sec ), minimum embedding dimension De Correlation Dimension (D2), and maximum Lyapunov exponent ( 1) for the stationary EEG segments (Data-set size: N = 2048 see text).

showed in table 5, it can be concluded that the dynamical behavior was more regular during the seizure. Summarizing, we found a good evidence that during the epileptic seizure there is a transition from a complex to a simple dynamical behavior. Furthermore, this result yields insights with respect to the theory of how epileptic seizures occur (i.e. as a synchronization of the disordered background EEG activity).

5.7 Conclusion
Deterministic chaos theory gives an alternative new type of analysis compared with the ones given by the traditional methods. However, the application of chaos methods, should be done with care in order to avoid misleading results. In this way a right election of parameters required for the calculations is critical. Chaos analysis has several prerequisites, some of them making impossible its application to the study of EEG signals. One of these prerequisites is the availability of long data recordings but on the other hand, data must be stationary. Consequently, chaos analysis cannot be applied in nonstationary short data recordings as in the case of event-related potentials. In these types of series, the dynamics of the system is changing completely in fractions of a second due to the e ect of the stimulus and therefore, it has no sense to de ne an attractor and to calculate any metric invariant from it. Since methods for calculation of D2 and 1 were developed for ideal signals (i.e. in nite, noise-free and stationary), then, due to the complex structure of the EEGs, results should be considered of relative relevance and, as suggested by Pritchard and Duke (1992), is more realistic to talk about Dimensionality instead of Correlation Dimension. By using this view, several groups reported changes in D2 and 1 values in di erent brain states and pathologies. Nevertheless, these measures seem not to be suit84

able for automatic detection processes or on-line analysis (Jansen, 1991). Furthermore, other methods turned out to be more suitable for these purposes (see Pjin et al., 1997 Gotman, 1990b). Since it was reported that ltered noise could lead to converging low D2 values, conclusions based on absolute values of D2 should be revised and validated. In this respect, the use of surrogate tests (Theiler et. al., 1992) was proposed. Furthermore, absolute results should be stable with respect to di erent election of the parameters necessary for calculations. This type of approach was mainly used for discriminating between deterministic chaos vs. random behavior of the EEGs. In this respect, as I will discuss in detail in section x7, the failure to validate converging low dimensional values of D2 with surrogate tests (then proving a low dimensional chaotic nature) implies that methods of Chaos are not suitable for answering this question rather than implying a random nature of the EEG signals. I would like to remark that I discussed the most used approach of Chaos analysis to EEG signals, namely, by means of the Correlation Dimension, Lyapunov exponents or Kolmogorov entropy. Although their results are not so promising as believed before, Chaos analysis is not limited to these invariants and alternative non linear approaches as for example the study of chaotic synchronization (Arnhold et al., 1999 Quian Quiroga et al., 1999c) could lead to interesting results.

85

6 Wavelet-entropy
6.1 Introduction
One very interesting question related with brain activity deals with its nature. Should brain signals be considered as noisy activity or as a deterministic phenomenon with some degree of order? and in the latter case, how can we quantify this degree of order? Several approaches were performed in order to shed light on this topic, especially after the introduction of the non-linear dynamics (Chaos) theory to the study of EEGs (see section x5). Another approach to this question is to consider the entropy. The concept of thermodynamic entropy is well known in physics as a measure of the order of a system (for a more detailed physical description see Feynman et al., 1964). Left side of g. 30 shows schematically this idea. In the upper graph (situation A), the molecules of a certain gas are isolated in the left side of a recipient because the division was just removed. After some time (situation B), molecules are subject to a free expansion then lling the whole recipient. Situation A corresponds to a more ordered state because the molecules are restricted to the left side therefore, situation A is associated with a lower entropy than situation B. Furthermore, for studying the nature of EEG signals it is very important to consider how these signals behave in several situations. Among these, the study of event-related potentials (i.e. the reaction of the ongoing EEG to an external or internal stimulation) could shed light on this question. As suggested by Basar (1980), ERPs are due to selective enhancements or synchronizations of the spontaneous brain oscillations of the ongoing EEG, thus giving a transition from a disordered to an ordered state (see section 1.3). Then, a method for measuring the entropy appears as an ideal tool for quantifying the \ordering" of the EEG signals upon stimulation. Recently, a measure of entropy de ned from the Fourier power spectrum, the spectral entropy (Powell and Percival, 1979), started to be applied to the study of brain signals (Inouye et al., 1991, 1993). An ordered activity (i.e. a sinusoidal signal) is manifested as a narrow peak in the frequency domain, thus having low entropy (see situation A in the right side of g. 30). On the other hand, random activity has a wide band response in the frequency domain, re ected in a high entropy value (situation B in the right side of g. 30). However, since the spectral entropy is based on the Fourier Transform, it has some disadvantages. As stated in previous chapters, Fourier Transform does not take into account the time evolution of the frequency patterns and furthermore, Fourier Transform requires stationarity of the signals and EEGs are highly non-stationary (Lopes da Silva, 1993a Mpitsos, 1989 Blanco et al., 1995a). 86

ENTROPY

Thermodynamics

Signal analysis

Situation A Order = low entropy

Fourier

Situation B Disorder = high entropy

Fourier

Figure 30: Entropy in thermodynamics and in signal analysis. These disadvantages of the Fourier Transform are partially resolved by using the Gabor Transform (Powell and Percival already de ned a time evolving entropy from the Gabor Transform by using a Hanning window). Nevertheless, as already discused in previous chapters, the selection of the window size is critical due to the Uncertainty Principle and this limitation becomes important when the signal has transient components localized in time as in ERPs. Wavelets have a varying window size, thus allowing a better time-frequency resolution for all the scales. Consequently, a further improvement to the spectral entropy is to de ne the entropy from the Wavelet transform. In the latter case, the time evolution of the frequency patterns can be followed with an optimal time-frequency resolution. In this chapter, I will rst introduce the de nition of the Wavelet-entropy and then I will describe its application to the study of ERPs (Quian Quiroga et al., 1999a,b Rosso et al., 1998). 87

6.2 Theoretical Background

Although the de nition of this time evolving entropy can be done from any timefrequency representation, as Gabor Transform or di erent type of wavelets, I will describe the method from the wavelet coe cients Ci j (where i denotes time and j are the di erent scales) obtained after applying the multiresolution decomposition method (see sec. x4.2.3). Once the coe cients Ci j are known, the energy for each time i and level j can be calculated as

Ei j = Ci2j

(51)

Since the number of components for each resolution level is di erent, I will rede ne the energy by calculating, for each level, its mean value in successive time windows ( t = 128ms) denoted by the index k which will now give the time evolution. Then, the energy will be:
+ t 1 iX Ei j Ek j = N
0

i=i0

(52)

where i0 is the starting value of the time window (i0 = 1 1 + t 1 + 2 t : : :) and N is the number of components in the time window for each resolution level. For every time window k, the total energy can be calculated as:

Ek =
and we can de ne the quantity

Ek j

(53)

kj pk j = E E k

(54)

as a probability distribution associated with the scale level j . Clearly, for each time P window k, j pk j = 1 and then, following the de nition of entropy given by Shannon (1948), the time varying Wavelet-entropy can be de ned as (for further details see Blanco et al., 1998a):

WS k = ;

pk j log2 pk j

(55)

88

6.3 Application to visual event-related potentials

6.3.1 Methods and Materials
In 9 voluntary healthy subjects (no neurological de cits, no medication known to a ect the EEG) two types of experiments were performed: 1. No-task visual evoked potential (VEP): subjects were watching a checkerboard pattern (sidelength of the checks: 50'), the stimulus being a checker reversal. 2. Oddball paradigm (NON-TARGET/TARGET stimuli): subjects were watching the same pattern as above. Two di erent stimuli were presented in a pseudorandom order. NON-TARGET stimuli (75%) were pattern reversal, and TARGET stimuli (25%) consisted in a pattern reversal with horizontal and vertical displacement of one-half of the square side length. Subjects were instructed to pay attention to the appearance of the target stimuli. In both cases, 200 stimuli were presented and the duration of each stimulus was 1 second. Recordings were made following the international 10 ; 20 system in seven di erent electrodes (F3, F4, Cz, P3, P4, O1, O2) referenced to linked earlobes. Data were ampli ed with a time constant of 1:5sec: and a low-pass lter at 70Hz. For each single sweep, 1sec: pre- and post-stimulus EEG were digitized with a sampling rate of 250Hz and stored in a hard disk. After visual inspection of the data, 30 sweeps free of artifacts were randomly selected for each type of stimuli (VEP, NON-TARGET and TARGET) for future analysis. A Wavelet Transform was applied to each single sweep using a quadratic B-Spline function as mother wavelet. The multiresolution decomposition method (Mallat, 1989) was used for separating the signal in frequency bands, de ned in agreement with the traditional frequency bands used in physiological EEG analysis. After a ve octave wavelet decomposition, the components of the following bands were obtained: 62 ; 125Hz, 31 ; 62Hz (gamma), 16 ; 31Hz (beta), 8 ; 16Hz (alpha), 4 ; 8Hz (theta) and the residues in the 0:5 ; 4Hz band (delta). For each subject the results of the wavelet decomposition of the 30 single sweeps were averaged, obtaining the mean components (Ci j ). Finally, from this coe cients the Wavelet-entropy was calculated as described in the previous section. In this case, since the WS is calculated from the averaged wavelet coe cients, only phase-locked oscillations will contribute to it, the others being cancelled (for the same data, an analysis of phase-locking in the alpha band was done in Quian Quiroga et.al., 1999d).

89

X1 0:15 X2 0:36 X3 0:53 Table 6: Wavelet entropy for a sinusoidal signal (X1 ), for a random signal (X3 ) and for 1 (X1 + X3 )). a sinusoidal signal with noise (X2 = 2
Statistical analysis Statistical analysis was performed by computing the mean entropy values in the second pre- and post-stimulus. Signi cance of entropy decreases were calculated for each electrode and stimuli type by comparing minimum pre- and post-stimulation with t-test comparisons.

Signal WS

6.3.2 Results
The Wavelet entropy (WS) was tested with 3 di erent time series. A: a pure sinusoidal signal with a frequency of 12 Hz (X1) B: a random signal (X3) and C: a signal composed of the sum of the two previous ones (X2). The random signal was obtained from a white noise generator. Table 6 shows the results of the WS for the three signals. As expected, the WS is higher for the noisy signal (broad-band spectrum) and minimum for the sinusoidal one (narrow-band spectrum). The event-related responses of one typical subject are shown in gure 31. Only the central and left electrodes are shown, having the right ones similar behavior. Left side of the gure corresponds to VEP, center to NON-TARGET and right side to TARGET stimuli. The P100 response is clearly visible upon all stimuli types at about 100 ms, best de ned in occipital locations. In the case of TARGET stimulation, a marked positive peak appears between 400 and 500 ms, according to the expected cognitive P300 response. Figure 32 shows the results of the band and total energies for the same subject. Only the lower frequency bands are plotted since the higher ones showed no relevant contribution to the total energy. All stimuli types show alpha and theta band increases in occipital locations at 100-200 ms, correlated with the P100-N200 complex. Only upon TARGET stimulation an increase in the posterior locations is visualized in the delta band at 500-600 ms, this increase being related with the cognitive response (P300). Results of the WS for the same subject are shown in gure 33. Decreases in the entropy are more signi cant in posterior sites upon TARGET stimulation, strongly 90

VEP

Non Target
-20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20

Target

F3 Cz
91

-20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20 -20 -10 0 10 20

P3 O1

-1sec.

1sec. -1sec.

1sec. -1sec.

1sec.

Figure 31: Evoked responses of one typical subject (y-axis values in V ).

VEP
500 500

NON-TARGET
500

TARGET

Alpha Theta Delta Total

F3
0 -1.0 500 -0.5 0.0 0.5 1.0 500 0 -1.0 -0.5 0.0 0.5 1.0

F3
0 -1.0 500 -0.5 0.0 0.5 1.0

Cz
0 -1.0 1000 -0.5 0.0 0.5 1.0 1000 0 -1.0 -0.5 0.0 0.5 1.0

Cz
0 -1.0 1000 -0.5 0.0 0.5 1.0

500

P3

500

P3

500

0 -1.0 1000 -0.5 0.0 0.5 1.0

0 -1.0 1000 -0.5 0.0 0.5 1.0

0 -1.0 1000 -0.5 0.0 0.5 1.0

500

500

O1
0 -1.0 -0.5 0.0 0.5 1.0 0 -1.0 -0.5 0.0 0.5 1.0

O1

500

0 -1.0 -0.5 0.0 0.5 1.0

Figure 32: Total and band energy for the recordings of the previous gure.

92

VEP
1 1

NON-TARGET
1

TARGET

F3
0 -1.0 1 -0.5 0.0 0.5 1.0 1 0 -1.0 -0.5 0.0 0.5 1.0

F3
0 -1.0 1 -0.5 0.0 0.5 1.0

Cz
0 -1.0 -0.5 0.0 0.5 1.0 0 -1.0 -0.5 0.0 0.5 1.0

Cz
0 -1.0 -0.5 0.0 0.5 1.0

93
1 1 1

P3
0 -1.0 1 -0.5 0.0 0.5 1.0 1 0 -1.0 -0.5 0.0 0.5 1.0

P3
0 -1.0 1 -0.5 0.0 0.5 1.0

O1
0 -1.0 -0.5 0.0 0.5 1.0 0 -1.0 -0.5 0.0 0.5 1.0

O1
0 -1.0 -0.5 0.0 0.5 1.0

Figure 33: Wavelet Entropy corresponding to the same recording.

correlated with the de nition of the P300 response. There are no decreases of entropy correlated with the P100 response. This is consistent with the distribution of energy described in g. 32, since the P100 peaks have a wide-band frequency composition corresponding to alpha and theta oscillations. On the other hand, the P300 response corresponded only to delta oscillations, thus having a lower entropy. It is very interesting to note that the results of the entropy are not directly related with the ones of the energy. This can be clearly seen, for example, by analyzing the TARGET response of electrode O1, where there is a high energy increase at about 200ms without signi cant changes in the entropy for this time. Considering the whole group of subjects, the grand average of the WS is shown in gure 34. As pointed out with the typical subject, posterior decreases are observable upon TARGET stimulation at about 600 ms, these decreases being clearly correlated with the de nition of the P300 response. On the contrary, P100 peaks produced no relevant decreases in the entropy since they corresponded to a wider range of frequencies. In order to verify statistically the decreases of entropy correlated with the P300 responses, pre- and post-stimulus mean entropy values were compared by using t-tests. As shown in gure 35, entropies are signi cantly decreased in posterior electrodes upon TARGET stimulation. Then, entropy decreases respond to a dynamical process related with the stimulation, rather than to random uctuations.

6.3.3 Discussion
Results of the calculation of the WS in di erent digitally generated signals con rmed the expected correlation between the WS and the complexity of them, being high for the random signal and low for the sinusoidal one. This shows that WS could be used as a measure for describing the behavior of EEG signals, since a noisy activity will have a broadband spectrum and consequently a high entropy, and on the other side, more ordered activity as a sinusoidal signal will have lower entropy. Then, WS gives a new approach to the measure and quanti cation of the order of a system, its physiological interpretation being very interesting due to its relation with tuning of cell groups involved in the generation of the EEG signal. Entropy in signal analysis The concept of entropy emerged last century as a useful state function applied to the study of the thermodynamic of gases (Feynman, 1964). Furthermore, entropy was related with the order of a system and its applications rapidly expanded to several disciplines due to its very interesting meanings. One important milestone was the introduction of the theory of communication (Shannon, 1948 see also Feynman, 1996) relating 94

VEP
1 1

NON-TARGET
1

TARGET

F3
0 -1.0 1 -0.5 0.0 0.5 1.0 1 0 -1.0 -0.5 0.0 0.5 1.0

F3
0 -1.0 1 -0.5 0.0 0.5 1.0

Cz
0 -1.0 -0.5 0.0 0.5 1.0 0 -1.0 -0.5 0.0 0.5 1.0

Cz
0 -1.0 -0.5 0.0 0.5 1.0

95
1 1 1

P3
0 -1.0 1 -0.5 0.0 0.5 1.0 1 0 -1.0 -0.5 0.0 0.5 1.0

P3
0 -1.0 1 -0.5 0.0 0.5 1.0

O1
0 -1.0 -0.5 0.0 0.5 1.0 0 -1.0 -0.5 0.0 0.5 1.0

O1
0 -1.0 -0.5 0.0 0.5 1.0

Figure 34: Grand average of the Wavelet Entropy.

Pre-stim Post-stim
1.0 0.5 0.0
VEP NT T

F3

1.0 0.5 0.0 1.0 0.5

F4

Cz

VEP

NT

P3
1.0 0.5 0.0
VEP NT T

0.0
VEP NT T

**

1.0 0.5 0.0

VEP

P4
*

NT

1.0

1.0 ** ** 0.5 0.0

VEP NT T VEP NT T

O1

0.5 0.0

O2
* p<.05 ** p<.01

Figure 35: T-test comparison of the pre- and post-stimulus entropy values. the infomation content of a message with the concept of entropy. The main idea is that a message composed by a deterministic signal, as for example a sinusoid, contains less information than a random one since it can be described just by its frequency. Shannon derived a formula for calculating the information content of a message (analog to the one used in the de nition of the WS in this work) and he called it information entropy, in relation with the concept of entropy in thermodynamics. Another important advance in the application of entropy to time signals was given by Powell and Percival (1979). Based on the wide of the peaks in the Fourier spectrum they de ned the spectral entropy as a measure of regularity. They applied the method to two potentials (a Henon-Heiles potential and a realistic molecular potential) showing entropy increases upon irregular behaviours of them. Crepeau and King Isaacson (1990) used this approach to study the Logistic map (a simple model of population dynamics) 96

and the Lorenz equations (a model of Rayleigh-Benard convection). By changing the control parameters of the models, they showed increases of the spectral entropy upon the disordering of the systems in their route to chaoticity. Inouye and coworkers (1991, 1993) introduced the spectral entropy to the analysis of EEG signals. From the spectral entropy they de ned an irregularity index and they reported in rest EEGs, more irregularity in anterior than in occipital areas. Furthermore, they reported a greater degree of EEG desynchronization during mental arithmetic in comparison with rest EEG. Stam and coworkers (Stam et al., 1993) instead, de ned an accelaration spectrum entropy from the second derivative of the signal and they reported di erences of this measure in dementia, Parkinson disease (Jelles et al., 1995) and during mental activation (Thomeer et al., 1994). Blanco et al. (1998a) proposed a further improvement by de ning the entropy from the Wavelet Transform due to its advantages over the Fourier transform. Among these, the most important one is that the time evolution of the entropy can be followed. Since the entropy is de ned from the time-frequency representation of the signal, the nearly optimal time-frequency resolution of the Wavelet Transform is crucial for having an accurate measure. This is particularly important in the case of event-related potentials, in which the relevant response is limited to a fraction of a second. Furthermore, Wavelet Transform lacks of the requirement of stationarity. I showed the application of the Wavelet-entropy (WS) to the study of ERPs. WS proved to be a very useful tool for characterizing the event-related responses, furthermore, the information obtained with the WS probed not to be trivially related with the energy and consequently with the amplitude of the signal. This means that with this method, new information can be accessed with an approach di erent from the traditional analysis of amplitude and delays of the event-related responses. WS as a measure of resonance in the brain Following the resonance theory, the ERP can be seen as an evoked synchronization, frequency stabilization, frequency selective enhancement and/or phase reordering of the ongoing EEG, and it should not be interpreted as an additive component to a noisy background EEG (see sec. x1.3). WS appears as an ideal tool for measuring the synchronization of the EEG oscillations upon stimulation. In this context, synchronization means that the group of cells involved in the generation of the response react to the stimulation tuned in frequency. That means, they produce a narrow band in the frequency domain and consequently they are correlated with a decrease in the entropy. Basar (1980) already mentioned the importance of the entropy for understanding the relation between the pre-stimulus ongoing EEG and the event-related responses by making an analogy with the concept 97

of magnetization in physics. WS is a natural measure of the resonance phenomena (see section x1.3). Resonance is related with a decrease of entropy, since only some of the spontaneous oscillations of the ongoing EEG will be enhanced, thus giving a more ordered frequency distribution than the broadband spectrum of the EEG. This relation was showed with the simulations described in table 1, in whose noisy (broadband) signals had higher entropy than the ones corresponding to ordered (narrowband) behaviors. Study of the P100-N200 complex In agreement with previous ndings (overview in Regan, 1989), the P100 peak was best de ned in occipital locations. Furthermore, its task independence (at least the one related with the TARGET stimuli) points towards a relation between this response and primary sensory rocessing. According to the resonance theory, it is very interesting to analyze the frequency characteristics of the di erent evoked peaks in order to understand the response mechanisms involved in its generation. The distribution of energy in the frequency domain showed that the P100-N200 responses corresponded to a wide range of frequencies mostly in the alpha and theta bands. These responses did not produce any signi cant decreases of entropy compared with the one of the normal ongoing EEG due to its wide-band frequency distribution of the involved generators. The involvement of alpha and theta oscillation in the generation of these peaks was already described by Stampfer and Basar (1985), who obtain a similar result by digitally ltering auditory evoked potential in humans. Study of the P300 response According to previous evidence (Regan, 1989 Polich and Kok, 1995 Basar-Eroglu et al., 1992), the P300 positive de ection was observed upon TARGET stimuli, best de ned in parietal and occipital electrodes at about 500-600 ms. These responses, traditionally related with a cognitive process, showed a signi cant decrease in the WS since they involved only delta generators. Although the correlation between the P300 and the entropy decreases was clear due to their common appearance only upon TARGET stimuli and their posterior localization, a di erence of about 100-200 ms was observed between them. This is because after the P300 there is a positive slow rebound in the delta range (see g. 3). Consequently, although the minimum of the P300 is at about 400-500 ms, the P300 complex can be viewed as a delta wave extending up to larger latencies. A similar observation was made by Stampfer and Basar (1985), who after ltering in the delta band the TARGET stimuli of an oddball paradigm in auditory ERP, described the presence of a negative de ection at about 600 ms as a continuation of the positive 98

P300 de ection. Due to the relation between the P300 and the decreases in the entropy, it can be tentatively assumed that the cognitive (P300) response involves a higher degree of order than the one related with the ongoing EEG and the one related with the P100 response. From this, we can postulate that the cognitive response is possibly related with a tuning of the EEG oscillations, thus giving a narrower frequency response re ected in a low entropy value. However, this conjecture should be veri ed with more subjects and with di erent experiments. Furthermore, since in this case the WS was de ned from the mean wavelet coe cients (after averaging the coe cients of the single trials) we can not discard the possibility of other oscillations being present in a non phase-locked way (therefore their contribution being cancelled when averaging). The importance of delta oscillations in the generation of the P300 was reported in several previous works. Basar-Eroglu et al. (1992) by using an auditory oddball paradigm in 10 subjects, pointed out the importance of delta oscillators in the generation of the P300, suggesting that they are related mainly with decision making and matching. Schurmann et al. (1995), also remarked the importance of delta oscillations in the generation of the P300, nding also delta enhancement upon TARGET responses in the single trials. Furthermore, they obtain better averaged responses by selectively averaging single trials with an enhanced delta response. A similar result was recently reported by Demiralp et al. (1999), who after applying a wavelet multirresolution decomposition to evoked responses, used the delta coe cients as discriminators between good and bad single trials.

6.4 Conclusions
Wavelet-entropy proved to be a very useful tool for characterizing the event-related responses. With this method a new approach to the measure of order/noise of a system can be achieved8 . Up to now, this type of descriptions were merely based on Chaos analysis. However, these methods have several prerequisites, some of them making their application to the study of EEG signals impossible. One of these prerequisites is the data length. Chaos analysis cannot be applied to short data recordings as in the case of event-related potentials. In this type of signals, the dynamics of the system is changing completely in fractions of a second due to the e ect of the stimulus, and then it has no sense to de ne an attractor and to calculate parameters as the Correlation Dimension,
8 We should remark that high WS values does not necesarily mean noise since for example chaotic deterministic systems also have a broadband spectra. However, analysis of how the EEG gets tuned in frequency after stimulation is directly related with the question of the deterministic/random nature of it as it will be further discussed in sec.7.1.3.

99

Lyapunov Exponent or Kolmogorov Entropy from them. Previous methods for measuring the spectral entropy from the Fourier Transform are also not suitable for the analysis of ERPs. This is because the Fourier Transform requires stationarity of the signal, and can not describe the time evolution of the frequencies. In this context, the de nition of a time-varying entropy allowed the study of the time evolution of the \ordering" of the ongoing EEG due to stimulation, thus making possible the correlation between changes in entropy of the signal and sensory/cognitive processes of the ERPs. The information obtained with the Wavelet-Entropy turned out not to be trivially related with the energy. This means that with this method, new information can be accessed with a di erent approach than the one obtained by making the traditional analysis of amplitude of the evoked responses. WS appears as a natural measure of order in EEG signals. This is very interesting in order to study synchronizations upon di erent stimuli as in the case of ERPs. Taking into account the resonance theory, WS can measure the degree of synchronicity of the cell groups involved in the di erent responses. Further implementations of the method are in progress in order to optimize the results by having more frequency bands. This would allow a more accurate de nition of the entropy and moreover will allow the de nition of the entropy for di erent frequency bands. However, we would like to remark that in the case of ERPs this is not so easy to achieve since a high time resolution is needed, thus limiting the frequency resolution due to the uncertainty principle.

100

7 General Discussion
In this chapter I will rst discuss how the results described in this thesis are related with several questions of neurophysiology, so far still unresolved with the traditional approaches. The joining of evidence obtained with the described methods sheds light on these topics and allows the conjecture of physiological mechanisms. In the second part of this chapter I will compare these methods, stressing their advantages and disadvantages when applied to the study of EEG signals.

7.1 Physiological considerations

7.1.1 Dynamics of Grand Mal seizures
Chaos analysis of epileptic seizures leads to the general result that during seizures a transition from a complex system to a simpler one takes place. On the other hand, by using the RIR de ned from the Gabor Transform, I showed and quanti ed a well de ned frequency behavior during seizures. Grand Mal seizures were dominated by alpha and theta frequencies. Delta oscillations decreased during them and had an abrupt increase correlated with the clonic phase. With Wavelet Packets, I showed with a better resolution the temporal evolution of these frequency patterns and it was possible to establish that the low frequency activity (3 ; 4Hz) related with the rhythmic contractions of the clonic phase, was in fact originated by the \slowing" of higher frequencies (at least 8 ; 9Hz). Then, during Grand Mal seizures there is a clear frequency dynamics: some seconds after the starting of the seizure alpha and theta activity dominates, these oscillations later becoming slower and when they are in the limit of the delta band (about 3 ; 4Hz) the clonic phase of the seizure starts and delta activity has an abrupt increase dominating the EEG recording. Moreover, it is reasonable to conjecture that the violent contractions of the clonic phase are the response to brain oscillations that are generated in higher frequencies, but owing to the fact that muscles cannot react so fast, muscle activity is then limited to a tonic contraction (muscular tension) until brain oscillations become slower and muscles are capable of contracting in resonance with them. The frequency pattern described is in agreement with studies in animals and with computer simulations. Furthermore, it would be very interesting to investigate possible causes of this behavior. Neuronal fatigue is one of the most plausible explanations. The ring of a neuron is produced as a response to excitatory inputs of neighboring neurons. This connection is done mainly by means of synaptical processes generated by neurotransmitters produced in the neurons. During an epileptic seizure there is an 101

abnormal ring of the neurons, re ected in high amplitude paroxysms as spikes or more generally in a marked increase of the overall amplitude of the EEG. Then, if the neurons are not capable of generating the necessary amount of neurotransmitters for sustaining this uncommon intensive ring, synaptical connections will decrease and excitatory input can not be transmitted. This process, known as neuronal fatigue, produces a change in the oscillatory behavior of cell groups, probably leading to the described \slowing" of the frequency patterns during Grand Mal seizures. It can be also conjectured that this frequency dynamics is due to a recruitment of GABA inhibitory channels, variation in the neurotransmitter balance, etc. Results in this respect should be validated and compared with analysis in intracranial recordings. I also showed with the mean and maximum band frequencies, de ned from the Gabor Transform, the presence of a low amplitude but well de ned peak in the delta band (1 ; 3Hz) during a seizure (referred to as a latent pacemaker). Then, based on the frequency dynamics previously described, it is possible to conjecture that oscillations in the range of the alpha or theta band characteristic of Grand Mal seizures, once they start to oscillate in the range of the delta band (after the \slowing" process), they provoke an abrupt increase of the delta activity due to a resonance phenomenon with the latent delta pacemaker. This resonance and massive synchronization would be responsible of the starting of the clonic phase of the seizure.

7.1.2 Event-related responses

Alpha responses to visual event-related potentials were studied with the Wavelet Transform. Since the sources and functions of alpha oscillations are still subject to discussion, the aim of this approach was to obtain more information about this topic by using a new and powerful method, namely wavelets. In this context, I showed that stimulation leads to alpha enhancements visualized upon all electrodes with some signi cant delays between the anterior and posterior locations. Enhancements were signi cantly higher in occipital locations, had short latency and were statistically independent of the stimulus type (one of them requiring a cognitive process), thus pointing towards a relation between alpha oscillations and sensory (i.e. \pre-cognitive") processing. Furthermore, the delay di erences between responses in di erent electrodes imply that event-related alpha oscillations have several sources, thus ruling out the hypothesis of a unique generator and propagation by volume conduction. It is interesting to remark that the resolution of the Wavelet Transform was crucial for achieving statistical signi cance of the results. However, it is important to mention that \alpha oscillations" is a concept that include several type of di erent processes and the previous interpretation should not be taken as exclusive or general description of them. 102

Wavelet analysis was also applied to study the gamma responses to bimodal stimulation (simultaneous auditory and visual stimulation). This analysis showed a signi cant enhancement of the responses upon bimodal stimulus in comparison with unimodal ones (auditory and visual separately). Then, it was possible to conjecture a relation between gamma oscillations and a fast process responsible of carrying the information that the two sensory perceptions of a bimodal stimulation correspond in fact to the same stimulus. Finally, a very interesting analysis of the event-related responses was accessed by the Wavelet-entropy. By seeing the ERP as a synchronization or selective enhancement of some of the ongoing EEG oscillations, the WS appears as a natural method for obtaining a quantitative measure of the ordering of the EEG spontaneous oscillations due to stimulation. Following this view, the P300 response, traditionally related with cognitive processes, seems to be related with a \tuned" response of EEG oscillations. On the other hand, P100 responses were related with oscillations not \tuned" in frequency, thus having a wider frequency composition.

7.1.3 Are EEG signals chaos or noise?

First reports of Chaos analysis on EEG signals showed convergence of the Correlation Dimension (D2 ) to small values, thus claiming that EEG dynamics correspond to low dimensional deterministic chaos. After the nding that ltered noise can also have convergent low dimensional D2 values, other works stressed the necessity of validation of the metric estimates by means of surrogates tests. In this direction, it was stated that the nature of EEG signals is indistinguishable from noise. In order to deal with this dispute, in the following I will discuss about noise as an inherent property of the system under study, and not about noise produced by the surrounding or by limitations of the measuring systems (ampli ers, etc.). In principle we can state that a system is random when we cannot predict its outcome, but in fact the concept of noise is an idealization since it depends on our ability for analyzing the system (I am excluding in this discussion problems related with quantum mechanics and the uncertainty principle). For example, if from the spin, the initial impulse, the mechanical laws, etc., we can calculate the evolution of a coin ipped in the air, then its outcome will not be random. In the past, the dynamics of air convections of the atmosphere, for example, was considered random, until Lorenz (1969) showed that it can be modeled by three di erential equations, thus being deterministic. Chaos theory has grown very fast, developing new methods that showed how many systems, in former times considered noise, have in fact a deterministic chaotic nature. One of the most used methods for distinguishing deter103

ministic chaos and noise is the D2. However, this method was in principle developed for stationary, noise-free and long data sets. Several e orts were made in order to adapt EEG signals to these requirements with di erent results, among them, singular value decomposition, alternative phase space reconstructions (multichanneling), ltering, etc. But, if it is impossible to distinguish EEG signals from noise with \chaos methods", is it because the EEG corresponds to a noisy phenomenon, or is it because the available methods are not suitable for analyzing the nature of the EEGs? It is like the question of the ipped coin. Is this phenomenon random, or is it that we don't have the ability, or a suitable method to study it? Probably neurophysiological evidence could give more straightforward evidence to resolve this dispute. In fact, EEG alpha oscillations are blockade with eyes opening external or internal stimulation lead to evoked responses that are reproducible upon similar conditions pre-stimulus EEG in uences evoked responses (Basar, 1980) EEG activity can be synchronized in sleep or in pathologies as epilepsy. In summary, EEG patterns have reproducible clear changes upon di erent conditions. Since all these experiments point towards a deterministic origin of the EEG, then, in my opinion, the question about a noisy origin of EEGs based on results obtained with methods as the D2 should be restricted to a discussion of the possibilities of these methods and not to a discussion about the nature of the signal.

7.2 Comparison of the methods

In principle, there is not a \best method" for quantitative analysis of EEG signals. The selection of the adequate method will depend on the type of signal to be studied and on the questions expected to be answered.

7.2.1 Fourier Transform vs. Gabor Transform

Fourier Transform gives a representation in the frequency domain. This allows the visualization of periodicities that would be di cult to observe from the EEG, especially when several rhythms occur simultaneously. Frequencies are grouped in bands, their total or relative power and their topographical distribution constituting the main quantitative method of analysis of the EEG. This procedure was adapted to several commercial systems and has been used as a diagnostical tool in medical centers. It is important to remark that the grouping in frequency bands is not arbitrary because it was shown that these bands can be related to di erent functions, sources and pathologies. However, Fourier Transform gives no information about the time of occurrence of the frequency patterns. Moreover, transients localized in time in the original signal will 104

a ect the whole spectrum and for this reason, in order to avoid spurious e ects in the Fourier spectrum, signals must be stationary. This is particularly important in the case of EEG signals due to the presence of artifacts. Artifacts are alterations (usually of high amplitude) of the ongoing EEG due to causes not related with the brain activity (e.g. blinking, head movements, etc.) and they give spurious e ects in the Fourier spectrum that can lead to misinterpretations. In the case of analyzing background EEGs, this is partially resolved by selecting small \artifact free" segments of data, later averaging the spectrum of the selected segments. However, this widely used procedure is very subjective because it requires the decision of what should be considered an appropriate segment to be analyzed (i.e. which segments are representative of the whole EEG?). An easy and intuitive way to obtain a time evolution of the frequency patterns is by making the Fourier spectrum of successive segments (\windows") of data, then plotting them as a function of time. This procedure is called the Short Time Fourier Transform or Gabor Transform and the plots obtained are called spectograms. The problem of stationarity is partially resolved by taking short windows. Spectograms give an elegant representation of the signal and they are suitable for visualizing large scale frequency variations (i.e. of the order of minutes or hours) as for example for studying sleep stages. However, as I showed in section x3, they are not suitable for analyzing epileptic seizures, in which the frequency patterns change in the order of seconds. In this context, the introduction of the band relative intensity ratio (RIR) and the mean and maximum band frequencies, allowed a more detailed study of the frequency behavior during Grand Mal epileptic seizures, as already summarized in section x7.1.1. Furthermore, with these quantitative parameters it was possible to make a statistical analysis of the frequency behavior in several scalp recorded seizures. Other interesting point to mention is that although in scalp recordings of Grand Mal seizures muscle activity obscures completely the EEG, it was possible to study the frequency patterns by leaving aside the high frequency components related with muscle artifacts, thus obtaining a very interesting quantitative frequency pattern that keeps \hidden" with the traditional analysis of EEG recordings.

7.2.2 Gabor Transform vs. Wavelet Transform

Gabor Transform gives an optimal representation of the EEG in the time-frequency domain. However, one critical limitation arises when choosing the size of the window to be applied due to the Uncertainty Principle. If the window is too narrow, the frequency resolution will be poor, and if the window is too wide, the time localization will be not so precise. In fact, frequencies can not be resolved instantaneously. Then, for slow processes a wide window will be necessary and for data involving fast processes, a narrow 105

window will be more suitable. Due to its xed window size, Gabor Transform is not optimal for analyzing signals having di erent ranges of frequencies. The main advantage of the Wavelet Transform is that the size of the window is variable, being wide when studying low frequencies and narrow when studying the high ones. Then, the time-frequency resolution is automatically adapted (see appendix xA.3), thus being an optimal method for analyzing signals involving di erent ranges of frequencies. Furthermore, due to their adapted window size, wavelets lacks of the requirement of stationarity. Wavelet Transform consists in making a correlation between the original signal and scaled versions of the same \mother function". This mother function can be chosen between a wide range of options each one having di erent characteristics that can be more or less appropriate depending on the signal to be analyzed. At this respect, BSpline functions were very suitable for analyzing EEG signals due to their compact support and smoothness. Successive correlations of the signal to be studied with scaled versions of the wavelet function (and their complementary function) can be arranged in a hierarchical scheme allowing the decomposition of the signal in di erent scales (frequency bands). This method, the multiresolution decomposition, allowed the study of the alpha responses to visual event-related potentials and the study of the gamma responses upon bimodal stimulation. The time-frequency resolution of wavelets was crucial for making physiological interpretations of the event-related responses (see section x7.1.2) because these results were based in statistically signi cant di erences in the amplitudes and time delays of the responses, di erences that in the latter case were in the order of 100 ms and would have been very di cult to resolve with other methods like the Gabor Transform. Furthermore, the access to discrete coe cients in the case of the Wavelet Transform allows a very easy implementation of statistical tests. As showed with alpha and gamma responses to ERPs, with the Wavelet Transform frequency behaviors can be resolved up to fractions of a second. On the other hand, with the election of an adequate window, Gabor Transform is more suitable for the analysis of signals with a more limited frequency content as shown with Grand Mal seizures, in which the interesting activity was limited to the lower frequencies of the EEG (up to 12:5Hz see section x3.4). Although with Gabor Transform the general characteristics of the frequency dynamics during the Grand Mal seizures was already visible (see section x7.1.1), by using an alternative decomposition of the signal based on the Wavelet Transform, the Wavelet Packets, it was possible to study with a better resolution the evolution of the frequency 106

peaks.

7.2.3 Wavelet Transform vs. conventional digital ltering

The multiresolution decomposition based on the Wavelet Transform has several advantages over conventional digital ltering based on the Fourier Transform (ideal lters). Moreover, due to the fact that the multiresolution decomposition method is implemented as a ltering scheme, it can be seen as a way to construct lters with an optimal time-frequency resolution. An important point to be mentioned is that the multiresolution decomposition has a powerful mathematical background in fact, it can be seen as a sequence of correlations between the original signal and dilatations and translations of a single \mother function" (and its complementary function, see section x4.2.3). Furthermore, the optimal mother function to be applied to a certain signal can be chosen based on its mathematical properties or just based on visual features that can be more or less suitable for the analysis of a certain signal (see sec. 4.2.4). Since wavelets have a varying window size adapted to each frequency range, the \ ltering" of some frequency bands does not a ect the morphology of the others. For example, it is well known that when ltering with Fourier based lters the high frequencies of the EEG, the morphology of the low frequencies is also a ected (e.g. in the case of a recording of an epileptic seizure, the shape of the spikes can be modi ed, thus obscuring important details). Moreover, in the case of ERPs, the use of a method based in wavelets avoids unwanted e ects as \ringing" (i.e. the spurious appearance of an stimulus related amplitude enhancement previous to stimulation). In general, it can be stated that the Fourier based ltering gives a more smooth signal than the one obtained by using the multiresolution decomposition due to the nearly optimal time-frequency resolution of the Wavelet Transform for every scale. In order to exemplify these advantages, in section 4.5.2 I showed with some selected sweeps how the multiresolution decomposition implemented with B-Spline functions leads to a better resolution of the event-related responses in comparison with an \ideal lter". Another interesting point is that the multiresolution decomposition gives discrete coe cients, thus making very easy the design and implementation of statistical tests. Similar type of analysis are more di cult to implement from digitally ltered signals. In this respect, the straightforward implementation of statistical tests plus the high time-frequency resolution of wavelets could be critical for obtaining signi cant results as showed for example in the analysis of ERPs (see sec. x4.5 and x4.6). 107

7.2.4 Chaos analysis vs. time-frequency methods (Gabor, Wavelets)

Since all the time-frequency methods described in this thesis are linear, due to the very complex non linear nature of EEG signals, a non-linear analysis as the one performed with the methods of Chaos theory could give new information, an alternative way of visualization, or at least a new way of quanti cation. Moreover, Chaos analysis can put constraints to the system, thus helping in making models. It can also give a useful way of data reduction, this for example being very useful for calculating cross-correlations between di erent EEG channels. However, since these methods have several prerequisites such as stationarity, small amount of noise, long data recordings, etc., they are very di cult to implement in the case of EEG signals and their results should be taken as relative values. In this respect, the validity of Chaos analysis for discriminating between a random or deterministic nature of EEG signals is subject to several criticisms as I described in section x7.1.3. Up to now, parameters such as the D2, 1 , etc., characterized general properties of the EEG, such as the complexity or chaoticity of the brain in di erent states and pathologies. However, the conclusions reached with these methods are still very vague in comparison with more precise conjectures about physiological processes accessed with time-frequency analysis, as described for example in section x7.1.1 in the analysis of epileptic seizures. In this respect, physiological interpretation from results obtained with time-frequency methods is very useful due to the relation of di erent brain oscillations with sources, functions and pathologies of the brain. Moreover, in many cases results reported with Chaos methods are easily visualized in the EEG or with other more simple methods. Then, it is reasonable to check if the results obtained with this approach are in fact new, by comparing them with traditional methods as simple correlations, Fourier analysis or the time-frequency methods described in this thesis. Despite all these di culties, new methods based on Chaos theory should be implemented in the analysis of EEG signals, this approach still being very promising for \visualizing" dynamical properties of the EEG that keep obscure to the linear methods.

7.2.5 Wavelet-entropy vs. frequency analysis

Wavelet-entropy gives new information about EEG signals in comparison with the one obtained by using frequency analysis or other standard methods. In fact, I showed that WS is independent of the amplitude or energy of the signal. Instead of giving an amplitude measurement, the WS shows how the energy is distributed in the di erent frequency bands (see section x6.3.2). WS gives a measure of the \order" of the signal, i.e. signals characterized by narrow peaks in the frequency domain are more ordered than 108

the ones having wide peaks (being the limit case a wide band spectrum corresponding to noise or to a chaotic system). Then, the concept of entropy is very interesting due to the fact that it can be associated to frequency tuning of the neuronal groups. Although in principle the entropy of the signal can be visualized from the Fourier spectrum (i.e. just by looking how narrow or wide are the peaks), the WS allows the following of its time evolution and furthermore, gives a reliable way of quanti cation. Moreover, the advantage of de ning the measure of entropy from the wavelet coe cients and not from other alternative time-frequency distribution as the Gabor Transform is that the resolution of the Wavelet Transform is crucial for analyzing the evolution of fast varying signals as in the case of event-related potentials.

7.2.6 Wavelet-Entropy vs. Chaos analysis

As stated in several parts of this thesis, ERPs can be considered as a selective enhancement, synchronization or in another words, an ordering of the spontaneous EEG oscillations. In this context, Wavelet Entropy appears as a natural and optimal method for measuring this evoked ordering of the EEG signals. Although using a completely di erent approach, and applied to di erent type of EEG signals, Chaos analysis was in principle used for answering similar type of questions. Parameters such as D2 or 1 give a measure of the complexity, chaoticity, and by extension give an idea of the order of the signal. In this context, converging low values of D2 for example, were used as a proof of a low dimensional, deterministic, ordered dynamic of the EEG signals in several situations. However as I already mentioned, chaos methods are very di cult to apply to EEG signals and furthermore, in many cases lead to pitfalls and wrong results. As discussed in section x7.1.3, it is very di cult to resolve disputes about the nature of EEG signals by using these methods and it is more reasonable to consider other physiological evidence as for example the response of the EEG to stimulation (ERP). However, Chaos methods are limited to the analysis of long and stationary EEG recordings, being impossible to implement them to the analysis of fast varying nonstationary signals as ERPs. On the other hand, WS is applicable to ERPs and appears as an ideal parameter for obtaining quantitative answers to these type of questions. In particular, I showed in section x6.3.2 that decreases of entropy after stimulation were correlated with an ordering of the brain rhythms involved in a cognitive process.

109

Conclusion
In this thesis I showed the application of several methods to the analysis of di erent type of EEG signals. Due to the high complexity of EEGs, these methods needed to be adapted or extended. Furthermore, it was possible to compare their abilities in reaching results that allow interesting physiological interpretations. The methods described complement the information obtained by the visual inspection of the EEG carried by trained electroencephalographers and furthermore, they give a quanti cation that allows the performance of statistical analysis. Moreover they can give an alternative way of visualization and in the best case the access to information that keeps \hidden" in the visual inspection of the EEG. In this context, I showed a dynamics of the frequency patterns during Grand Mal seizures. Seizures were dominated by alpha and theta rhythms, later becoming slower with the starting of the clonic phase. Moreover, Chaos analysis showed a transition to a simpler system during the seizures. I also showed a distributed origin of event-related alpha oscillations, this ones being related with primary sensory processing. It was possible to conjecture a relation between gamma oscillations and a process responsible of carrying out the information that two sensory perceptions of a bimodal stimulation correspond in fact to the same stimulus. Responses to unexpected (TARGET) stimulation, traditionally related with cognitive processing, showed a \tuning" in their frequency composition in comparison with the ongoing EEG. All these results give a valuable contribution in understanding the brain dynamics. However, despite all the advances due to the development of new techniques and experiments, is still very little what we know about this topic. The attempts to understand the dynamics of the brain by analyzing the EEG is like trying to understand the conversations occurring in a building by analyzing a sound recorded from far away. Di erent stages and apartments are making di erent tasks, and we are not able to get inside and see what is going on. The EEG, the sound recorded from far away, is still one of our main tools to access to one of the most unknown and complex systems in nature, one of the still elusive \treasures" of science. In this context, we must design clever experiments and we are obliged to develop a great variety of methods and improve them up to the limits of their abilities in order to learn more about the behavior of the brain. And that is what makes the study of the brain so fascinating!

110

A Time-frequency resolution and the Uncertainty Principle

In this section after the introduction of some basic concepts from signal analysis, I will show the proof of the Uncertainty Principle. Then, I will describe its application to the Fourier, Gabor and Wavelet Transform, stressing the advantages of each method related with time-frequency localization properties.

A.1 Preliminary concepts

1 x2 (t)dt = 1) with a corresponding Lets consider a normalized signal x(t) (i.e. ;1 Fourier Transform X (!) 9. The energy density can be written as jx(t)j2 in the time domain or as jX (!)j2 in the frequency domain. Since the total energy or intensity should be the same in both domains, we have the following relation (Parceval's theorem)
R Z 1 1 I= dt = 2 jX (!)j2 d! (56) ;1 ;1 Considering the energy densities per time and frequency, the average time can be de ned as: Z

jx(t)j2

<t>=
and the average frequency as:

;1 1 ;1

t jx(t)j2 dt ! jX (!)j2 d!

(57)

<!>=

(58)

Furthermore, the mean frequency < ! > can be calculated without the previous computation of the Fourier spectrum, just by using the following relation (see demonstration in Cohen, 1995 pp:11)

<!>=

1 ;1

jX (!)j2 d! =

1 ;1

d x(t) dt x (t) 1 i dt

(59)

were denotes complex conjugation. From the energy densities we can also de ne the second order moments as:

< t2 > =

1 ;1

t2 jx(t)j2 dt

(60)

9 for simplicity I will use in this section the angular frequency ! = 2 f

111

d x(t) 2 dt (61) ;1 ;1 dt where we used eq. 59. Finally, the time localization, or duration, can be de ned by means of the standard deviation as: < ! 2 >=
Z

!2

jX (!)j2 d! =

2=

1 ;1

(t; < t >)2 jx(t)j2 dt = < t2 > ; < t >2

(62)

and the frequency localization, or bandwidth, as:

2 !=
Z

1 ;1

(!; < ! >)2

jX (!)j2 d! =

1 ;1

1 d ; < ! > x(t) 2 dt i dt

(63)

A.2 Uncertainty Principle

Theorem p For a normalized signal x(t), if tx(t) ! 0 for jtj ! 1, then
t !

Proof Due to the fact that the mean time and the mean frequency can be changed by using a simple transformation (Cohen, 1995), lets assume that < t >= 0 and < ! >= 0. Then, from eq. 62 and eq. 63 we obtain:
t

1 2

or

1 4

(64)

2=

;1

t2 jx(t)j2 dt
Z

(65)

2 !=

1 ;1
Z

!2
1

jX (!)j2 d! =
t2

1 ;1

and therefore
2

d x(t) 2 dt dt

(66)

2 d x(t) dt dt t ;1 ;1 dt From the Schwartz inequality 10 , by taking f = tx and g = s0 we obtain: 2 !=

jx(t)j2

(67)

2 d dt x(t) dt ;1 ;1 dt and resolving the right hand integrand

Z

t2

jx(t)j2

1 ;1

2 d tx(t) dt x(t) dt

(68)

10 R 1 jf (x)j2 dx R 1 jg(x)j2 dx ;1 ;1

R1

;1 f (x)g(x) dx

112

Z 1 2 (t) +1 1 Z 1 d 1 d tx 1 2 tx(t) dt x(t) dt = 2 t dt x (t) dt = 2 ; 2 x2 (t) dt = ; 2 ;1 ;1 ;1 ;1

replacing in eq. 68 and then in eq. 67, the uncertainty principle is proved.

A.3 Time-frequency resolution of the Fourier, Gabor and Wavelet Transform

Fourier Transform is based in making a correlation between the original signal and complex sinusoidal functions (see eq. 3). Since the Fourier Transform of these functions is:

x(t) = ei! t ! X (!) = 2 (! ; !0)

0

(69)

then, the sinusoidal \mother functions" of the Fourier Transform are perfectly localized in the frequency domain. However, they are not localized in time. As stated in sec. x3.2, Gabor Transform consist in making a correlation between the signal and amplitude modulated complex sinusoids, by using a windowing function. Gabor (1946) suggested the use of a Gaussian function as window due to its simultaneous localization in time and frequency domains. In fact, the Fourier Transform of a Gaussian function is also a Gaussian:

x(t) = g

(t) = e; t2

! X (!) =

2 ;! 4

(70)

1 , thus Moreover, for a Gaussian function the inequality 68 is an equality and ! t = 2 having Gaussian functions the best time-frequency resolution allowed by the uncertainty principle. In the case of the mother functions of the Gabor Transform, i.e. g e;i!t where g is the Gaussian windowing function, the equality in eq. 68 also holds, thus having an optimal time-frequency resolution. The frequency resolution (bandwidth) of a normalized Gaussian function can be calculated by using eq. 63:

(71) 2 and in the case of the functions of the Gabor Transform, g e;i!t , by applying eq. 59 and eq. 63 we obtain for one of these mother functions (! = !0) the same bandwidth but localized in the frequency !0, i.e.:
!

2=

113

Figure 36: Time-frequency resolution of the Gabor Transform. (72) 2 Then, in the case of Gabor Transform, the frequency resolution depends on the window wide, corresponding to small values of (wide windows see eq. 16) a better frequency resolution (but a worst time resolution see also discussion of the window size in sec. x3.2). analyzing The area determined by the frequency window multiplied by the time window is called time-frequency window. Plotting of the time-frequency window is an easy way to visualize the time-frequency resolution: its wide represents the time resolution, its height represents the frequency resolution and its area represents the time-frequency resolution, this last one having a lower bound determined by the uncertainty principle. In the case of the Gabor Transform, once the window is xed, the frequency resolution is the same for all the frequencies (and therefore the time resolution too see g. 36). Then, for low frequency signals, a wide window will be suitable and in the case of high frequencies, a narrow window will do the job. However, due to its xed window size, Gabor Transform is not suitable for analyzing signals with di erent ranges of frequencies. On the other hand, in the case of the Wavelet Transform, the size of the window is adapted, thus giving an optimal resolution for all frequencies (see g. 37). The Wavelet Transform consists in making a correlation between the original signal
!

< ! > = !0

2 =

114

Figure 37: Time-frequency resolution of the Wavelet Transform. and dilatations (contractions) of the same mother function (t)

t;b (73) a were a denotes the di erent scales of dilatation (contraction) and b denotes the time localization of the mother function. Let us denote by ! ~ and by ! the mean frequency and the bandwidth of the mother function (t) and by ! ~ (a) and ! (a) the ones of the scaled functions. Then, by applying eq. 59 and eq. 63 we obtain:
a b (t)

= jaj;1=2

~ 2 (a) = 1 2 (74) ! ~ (a) = ! ! a a2 ! Then, equation 74 shows the main feature of the Wavelet Transform. Wavelets provide a time-frequency window which automatically narrows when studying high frequencies and widens when analyzing low frequencies, but always keeping the area of the timefrequency window constant, as shown in g. 37.

115

116

References
Abarbanel HDI, Brown R, Sidorowich JJ and Tsimring LS. Rev. Mod. Phys. 65, 1331, 1993. Abraham R and Shaw C. Dynamics - The geometry of behavior. Part 2: Chaotic behavior. Aerial Press, California, 1983 Achermann P, Hartmann R, Gunzinger A, Guggenbuhl W and Borbely A. All night sleep EEG and arti cial stochastic control signals have similar correlation dimensions. Electr. Clin. Neurophysiol., 1994a 90: 384-387. Achermann P, Hartmann R, Gunzinger A, Guggenbuhl W and Borbely A. Correlation dimension of the human sleep electroencephalogram: cyclic changes in the course of the night. Eur. J. Neurosc. 1994b 6: 497-500. Ademoglu A, Micheli-Tzanakou E and Istefanopulos Y. Analysis of pattern reversal visual evoked potentials by spline wavelets. IEEE Trans.Biom.Eng., 1997 44:881-890 Adrian ED. A erent discharges to the cerebral cortex from peripheral sense organs. J Physiol, 1941, 100:159-191. Akay M, Akay YM, Cheng P and Szeto HH. Time-frequency analysis of the electrocortical activity during maturation using wavelet transform. Biol. Cybern., 1994 71: 169-176. Albano, M., Muench, J., Schwartz, C., Mees, A., and Rapp, P., Singular Value Descomposition and the Grassberger-Procaccia Algorithm. Phys. Rev. A, 38, 0, 1988. Andersen P, Andersson SA. Physiological Basis of the Alpha Rhythm. New York: Appleton-CenturyCrofts, 1968. Arnhold J, Grassberger P, Lehnertz K and Elger C. A robust method for detecting interdependencies: Application intracranially recorded EEG. Physica D, 1999 (in press). Arroyo S, Lesser R, Gordon B, Fisher R and Uematsu S. Subdural electrodes. In: Niedermeyer, E., and F. H. Lopes da Silva (eds). Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993, pp: 701-711. Babloyantz A, Salazar JM and Nicolis C. Evidence of chaotic dynamics of brain activity during the sleep cycle. Phys. Lett. 3:152-156 1985 Babloyantz A. Evidence of chaotic dynamics of brain activity during the sleep cycle. In: G Meyer-Kress (Ed.), Dimesions and entropies in chaotic systems. Springer, Berlin, 1986, pp:241-245. Babloyantz A and Destexhe A. Low-dimensional chaos in an instance of epilepsy. Proc. Natl. Acad. Sci. 83: 3513-3517 1986. Babloyantz A. Chaotic Dynamics in Brain Activity. In: Erol Basar (ed.) Dynamics of Sensory and Cognitive Processing by the Brain. Springer Verlag, Berlin 1988. Bal T, Von Krosigk M and McCormick D. Synaptic and membrane mechanisms underlying synchronized oscillations in the ferret LGNd in vitro. J. Physiol. Lond. 1995a, 483:641-663. Bal T, Von Krosigk M and McCormick D. Role of the ferret perigeniculate nucleus in the generation of synchronized oscillations in vitro. J. Physiol. Lond. 1995b, 483:665-685.

117

Bartnik EA, Blinowska KJ and Durka PJ. Single evoked potential reconstruction by means of wavelet transform. Biol. Cybern., 1992 67: 175-181 Basar E. EEG-Brain dynamics. Relation between EEG and brain evoked potentials. Amsterdam: Elsevier, 1980. Basar E and Stampfer H. Important associations among EEG-dynamics, event-related potentaials, short-term memory and learning. Int. J. Neurosc., 1985 26: 161-180 Basar E. and Bullock T (Eds.). Brain dynamics. Springer Verlag, Berlin, 1989. Basar E (Ed). Chaos in Brain Function, Springer-Verlag, Berlin, 1990. Basar E, Basar-Eroglu C, Parnefjord R, Rahn E, Schurmann M. Evoked potentials: ensembles of brain induced rhythmicities in the alpha, theta and gamma ranges. In: Basar, E, Bullock TH, eds. Induced rhythms in the brain. Boston: Birkhuser, 1992. Basar E and Schurmann M. Alpha rhythms in the brain: functional correlates. News Physiol Sci, 1996 11: 90-96 Basar E, Schurmann M, Basar-Eroglu C and Karacas S. Alpha Oscillations in brain Functioning: an Integrative Theory. In Brain Alpha Activity - New aspects and Functional Correlates. E Basar, R Hari, FH Lopes da Silva and M Schurmann (eds.). Int. J. Psychophysiology, special issue, 1997 26: 5-29. Basar E and Quian Quiroga R. Chaos in brain function. In: Brain Oscillations: Principles and approaches by E. Basar. Springer Verlag, 1998. Basar E, Rahn E, Demiralp T and Schurmann M. Spontaneous EEG theta activity controls frontal visual evoked potential amplitudes. Electr. Clin. Neurophysiol., 1998 108: 101-109. Basar E. Brain function and oscillations. Volume I: Brain oscillations: principles and approaches. Berlin-Heidelberg-New York: Springer, 1998. Basar E. Brain function and oscillations. Volume II: Integrative brain function: Neurophysiology and cognitive processes. Berlin-Heidelberg-New York: Springer, 1999. Basar E, Demiralp T, Schurmann M, Basar-Eroglu C and Ademoglu A. Oscillatory brain dynamics, wavelet analysis and cognition. Brain and language, 1999 66: 146-183. Basar-Eroglu C, Basar E and Schmielau F. P300 in freely moving cats with intracranial electrodes. Int. J. Neuroscience, 1991 60: 215-226. Basar-Eroglu C, Basar E, Demiralp T and Schurmann M. P300-response: posible psychophysiological correlates in delta and theta frequency channels. A review. Int. J. Psychophysiology, 1992 13: 161-179. Basar-Eroglu C, Kolev V, Ritter B, Aksu F and Basar E. EEG, auditory evoked potentials and evoked rhythmicities in three-year-old children. Int. J. Neurosci, 1994 75: 239-255. Basar-Eroglu C, Struber D, Schurmann M, Stadler M and Basar E. Gamma-band responses in the brain: a short review of psychophysiological correlates and functional signi cance. Int. J. Psychophysiology, 1996a 24: 101-112. Basar-Eroglu C, Struber D, Kruse P, Basar E and Stadler M. Frontal gamma-band enhancement during multistable visual perception. Int. J. Psychophysiology, 1996b 24: 113-126.

118

Berger, H. Uber das Elektrenkephalogramm des Menschen. Arch. Psychiat. Nervenkr. 1929 87:527570. Bertrand O, Bohorquez J and Pernier J. Time-frequency digital ltering based on an invertible wavelet transform: an application to evoked potentials. IEEE Trans. Biomed. Eng., 1994 41: 77-88. Besthorn C, Sattel H, Geiger-Kabisch C, Zerfass R and Forstl H. Parameters of EEG dimensional complexity in Alzheimer's disease. Electr. Clin. Neurophysiol., 1995 95: 84-89. Blanco S, Garc a H, Quian Quiroga R, Romanelli L and Rosso OA. Stationarity of the EEG series. IEEE Eng. Med. Biol., 1995a, 14: 395-399. Blanco S, Quian Quiroga R, Rosso OA and Kochen S. Time-frequency analysis of electroencephalogram series. Phys. Rev. E, 1995b, 51 (3): 2624-2631. Blanco S, Kochen S, Quian Quiroga R, Riquelme L, Rosso OA and Salgado P. Characterization of epileptic EEG time series (I): Gabor transform and nonlinear dynamic methods. In: E M Fernandez Verdaguer and C A D'Attellis (Eds.), Wavelet theory and harmonic analysis in applied sciences. Birkhuser, 1996a. Blanco S, D'Atellis, Isaacson, Rosso OA, Sirne RO. Time-frequency analysis of electroencephalogram series (II): Gabor and Wavelet Transform. Phys. Rev. E, 1996b, 54 (6): 6661-6672. Blanco S, Kochen S, Rosso OA and Salgado P. Applying time-frequency analysis to seizure EEG activity. A method to help to identify the source of epileptic seizures. IEEE Eng. Med. Biol., 1997, 16: 64-71. Blanco S, Figliola A, Quian Quiroga R, Rosso OA and Serrano E. Time-frequency analysis of electroencephalogram series III: information transfer function and wavelets packets. Phys. Rev. E, 57(1):932940 1998a. Blanco S, Creso J, Figliola A, Quian Quiroga R and Rosso OA. Nonlinear Dynamical Analysis of Scalp EEG Epileptic Series. In: Instabilities and Non-Equilibrium Structures VI. E Tirapegui and J Martinez (eds.). Kluwer Academic Press 1998b. Brazier MAB. Spread of seizure discharges in epilepsy: anatomical and electrophysiological considerations. Exp. Neurol., 1972 36:263-272. Broomhead, D. and King, G. Physica D, 20:217 1986. Bullock TH and McClune MC. Lateral coherence of the electrocorticogram: a new measure of brain synchrony. Electroenceph. Clin. Neurophysiol 1989 73:479-498. Bullock TH, McClune MC, Achimowicy JZ, Iragui-Madoz VJ, Duckrow RB and Spencer SS. Temporal uctuations in coherence of brain waves. Proc. Natl. Acad. Sci. USA 1995a 92:11568-11572. Bullock TH, McClune MC, Achimowicy JZ, Iragui-Madoz VJ, Duckrow RB and Spencer SS. EEG coherence has structure in the millimeter domain: subdural and hippocampal recordings from epileptic patients. Electr. Clin. Neurophysiol., 1995b 95:161-177. Chui C. An introduction to wavelets. Academic Press, San Diego, 1992. Clark I, Biscay R, Echeverria M and Virues T. Multiresolution decomposition of non-stationary EEG signals: a preliminary study. Compt. Biol. Med. 1995, 25: 373-382. Cohen L. Time-frequency analysis. Prentice-Hall, New Jersey, 1995.

119

Cooley WJ and Tukey JW. An algorithm for the machine calculation of complex Fourier series. Math. Comput., 1965, 19: 297-301. Crepeau JC and King Isaacson L. On the spectral entropy behavior of self-organizing processes. J. Non-equilib. Thermodyn., 1990 15: 115-126. Darcey T M and Williamson P D. Spatio-temporal EEG measures and their application to human intracranially recorded epileptic seizures. Electr. Clin. Neurophysiol., 1985, 61 (6): 573-587. Demiralp T and Basar E. Theta rhythmicities following expected visual and auditory targets. Int. J. Psychophysiology, 1992 13: 147-160. Demiralp T, Ademoglu A, Schurmann M, Basar-Eroglu C and Basar E. Detection of P300 waves in single trials by the wavelet transform. Brain and language, 1999 66: 108-128. Du y FH, Burch el JL and Lombroso CT. Brain electrical activity mapping (BEAM): a method for extending the clinical utility of EEG and evoked potential data. Ann. Neurol., 1979 5: 309-321. Du y FH. Tophographic mapping of brain electrical activity. London: Butterworths 1986. Dumermuth G and Molinari L. Spectral analysis of EEG background activity. In: A. Gevins and A. Remond (Eds.), Handbook of electroencephalography and clinical neurophysiology, Vol. I: Methods of analysis of brain electrical and magnetic signals. Elsevier, Amsterdam, 1987, pp: 85-130. Eckhorn R, Bauer R, Jordan W, Brosch M, Kruse W, Munk M and Reitboeck H. Coherent oscillations: A mechanism of feature linking in the visual cortex? Biol. Cybern., 60: 121-130, 1988. Eckman, J., Kamphort, S., Ruelle, D. Recurrence Plots of Dinamical Sistems. Europhys. Lett. 1987 4: 973-977. Elbert T, Ray W, Kowalik Z, Skinner J, Graf K and Birmbaumer N. Chaos and physiology: deterministic chaos in excitable cell assemblies. Phys. Rev., 1994 74: 1-47. Feynman R, Leighton R and Sands M. The Feynman lectures on physics. Addison-Wesley, Reading, MA, 1964. Feynman R. Feynman lectures on computation. A Hey and R Allen (eds.). Addison-Wesley, Reading, MA, 1996. Frank GW, Lookman T, Nerenberg MAH and Essex C. Chaotic Time Series Analysis of Epileptic Seizures. Physica D, 46: 427-438 1990. Gabor D. Theory of communication. J. Inst. Elec. Eng., 1946, 93: 429-459. Galambos R. A comparison of certain gamma band (40-Hz) brain rhythms in cat and man. In: Basar E, Bullock TH (eds.) Induced Rhythms in the Brain. Boston Basel Berlin: Birkhuser, 1992. Gallez D and Babloyantz A. Predictability of human EEG: a dynamical approach. Biol. Cybern. 64: 381-391 1991. Garc a H and Quian Quiroga R. La utilidad del EEG cuanti cado en Neuro siologia Clinica. Archivos de Neurologia, Neurocirugia y Neuropsiquiatria, 1998 2: 34-43. Gasser T. EEG Informatics: a Didactic Review of Methods and Applications of EEG Data Processing, A. Remond (Eds), Elsevier, New York, pp 37-55, 1977.

120

Gastaut H and Broughton R. Epileptic seizures. Charles C. Thomas, Spring eld IL, 1972. Gersch W and Goddard GV. Epileptic focus location: spectral analysis method. Science 1970, 169: 701-702. Gevins AS. Overview of computer analysis. In: A. Gevins and A. Remond (Eds.), Handbook of electroencephalography and clinical neurophysiology, Vol. I: Methods of analysis of brain electrical and magnetic signals. Elsevier, Amsterdam, 1987, pp: 31-83. Golomb D, Wang XJ and Rinzel J. Propagation of spindle waves in a Thalamic Slice model. Journal of Neurophysiology 1996, 75: 750-769. Gotman J, Ives JR and Gloor P. Frequency content of EEG and EMG at seizure onset: possibility of removal of EMG artefact by digital ltering. Electr. Clin. Neurophysiol., 1981, 52(6): 626-639. Gotman J. Measurement of small time di erences between EEG channels: method and application to epileptic seizure propagation. Electr. Clin. Neurophysiol., 1983 56:501-514. Gotman J, Ives JR, Gloor P. Long-term monitoring in epilepsy. Electr. Clin. Neurophysiol. Suppl. 37, 1985. Gotman J. Interhemispheric interactions in seizures of focal onset: data from human intracranial recordings. Electr. Clin. Neurophysiol 1987 67:120-133. Gotman J. The use of computers in analysis and display of EEG and evoked potentials. In: D. D. Daly and T. A. Pedley (eds.), Current practice of clinical electroencephalography. Raven Press Ltd., New York, 2nd ed., 1990a, pp: 51-83. Gotman J. Automatic seizures detection: improvements and evaluation. Electroenceph. Clin. Neurophysiol. 1990b 54: 530-540 Grassberger, P. and Procaccia, I. Physica D, 9:189 1983. Gray C and Singer W. Stimulus-speci c neuronal oscillations in orientation columns of cat visual cortex. Proc. Natl. Acad. Sci. USA, 1989 86: 1698-1702. Gray C, Konig P, Engel A and Singer W. Oscillatory responses in cat visual cortex exhibit intercolumnar synchronization which re ects global stimulus properties. Nature, 338: 335-337, 1989. Grossmann and Morlet. Decomposition of Hardy Functions into square integrable wavelets of constant shape. SIAM J. Math. Anal., 1984 15: 723-736. Iasemidis L, Sackellares J, Zaveri H and Williams W. Phase space topography and the Lyapunov exponent of electrocorticograms in partial seizures. Brain Topography, 1990 2: 187-201. Iasemidis LD and Sackellares JC. The evolution with time of the spatial distribution of the largest Lyapunov exponent on the human epileptic cortex. in: DW Duke and WS Pritchard (eds) Measuring Chaos in the Human Brain. World Scienti c, Singapore, 1991 pp: 49-82 Inouye T, Shinosaki K, Sakamoto H, Toi S, Ukai S, Iyama A, Katzuda Y and Hirano M. Quanti cation of EEG irregularity by use of the entropy of the power spectrum. Electr. Clin. Neurophysiol., 1991 79: 204-210. Inouye T, Shinosaki K, Iyama A and Matzumoto Y. Localization of activated areas and directional EEG patterns during mental arithmetic. Electr. Clin. Neurophysiol., 1993 86: 224-230.

121

Jansen BH. Quantitative analysis of electroencephalograms: is there chaos in the future?. Int. J. Biomed. Comput., 27: 95-123 1991. Jelles B, Achtereekte H, Slaets J and Stam C. Speci c patterns of cortical dysfunction in dementia and Parkinson's disease demostrated by the acceleration spectrum entropy of the EEG. Clin. Electroencephalogr., 1995 26: 188-192. Jenkins GM and D. Watts D: Spectral Analysis and its Applications, Holden-Day, San Francisco, 1968. John ER, Harmony T and Valdes-Sosa P. The use of statistics in electrophysiology. In: A. Gevins and A. Remond (Eds.), Handbook of electroencephalography and clinical neurophysiology, Vol. I: Methods of analysis of brain electrical and magnetic signals. Elsevier, Amsterdam, 1987, pp: 497-540. Kaiser G. A friendly guide to wavelets. Birkhauser, Boston 1994. Kaplan PW, Lesser RP. Long-term monitoring. In: Daly DD, Pedley TA, eds. Current practice of clinical electroencephalography. 2nd edition. New York: Raven Press, 1990: 513-534. Kennel M. B., Brawn R. and Abarbanel H. D. I. Determining embedding dimension for phase-space reconstruction using a geometrical construction. Phys. Rev. A 45: 3403-3411 1992. Klimesch W. EEG-alpha rhythms and memory processes. In Brain Alpha Activity - New aspects and Functional Correlates. E Basar, R Hari, FH Lopes da Silva and M Schurmann (eds.). Int. J. Psychophysiology, special issue, 1997 26: 319-340 Kolev V and Schurmann M. Event-related prolongation of induced EEG rhythmicities in experiments with a cognitive task. Int. J. Neurosci, 1992 67: 199-213. Kolev V, Basar-Eroglu C, Aksu F and Basar E. EEG rhythmicities evoked by visual stimuli in threeyear-old children. Int. J. Neurosci., 1994 75: 257-270. Kolev V, Demiralp T, Yordanova J, Ademoglu A and Isoglu Alkac U. Time-frequency analysis reveals multiple functional components during oddball P300. Neuroreport 1997, 8: 2061-5. Krystal AD and Weiner RD. The largest Lyapunov exponent of the EEG in ECT seizures. in: DW Duke and WS Pritchard (eds) Measuring Chaos in the Human Brain. World Scienti c, Singapore, 1991 pp: 113-127 Lehmann D. Multichannel topography of human alpha EEG elds. Electr. Clin. Neurophysiol 1971 31:439-449. Lehmann D. Principles of spatial analysis. In: A. Gevins and A. Remond (Eds.), Handbook of electroencephalography and clinical neurophysiology, Vol. I: Methods of analysis of brain electrical and magnetic signals. Elsevier, Amsterdam, 1987, pp: 309-354. Lehmann D and Koenig T. Spatio-temporal dynamics of alpha brain electric elds, and cognitive modes. In Brain Alpha Activity - New aspects and Functional Correlates. E Basar, R Hari, FH Lopes da Silva and M Schurmann (eds.). Int. J. Psychophysiology, special issue, 1997 26: 99-112. Lehnertz K and Elger C. Spatio-temporal dynamics of the primary epileptogenic area in temporal lobe epilepsy characterized by neuronal complexity loss. Electr. Clin. Neurophysiol., 1995 95: 108-117. Lehnertz K and Elger C. Can epileptic seizures be predicted? Evidence from nonlinear time series analysis of brain electrical activity. Phys. Rev. Lett., 1998 80: 5019-5022.

122

Lopes da Silva FH, van Lierop THMT, Schrijer CF, Storm van Leeuwen W. Organization of thalamic and cortical alpha rhythms: spectra and coherences. Electr. Clin. Neurophysiol., 1973a 35: 626-639. Lopes da Silva FH, van Lierop THMT, Schrijer CF, Storm van Leeuwen W. Essential di erences between alpha rhythms and barbiturate spindles: spectra and thalamo-cortical coherences. Electr. Clin. Neurophysiol., 1973b 35: 641-645. Lopes da Silva FH, Storm van Leeuwen W. The cortical source of the alpha rhythm. Neurosci. Lett., 1977 6: 237-241. Lopes da Silva FH. Neural mechanisms underlying brain waves: From neural membranes to networks. Electr. Clin. Neurophysiol 1991, 79: 81-93. Lopes da Silva F H. EEG analysis: theory and practice. In: E. Niedermeyer and F. H. Lopes da Silva (Eds.), Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993a, pp: 1097-1123. Lopes da Silva F H. Computer-assisted EEG diagnosis: Pattern recognition and brain mapping. In: E. Niedermeyer and F. H. Lopes da Silva (Eds.), Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993b, pp: 1063-1086. Lorenz E. Atmospheric predictability as revealed by naturally occurring analogies. J. Atmos. Sci., 1969 26: 636. Lutzenberger W, Preissl H and Pulvermuller F. Fractal dimension of electroencephalographic time series and underlying brain processes. Biol. Cybern., 1995 73: 477-482. Mallat S. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Analysis and Machine Intell., 1989 2 (7): 674-693. Maltseva IV and Masloboev YP. Alpha rhythm parameters and short-term memory span. In Brain Alpha Activity - New aspects and Functional Correlates. E Basar, R Hari, FH Lopes da Silva and M Schurmann (eds.). Int. J. Psychophysiology, special issue, 1997 26: 369-380. Mason C and Kandel E. Central visual pathways. In E. Kandel, J. Schwartz and T. Jessell (Eds.) Principles of neural science. Elsevier, New York 1991, pp: 420-439. Maurer K and Dierks T. Atlas of brain mapping. Springer-Verlag, Berlin 1991. Mees, A., Rapp, P. and Jennings, L. Phys. Rev. A 36, 340 1987. Meierkord H, Will B, Fish D, Shovone S. The clinical features and prognosis of pseudoseizures diagnosed using video EEG telemetry. Neurology 1991 41: 1643-1646. Molle M, Marshall L, Lutzenberger W, Pietrowsky R, Fehm H and Born J. Enhanced dynamic complexity in the human EEG during creative thinking. Neurosc. Lett., 1996 208: 61-64. Mpitsos, GJ. Chaos in Brain Function and the Problem of Nonstationarity: A Commentary. In: E. Basar and T.H. Bullock (Eds.) Brain Dynamics. Springer-Verlag, Berlin, 1989, pp: 521-535. Niedermeyer E. The normal EEG of the waking adult. In: E. Niedermeyer and F.H. Lopes da Silva (eds). Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993a, pp: 131-152.

123

Niedermeyer E. Depth Electroencephalography. In: E. Niedermeyer and F. H. Lopes da Silva (eds). Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993b, pp: 661-677. Niedermeyer E. Epileptic Seizure Disorders. In: E. Niedermeyer and F. H. Lopes da Silva (eds). Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993c, pp: 461-564. Nuwer MR, Lehmann D, Lopes da Silva FH, Matsuoka S, Sutherling W and Vivert JF. IFCN guidelines for topographic and frequency analysis of EEGs and EPs. Report of and IFCN committee. Electr. Clin. Neurophysiol 1994 91:1-5. Osborne AR and Provenzale A. Finite correlation dimension for stochastic systems with power-law spectra. Physica D, 1989 35: 357-381. Ott E, Sauer T and Yorke J. Coping with Chaos. John Wiley and Sons, Inc. New York, 1994. Palus, M. y Dvorak, I. Singular-Value Decomposition in Attractor Reconstruction: Pitfalls and Precautions. Physica D, 55, 221-234 1992. Pedley TA and Traub RD. Physiological basis of the EEG. In: D. D. Daly and T. A. Pedley (eds.), Current practice of clinical electroencephalography. Raven Press Ltd., New York, 2nd ed., 1990, pp: 107-137. Pesin Y. Lyapunov characteristic exponents and smooth ergodic theory. Usp. Math. Nauk, 32: 55-68 1977. Petsche H, Kaplan S, von Stein A and Filz O. The possible meaning of the upper and lower alpha frequency ranges for cognitive and creative tasks. In: Brain Alpha Activity - New aspects and Functional Correlates. E Basar, R Hari, FH Lopes da Silva and M Schurmann (eds.). Int. J. Psychophysiology, special issue, 1997 26: 77-97. Pfurtscheller G and Lopes da Silva FH (eds.). Functional Brain Imaging. Toronto: Huber 1988. Pijn JPN. Quantitative evaluation of EEG signals in epilepsy. Nonlinear associations, time delays and nonlinear dynamics. PhD Thesis, Universiteit van Amsterdam, 1990. Pijn JPN, Van Neerven J, Noest A and Lopes da Silva FH. Chaos or noise in EEG signals dependence on state and brain site. Electr. Clin. Neurophysiol., 1991, 79: 371 Pjin JP, Velis DN, van der Heiden M, DeGoede J, van Veelen C and Lopes da Silva FH. Nonlinear dynamics of epileptic seizures on basis of intracranial EEG recordings. Brain Topography 9: 249-270 1997. Polich J and Kok A. Cognitive and biological determinants of P300: an integrative review. Biol. Psychol., 1995 41: 103-146. Porter RJ, Sato S. Prolonged EEG and video monitoring in the diagnosis of seizure disorders. In: E. Niedermeyer and F.H. Lopes da Silva, eds. Electroencephalography. 3rd edition. Baltimore: Williams and Wilkins, 1993: 729-738. Powell G and Percival I. A spectral entropy method for distinguishing regular and irregular motion of Hamiltonian systems. J. Phys. A: Math. Gen., 1979 12: 2053-2071.

124

Pradhan N, Sadasivan P, Chatterji S and Narayana Dutt D. Patterns of attractor dimensions of sleep EEG. Comput. Biol. Med., 1995 25: 455-462. Principe JC and Lo PC. Towards the determination of the largest Lyapunov exponent of EEG segments. in: DW Duke and WS Pritchard (eds) Measuring Chaos in the Human Brain. World Scienti c, Singapore, 1991 pp: 156-166 Pritchard WS and Duke DW. Measuring chaos in the brain: a tutorial review of nonlinear dynamical EEG analysis. Int. J. Neurosci. 67:31-80 1992. Pritchard W, Duke D, Coburn K, Moore N, Tucker K, Jann M and Hostetler R. EEG-based, neural-net predictive classi cation of Alzheimer's disease versus control subjects is augmented by non-linear EEG measures. Electr. Clin. Neurophysiol., 1994 91: 118-130. Pulvermuller F, Lutzenberger W, Preiel H and Birbaumer N. Spectral responses in the gamma band: physiological signs of higher cognitive processing? Neuroreport, 6:2059-2064, 1995. Qian S and Chen D. Joint time-frequency analysis: methods and applications. Prentice Hall NJ, 1996. Quian Quiroga R., Pirra L., Podesta C., Leiguarda R. and Rabinowicz A. Time Distribution of Epileptic Seizures during Video-EEG monitoring. Implications for health insurance system in developing countries. Seizure 1997a, 6: 475-477. Quian Quiroga R, Blanco S, Rosso OA, Garc a H and Rabinowicz A. Searching for Hidden Information with Gabor Transform in Generalized Tonic-Clonic Seizures. Electr. Clin. Neurophysiol., 103: 434-439, 1997b. Quian Quiroga R and Schurmann M. Wavelet analysis of visual evoked potentials: alpha responses. In: Recent advances in human neurophysiology. I. Hashimoto and R. Kakigi (eds.). Elsevier Science 1998, pp: 450-454. Quian Quiroga R and Schurmann M. Functions and sources of event-related EEG alpha oscillations studied with the Wavelet Transform. Clin. Neurophysiol., 1999 110: 643-655. Quian Quiroga R, Rosso OA and Basar E. Wavelet-entropy: a measure of order in evoked potentials. EEG Suppl. 49: 298-302 1999a. Quian Quiroga R, Rosso OA Schurmann M and Basar E. Wavelet-entropy in cognitive potentials: a new method shows frequency tuning of EEG-oscillations. Submitted to Electr. Clin. Neurophysiol., 1999b. Quian Quiroga R, Arnhold J and Grassberger P. Driver-Response relationships cannot be learned from synchronization patterns. Submitted to Phys. Rev. E, 1999c. Quian Quiroga R, Basar E and Schurmann M. Phase locking of event-related alpha oscillations. Proceedings of the interdisciplinary Workshop: Chaos in Brain? World Scienti c 1999d, in press. Regan D. Human brain electrophysiology. Evoked potentials and evoked magnetic elds in science and medicine. Elsevier, Amsterdam, 1989. Reilly E. EEG Recording and Operation of the Apparatus. In: E. Niedermeyer and F.H. Lopes da Silva (eds). Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993, pp: 104-124

125

Rohrbaugh J, Varner J, Paige S, Eckardt M and Ellingson R. Auditory and visual event-related perturbations in the 40Hz auditory steady-state response. Electr. Clin. Neurophysiol., 1990 76: 148-164. Roschke J and Aldenho J. The dimensionality of human's electroencephalogram during sleep. Biol. Cybern., 1991 64: 307-313. Roschke J, Fell J and Beckmann P. The calculation of the rst positive Lyapunov exponent in sleep EEG data. Electr. and Clin. Neurophysiol. 86: 348-352 1993. Roschke J, Mann K and Fell J. Nonlinear EEG dynamics during sleep in depression and schizophrenia. Int. J. Neurosci. 75: 271-284 1994. Rosenstein M, Collins J and De Luca C, A practical method for calculating largest Lyapunov exponents from small data sets. Physica D 65: 117-134 1993. Rosenstein M, Collins J and De Luca C, Reconstruction expansion as a geometry-based framework for choosing proper delay times. Physica D, 73: 82 1994. Rosso OA, Quian Quiroga R, Blanco S, Figliola A and Basar E. Wavelet-Entropy applied to Brain Signal Analysis. Proceedings of the IX European Signal Processing Conference, 1998, vol.IV, pp: 2445-2448. Sakowicz O, Quian Quiroga R, Schurmann M and Basar E. Oscillatory gamma-range responses following bimodal stimulation - Wavelet decomposition of human event-related potentials. (1999, to be submitted) Samar VJ, Swartz KP and Raghveer MR. Multiresolution analysis of event-related potentials by wavelet decomposition. Brain and cognition 1995, 27: 398-438. Schi S.J., Aldrouby A, Unser M and Sato S. Fast wavelet transformation of EEG. Electr. Clin. Neurophysiol., 1994a, 91: 442-455. Schi S, Milton J, Heller J and Weinstein S. Wavelet transforms and surrogate data for electroencephalographic spike and seizure localization. Optical Engineering 1994b, 33: 2162-2169. Schreiber T. Detecting and analysing nonstationarity in a time series using nonlinear cross predictions. Phys. Rev. Lett., 1997, 78: 843-846. Schurmann M, Basar-Eroglu C, Kolev V and Basa E. A new metric for analyzing single-trial eventrelated potentials (ERPs): application to human visual P300 delta response. Neuroscience Letters, 1995 197: 167-170. Schurmann M, Basar-Eroglu C and Basar E. Gamma responses in the EEG: elementary signals with multiple functional correlates. Neuroreport, 1997 8: 531-534. Schutt A and Basar E. The e ects of acetylcholine, dopamine and noradrenaline on the visceral ganglion helix pomatia - II. Stimulus evoked eld potentials. Comp. Biochem. Physiol., 1992 102C: 169-176 Schutt A, Basar E and Bullock TH. The e ects of acetylcholine, dopamine and noradrenaline on the visceral ganglion helix pomatia - I. Ongoing compound eld potentials of low frequencies. Comp. Biochem. Physiol., 1992 102C: 159-168 Senhadji L, Dillenseger JL, Wendling F, Rocha C and Linie A. Wavelet analysis of EEG for threedimensional mapping of epileptic events. Annals of Biom. Eng. 1995, 23: 543-552.

126

Serrano E, Ph.D. Thesis, Department of Mathematics, FCEyN, Universidad de Buenos Aires, Argentina, 1996. Shannon CE. A mathematical theory of communication. Bell System Technol. J, 1948, 27:379-423 623-656. Sheer D and Schrock B. Attention. In: Hannay J (ed.), Experimental techniques in human neuropsychology, pp: 95-137. Oxford University Press, New York, 1986. Shepherd, G. Neurobiology. Oxford University Press, New York 1988 Shuster HG. Deterministic Chaos. VCH Publishers, New York, 1988. Speckermann EJ and Elger CE. Introduction to the neurophysiological basis of the EEG and DC potentials. In: E. Niedermeyer and F. H. Lopes da Silva (Eds.), Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993, pp: 15-26. Stam CJ, Tavy D and Keunen R. Quanti cation of alpha rhythm desynchronization using the acceleration spectrum entropy of the EEG. Clin. Electroencephalogr., 1993 24: 104-109. Stam CJ, Jelles B, Achtereekte HAM, Rombouts SARB, Slaets JPJ and Keunen RWM. Investigation of EEG non-linearity in dementia and Parkinsons disease. Electr. Clin. Neurophysiol., 95: 309-317 1995. Stampfer HG and Basar E. Does frequency analysis lead to better understanding of human event related potentials. Intern. J. Neuroscience, 1985 26:181-196. Steriade M, Gloor P, Llinas RR, Lopes da Silva FH and Mesulam MM. Basic mechanisms of cerebral rhythmic activities. Electr. Clin. Neurophysiol 1990, 76: 481-508. Steriade M. Cellular substrates of brain rhythms. In: E. Niedermeyer and F. H. Lopes da Silva (Eds.), Electroencephalography: Basic Principles, Clinical Applications and Related Fields. Baltimore: Williams and Wilkins 3rd ed., 1993, pp: 27-62. Storm van Leewen W, Wieneke G, Spoelstra P and Versteeg H. Lack of bilateral coherence of mu rhythm. Electr. Clin. Neurophysiol., 1978 44:140-146. Strang G. and Nguyen T. Wavelets and lter banks. Wellesley-Cambridge Press, Wellesley, 1996. Takens, F. Dynamical Systems of Turbulence. D. Rand and L. Young (eds.). Springer,Berlin 1981. Thakor NV, Xin-rong G, Yi-Chun S and Hanley DF. Multiresolution wavelet analysis of evoked potentials. IEEE Trans. Biomed. Eng., 1993 40: 1085-1094. Thatcher RW, Krause PJ and Hrybyk M. Cortico-cortical associations and EEG coherence: a twocompartmental model. Electr. Clin. Neurophysiol 1986 64:123-143. Theiler, J. Spurious dimensions from correlation algorithms applied to limited time series data Phys. Rev.A, 34:2427 1986. Theiler J, Eubank S, Longtin A, Galdrikian B and Doyne Farmer J. Testing for nonlinearity in time series: the method of surrogate data. Physica D, 1992 58: 77-94. Thomeer E, Stam C and van Woerkom T. EEG changes during mental activation. Clin. Electroencephalogr., 1994 25: 94-98.

127

Tiitinen H, Sinkkonen J, Reikainen K, Alho K, Lavikainen J and Naatanen R. Selective attention enhances the auditory 40Hz transient response in humans. Nature, 364:59-60, 1993. Unser M, Aldroubi A and Eden M. On the asymptotic convergence of B-Spline wavelets to Gabor functions. IEEE Trans. Inf. Theory., 1992 38: 864-872. Unser M, Aldroubi A and Eden M, Signal Processing 30: 141 1993. Unser M and Aldroubi A. A review of wavelets in biomedical applications. Proc. IEEE 1996, 84: 626-638. Unser M. Ten good reasons for using spline wavelets. Proc. SPIE Vol. 3169, Wavelets Applications in Signal and Image Processing V 1997, pp:422-431. Wallenstein GV and Nash AJ. Does the Largest Lyapunov Exponent Quantify the Dynamical Complexity of the Event-Related Potential? An Algorithm Using Adaptive Step-Size Control for Variable Propagation Times. in: DW Duke and WS Pritchard (eds) Measuring Chaos in the Human Brain. World Scienti c, Singapore 1991, pp: 232-256. Walter D, Rhodes J, Brown D and Adey W. Comprehensive spectral analysis of human EEG generators in posterior cerebral regions. Electr. Clin. Neurophysiol., 1966 20: 224-237. Wang XJ, Golomb D and Rinzel J. Emergent spindle oscillations and intermittent burst ring in a thalamic model: Speci c neuronal mechanisms. Proc. Natl. Acad. Sci. USA 1995, 92:5577-5581. Weaver HJ. Theory of discrete and continuous Fourier analysis. Wiley-Interscience, 1989. Wolf A, Swift J, Swinney H and Vastano J. Determining Lyapunov Exponents from a time series. Physica D, 16: 285-317 1985. Zhang J and Zheng C. Extracting evoked potentials with the singularity detection technique. IEEE Eng. Med. Biol Sept/Oct 1997, pp: 155-161.

128

Biographical sketch
21.03.1967 Nov. 1984 March 1993 March 1991 - March 1994 March 1994 - April 1996 March 1989 - April 1993 April 1993 - May 1996 May/August 1994 April 1993 - May 1995 May 1995 - May 1996 June 1996 - May 1998 June 1998 - August 1998 Nov. 1998 born in Buenos Aires Argentina nished the high school studies nished the studies on Physics at the University of Buenos Aires second class teaching assistant at the Dept. of Physics - University of Buenos Aires rst class teaching assistant at the Dept. of Physics - University of Buenos Aires second class teaching assistant at the Ciclo Basico Comun University of Buenos Aires rst class teaching assistant at the Ciclo Basico Comun University of Buenos Aires given advanced course over Physics applied to Anesthesiology at the Hospital Italiano scholar at the department of Neurophysiology - Institute of Neurological investigations (FLENI), Argentina scholar at the department of Epilepsy - Institute of Neurological investigations (FLENI), Argentina guest scientist at the Institute of Physiology - Medical University of Lubeck, Germany. Supported by the BMBF guest scientist at the Institute of Physiology - Medical University of Lubeck, Germany. Supported by the Med. Univ. Lubeck post-doc position at the Forschungszentrum Julich, Germany

129