High-frequency electroencephalography(hf-EEG): Non-invasive detection of

spike-related brain activityvorgelegt von

Diplom-IngenieurTommaso Fedele

aus Berlin

von der Fakultät IV - Elektrotechnik und Informatikder Technischen Universität Berlin

zur Erlangung des akademischen Grades

Doktor der Naturwissenschaften- Dr. rer. nat. -

genehmigte Dissertation

Promotionsausschuss:

Vorsitzender: Prof. Dr. B. BlankertzBerichter: Prof. Dr. K.-R. MüllerBerichter: Prof. Dr. G. CurioBerichter: Dr. M. BurghoffBerichter: Dr. V. Nikulin

Tag der wissenschaftlichen Aussprache: 26.06.2014

Berlin 2014

Acknowledgments

I would need another PhD thesis to properly thank all the people who walked with me

on the long path that brought me here. My first thought goes to my family, my mother

Cristina, my father Matteo, who taught me hard working while keeping critical vision of

myself and of the world around me, and were close to me from distance together with my

sister Francesca, and her husband Anes, always ready to offer their support. A special

thanks goes also to my grandparents Enza and Renato, who guided my first steps into

education, hoping that in some way they could share with me the joy of this moment.

I thank my supervisors, Gabriel Curio and Klaus Robert Müller, who supported me

with their experience, leaving me free to explore and showing me an inspiring and open

approach to scientific research that I will bring with me for my future. I thank the people

who trusted me at the start of this adventure in the framework of the Bernstein Focus Neu-

rotechnology Berlin (BNFT-B), Gabriel Curio and Martin Burghoff, not only for teaching

me about neuroscience and technology, but also for making me to feel at home in these

years between the group of Neurophysics, at the Campus Benjamin Franklin (CBF) in

Berlin and the PTB (Physikalisch-Technische Bundesanstalt, Berlin). In CBF, I could

enjoy the presence and support of spectacular colleagues, like Vadim Nikulin, Bartosz

Telenczuk, Friederike Hohlefeld, Katherina von Carlowitz-Ghori, Gunnar Waterstraat,

Zubeyir Bayraktaroglu, Jewgeni Kegeles, Natalie Schaworonkow and Irene Sturm.

In PTB, I felt very fortunate to work under the patient guide of Hans-Jürgen Scheer,

enjoying and learning electronic design. A special thank goes to Antonino Cassara and

Rainer Körber, who shared with me long recording sessions accompanied by enlightening

discussions.

Moreover, I would like to thank people I met in the course of these years, who have

been very important for my professional and personal development: in the context of

BNTF-B and Bernstein Centre for Computational Neuroscience: Jan Mehnrt, Christoph

Schmitz, Benjamin Blankertz, Sven Dähne, Matthias Treder, Stefan Haufe, Anne Por-

badnigk, Claudia Sannelli, Vanessa Casagrande, Julia Schaeffer, Alex Susehmil, Philipp

i

Chaput (just to name a few of them) for sharing discussions, suggestions, collaborations

and not less importantly bureaucratic issues.

All this work would have never reached an end if I had not received fundamental sup-

port from old and new friends. I start with the ’Berliners’, hoping to not forget anyone. A

special thanks goes to Janine, who patiently supported me in this last year; of course the

Italian ’gang’, composed by Fabrizio, Alessandro, Massimo, Manolo, Benito, and Mattia;

the people from the glorious Nil Kiosk basketball team: captain Sven, coach Stefano (and

Itxa), Alberto L., Alberto S., Beni, Flo, Gio, Luca, Matteo, Oliver, Pan, Sirio, Toni, and

still Laura Groom and Stefano Larsen, Federica, and Annika; the locals, Merle, Rene’,

Gesine and many many others; still, among the people scattered around the world, I want

to remember Dan, in Canada and Simon, in Australia, Anna and Evaristo in Minneapo-

lis, Andrew in Los Angeles; and back to Italy, a thought to the people who, beside the

distance in space and time, I had very close during these years: Marco, Enrica and Fabio

from Villa Rosa, and finally Simone, Riccardo, Nando, Marco, Alessandro and Valerio

from my beloved Rome!

ii

Abstract

Brain dynamics generate electric fields, whose projections can be recorded at the level

of the scalp. Electroencephalography (EEG), given its low-cost, portability, and mil-

lisecond range temporal resolution, is the more widespread non-invasive technology in

use for the investigation and the monitoring of neurophysiological activity. The com-

plex ensemble of ongoing neural electro-chemical interactions relies on action potentials

propagation and synaptic transmission in a variety of cortical and subcortical structures.

Spatial extension and temporal synchronization of these events define their non-invasive

detectability, quantified in terms of Signal-to-Noise Ratio (SNR). In particular, slower

potentials belong to larger neural substrates, and express higher SNR, while faster events

are more localized and often buried by noise. For this reason, standard EEG recordings

(<100 Hz) mainly reflects mass post-synaptic potentials, which are the input of the neu-

ral networks processing, but generally miss correlated spiking activity, representing the

net computational output. Recent intracranial EEG recordings revealed that frequencies

above 100 Hz convey signals highly informative for application scenarios as movement

decoding for BCI, dissociating spatial attention from movement preparation in motor

cortex, and focus localization in neocortical epilepsy. While such novel neurophysiolog-

ical concepts are advancing rapidly, they are compromised by a progressively decreasing

SNR for higher frequencies. Nevertheless, it was shown that bursts in the range of 600

Hz, mimicking spiking activity, can be isolated in somatosensory evoked potential (SEP)

non-invasive recordings in healthy humans by means of median nerve stimulation. These

fast oscillatory patterns represent an excellent workhorse for the improvement of non-

invasive detection of human high-frequency EEG. The aim of this thesis is to analyze

the factors hindering the high-frequency neural signatures, systematically breaking down

their contribution along the measurement system chain, from the sensor applied to the

scalp to the recording system requirements. The detection and characterization of high-

frequency EEG components will be pursued by integrating the physiological paradigm of

600 Hz SEP bursts with the recent progress in low-noise amplifier technology, and multi-

iii

variate data analysis of scalp potential distribution, in order to achieve a novel integrated

neurotechnology for the noninvasive monitoring of cortical population spikes.

iv

Zusammenfassung

Das Gehirn erzeugt elektrische Felder, deren Potentiale an der Kopfoberfläche aufgenom-

men werden können. Elektroenzephalographie (EEG) ist die meist verbreite Technologie,

die für die Aufnahme solcher Signale verwendet wird, da sie die folgenden Vorteile mit

sich bringt: niedrige Kosten, Mobilität und eine hohe zeitliche Auflösung (im Millisekunden-

Bereich). EEG wird dabei verwendet um neurophysiologische Aktivität zu erforschen

und zu überwachen. Komplexe Ensembles ständig andauernder, elektro-chemischer In-

teraktionen sind auf die Verbreitung von Aktionspotentialen und die synaptischen Trans-

missionen in einer Vielzahl von kortikalen und subkortikalen Strukturen angewiesen.

Die räumliche Ausbreitung und die zeitliche Synchronisation dieser Ereignisse definieren

dabei ihre Erfassungsgrenze bei nicht-invasiven Messungen. Diese ist anhand des Signal-

Rausch-Verhältnisses (Signal-Noise-Ratio, SNR) quantifizierbar. Spezieller lässt sich

sagen, dass langsamere Potentiale von größeren neuronalen Zusammenhängen stammen

und höhere SNR auslösen. Wohingegen schnelle Ereignisse stärker lokalisiert sind und

sehr häufig im Rauschen verborgen bleiben. Aus diesem Grund reflektieren klassischen

EEG-Messungen (<100 Hz) hauptsächlich massenhafte, post-synaptische Potentiale, die

den Input für die Verarbeitung durch neuronale Netzwerke darstellen, aber im Allge-

meinen keine korrelative Spike-Aktivitäten aufzeigen, die dem eigentlich Output der

computativen Netzwerke entsprechen würde. In letzter Zeit w rde mit intrakraniellen

EEG-Aufnahmen gezeigt, dass Frequenzen über 100 Hz hoch informative Signale en-

thalten, die u.a. für Anwendungsszenarien wie Bewegungsdekodierung für Computer-

Gehirn-Schnittstellen (Brain Computer Interfaces, BCIs), dissoziativer räumlicher Aufmerk-

samkeit von Bewegungsvorbereitungen im motorischen Cortex und die Lokalisierung

des Anfallherdes bei Epilepsie verwendet werden können. Während sich solche neuen

neurophysiologischen Konzepte schnell weiterentwickeln, sind sie dennoch durch die

Verringerung der SNR bei höheren Frequenzen beeinträchtigt. Nichtsdestotrotz konnte

gezeigt werden, dass Ausbrüche im Bereich von 600 Hz, die Spike-Aktivität nachahmen,

in somatosensorisch evozierten Potentialen (Somatosensory Evoked Potentials, SEP) auch

v

bei nicht-invasiven Messungen an gesunden Menschen isoliert werden können. Dies

wurde anhand von Stimulierung der Mittelarmnerven gezeigt. Die dabei entstehenden,

schnell oszillierenden Muster repräsentieren ein exzellentes Arbeitspferd für die Weiter-

entwicklung der nicht-invasiven Detektion im menschlichen Hoch-Frequenz-EEG. Das

Ziel der vorliegenden Arbeit ist es, Faktoren zu finden, die eine Detektion der hochfre-

quenten neuronalen Signaturen behindern. Dabei wird ihr Beitrag systematisch entlang

des Messsystems analysiert: von den Sensoren, die auf dem Kopf angebracht werden,

bis hin zu den Voraussetzungen an das Messinstruments. Die Detektion und Charak-

terisierung der hochfrequenten EEG-Komponenten wird durch die Zusammenführung

des physiologischen Paradigmas der 600-Hz-SEP mit den jüngsten Entwicklungen in der

rauscharmen Verstärkertechnologie verfolgt. Durch die zusätzliche Integration von mul-

tivariater Datenanalyse der Potentialen an der Kopfoberfläche entsteht eine neue Neu-

rotechnologie um nicht-invasiv kortikale Populationen von Spikes zu beobachten.

vi

Vita and Publications

Vita

1981 born in Rome, Italy2000 Abitur at Liceo Classico Statale Aristofane, Rome, Italy2000-2006 Bachelor and Master in Biomedical Engeneering at University

of Tor Vergata, Rome, Italy2006 Atmel, development of software and hardware for DSP embedded

application, Rome, Italy2005-2007 Scientific collaborator for the project Energy Management at the

Mechanical Engineering Department of the University of Tor Vergata,Rome, Italy

2008 April 2009 Scientific collaborator for the project Muscular and nervous signalsmeasurement and analysis in human, at the Motor NeurophysiologyDepartment of Santa Lucia Foundation, Rome, Italy

since April 2009 Scientific collaborator in the project B1 of the Bernstein Focus forNeuro-Technology, Berlin (BFNT-B): High frequency electroencephalography(hf-EEG): An emerging neurotechnology for noninvasive detectionof spike related brain activity, Charite, Berlin, Germany

vii

Journal Articles

Scheer, H.J., Fedele, T., Curio, G., Burghoff, M. (2011): Extension of non-invasive EEGinto the kHz range for evoked thalamocortical activity by means of very low noiseamplifiers. Physiol Meas 32, 73:77.

Fedele, T., Scheer, H.J., Waterstraat, G., Telenczuk, B., Burghoff, M., Curio, G.(2012):Towards noninvasive multi-unit spike recordings: mapping 1 kHz EEG signals overhuman somatosensory cortex. Clin Neurophysiol 123, 2370:2376.

Waterstraat, G., Telenczuk, B., Burghoff, M., Fedele, T., Scheer, H.J., Curio, G.(2012):Are high-frequency (600 Hz) oscillations in human somatosensory evoked poten-tials due to phase-resetting phenomena? Clin Neurophysiol 123, 2064:2073.

Fedele, T., Scheer, H.J., Burghoff, M., Waterstraat, G., Nikulin, V., Curio, G. (2013): Dis-tinction between added-energy and phase-resetting mechanisms in non-invasivelydetected somatosensory evoked responses. Conf Proc IEEE Eng Med Biol Soc,1688:1691.

Fedele, T., Scheer, H.J., Burghoff, M., Curio, G., Körber, R. (2014a): Dedicated lownoise EEG MEG systems for 1 kHz SEP detection. (submitted to "Physiol Meas")

Fedele, T., Scheer, H.J., Burghoff, M., Waterstraat, G., Nikulin, V., Curio, G., (2014b):Canonical Correlation Average Regression: source space analysis in critical SNRconditions. (in preparation for NeuroImage)

Fedele, T., Scheer, H.J., Burghoff, M., V., Waterstraat, G., Curio, G. (2014c): Low-noisebio-amplifier technology optimized for high-frequency EEG detection. (under re-vision for Journal of Neural Engineering)

Nikulin, V.*, Fedele, T.*, Mehnert, J., Noack, C., Lipp, A., Steinbrink, J., and Curio, G.,(2014d): Monochromatic ultra slow oscillations in the human electroencephalo-gram. NeuroImage vol. 97C p. 71-80.

Fedele, T.*, Waterstraat, G.*, Curio, G. (2014): Recording human cortical populationspikes non-invasively - an EEG tutorial. (under revision for Journal of NeurosciMethods)

Hilschenz, I., Körber, R, Scheer H.J., Fedele, T., Albrecht, H.H., Cassará A.M., Hartwig,S., Trahms, L., Haase, J., Burghoff, M., 2013. Magnetic resonance imaging atfrequencies below 1 kHz. Magn Reson Imaging 31, 171:177.

Conference Abstracts

Fedele T., Scheer H.J., Waterstraat G., Telenczuk, B., Burghoff, M.., Curio, G. (2011):Towards noninvasive multi-unit spike recordings: Mapping 1 kHz EEG signalsover human somatosensory cortex. 14th European Congress on Clinical Neuro-physiology (ECCN), Rome, Italy.

viii

Fedele, T, Scheer, H.J., Burghoff, M., Curio, G. (2012): A novel low-noise system fornon-invasive high-frequency EEG recordings, World Congress on Medical Physicsand Biomedical Engineering, Beijing, China.

Cassará, A.M., Körber, R., Hilschenz, I., Höfner, N., Voigt J., Fedele, T., Burghoff, M.,Maraviglia, B. (2012): Toward neuronal current spectroscopy at Ultra-Low fieldNMR. Biomed Tech (BMT), Leipzig, Germany.

**Fedele, T., Scheer, H.J., Waterstraat, G., Telenczuk, B., Burghoff, M., Curio, G. (2012):A novel low-noise system for non-invasive high-frequency EEG recordings. Posterpresented at BBCI Workshop, Berlin, Germany.

Nikulin, V.*, Fedele, F.*, Mehnert, J., Noack, C., Lipp, A., Steinbrink, J., and Curio,G. (2012): Monochromatic ultra slow oscillations in the human electroencephalo-gram, Biomag, Paris, France.

Fedele, T. , Scheer, H.J., Rainer, K., Burghoff, M., Curio, G. (2012): Sensitivity of low-noise EEG and MEG systems at 1 kHz, Biomag, Paris, France.

Cassará AM, Körber R, Hilschenz I, Höfner N, Voigt J, Fedele T, Burghoff M, Mar-aviglia B. (2012): Toward neuronal current spectroscopy at Ultra-Low field NMR,Biomag, Paris, France.

Nikulin, V.*, Fedele, F.*, Mehnert, J., Noack, C., Lipp, A., Steinbrink, J., and Curio,G. (2013): Monochromatic ultra slow oscillations in the human electroencephalo-gram, Human Brain Mapping, Seattle, USA.

Fedele., T , Scheer, H.J., Burghoff, M., Waterstraat, G., Nikulin, V., Curio, G. (2013):Non-invasive detection of cortical population spikes: Functional discrimination ofpre- vs. postsynaptic components in SEP at 1 kHz , Human Brain Mapping, Seattle,USA.

*These authors contributed equally to the work.**awarded from BBCI 2012 posters committee

Talks

A novel low-noise system for non-invasive high-frequency EEG recordings. 2012, World

Congress on Medical Physics and Biomedical Engineering, Beijing, China.

A novel low-noise system for non-invasive high-frequency EEG recordings / Monochro-

matic ultra-slow oscillations in the human electroencephalogram. 2013, McGill

Institute, Montreal, Canada.

A novel low-noise system for non-invasive high-frequency EEG recordings / Monochro-

matic ultra-slow oscillations in the human electroencephalogram. 2013, Washing-

ton University, Seattle, USA.

ix

A novel low-noise system for non-invasive high-frequency EEG recordings / Monochro-

matic ultra-slow oscillations in the human electroencephalogram, 2013, Minnesota

University, Minneapolis, USA.

A novel low-noise system for non-invasive high-frequency EEG recordings / Monochro-

matic ultra-slow oscillations in the human electroencephalogram, 2013, Guest Lec-

ture at MPI, Leipzig, Germany.

Teaching

HFO High frequency oscillations. Practical Session. BBCI Summer School, 2012,

Berlin.

MATLAB for MedNeuro, 2013, Charite, Berlin.

x

Contents

1 Introduction 11.1 Scientific Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Fundamentals 72.1 Neurophysiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.1 EEG signal generation . . . . . . . . . . . . . . . . . . . . . . . 72.1.2 High frequency oscillations (HFO): definition of a spectral range . 82.1.3 Physiological HFO . . . . . . . . . . . . . . . . . . . . . . . . . 92.1.4 HFO in epilepsy . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.2 Biophysical model for feasibility study . . . . . . . . . . . . . . . . . . . 162.3 Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.3.1 Time-Frequency Transform . . . . . . . . . . . . . . . . . . . . 202.3.2 Spectral filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3.3 The forward model . . . . . . . . . . . . . . . . . . . . . . . . . 232.3.4 Source space analysis . . . . . . . . . . . . . . . . . . . . . . . . 24

2.4 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.4.1 Somatosensory system . . . . . . . . . . . . . . . . . . . . . . . 272.4.2 SEP recording protocol . . . . . . . . . . . . . . . . . . . . . . . 28

3 High frequency SEP (hf-SEP) - kappa band 313.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.2 Detection of kappa band components by means of low-noise EEG ampli-

fier technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.2.1 Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.3 Multichannel EEG kappa components characterization . . . . . . . . . . 413.3.1 Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.4 Low noise EEG/MEG recordings of kappa band components . . . . . . . 473.4.1 Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

xi

Contents

3.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.5.1 On the low noise hf-EEG . . . . . . . . . . . . . . . . . . . . . . 523.5.2 On the low noise combined hf-MEG/EEG . . . . . . . . . . . . . 55

4 Towards single trial resolution 594.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4.2.1 Simulation settings . . . . . . . . . . . . . . . . . . . . . . . . . 634.2.2 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . 644.2.3 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.3.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.3.2 Sensor space analysis of SEP data . . . . . . . . . . . . . . . . . 704.3.3 Source space analysis of SEP data . . . . . . . . . . . . . . . . . 704.3.4 Source reconstruction on patterns . . . . . . . . . . . . . . . . . 734.3.5 Latency analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 764.3.6 Time envelope analysis . . . . . . . . . . . . . . . . . . . . . . . 78

4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5 Low-noise bio-amplifier technology 855.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

5.1.1 Single Ended versus Differential Inputs . . . . . . . . . . . . . . 865.1.2 Common Mode and Isolation Mode . . . . . . . . . . . . . . . . 875.1.3 Interference model . . . . . . . . . . . . . . . . . . . . . . . . . 885.1.4 Low-noise bio-amplifier: biophysical model and design . . . . . . 905.1.5 Noise calculation from the input stage . . . . . . . . . . . . . . . 93

5.2 The new low-noise amplifier design . . . . . . . . . . . . . . . . . . . . 965.2.1 Electro-technical model for the interference estimation . . . . . . 965.2.2 Stray capacitances estimation . . . . . . . . . . . . . . . . . . . 101

5.3 Set up implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1035.3.1 First single channel prototype . . . . . . . . . . . . . . . . . . . 1035.3.2 HF-EEG amplifier . . . . . . . . . . . . . . . . . . . . . . . . . 104

5.4 EEG Recordings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065.4.1 Inside/outside of the shielded room . . . . . . . . . . . . . . . . 1065.4.2 Neurophysiological SEP recordings in a clinical environment . . . 108

5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6 Summary and Conclusion 117

A Appendix 121

xii

Contents

A.1 Experimental Sessions . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

List of Equations 125

List of Figures 127

List of Tables 129

Bibliography 131

xiii

1. Introduction

The human brain is a highly interconnected network of about 1012 neurons whose collec-

tive activity generates complex human behaviours, from sensorimotor response to con-

sciousness. This fascinating machine can be partially accessed by placing electrodes on

the head and recording the far field electric potential projections available on the scalp.

The detected signal represents the blurry sight of intense biochemical and electrical in-

teractions, responsible for the neuronal communications, as post-synaptic and spiking

activities. However, the detectability of these phenomena is strictly dependent on the

amount of neurons involved and on their level of synchronization: such detectability can

be generically expressed in terms of Signal-to-Noise Ratio (SNR), addressing as Noise

the contributions to the Signal not generated by the source of interest. Standard EEG

recordings (< 100 Hz) primarily reflect mass post-synaptic potentials, rather than spikes,

which are the basic output of neural processing. Since not all synaptic inputs lead to an

initiation of action potentials, measurements of summed post-synaptic potentials alone

cannot show the net computational effect on neuronal output. Standard EEG methods

do not, therefore, provide definitive conclusions about the contribution of neuromodula-

tory, feedforward and feedback connections to neural processing, and may even confound

excitation and inhibition [Speckmann and Elger, 2005].

It has been shown that the brain generates spectral components ranging from 0.05

to 1000 Hz [Klostermann et al., 2002, Hanajima et al., 2004, Buzsaki and Draguhn,

2004] with Signal-to-Noise Ratio dependent on the physical architecture of neuronal net-

works [Gyorgy Buzsaki, 2011]: larger neural populations express oscillations at slower

frequency, with higher magnitude. Thus, since faster oscillations originate from smaller

neural substrates, they are bounded to a lower SNR, as stated by the 1/f nature of EEG

spectral power estimation [Freeman et al., 2003].

EEG scalp measurements cover mostly post-synaptic synchronized neural activity, while

action potentials are reported and analysed on the basis of invasive recordings in animals

and patients. Nevertheless, it was shown that High Frequency Oscillations (HFOs) in

1

1. Introduction

the range of 400-1000 Hz, mimicking spiking activity, can be recorded non-invasively

from the somatosensory cortex in healthy humans by means of median nerve stimula-

tion [Curio, 2005, Ozaki and Hashimoto, 2011]. Moreover, the increasing interest in fast

oscillatory activity is motivated by the correlation between HFOs and epileptic focus lo-

calization, looking at HFOs as biomarker for epilepsy [Engel, 2011].

The investigation of these neurophysiological signatures is severely impaired by the

low SNR. Nevertheless, if high frequency somatosensory evoked potentials (SEP) appear

on the scalp as ripples of few hundreds of nV, at the same time they present an high level of

synchronization, as they can be isolated by averaging across a large number of repetition.

In the case of epileptic ripples, the detection power is constrained to the single event, and

signatures above 250 Hz can be observed only by post-surgical invasive recordings.

Given phase-locked nature of the SEP, burst-like activity typically around 600 Hz

(named sigma-burst, [Curio, 2005], arising at 15-30 ms after the stimulus onset, has

been extensively investigated with commercial EEG recording system. The sigma burst

sources could be characterized according to their spatial localization, refractory proper-

ties, modulation effect of sleep, arousal, attentional states, just to name some experimental

protocols. Clinical recordings pointed out the variability between healthy human subjects

and patients affected by Parkinson, cervical dystonia, myoclonus, multiple sclerosis, mi-

graine and epilepsy. Moreover, simultaneous recording of epidural EEG and single units

in the primary somatosensory cortex of awake behaving monkeys proved that part of

the 600 Hz SEP components detectable in the macro-EEG are generated through highly

synchronized cortical population spikes [Baker et al., 2003]. Notably, in vitro [Steriade,

2001] and in vivo [Bragin et al., 1999, Ikeda et al., 2002] studies have shown the pres-

ence of even faster oscillatory patterns, up to 1 kHz. Thus, the opportunity to record EEG

correlates for repetitive population spikes in a non-invasive framework was demonstrated.

However, in order to reach single trial visibility, systematic breakdown of noise contri-

butions was needed. To this end, it was shown that, above a few hundreds Hz, the noise

is not strictly dependent on biological background activity, but is related to the intrinsic

properties of the recording system [Scheer et al., 2006]. Partial overcoming of these tech-

nical issues led to the achievement of a higher SNR, enlarging the spectral window of

observation to components projected on the scalp as tiny as tens of hundreds of nV, at

around 600 Hz. The implementation of low-noise setup, designed to be operated in an

electro-magnetically shielded environment, allowed to reach single trial visibility of the

2

1.1. Scientific Contribution

sigma burst [Waterstraat et al., 2012].

It remains unexplored whether this highly demanding recording performance can be ef-

fectuated also in a clinical environment, where extrinsic noise contributions could severely

affect the quality of the measurement. Also, the multivariate nature of the EEG has not

been yet fully exploited, and further directions of improvement in terms of augmented

SNR as well as faster component detectability are still uncovered.

1.1. Scientific Contribution

The aim of this work is to provide optimized tools for the investigation of non-invasively

detected high frequency EEG oscillations. The presented contributions have been achieved

traveling on three intersecting routes:

• High frequency scalp EEG detectability up to and above 1 kHz by means of low

noise technology in an electromagnetically shielded environment.

• Characterization of high frequency multivariate EEG recordings by machine learn-

ing approaches targeted to describe the spike-like activity generative mechanisms.

• Design and implementation of a low-noise portable EEG system, and realization of

optimal recording in a clinical environment.

The complete measurement chain, from the recording protocol, to the EEG system and

to the data analysis strategies has been investigated and tuned, in order to empower, at

different levels, the high frequency oscillations detectability. Thus, this work addresses

simultaneously aspects relative to signal detectability, from HFOs neurophysiology to

biophysical factors influencing their visibility, together with technological aspects, char-

acterizing the recording set up sensitivity. On the neurophysiological side, the presence

of SEP components even faster than the sigma burst is described. These ripples, spiking

at about 1 kHz, are named kappa burst. The differentiation among the diverse high fre-

quency contributions is achieved by optimal decomposition approaches, previously tested

in the context of a theoretical framework describing the detectability of short-timed os-

cillatory activity. On the technological side, the project and the realization of a new low-

noise recording system, suitable to be operated outside the electromagnetically shielded

room, has allowed the achievement of an ideal measurement condition also in clinical

environment.

3

1. Introduction

1.2. Outline of the Thesis

Following this introduction, the first chapter of the thesis gives an overview of the EEG

signal generation, enumerating the most significant contributions to the knowledge of

HFOs in human and animal studies, from high frequency SEP studies to the the latest

findings in epileptic HFOs. In the second part general notions of signal processing and

machine learning mathematical approaches are illustrated. In the last part the experimen-

tal protocol is described.

Chapters 3-5 contain the candidate personal contribution in planning and conducting

experiments, perform neurophysiological data analysis, and actively participate to the de-

velopment and realization of an in-house portable recording system, optimized for high

frequency EEG components detection. The data presented in chapters 3 and 4 were

collected during experiments conducted in the facilities offered by PTB (Physikalisch-

Technische Bundesanstalt, Berlin), with an available low-noise multichannel EEG system

designed to be operated in an electromagnetic shielded environment. Active participation

to initial high-frequency EEG recordings, in terms of set-up implementation and data

analysis, constituted a fundamental preliminary introduction for the understanding of the

biophysical scenario characterizing HFOs detection. The opportunity to learn from lead-

ing experts in the field, allowed the candidate to further develop subsequent progresses

in HFOs characterization (chapters 3-4). In chapter 5 the technological achievements

of the research activity are described: this section of the PhD project was conducted in

parallel to the work described in chapters 3-4, but it is placed afterwards because hard-

ware technical certification and ethical committee approval allowed its usage in a clinical

environment only in the later moment. The design, the test, and the realization of the

new low-noise portable EEG system presented here were performed in collaboration with

PTB. In this moment the new hardware is available in the department of neurophysiol-

ogy in Campus Benjamin Franklin, Berlin, in order to be used for SEP recordings in a

non-electromagnetically shielded environment, and in particular for non-invasive HFOs

detection in epilepsy patients.

In chapter 3 experiments conducted to demonstrate and characterize the kappa burst

are described. In the first section optimal low-noise setup criticality (noise figure of 2.7

nV/√

Hz) is shown and its performance is compared with a higher noise system, in the

framework of spontaneous EEG activity and SEP recordings. The extension of the opti-

mal low-noise system to a multichannel device is utilized to characterize and disentangle

4

1.2. Outline of the Thesis

high frequency EEG topographic maps for the sigma and kappa bursts. In the last part,

simultaneous low-noise EEG/low-noise MEG SEP recordings are presented, and elec-

tric and magnetic projections are compared, in order to demonstrate the complementary

contribution of the combined low-noise set up.

In chapter 4, SEP data decomposition in source space is performed through diverse

mathematical techniques, such as Common Spatial Pattern and Canonical Correlation

Analysis based approaches. The algorithms are described and their performance evalu-

ated through a series of simulations, mimicking different biological mechanisms for the

generation of high frequency oscillations. The results of the computational framework

are compared with the performance on experimental data, in terms of SNR and requested

number of trials. The experimental protocol design envisages here two median nerve

stimulation rate, at 1 and 8 Hz, differently affecting the spiking activity refractory prop-

erties. The analysis of the spatial and temporal features expressed by sigma and kappa

bursts in source space is provided, showing the virtue of the analytical approach and

suggesting the presence of different mechanisms for the two high-frequency components

generation.

Chapter 5 describes the design, realization and testing of the low-noise portable EEG

system. Starting with a general overview of the technical issues constraining a bio-

amplifier design, the criticality of power line rejection and input noise are discussed.

The new design is then presented within a computational physical model of the interfer-

ence rejection. Consecutive hardware implementations have been tested to evaluate the

performance inside and outside the electromagnetically shielded room. The components

of the final prototype are described, and the quality of recordings performed in a clinical

environment is illustrated.

The final chapter summarizes the achieved results, giving a neurophysiological and

technological vision. The potential application of the realized portable highly sensitive

device in a clinical environment is discussed, focusing mainly on neocortical epileptic

HFOs.

5

2. Fundamentals

In this chapter, we introduce concepts and tools which are necessary to the understanding

of the thesis. Starting from a neurophysiological introduction on the EEG signal gen-

eration, and a short review in the context of HFOs, a biophysical scenario, theoretical

background for signal processing are provided. Finally, the experimental protocol is de-

scribed. This widely general platform of notions provides to the reader a starting point

to identify aspects belonging to diverse fields of interest. The multidisciplinary character

of this work must not be an obstacle to the complete comprehension of the targets of the

investigation, methodological choices and achieved results.

2.1. Neurophysiology

2.1.1. EEG signal generation

In this section, we briefly discuss the mechanisms underlying the transformation of cere-

bral electrical activity into EEG potentials and describe the two most important neuro-

physiological phenomena observed in EEG signals.

Information processing in the brain takes place in approximately one-hundred bil-

lion interconnected neurons, which are specialized cells that consist of a cell body (the

soma), dendrites, an axon and an enclosing membrane. The electroencephalographic

signal arises as a result of synchronous activity of large populations of neurons with

similar spatial orientation. Following Baillet [Baillet et al., 2001] and Wolters and de

Munck [Wolters and Munck, 2007], this process can be summarized as follows. Neurons

are electrically charged through transport proteins that pump ions across their membranes.

An axonal potential leads to the generation of excitatory postsynaptic potentials (ESPs)

at the apical dendritic tree, which causes the dendrite to release ions through its mem-

brane. The resulting depolarization of the membrane establishes an electrical potential

difference between the apical dendrite and the non-excited cell soma and basal dendrites.

7

2. Fundamentals

This generates current flowing through the intracellular space of the neuronal dendritic

trunk, called primary currents. The electric charges conservation implies that there is also

current flow in the opposite direction. The respective currents close the loop through the

extracellular space, and are called secondary currents. In certain cerebral structures, there

exist large populations of equally-aligned neurons. If these neurons are synchronously ac-

tivated, their primary currents add. The corresponding secondary currents, which spread

over the whole volume conductor, are strong enough to be measurable as scalp potentials.

It is possible to mathematically model the propagation of secondary currents for a given

(primary) current source and volume conductor model using the fact that all currents are

passive in the frequency ranges of interest. In general, the electric potential observed

at the scalp surface is more widespread the deeper the generating source is, while it is

stronger, the more neurons are acting synchronously, the more similar their spatial align-

ment is and the more superficially they are located. Pyramidal cortical neurons are the

likely main contributors to EEG potentials, because they are superficially located and

spatially similarly aligned (perpendicular to the cortical surface). However, ESPs (as

well as IPSP), offer only a partial information of the information processing expressed by

neural population networks. Part of this information is carried by action potentials, rep-

resenting the net computational output of the ensemble of biochemical interactions. The

detectability of these phenomena is strictly dependent on the amount of neurons involved

and on their level of synchronization: such detectability can be generically expressed in

terms of Signal-to-Noise ratio (SNR). Standard EEG recordings (< 100 Hz) primarily

reflect mass post-synaptic potentials, rather than spikes. In this context high frequency

SEP represent a unique opportunity to directly investigate action potentials contribution

to the scalp EEG, offering the opportunity to partially bridge the gap between invasive

and non-invasive neurophysiological recordings.

2.1.2. High frequency oscillations (HFO): definition of a spectral range

In the EEG community there is always some ambiguity when the term high-frequency is

used. According to the literature, it could be related to low (>30 Hz) and high (>80 Hz)

gamma, to ripples identified in animal studies (around 100-200 Hz), or spike-like activity

(> 300 Hz). Before delving into the description of high-frequency oscillations (HFO)

findings, it is important to define the spectral range of interest addressed in this review.

Starting from the sources spectral profile, we retrace the biophysical limitations and the

8

2.1. Neurophysiology

technological aspects characterizing their non-invasive detection. The main interest is to

discuss the opportunity to record highly synchronized spiking activity, in the range of

>250 Hz. In vitro [Steriade, 2001] and in vivo [Bragin et al., 1999, Jones et al., 2000,

Ikeda et al., 2002, Baker et al., 2003] studies have demonstrated the presence of such

fast oscillatory patterns. A traditional scalp EEG spectrum, is typically characterized by

a 1/f trend, reaching a noise floor at around 250 Hz. It was shown that in this spectral

range the noise is not strictly dependent anymore by background activity, but is related to

instrinsic properties of the recording system [Scheer et al., 2006]. Partial overcoming of

these technical issues leads to the achievement of a higher Signal-to-Noise (SNR) Ratio,

enlarging the spectral window of observation to components projected on the scalp as tiny

as tens of nV at 1 kHz [Fedele et al., 2012].

Therefore, the interest in fast neurophysiological activity together with the biophys-

ical constrains for its detection define a spectral range of interest typically unusual for

non-invasive EEG recordings in the humans, as rhythmic events above 400 Hz. In the

following paragraphs aspects of physiological and pathological high frequency electroen-

cephalography (hf-EEG) recordings are addressed, a description of their properties and

the hypothesis of their generative mechanisms are provided in order to shed light on the

realistic opportunity to detect them in a non-invasive fashion.

2.1.3. Physiological HFO

The EEG is the main clinical non-invasive tool for recording human brain activity at

high temporal resolution. Standard EEG frequency range (< 100 Hz) recordings reflect

mass post- synaptic potentials, the input of the neural computation, mostly precluding

access to spike related activity, which is the output. In this sense EEG measurements

are in principle restricted to the extraction of partial information, a low-passed filtered

estimation of the investigated underlying neural processes. Looking at it by a purely

biophysical point of view, even excitation and inhibition cannot be disentangled, as they

could express the same postsynaptic far field patterns [Speckmann and Elger, 2005].

Nevertheless non-invasive EEG is capable of recording human brain high frequency

activity (> 400 Hz), a spectral range belonging to fast action potentials. Such short-

timing activity is characterized by very low amplitude (tens to few hundreds of nanoVolts

peak-to-peak) at the level of the scalp, critically detectable in terms of SNR. In particular

somatosensory evoked potentials, by means of peripheral nerve stimulation, have been

9

2. Fundamentals

recorded and isolated by averaging a large amount of trials.

In the mid-seventies, clinical neurophysiologists reported for the first time about few

small notches overlying the N20 peak of Somatosensory Evoked Potential (SEP) follow-

ing median nerve stimulation, that we now generally name somatosensory HFOs [Cracco

and Cracco, 1976]. Some ten years later, the introduction of digital band pass filtering

allowed quantitative analysis on the distinction between the high frequency ripple band

(>400 Hz) and the slower N20 peak (centered below 100 Hz). We refer to these ripples as

sigma-band HFO, as their central frequency lies around 600 Hz [Curio, 2005]. Magnetic

multichannel recordings allowed the co-localization of HFO and N20m in the primary

somatosensory cortex [Curio et al., 1994], characterized by a typical somatotopic spatial

arrangement [Curio et al., 1997]. Nonlinear recruitment at increasing stimulus inten-

sity [Klostermann et al., 1998] and short term variability verified with deep brain elec-

trodes [Klostermann et al., 2000, Klostermann et al., 2002] led to a marked distinction

between HFO and lower frequency response.

The investigation was focused not only on the distinction of different spectral compo-

nents, but also to disentangle earlier and later contributions inside the ripple band. Source

reconstruction studies enriched the spatiotemporal characterization of early (pre-synaptic)

and late (post-synaptic) components with respect to the N20 peak of cortical pyramidal

synaptic origin [Nakano and Hashimoto, 1999, Haueisen et al., 2000, Haueisen et al.,

2001, Gobbelé et al., 2004] . In agreement with this definition, studies on human healthy

subjects and patients served to describe pre- and post-synaptic features with respect to

brain states and protocols parameters. While the first part of the burst remained mostly

stable in power and timing, composite behavior was observed in the later part: it fades

with increasing stimulation frequency [Emori et al., 1991, Gobbelé et al., 1999, Kloster-

mann et al., 1999, Mackert et al., 2000, Urasaki et al., 2002]; it is sensitive to sleep-

wake cycle [Yamada et al., 1988,Hashimoto et al., 1996], and distinct amplitude modula-

tion by slight natural or benzodiazepine induced vigilance fluctuations arousal [Gobbelé

et al., 2000, Haueisen et al., 2000, Klostermann et al., 2000]; it is enhanced by hyper-

ventilation [Mochizuki et al., 1999] and inhibitory sequences of transcranial magnetic

stimulation [Restuccia et al., 2007, Murakami et al., 2008], and diminished by tactile

or motor interference [Hashimoto et al., 1999, Klostermann et al., 2001, Tanosaki et al.,

2002,Gobbelé et al., 2003b]; it distinctly correlates with development and aging [Nakano

and Hashimoto, 2000].

10

2.1. Neurophysiology

Extracellular or cortical recording of HFO activities were performed in vivo in animal

models. Administration of specific antagonists led to different results: while Ikeda [Ikeda

et al., 2002] reported on an abolished later part of the burst after kynurenic acid admin-

istration in guinea pigs, Jones [Jones et al., 2000] describe an increase in the number

of bursts at subconvulsive concentrations of bicuculline methiodide in epileptic rats. In

the first case hypothesis for the employment of GABA inhibitory neurons was proposed,

while in the latter population spiking in pyramidal cells was postulated. Co-recordings in

macaque of scalp EEG and single unit activity has allowed to identify the relation between

spikes timing and the corresponding hf-EEG epidural components: the peaks of the pop-

ulation peri-stimulus time histogram (PSTH) calculated from those responses align with

peaks of the averaged hf-EEG [Baker et al., 2003]. Quantitative analysis on single neuron

spiking pattern clustering and corresponding scalp EEG SEP subaverages shows partial

covariation between action potentials and epidural hf-EEG [Telenczuk et al., 2011].

Nowadays, it is still not clear which neural populations are responsible for such a non-

invasively detectable high frequency contribution. The main hypothesis, coming from the

results outlined above, is that the early pre-synaptic component is generated by thalomo-

cortical fibers arriving in area 3b while the later relates to GABA feedforward interneu-

rons network [Ozaki and Hashimoto, 2011], whose activity negatively correlates with

lower frequency response of pyramidal post-synaptic origin, as the N20. Thus, it is not

the specific target of this work to favor a specific explanation. The main interest here

is to state that it is possible to record spiking activity non-invasively, shedding light on

underlying neural mechanisms otherwise accessible with demanding set-up, preparation,

and concomitant reference to animal studies.

The potential role of high frequency components as fast as f>400 Hz in humans have

been related also to pathological conditions, where MN stimulation could be performed

on patients. Neural activity generating SEP - HFO can be modulated by various move-

ment disorders: cortical HFOs were increased in patients with Parkinson disease [Mochizuki

et al., 1999, Inoue et al., 2001], and prolonged in patients with myoclonus epilepsy

[Mochizuki et al., 1999] or benign rolandic epilepsy [Kubota et al., 2004]. Heteroge-

neous variability is observed in cortical myoclonus [Liepert et al., 2001, Alegre et al.,

2006]. Migraine Patients showed decreased both sub- and intra-cortical somatosensory

HFOs [Sakuma et al., 2004,Coppola et al., 2005]. In schizophrenia an imbalance between

excitatory and inhibitory regulation in thalamocortical systems was suggested. Consis-

11

2. Fundamentals

tently with this idea a reciprocal relation between the decreased HFOs and enhanced N20

potential gives further support to the interneurons hypothesis [Norra et al., 2004,Waberski

et al., 2004]. In subjects affected by multiple sclerosis the bursts were prolonged [Rossini

et al., 1985], whereas in some cases a decrease in N20 was accompanied by intact bursts

expression [Gobbelé et al., 2003a].

In the recent years great interest has raised in the detection of spontaneous HFO in

epileptic patients (250-1000 Hz) because there is converging evidence that these features

could be used as a biomarker for the presurgical identification of epileptic generative

area [Zijlmans et al., 2012]. Given the impact of epilepsy on human population, the

extensive amount of data, studies and results, we reserved to this pathological condition

a more detailed description.

2.1.4. HFO in epilepsy

The interest in the non-invasive detection of high frequency neurophysiolocal signatures

relates, other than to the investigation of the diverse components of the SEP, to patho-

physiological conditions characterized by fast abnormal neuronal activity. In paragraph

2.1 some examples have been reported. Here we describe another situation where HFO

play an important role: the epileptic fast ripples. In last 10-15 years, substantial progress

has been made in recording pathophysiological oscillations in the range of 250-1000 Hz,

emitted in proximity of epileptogenic brain areas and in recognizing these so-called fast

ripples as potential biomarker for the presurgical identification of the seizure onset zone

(SOZ). For a complete review of the available results from animal and patients studies,

as well as for the speculative hypothesis of the generative mechanisms we suggest more

specific and extensive reviews [Bragin et al., 2010, Jacobs et al., 2012, Zijlmans et al.,

2012]. Here we recall the main achievements of the recent research, looking in particular

at sources intensity, location and spatial extension, as well as to recording techniques, in

order to discuss the opportunity to record such HFOs noninvasively from the scalp.

Invasive Recordings

In the early 90s a few patient studies focused the attention of the scientific community

on the presence of fast rhythmic activity (>100 Hz) [Allen et al., 1992,Fisher et al., 1992]

at seizure onset. More systematic recordings on animal models and epileptic patients

described the presence of HFOs up to 500 Hz. A distinction between ictal events, Ripples

12

2.1. Neurophysiology

(R: 80-250 Hz) and Fast Ripples (FR: 250-500 Hz) was made for the first time [Bragin

et al., 1999]. The group at UCLA, through microelectrode recordings (platinum iridium

flexible microelectrodes, 40 µm diameter), pointed out the distinctive nature of FR with

respect to Ripples, in terms of spatial patterns [Bragin et al., 2002] and modulation due to

sleep stages [Staba et al., 2004]. The bipolar local derivation of this specific set-up made

possible the extraction of FR with amplitudes up to 500 µVpp (microVolt peak-to-peak).

Macro-EEG extensive recordings on patients were performed by McGill Centre in

Montreal by frameless stereotactic setup [Olivier et al., 1994], combining deep and epidu-

ral contacts: deep electrodes provided up to 9 contacts interspaced by 0.5 cm with effec-

tive surface area of 0.8 mm2 (2 kHz sampling frequency, low-pass filter at 500 Hz). They

isolated HFOs in R and FR ranges during seizures [Jirsch et al., 2006, Urrestarazu et al.,

2006]; since macroelectrodes imply a broader source spatial averaging [Chatillon et al.,

2013], the amplitude of the detected HFO drops consistently to tens of µVs; neverthe-

less, the achieved resolution still offers the opportunity to detect and characterize HFO

in relation to spikes and seizure onset zone (SOZ). Ripples occurrence was classified in

relation to interictal slower events [Urrestarazu et al., 2007], showing only partial de-

pendence from the more pronounced spiking activity. Quantitative spatial distribution

analysis of spikes, Ripples and Fast Ripples demonstrate that the localization of interic-

tal HFO highly correlates with the SOZ [Jacobs et al., 2008, Crepon et al., 2010], and

FR in particular, given their extremely localized nature, could serve as a biomarker more

reliably than spike-generating areas, also in absence of evidence for lesions [Andrade-

Valenca et al., 2012], even if they could be indicative for cases of dysplasia [Kerber et al.,

2013]. Notably, 500 Hz seems to not represent and upper limit, at least for hippocampal

networks, capable to produce HFO as fast as 800 Hz [Kobayashi et al., 2010]. FR stand

out not only for their spectral properties, but also because their occurrence increases in

response to anticonvulsant medication reduction [Zijlmans et al., 2009], while spikes re-

mains stable, and show more consistent behaviour during non-REM sleep [Bagshaw et al.,

2009]. Post-surgical outcome evaluation further confirmed HFO clinical value [Akiyama

et al., 2006, Jacobs et al., 2010, Nariai et al., 2011] even if this cannot be generalized to

all epileptic areas [Haegelen et al., 2013]. Unfortunately, even if HFO seem to represent

a key element in terms of SOZ localization, the investigation of temporal predictive pat-

terns has not led to the same outstanding results: an increase of HFO was observed a few

seconds before seizure onset [Khosravani et al., 2009], but more extensive analysis could

13

2. Fundamentals

not point out a significant systematic change in the minutes preceding the seizure [Jacobs

et al., 2009]. This could be due to SNR, electrodes placements, type of epilepsy, and still

need further clarification.

By a technical point of view, studies with micro- and macro-electrodes could demon-

strate the presence of HFO in the 250-500 Hz range. Micro-electrodes offer extremely

detailed information on the spatial location, and a higher SNR, while macro-electrodes

recordings are possibly affected by source spatial averaging, even if evidence for de-

tectability has been provided. The main issue is to understand whether we are looking at

the same neural features.

In order to bridge the gap across different spatial scales, complementary studies were

performed: eight hybrid depth electrodes (surface area 9.4m2 and impedance 200-500 Ω)

were combined together with 27 microwires (40 µm diameter and impedance p to 1 kΩ)

in wideband recording (32 kHz) from human medial temporal lobe in patients [Worrell

et al., 2008]. As could be expected, significant differences in electrodes surface area (10−3

m2 versus 9.3 m2) lead to qualitatively different results. In particular, faster components

spatially confined to smaller areas are consistently recorded at a smaller spatial scale. In

a later study FR with an amplitude of 20 µVpp were recorded with implanted micro-

electrode (µEEG) and compared with signals from macrolectrodes placed as closed as

400 µm [Schevon et al., 2009]. Also in this case HFO recordings at different scales

depict diverse temporal patterns, suggesting only partial overlapping of HFO underlying

neural generators.

The great majority of the results on epileptic HFO refer to deep brain areas, typically

in the medial temporal lobe. Nevertheless, evidence for also on FR range oscillations

is reported also for neocortical epilepsy. Subdural µelectrode (impedance <100 Ω, 4.15

m2 effective area) recordings band-passed at 200-500 Hz could isolate 100 µVpp FR

[Cho et al., 2012], even if their occurrence was limited respect to deeper structures and

possibly only in relation to visible lesions. Whether this is due to the nature of the neural

population, the type of epilepsy and the electrode size has not been addressed yet. A

similar setup [Usui et al., 2010] allowed isolation of FR in the range of 1000 Hz (tens

of µVs) in cortical regions, distinct from the FR range in terms of onset and duration.

Intraoperative ECoG in a large young patients population demonstrated the diagnostic

utility of interictal HFOs above 250 Hz (mean duration of 30 ms, mean frequency of

300 Hz) detected from neocortical sites: the role of interictal HFOs as an independent

14

2.1. Neurophysiology

predictor was confirmed by good post-surgical outcome [Wu et al., 2010], particularly if

the ripples were superimposed to ictal events [Wang et al., 2013].

Evidence for HFO in the frequency range of interest of this work is limited to invasive

recordings, mainly from deep brain structures, but also from the neocortex. Nevertheless,

by a biophysical point of view it is important to keep track of the evidence of slower

spectral components recorded on the scalp, in order to discuss the potential and the limits

to record faster and possibly more localized HFO sources.

Non-invasive recordings

Evidence for HFO in the frequency range of interest of this review is limited to invasive

recordings, mainly from deep brain structures, but also from the neocortex. Nevertheless,

by a biophysical point of view it is important to keep track of the evidence of slower

spectral components recorded on the scalp, in order to discuss the potential and the limits

to record faster and possibly more localized HFO sources.

Ictal

In this context, the group of Child Neurology of Okayama University Hospital, Japan,

provides several studies on the detection of gamma ictal components in relation to spasms

[Kobayashi et al., 2004] in the 50-100 Hz range and with appp amplitude from 50 to

100 µVpp. Gamma peaks spatiotemporal features could be characterized in the time-

frequency domain and related to preceding beta peaks in terms of latency and central

frequency [Inoue et al., 2008]. It was addressed that ictal EEG gamma rhythms during

tonic seizures indicate common generative mechanisms with epileptic spasms, resulting

in desynchronization at seizure onset [Kobayashi et al., 2009, Kobayashi et al., 2013]. A

681 young patients study, with overnight EEG monitoring [Wu et al., 2008], identified

in paroxysmal activity events (max 70 Hz, 100 µVppp) during REM sleep a specific

indicator of ictal sites and seizure severity.

Interictal

Scalp gamma feature recordings ( 120 Hz central frequency, 10-20 µVpp, over a base-

line of around 1-5 µV) was reported in epileptic children during slow-wave-sleep pat-

terns [Kobayashi et al., 2010]: no clear correlation with other pathological parameters

was provided.

15

2. Fundamentals

The detection of slower oscillatory patterns (< 100 Hz, ca 10 µVpp) in idiopathic

partial epilepsy [Kobayashi et al., 2011] and in focal epilepsy [Andrade-Valenca et al.,

2011] was indicative for the SOZ identification, more reliably than interictal spikes, and

independently from the subject age and the sleep stage. In particular, in [Andrade-Valenca

et al., 2011], the issue of the biophysical nature of faster components was addressed:

even if the sampling frequency was set to 600 Hz, it was speculated that scalp EEG it is

not capable to detect higher contribution, addressing the physical limitation not to tissue

attenuation, but to the limited brain area involved in neural source synchronization.

2.2. Biophysical model for feasibility study

Neuronal populations generate locally detectable HFO patterns. Our target is to identify

the biophysical constrains to be satisfied in order to achieve the non-invasive detection of

these features with scalp EEG.

Thus, on one side, it is crucial to minimize every possible contribution affecting the

recording systems sensitivity. This relates to technical design and physical issues, and

their match to physiological spectral properties of the sources of interest.

On the other hand, realistic expectation of generators projections have to be computed,

taking into account the main biophysical parameters influencing electromagnetic field

propagation on the scalp surface as a measurable biopotential.

While there is ample evidence from invasive measurements for oscillations in the HFO

range, their stable detection at the scalp appears to be hindered by noise. Systematic

characterization of noise sources in the HFO frequency range has isolated three main

contributions: biological background noise, impedance thermal noise, electronic noise

of the measurement system [Scheer et al., 2006]. In the traditional EEG spectral bands

(<100 Hz) the biological background noise highly dominates, shaped by its 1/f trend, but

in the HFO range it appears to be comparable to the other two contributions. For this

reason decreasing the impact of the extrinsic (impedance of skin-gel-electrode interface)

and intrinsic (amplifier noise at the input) technical factors enables the noise floor of the

measurement system to be lowered, providing a broader spectral window of observation.

In Figure 2.1 [Scheer et al., 2006] the comparison of EEG spectra in relaxed condition

(free of muscular artefacts) performed with a traditional set up and an optimized one is

reported. In particular, in the optimized setup the impedance is kept as low as 1 kΩ,

16

2.2. Biophysical model for feasibility study

Figure 2.1.: Amplitude Density Spectrum of Electroencephalographic recordings with thesubject in a relaxed and task free condition. Power components for differ-ent spectral ranges are shown. Comparison between commercial amplifierrecording (black), low-noise amplifier recording (green), low-noise amplifiernoise figure (blue).

17

2. Fundamentals

while the electronic noise at the input of the preamplifier is 4.8 nV/√

Hz, 5 to 6 times

smaller than the typical value implemented on commercial EEG and neurophysiological

acquisition systems. This custom-made setup has been used to demonstrate the detection

and characterization of high frequency SEP up to 1 kHz [Scheer et al., 2011, Fedele

et al., 2012]. In order to stress this point further, a simple theoretical computation of the

recording setup noise floor can be calculated by equation 2.1

noise f loor =√

BW · (e2ampl + e2

th) (2.1)

where BW is the signal Bandwidth in [Hz], eampl is the amplifier input noise, and eth is

thermal noise from the impedance defined as

eth =√

4kT R (2.2)

with k being the Boltzmann’s constant in J/K, T the temperature of the impedance in

K, R the real part of the impedance. Considering an amplifier noise of 3 nV/√

Hz (J-FET

input technology), an impedance of 1 kΩ, and a bandpass 100 Hz wide, the resulting

noise floor is 50 nVRMS (nanoVolt Root-Mean-Square). Assuming a Gaussian noise dis-

tribution, the impact of the system is statistically confined in a 300 nVpp range. This

means that a fast ripple event presenting a wavelet above 300 nVpp could in principle

be detected on the scalp. Even if the non-invasive access to spike-like activity has been

demonstrated, the great limitation of these studies is the high number of trials needed to

clearly isolate the spatiotemporal pattern of interest. SEP scalp recordings rely on thou-

sands of repetitions that, given the phase-locked nature of the response, can be averaged

offline. Assuming the noise to be uncorrelated, averaging allows an increase of the SNR

proportionally to the square root of the number of trials. Nevertheless, single trial infor-

mation can be extracted under certain conditions: focussing on the sigma band (400-900

Hz), a theoretical framework relating the required SNR to the recording system noise and

the number of trials has been proposed, defining the ideal recording conditions for the

detection of single events from scalp macro EEG [Waterstraat et al., 2012]. In this sense

the low-noise technology plays a central role in improving the sensitivity to HFO [Scheer

et al., 2011].

By a macroscopic point of view, it has been clarified that the low-pass filtering proper-

18

2.2. Biophysical model for feasibility study

ties of the scalp at high frequencies can be neglected, at least in for a spectral band lower

than 2000 Hz [Pfurtscheller and Cooper, 1975, Oostendorp et al., 2000]. Still, power

dampening for high-frequency must be explained for non-invasive measurements.

Recently, systematic set of simulations, based on neural mass models active in the

beta band, explored the role of several biophysical factors as source-electrode distance,

skull conductivity, neural synchronization, source cortical area and background activity

with respect to scalp detectability [Cosandier-Rimele et al., 2012]. Even if connectivity

patterns and artifacts were not directly taken into account, the prominent role of back-

ground activity has been outlined. In addition, while the skin-skull discontinuity in con-

ductivity reflects a minor effect, synchronization and source spatial extension critically

define detectability ranges. Experimental results obtained with simultaneous EcoG and

scalp EEG [Tao et al., 2007, Andrade-Valenca et al., 2011] pointed out that a synchro-

nized area of 5-10 cm2 is required to provide the needed SNR at the level of the scalp.

In the HFO spectral range, the limited influence of the biological background and aug-

mented sensitivity of the low-noise technology allow to access the activity of the ensem-

ble of synchronized neural units: spatially and temporally coherent activation produces,

in turn, large-amplitude macroscopic oscillations. Notably, realistically shaped three-

dimensional single-neuron models quantitatively describe the impact of action potentials

on non-invasive recordings [Murakami and Okada, 2006]. Action potentials generated at

the level of the neural soma propagate antidromically along the dendritic tree (sAP) and

along the axon (aAP). Considering respectively dipolar and quadripolar current source

configurations [Nunez and Srinivasan, 2006], the number of synchronous events neces-

sary to achieve a far field potential of 300 nV at the scalp could be approximated in terms

of far field as

Φdipole(r,θ)≈Idcosθ

4πσr2 (2.3)

Φquadrupole(r,θ)≈Id2

32πσr3 (3cos2θ −1) (2.4)

With Φ: far field potential, r: distance from the source, I: current amplitude, d: source-

sink distance, θ : angle respect to dipole axis (θ = 0 in this computation), σ : medium

19

2. Fundamentals

conductivity. In this way, at least the order of magnitudes of specific contributions can be

estimated (Table 2.1).

PSCs sAP aAPSurface Potential (nV)- Cortex (2.5 mm) 445 26 15- Scalp (1.5 cm) 3 0.18 0.017

Ratio cortex/scalp 144 144 864

APs for Φscal p = 300nV pp 96 1685 16958

Table 2.1.: : Comparison of contributions from post-synaptic currents (PSCs), somaticaction potentials (sAP) and axonal action potentials (aAP) to electric potentialmeasured at cortical surface and scalp. Last row represents an approximatenumber of action potentials required to generate a scalp potential of 200 nVppby each of the mechanisms. Conductivity dampening factor at the skull is0.25.

In addition, the contribution of each source decays with the inverse of its frequency.

Biophysical models of tissue reactivity and ion diffusion effect account for the effect of

frequency dependent medium properties on the 1/f spectral trend at lower frequencies (f

< 100 Hz), while in the HFO the conductivity can be considered almost constant, and

polarization effects can be neglected (f<1 kHz, [Logothetis et al., 2007]). An alterna-

tive explanation attributes the low-pass filtering to dendritic trees [Nunez and Srinivasan,

2006, Lindén et al., 2010, Leski et al., 2013]: the distance at which the synaptic currents

can penetrate a dendritic tree declines with the frequency of the input [Koch, 2004]. As a

result, the separation between the current sink and source becomes smaller with increas-

ing frequency and the resulting far-field potential at high-frequencies is attenuated.

2.3. Signal Processing

2.3.1. Time-Frequency Transform

In order to describe the power distribution along the temporal and spectral domain, time-

frequency data representation has been used. Signal projection over the timeâfrequency

can be obtained with diverse algorithmic approaches. In this work, we opted for the

20

2.3. Signal Processing

Stockwell transform (S-transform; [Stockwell et al., 1996]). The S transform is a gener-

alization of the short-term Fourier transform (STFT), offering frequency dependent tem-

poral resolution and thus optimizing the time localization in each frequency bin while

maintaining the properties of the Fourier spectrum, such as absolute reference to phase.

Given a time signal h(t), its continuous S-transform is

S(τ, f ) =∫

∞

−∞

h(t)| f |√2π

e−(t−τ)2 f 2

2 e−i2π f dt (2.5)

A voice S(τ , f 0) is defined as a one dimensional function of time for a constant fre-

quency f 0, which shows how the amplitude and phase for this exact frequency changes

over time. A local spectrum S(τ 0, f ) is a one dimensional function of frequency for a

constant time t0.

The Gaussian window is chosen for several reasons: 1) it uniquely minimizes the

quadratic time-frequency moment about a time-frequency point, 2) it is symmetric in time

and frequency - the Fourier transform of a Gaussian is a Gaussian, 3) a Gaussian function

does not present side lobes (a local maxima in the absolute value of the S-transform is not

an artifact). However, as is the case with Power Spectral Estimation, any desired window

may be employed. The derivation of the S-Transform is here shown from two starting

point: the Short Time Fourier Transform (STFT), and the Wavelet Transform (WT). The

Fourier spectrum is defined, for some window function g(t) as

H( f ) =∫

∞

−∞

h(t)g(t)e−i2π f dt (2.6)

Then we can identify the connection to the S-transform by choosing the following

Gaussian window function, with normalization factor inversely proportional to the fre-

quency.

g(t) =| f |√2π

e−t2 f 2

2 (2.7)

Allowing the Gaussian window function to translate in time by the quantity τ , substi-

tuting 2.7 in 2.6, we obtain 2.5, the S-Transform revised as a STFT with a frequency

dependent window. The same holds for the construction of a general WT, as

21

2. Fundamentals

W (τ,d) =∫

∞

−∞

h(t)1√d

ψ(t− τ)

ddt (2.8)

Choosing a mother wavelet ψ with a dilation function d inversely proportional to the

frequency, and characterized by a phase correction factor such that

Ψ((t− τ) f ) = e−(t−τ)2 f 2

2 e−i2π f (τ−t) (2.9)

we obtain the S-Transform as a special case of the WT. Here, two important obser-

vation have to be made: the normalization is chosen frequency dependent, and the phase

correction allows to separate amplitude and phase computation. During the iterative com-

putation, this structure does not imply any shift in the oscillatory exponential kernel char-

acterizing the phase, so that remains absolutely referenced to the initial point of the time

series.

The practical implementation consists of the following steps: the Fourier transform

of the time series h(t) is computed. Then, the spectrum is the shifted, passing to the

next voice frequency. The Gaussian, shaped in base to the frequency, is multiplied to the

spectrum, and the result is translated in time by IFFT (Inverse Fast Fourier Transform)

[Stockwell et al., 1996].

2.3.2. Spectral filters

It is of outmost interest to visualize time-trends in specific spectral ranges. A time-domain

representation of the signal in the frequency range of interest can be obtained by spectral

filtering, or band-pass filtering if the frequency range is contiguous. Filtering consists in

multiplying the input time signal with a series of coefficients, optimized to extract a spe-

cific spectral content. Among the diverse possible implementation we chose causal filters

of Butterworth type, because they provide maximal flat response in the band of interest

(passband), while suppressing the remaining (stopband) frequencies [Butterworth, 1930].

The equation is of the form

(t) =1a0

(N

∑n=0

bnx(t−q)−N

∑n=1

any(t−q)

)(2.10)

22

2.3. Signal Processing

Where the signal x(t), the filtered signal y(t), N the filter order and bn and an filter

coefficients (computed with the Matlab function butter.m). Zero-phase digital filtering is

achieved by processing the input data in both the forward and reverse directions [Oppen-

heim et al., 1999].

2.3.3. The forward model

The basic macroscopic model of EEG generation [Nunez and Srinivasan, 2006] considers

the tissue as a resistive medium considering only effects of volume conduction, neglecting

the marginal capacitive effects [Pfurtscheller and Cooper, 1975,Oostendorp et al., 2000].

In this sense, the source propagation to the scalp is instantaneous, and its biophysical

contribution can be estimated in agreement with the quasi-static approximation for the

electric field propagation. A single current source s(t) contributes linearly to the scalp

potential

x(t) = as(t) (2.11)

where the propagation vector a represents the individual coupling strengths of the

source s(t) to the N surface electrodes. In general, the propagation vector a depends

on three factors; the conductivity of the intermediary layers (brain tissue, skull, skin); the

spatial location and orientation of the current source within the brain; and the impedances

and locations of the scalp electrodes. In order to model the contribution of multiple source

signals to the surface potential, the propagation vectors of the individual sources are ag-

gregated into a matrix A and the overall surface potential results in

x(t) = As(t)+n(t) (2.12)

This model also incorporates an additive term n(t): [1 x nel], which comprises any

contribution not described by the matrix A. Although some of the originating sources

might be of neocortical origin, n(t) is conventionally conceived as noise, emphasizing

that those activities are not subject of the investigation. The propagation matrix A is often

called the forward model, as it relates the source activities to the signals acquired at the

23

2. Fundamentals

different sensors. In this regard, the propagation vector a of a source s(t) is often referred

as the spatial pattern of s(t), and can be visualized by means of a scalp map. In the chapter

IV an implementation of the forward model as proposed by Nolte and Dassios [Nolte and

Dassios, 2005] will be used to simulate high frequency EEG source propagation over the

scalp.

2.3.4. Source space analysis

The EEG can be considered as a multivariate time signal shaped by the projection of un-

derlying neural currents to the scalp electrodes. In this sense here we refer to the EEG

recording as the sensor space, X , and to the generators as the source space, S. Abstracting

from the sources specific spatial location, we can revise the sensor space as the linear

combination A of the source space. Since the inverse model solution is a mathematically

ill-posed problem, and a correct spatial source estimation is highly demanding computa-

tionally and experimentally, we restrict our analysis to a subspace of solutions, as many

as the number of electrodes used, such that

X = AT S (2.13)

where W = A−1 : [nel x nel]. Each source, given its power, location and orientation, is

related to the scalp EEG signal by a specific spatial signature, expressed by the columns

of A. At the same time, the contribution of each channel to the estimation of the single

source is parameterized by a set of coefficients, expressed by each column of W . We call

then A the ensemble of spatial patterns, and W the ensemble of spatial filters. Filters and

corresponding patterns can be estimated on a time interval by assuming specific statistical

properties of the signal covariance matrix. Here we recall two algorithmic approaches,

which will be used in the Chapter IV for the source space analysis of the high frequency

EEG signal.

Spatial filters can be extracted by a different mathematical approach. Here we will not

focus on the mathematical details of the matrices decomposition. I will formulate the

optimization criteria for two approaches that will be used in this work, in order to discuss

their limitation and their advantages.

24

2.3. Signal Processing

Common spatial pattern (CSP)

The CSP algorithm extracts components maximally differing in power contained in

the same data set at two different time intervals, defining different conditions. More

specifically, CSP filters maximize the variance of the spatially filtered signal under one

condition while minimizing it for the other condition. Being Σ(+) and Σ(−) the covariance

matrices of the data set (+) and data set (-) belonging to two different conditions, then the

CSP analysis is given by the simultaneous diagonalization of the two covariance matrices

W TΣ(+)W = Λ

+

W TΣ(−)W = Λ

−(2.14)

where Λ is diagonal and the scaling of W is commonly determined such that Λ+ +

Λ− = I (Fukunanga, 1990). Technically this can be achieved by solving the generalized

eigenvalue problem

Σ(+)w = λΣ

(−)w (2.15)

Then Equation 2.14 is satisfied for W having as columns the generalized eigenvectors

w j ( j = 1, . . . ,C) of Equation 2.15 (as column vectors) and for λ the diagonal matrix of

generalized eigenvalues. The eigenvectors give the direction of rotation of the covariance

matrices, while the eigenvalue express the power associated to each component. The

discrimination between two conditions depends on the value of the elements of λ . Each

λ j = λ(+)

j / λ(−)j and corresponds to the components power ratio.

Another way to look at CSP as a discriminative technique between two classes, is to

reformulate the problem as follows:

Sd = Σ(+)−Σ

(−)

Sc = Σ(+)+Σ

(−)(2.16)

where Sd contains the power difference between the two classes, while Sc relates to

their common features. Since the covariance matrices are a descriptive statistics of the

power structure of a data set, we can think of our optimization criteria as a power ratio

maximization expressed in terms of GEVD (Generalized EigenValue Decomposition):

25

2. Fundamentals

maxwwT SdwwT Scw

(2.17)

In Chapter IV CSP will be formulated in this way, to extract the high-frequency com-

ponents of interest in contrast to the corresponding band pass filtered baseline.

Canonical Correlation Analysis (CCA)

CCA algorithm extracts components maximally correlated in time between two multi-

dimensional variables. A set of specific spatial filter is identified for each variable, and a

correlation coefficient is assigned to each of the extracted components. In other words, it

finds the basis in which the correlation matrix between the variables is diagonal and the

correlations on the diagonal are maximized.

For variables X and Y , the following optimization problem is formulated:

maxwx,wyρ =wT

x XY T wy√wT

x XXT wxwTy YY T wy

=wT

x Cxywy√wT

x CxxwxwTy Cyywy

(2.18)

Where wx, wy, are two basis vector mutually maximizing the correlation ρ , and Cxy, Cxx,

Cyy are the cross-correlation and the two covariance matrices, respectively. It is possible

to demonstrate that to solve equation 4.7 is equivalent to solve one of the two following

EDV problems:

C−1xx CxyC−1

yy Cyxwx = ρ2wx

C−1yy CyxC−1

xx Cxywy = ρ2wy

(2.19)

which lead to complementary solutions, given the relation:

Cxywy = ρλxCxxwx

Cyxwx = ρλyCyywy

(2.20)

In chapter IV, CCA will be used to exploit the correlation across trials.

Similarity Index

Given two scalar vectors, identifying, for example, two spatial patterns P1 and P2 for

26

2.4. Experimental Protocol

different data subsets belonging the same dataset, we define similarity index σ as their

scalar product:

σ =< P1,P2 > (2.21)

If σ is greater than 0.85, the extracted set of coefficient computed on the complete

dataset is considered reliable to conduct further analysis.

2.4. Experimental Protocol

Most of the analyses and results presented in the thesis will concern the somatosensory

system of humans. Here, we will briefly introduce the essential notions about the system

anatomy and experimental methods used to record neuronal activity.

2.4.1. Somatosensory system

The somatosensory system is a part of the nervous system responsible for sensation of

touch, temperature, pain and body position (proprioception). Sensory information is

transduced into the electrical activity of neurons by specialized receptors in skin and

muscles. For example, skin mechanoreceptors transduce information about fine touch

and transmit it to neurons in the dorsal root ganglion of the spinal cord. The neurons carry

the information up the spinal cord and form the first synapse either in the cuneate nucleus

(fibers from upper body) or gracile nucleus (fibers from lower body) of the medulla ob-

longata. Axons of the postsynaptic neuron cross the midline of the brain and proceed at

the contralateral side in a fiber bundle called the medial lemniscus towards the thalamus.

In the ventral posterior lateral nucleus of the thalamus the axons contact thalamocortical

neurons that send their afferents to Brodmann area 3b of the cortex, typically located in

the posterior bank of the central sulcus. Throughout the somatosensory pathway the sen-

sory information is represented in somatotopic fashion, which means that anatomically

close neurons carry information about neighboring areas of the body. The radial part of

the palm and the palmar surface of the thumb, index and ring fingers are innervated by the

median nerve which sends information to the spinal cord. The representation is preserved

in the medulla and in the thalamus and forms a part of the ordered representation of the

body in the cortex (the so-called homunculus).

27

2. Fundamentals

2.4.2. SEP recording protocol

The median nerve represents one of the main inputs of the somatosensory system. Its

peculiarity lies in its bidirectional nature, as it includes afferent and efferent fibers, the

first giving rise to the palmar and digital cutaneous branches, the second innervating flexor

muscles in the anterior compartment of the forearm muscles. Its receptive field can be

reached by placing anodic and cathodic terminals of a current generator at the level of

the wrist, and its receptive field can be evaluated by the rhythmic contraction of M. flexor

pollicis longus, in response to the stimulus train pulses. Typical motor thresholds are

found around 3-5 mA, for pulses duration of 200 µs. The presented data were collected

by optimizing the stimulation contacts position, and tuning the current level at 1.5 x Motor

Threshold. The recordings described in the thesis are listed in A.1. They are labeled with

a Latin number, which is used in the dissertation, in order to help the reader to get oriented

among different experimental sessions.

28

2.4. Experimental Protocol

Figure 2.2.: Experimental protocol description for SEP recordings by means of mediannerve stimulation.

29

3. High frequency SEP (hf-SEP) -kappa band

In this chapter the detection and characterization of the high frequency SEP components

above 1 kHz is presented. The technological achievements necessary for a stable detec-

tion are discussed, and neurophysiological findings of the hf-frequency scalp signatures

are described. The presented results are adapted from [Scheer et al., 2011, Fedele et al.,

2012].

3.1. Introduction

Historically, the first electroencephalogram (EEG) was designed to write traces on paper,

and this made it limited in the frequency response up to about 80 Hz. The power of the

spontaneous EEG is mainly concentrated within this frequency band, and it had been as-

sumed that the scalp EEG beyond this limit was not of scientific or clinical relevance.

Exploration of signals during the past 25 years, driven by digital EEG and powerful ac-

quisition and signal processing techniques, revealed that the true frequency range of the

EEG is much broader and comprises signal frequencies from near DC up to more than 1

kHz [Niedermeyer and Sherman, 2001, Vanhatalo et al., 2005].

High-frequency EEG signals are mostly part of evoked potentials, which are generated

by the brain upon auditory, somatosensory or visual stimulation. These signals have

amplitudes not more than a few micro volt or even less, which is far below the level of the

spontaneous background EEG observed inside the traditional bandwidth of 0.1-80 Hz.

In this frequency band, digital band-pass filtering and averaging of epochs are the most

popular tools to extract the evoked responses from the recorded data, while the role of

amplifier noise is more or less disregarded. We explore here the content of less traditional

frequency range, looking at responses as fast as 600 Hz to 1 kHz, describing how the latest

technological development could potentially open the way to characterize non-invasively

31

3. High frequency SEP (hf-SEP) - kappa band

physiological generators, that would remain otherwise undetected.

In particular, somatosensory evoked human EEG responses contain contributions in

different temporal and spectral domains. The early low-frequency thalamic (P16) and

cortical (N20) peaks are overlapped by smaller high-frequency contributions (around

600 Hz, with amplitude of few hundreds of nanoVolt peak-to-peak: [Cracco and Cracco,

1976]; MEG correlates: [Curio et al., 1994, Hashimoto et al., 1996]). These repetitive

spike-like wavelets, designated here as sigma-burst, represent the opportunity to record

and monitor non-invasively highly synchronized and rapidly repeating population spikes

generated possibly by cuneothalamic and thalamocortical relay cells, cortical bursting

pyramidal cells, and fast-spiking inhibitory interneurons (overview: [Curio, 2005]): Early

observations of these wavelets were interpreted as volume-conducted far-field potentials

generated in diencephalic structures [Cracco and Cracco, 1976]; later studies identified

the location of sigma-burst components in area 3b of human primary somatosensory cor-

tex [Curio et al., 1994, Hashimoto et al., 1996]. While low-frequency EEG and MEG

(<100 Hz) are mainly generated by postsynaptic mass excitation [Okada et al., 1997], it

was proposed that high-frequency EEG/MEG reflects spiking activity [Curio et al., 1994].

This hypothesis is supported by simultaneous macroscopic high-frequency EEG and inva-

sive single-cell micro-recordings in awake macaque monkeys [Baker et al., 2003] which

have shown that the macroscopic sigma-burst is coincident with synchronous spike bursts

of neurons in the primary somatosensory cortex. Further non-invasive characterization of

sigma-burst generators revealed a typical somatotopic arrangement [Curio et al., 1997],

nonlinear recruitment with increasing stimulus intensity [Klostermann et al., 1998], non-

linear refractory behaviour with increasing stimulus frequency [Klostermann et al., 1999],

and a distinct amplitude modulation by slight natural or benzodiazepine induced vigi-

lance fluctuations [Gobbelé et al., 2000, Haueisen et al., 2000]. Furthermore, subcortical

sigma-burst components were identified in EEG [Gobbelé et al., 2000] and in intratha-

lamic recordings from tremor patients with deep brain electrodes [Klostermann et al.,

2002]. Also burst components linking lower (200 Hz) and upper (600 Hz) frequency

domains were characterized [Haueisen et al., 2001], as well as differential modifications

of bursts and low-frequency SEP components during sensorimotor interactions [Gobbelé

et al., 2003b]. Moreover, clinical studies showed burst alterations in basal ganglia disease,

multiple sclerosis, and schizophrenia [Waberski et al., 2004].

Notably, fast oscillations are not limited to the sigma-burst range. In particular, os-

32

3.2. Detection of kappa band components by means of low-noise EEG amplifier technology

cillations at and above 1 kHz were found intrathalamically in patients with Parkinson’s

disease [Klostermann et al., 2002,Hanajima et al., 2004], reflecting locally restricted near-

field activity. Oscillations faster than sigma-bursts have been observed also epidurally

in spinal cord and dorsal column nuclei measurements [Insola et al., 2008, Insola et al.,

2010], and at the cortical level in epileptic patients fitted with subdural electrodes [Sakura

et al., 2009]. Thus, while there is ample evidence from invasive measurements for oscil-

lations in the kHz range, their stable detection at the scalp appears to be hindered by

noise. Noise in EEG measurements can have various origins, primarily as background

neural activity, but also as thermal fluctuations in the amplifier or at the electrodeâskin

interface. Also spontaneous brain and myoelectric activity are noise sources of their own

that interfere with the signal of interest. At frequencies below 100 Hz, the signal-to-noise

ratio (SNR) is dominated by spontaneous brain activity and practically independent of

the EEG amplifier, while in the frequency range above 400 Hz, amplifier noise becomes

increasingly more important [Scheer et al., 2006].

This chapter deals with the possible improvement of the signal quality of high-frequency

SEPs by the use of amplifiers having a significantly lower intrinsic noise. In addition,

we show both, time-domain and spectral SEP features, and demonstrate that sigma- and

kappa-bursts express specific scalp voltage patterns, pointing to partially different cortical

and subcortical generators.

3.2. Detection of kappa band components by means oflow-noise EEG amplifier technology

3.2.1. Settings

The quality of the recordings for hf-EEG components has been tested comparing the per-

formance of two bioamplifiers with different input noise level, acquiring somatosensory

evoked response following median nerve stimulation (Recording session I, A.1).

SEP measurements were performed inside an electrically and magnetically shielded

room made by Vakuumschmelze (AK3b), originally designed for MEG. Inside this room

magnetic and electric fields generated by electric power lines or other noise sources are

reduced by several orders of magnitude. The shielded room provides an ideal environ-

ment for the exploration of extremely weak signals. Sintered Ag/AgCl ring electrodes

with abrasive paste were used and placed on the scalp of a healthy and awake volunteer

33

3. High frequency SEP (hf-SEP) - kappa band

using a cap, which has holders at positions according to the 10/20 system. The mon-

tage examined here is C3-F3, with the ground electrode placed on the forehead. The

impedance at the electrodeâskin interface C3-F3 was set to be below 1 kΩ (real part ≈700 Ω at 1 kHz) and was checked repeatedly to be nearly stable during sequential SEP

measurements. It proved essential that the volunteer was comfortable resting on a bed

and requested to keep a relaxed posture so as to minimize interference from electromyog-

raphy (EMG). An electrical monophasic square wave constant-current stimulus of 200 µs

width and 5.2 Hz repetition rate was applied transcutaneously to the median nerve at the

right wrist at about 1.5 times motor threshold. Evoked responses were acquired within a

bandwidth of 0.16-2000 Hz (total gain 10 000, ADC rate 5 kHz, resolution 20 bits).

Usually conventional commercial multichannel EEG systems are used to study high-

frequency EEG. In spite of the offered option to record signals in the kHz range, the

amplifier noise spectral density en is rarely published by the manufacturers of EEG de-

vices. Sometimes it can be estimated from a specified root-mean-square (RMS) voltage

E given in a bandwidth BW according to

en =E√BW

(3.1)

Actual guidelines of the American Clinical Neurophysiology Society (2006) on evoked

potential measurements state that amplifier noise must not exceed 28 nV /√

Hz; typical

commercial multichannel EEG apparatus present noise in this order. To investigate the

influence of amplifier noise on the total SNR, we used two different preamplifier ver-

sions: amplifier I with a voltage noise density of 4.8 nV /√

Hz and amplifier II with 12

nV /√

Hz. The EEG preamplifiers used in this study were developed in-house by PTB

and are an integrated part of the MEG/EEG measurement system with 32 EEG channels

in total. Amplifier II with a voltage noise of 12 nV /√

Hz is a standard design using in-

strumentation amplifiers and one common low-noise buffer for the reference electrode. A

simplified circuit diagram of the lower noise design amplifier I is depicted in Figure 3.1.

The circuit principle originally was proposed and discussed by [Virtanen et al., 1997].

For our application in high-frequency EEG, it was realized by use of lower noise compo-

nents. Virtual absence of any power line interference was achieved by use of rechargeable

batteries for power supply. The total noise of the whole amplifier is usually determined

by the components of the input stage if its gain is high enough [Netzer, 1981]. Our input

34

3.2. Detection of kappa band components by means of low-noise EEG amplifier technology

stage has a gain of 25 and noise of one channel is mainly generated by input amplifiers

µA and Reference amplifier uREF and the thermal noise of the resistors involved. Current

noise terms can be neglected because of field effect transistor inputs of the used amplifiers

(AD743, Analog Devices). Therefore, the input-related voltage noise density unoise can

be estimated as

enoise =

√u2

A +u2REF +4kT R2(1+

R2

R1) (3.2)

From data sheet of AD743, the voltage noise density is uA = uREF = 3.2 nV /√

Hz at 1

kHz. With the resistor values shown being at absolute temperature T ≈300 K and Boltz-

mann constant k, one gets 4.8 nV√

Hz. The whole amplifier comprises four identical

eight-channel modules as shown in Figure 3.1 (reference inputs are in parallel). With

this arrangement, the amplifier noise of the 32-channel system can be reduced on demand

by channel averaging at the expense of available channels [Scheer et al., 2006]. To this

end, N = 4 individual inputs, i.e. one of each module, can be wired in parallel by the

user. An average over the corresponding outputs then will leave the signal unchanged,

while the amplifier noise as an average over uncorrelated noise sources is halved (noise

≈ 1/√

N). Thereby we can use 32 channels with an amplifier noise of 4.8 nV /√

Hz or 8

channels with theoretically 2.4 nV /√

Hz. In the present version, we found 2.7 nV /√

Hz

because of the thermal noise generated by an electronic analog switch being in series

connection for all reference amplifiers. The switch is for calibration purposes and could

be replaced by a lower noise type in future designs.

We sequentially measured the SEP response by use of one of the two amplifiers but

without changing montage and preparation of the electrodes. In the experiment with

amplifier I, four of its channels were connected to electrode C3 in parallel. In this way, we

simultaneously realized an effective amplifier noise of 2.7 nV /√

Hz after averaging of the

corresponding four outputs. An ensemble of 6000 trials was recorded. Signal processing

including averaging and zero phase shift digital band-pass filtering was performed offline.

The input voltage noise densities of the amplifiers given above were measured with

their inputs shortcut to internal ground by use of a FFT analyzer (Agilent model 35670A)

connected to the analog amplifier output stage. It was also proved that amplifier current

noise is negligible with 10 kΩ resistors from inputs to ground. Electrode impedances in

the experiments were well below this value.

35

3. High frequency SEP (hf-SEP) - kappa band

Figure 3.1.: Simplified circuit diagram of the eight-channel low-noise amplifier unit.

36

3.2. Detection of kappa band components by means of low-noise EEG amplifier technology

3.2.2. Results

An FFT analysis of total noise without stimulation is depicted in Figure 3.2 where the

amplifier inputs are connected to the electrodes C3-F3. To illustrate the interfering effect

of the muscle activity on in spectral terms the subject was asked for clenching of jaw

muscles, when connected to Amplifier II. As a result, this spectrum is dominated by the

muscle activity and strongly overlaps the interesting high-frequency range, as shown by

the blue trace. Therefore, to minimize EMG contamination, it proved essential that the

subject must be relaxed as good as possible to take maximum advantage by the use of a

very low noise amplifier. In an ideal artifact-free, relaxed condition, the corresponding

spectra show their typical 1/f trend. Both traces are flat in the high-frequency part, which

implies that they are dominated by constant distributions of the amplifier and thermal

(white) noise of the inter-electrode impedance (≈3.5 nV /√

Hz).

Amplifier voltage noise in nV/√

Hz12 4.8 2.7

Total noise in nV, BW = 423-800 Hz 5.7 (3.9) 3.5 (3.1) 3.2 (3.0)Total noise in nV, BW = 800-1400 Hz 4.8 (3.7) 1.7 (2.1) 1.5 (1.8)

Table 3.1.: RMS voltage of background noise at site C3-F3 calculated from averaged SEPtime series. Values in brackets are found by numerical integration of FFT data.Parameters: bandwidth (BW) and noise of the used amplifier.

Theoretically calculation of both these contributions matches the white noise level of

the measured spectra. These values define the sensitivity of the measurement system. In

the SEP measurement we are interested in picking up tiny contributions at the level of the

scalp, and in comparing the performances of the two system with respect to the detected

signal, on the base of their respective SNR. Since we expect the signal of interest to

appear between 10 and 30 ms after the stimulus onset (signal temporal window), and

every trial lasts almost 200 ms, it reasonable to consider, in absence of artifacts, a time

window ranging from 40 to 180 ms post-stimulus as a good representation of the overall

background noise (noise temporal window). We can reasonably assume this noise to have

a Gaussian distribution and be uncorrelated in time, allowing to perform averaging and

isolate the phase locked components. It is equivalent in this sense to estimate the power

density from several noise windows from ongoing recording, or from the averaged signal

noise window, back-multiplying it by the âN, with N being the number of averages.

37

3. High frequency SEP (hf-SEP) - kappa band

Figure 3.2.: FFT of background noise acquired at site C3-F3 without stimulation. 1: con-dition with clenching of jaw muscles. 2: relaxed using the 12 nV /

√Hz

amplifier. 3: relaxed using the 4.8 nV /√

Hz amplifier.

38

3.2. Detection of kappa band components by means of low-noise EEG amplifier technology

Table 3.1 shows the root-mean-square RMS voltage of background noise calculated

from the averaged time series of the SEP experiment in the noise temporal window, which

was free of evoked hf-SEP. For comparison, the values in brackets are found by numerical

integration of FFT data shown in Figure 3.2 (red and black traces) normalized by the

square root of number of averages (N = 6000). While the data used for the FFT were

acquired sequentially and without stimulation, these values were found comparable with

the baseline from SEP data.

An improvement of a factor of two (in power) can be observed passing by the 12 nV/√

Hz to the 4.8 nV/√

Hz amplifier. Apparently not significant improvement is given by

the 2.7 nV /√

Hz configuration: in this case we had an impedance value of approximately

1 kΩ, which dominates the noise figure, nullifying the benefit of having a lower amplifier.

With both amplifiers, wide-band SEP data (N = 6000 averages) were found to contain os-

cillatory components directly visible at the rising edge of the N20 response, starting with

a latency of 15 ms after the stimulus (Figure 3.3, top). Practically no difference in the

wide-band SNR can be recognized between the two amplifiers used. For separation of the

well-known 600 Hz wavelet, we used a band-pass filter from 423 to 800 Hz (Figure 3.3,

middle). Already, here an improvement in the SNR by use of the lower noise amplifier

can be seen. Taken into account the results from the measurements with implanted elec-

trodes [Klostermann et al., 2002], we looked for the frequency components in the kHz

range.

Actually, after filtering between 800 and 1400 Hz, the present scalp data acquired

with the lower noise amplifier (I) clearly show a 950 Hz centered wavelet (Figure 3.3,

bottom left), with its latency and frequency being comparable to the results measured by

implanted deep brain electrodes. Also, when using the higher noise amplifier (II), a very

short wavelet at ≈950 Hz seems to arise. It is important to notice here, beside the quality

of the oscillatory component in the signal window, how the distinctive performance of a

low-noise approach provides the opportunity to lower the noise contribution outside of

the signal window to an almost flat line.

39

3. High frequency SEP (hf-SEP) - kappa band

Figure 3.3.: Comparison of measurement data. Left column: 2.7 nV /√

Hz amplifier (fourchannels with 4.8 nV /

√Hz in parallel). Right column: 12 nV /

√Hz ampli-

fier. Top: wide-band (0.1-2000 Hz) SEP signal (6000 averages), recorded atC3-F3. Middle: band-pass-filtered 423-800 Hz. Bottom: band-pass-filtered800-1400 Hz.

40

3.3. Multichannel EEG kappa components characterization

3.3. Multichannel EEG kappa components characterization

3.3.1. Settings

Extremely low amplitude oscillatory somatosensory evoked EEG signals at approximately

950 Hz, previously detected only using implanted electrodes, are measured with scalp

electrodes by means of a very low noise amplifier. Accepting 4.8 nV /√

Hz as a gold-

standard in terms of system technology, a multichannel system has been realized, in order

to further map and characterize kHz range components, and distinguish them from the

well-known 600 Hz SEP ripple (Recording session II, A.1).

SEP measurements were performed in six healthy subjects, inside an electrically and

magnetically shielded room (Vakuumschmelze AK3b). Sintered Ag/AgCl ring electrodes

were placed using a fabric cap with holders at positions according to the 10/20 system.

For three subjects 29 EEG electrodes were distributed over the scalp; for the other three

subjects 10 electrodes were used, focused over the pericentral primary sensorimotor areas

bilaterally. The reference was placed on the nasion, and ground on Fz. The impedances

at all electrode-skin interfaces were carefully prepared using standard abrasive paste until

for all electrodes values at or below 1 kΩ were achieved; impedance stability below 1 kΩ

was checked and confirmed repeatedly during all measurements. During the recordings

the subject was comfortably resting on a bed in a relaxed position. A constant-current

square wave electrical stimulus of 200 µs width, 5-8 mA (1.5x motor threshold) and 4.3

Hz repetition rate was applied transcutaneously to the median nerve at the left wrist. Data

were acquired using a bandwidth of 0.16 to 2000 Hz (total gain 10000, ADC rate 5 kHz,

resolution 20 bits). We utilised a low-noise custom-made amplifier, with a white noise

level of 4.8 nV/√

Hz [Scheer et al., 2006]. Noise spectra were obtained with the amplifier

input connected either to ground or to the relaxed subject without stimulation, using an

FFT analyzer (Agilent model 35670A) connected to the amplifier analog output.

The data analysis was performed on averages of about 5000 trials in order to isolate

stimulus-locked low-amplitude response components. Stimulus artefacts were removed

in single-trials in each channel by cubic interpolation from -5 to + 5 ms around the stim-

ulus onset. Trials exceeding three times the average standard deviation in the frequency

range 200-2000 Hz were rejected. Frequency analysis by means of S-Transform [Stock-

well et al., 1996] was performed to identify distinct spectral components, and the obtained

time-frequency representations were normalized with respect to the prestimulus (-40 to -

41

3. High frequency SEP (hf-SEP) - kappa band

10 ms) mean power at each frequency bin thereby rendering visible wide power variations

in the spectral range spanning from near-DC to 2 kHz (hardware anti-aliasing frequency

low-pass filter). The averaged signal was filtered in three distinct frequency bands (sigma,

kappa1, kappa2) as determined from the S-transform power distribution. In these indi-

vidually determined frequency bands, scalp map voltage distributions allowed to charac-

terize the spatiotemporal evolution of different evoked patterns in the subjects fitted with

29 electrodes. The filter and S-Transform impulse response were tested using artificial

signals. Thereby, possible artefact contributions could be excluded; in particular, filter

ringing around the stimulus event was avoided by the cubic interpolation.

3.3.2. Results

For the analysis a signal window was defined (10-50 ms post-stimulus) as well as a base-

line window (-40 to -10 ms). For each subject the power spectral density of the averaged

signal of all channels was determined for the two windows in order to individually delin-

eate the corner frequencies for the bandpass filtering. In all subjects increases in power

spectral density were observed in the N20 range (up to 200 Hz), in the sigma range (400-

900 Hz) and, additionally, further distinct components at and also above 1 kHz. Subjects

1 and 4 expressed a prominent component at 1400 Hz, subjects 2, 3, 5 and 6 at 1200 Hz

(PSD, not shown).

In agreement with previous studies, a sigma-burst band was defined between 400 and

800 Hz, and two bands were added for the components at and above 1 KHz (bandpass

corner frequencies for each subject are reported in Tab.2). In order to prevent overlap

between different components these separation bands were optimized for each subject.

We present raw data referenced to the nose electrode.

Subject Sigma Band (Hz) Kappa1 Band (Hz) Kappa2 Band (Hz)Subject 1 450-750 850-1150 1250-2000Subject 2 400-700 800-1000 1150-2000Subject 3 400-700 800-1150 1250-2000Subject 4 500-850 900-1050 1120-2000Subject 5 450-700 800-1100 1200-2000Subject 6 450-750 810-1050 1150-2000

Table 3.2.: Filter settings for each subject, relative to sigma, kappa1 and kappa2 spectralranges.

42

3.3. Multichannel EEG kappa components characterization

(a)

(b)

Figure 3.4.: hf-SEP Time-Frequency representation (TFR) and spectral ranges selection.(a) Three subjects fitted with 29 electrodes. Upper panel: Normalized time-frequency S-Transform. Along the t = 0 axis color-coded digital filter bandsare indicated. Bottom panel: SEP time curves in complementary frequencybands color-coded as in the upper panel (for numerical values cf. Table 3.2);(b) hf-SEP for three subjects fitted with 10 electrodes. Same panel layout asin Figure 3.4a.

43

3. High frequency SEP (hf-SEP) - kappa band

To better localize these components in the temporal domain a frequency analysis was

performed (Figure 3.4) by means of S-Transform [Stockwell et al., 1996], which allowed

to identify three distinct frequency ranges with increased amplitude, slightly different for

each subject (Table 3.2); the corresponding time curves of a representative pericentral

channel contralateral to the stimulated median nerve are displayed in Figure 3.4. Fig-

ure 3.4 exhibits the corresponding time evolution of characteristic voltage scalp maps

so that a spatiotemporal assessment of the high-frequency somatosensory evoked poten-

tial (hf-SEP) becomes possible. Exploratory analysis showed that while the improved

sensitivity achieved here permitted to identify and map high-frequency components at

and above 1 kHz, further increases of SNR would be required to warrant a stable in-

verse multi-source reconstruction; accordingly, in the following topographic descriptions

are reported that can be obtained from the available spatio-temporal characterisation of

the hf-SEP as shown in Figure 3.5 and 3.6. In the full-band traces two low-frequency

peaks were identified, P16 (onset ranging from 15.0 to 16.8 ms) with a widely distributed

monopolar scalp pattern and N20 (from 19.2 to 21.8 ms) with a typical dipolar pattern

(data not shown), consistent with a deep thalamic P16 and a cortical area 3b N20 gener-

ator. In the sigma-band (400-800 Hz) two main components were encountered, an early

monopolar one (onset from 12.2 to 15.4 ms), showing two or three peaks with inter-peak

intervals of about 1.8 ms, and a later dipolar component (similar to the N20 map) with

onset from 18.8 to 21.8 ms featuring four to five peaks (inter-peak intervals 1.6 ±0.13

ms, mean ±s.d.), thereby reproducing the typical spatiotemporal sequence of a leading

thalamic sigma-burst generator and a trailing cortical source [Gobbelé et al., 2000]. In

the kappa-band (800-2000 Hz) the early onset monopolar pattern indicated a subcortical

deep radial source, pointing towards C4-CP4 and starting between 11.8 and 14.6 ms. The

inter-peak-interval (IPI) across all subjects for the major peaks is IPI = 1.12 ±0.1 ms

(mean ±s.d.). Around 20 ms this component became more focal over the somatosensory

area contralateral to the stimulus, lasting until between 21.4 and 23.4 ms. Consistently

with the time-frequency analysis, different spatial pattern in the kappa band can be dis-

tinguished: kappa1 (800-1150 Hz) and kappa2 (1100-2000 Hz). Figure 3.6 demonstrates

the pattern stability in the kappa1-band with a repetition period of 1 ms corresponding

to a frequency of 1 kHz. Specifically, the alternation between a focal monopolar and a

narrowly spaced tripolar pattern points to a superficial and complex (radial and tangen-

tial) multiple sources generator. No repeating pattern was identified for the kappa2 band

44

3.3. Multichannel EEG kappa components characterization

Figure 3.5.: Spatiotemporal pattern stability for (a) kappa1 band (850-1150 Hz) and (b)kappa2 band (1250-2000 Hz) for subject 1.

(Figure 3.6 b); notably, there was always an alternation between a monopolar and a more

complex pattern, but here the temporal binning is not a perfect divider of one main peri-

odicity. Similar results were observed for all subjects with individually slightly varying

different spatiotemporal patterns.

Thus, while the sigma-burst can be interpreted by two (monopolar and dipolar) sources,

the kappa components apparently have more complex origins (Figure 3.6). Both sigma

and kappa wavelets have their onset around 13 ms, corresponding to a deep radially ori-

ented component with a monopolar scalp pattern. In contrast, the cortical kappa band

generators were evidently more complex than the superficial sigma source in area 3b,

possibly reflecting spike activities from multiple generators in the primary somatosen-

sory hand area with different soma and/or axon orientations.

45

3. High frequency SEP (hf-SEP) - kappa band

Figure 3.6.: Kappa band and sigma band reflects at least partially different neural sub-strates. Superimposed time curves and selected scalp maps at identical laten-cies for sigma band (a) and kappa1 band (b) for subject 1.

46

3.4. Low noise EEG/MEG recordings of kappa band components

3.4. Low noise EEG/MEG recordings of kappa bandcomponents

3.4.1. Settings

We recorded simultaneously SEP/SEF in five human healthy subjects inside an electro-

magnetically shielded room, combining a low-noise EEG custom-made amplifier (4.7

nV/√

Hz, [Scheer et al., 2006]) and low-noise custom-made single-channel MEG (two

different data set measured with two different setup, as specified below).

Ag/AgCl sintered electrodes were placed on F3 and C3’ position of the 10/20 system,

with reference and ground on nose and left cheek, respectively. Impedance values were

kept below 3 kΩ. The MEG system consisted of a single stage DC SQUID current sensor

with direct read out [Drung et al., 2007] coupled to a first order axial gradiometer. It is

immersed in liquid helium and operated in a low-noise Dewar. The warm-cold distance

was 8.8 mm when measured at room temperature. The Dewar position was optimized by

climbing the N20m amplitude gradient when moving stepwise laterally from C3 towards

the left temporal lobe. The measurements were carried out in a magnetically shielded

room, which consists of 2 layers of mu-metal and one layer of eddy-current shielding

(PTB, Berlin). The data presented here have been acquired during two separate experi-

mental sessions, realized performed with two different MEG setups. characterizing the

single channel SQUID system. In the first case, a first -order axial gradiometer (diameter

20 mm, baseline 120 mm) with a total noise level of 1.62 fT/√

Hz at 1 kHz referred to

the bottom pick-up loop was connected to the input of the SQUID current sensor. In the

second session an augmented low noise MEG setup was used: since the larger diameter

and the large baseline make the sensor suitable for the detection of deep sources, we im-

plemented a gradiometer with 45 mm diameter and 120 mm baseline, improving the noise

level to 0.50 fT/√

Hz at 1 kHz. The sensor is immersed in liquid helium and operated in

a low-noise dewar with a special superinsulation. We refer in the following to the setup

used in the first session as standard MEG and to low-noise MEG for the one adopted

in the second session (recording session IIIa and IIIb A.1, respectively). An electric

square-wave pulse, 200 µs wide, was delivered on the right wrist in correspondence of

the Median nerve with stimulation frequency of 4.1-4.3 Hz. The current was set a 1.5

motor threshold (7-13 mA) with sampling frequency of 5 kHz and antialiasing low-pass

filter cut-off at 2 kHz. After optimal placement of the dewar over the scalp, one minute

47

3. High frequency SEP (hf-SEP) - kappa band

resting state was recorded to verify the noise level of the combined system, followed by

2 right side stimulation sessions of 10 minutes for the experiment I for a total amount of

6000 trials, and 4 right side stimulation sessions of 10 minutes for the experiment I for a

total amount of 12000 trials. Subjects were instructed to remain awake with eyes open.

Responses were averaged and evaluated by time-frequency Stockwell transform [Stock-

well et al., 1996]; three different bands (N20, sigma and kappa) were identified and fil-

tered in order to compare the source projections in terms of electric and magnetic fields.

The Signal-to-Noise-Ratio (SNR) was calculated as ratio between root mean squared

(RMS) bandpass-filtered averaged evoked response in the temporal window of interest

(15-25 ms post-stimulus) and the RMS of a 10 ms baseline segment taken from the pres-

timulus period.

The energy of pre- and post-synaptic components was analyzed in terms of root-mean

square (RMS) for sigma and kappa bands. The pre- and post-synaptic components were

identified as the content of the 5 ms time trend preceding and following the N20 peak. A

reference to the RMS of the baseline is also calculated. Mean and standard error across

subjects is provided. The statistical difference between high frequency pre- and post-

synaptic RMS distributions has been tested with Wilcoxon Rank-Sum Test.

3.4.2. Results

Noise figures of the two setups are compared in Figure 3.7. Power spectral density of 3

minute resting shows the expected difference between the two MEG setup, with the low

noise MEG used for lying on 0.50 fT/√

Hz at 1 kHz, a quarter of the noise reached with

the standard EEG (around 1.62 fT/√

Hz at 1 kHz). The EEG system is roughly the same

for both sessions. The apparent baseline separation has to be referred to a difference in

the impedance level: while in standard MEG session was possible to achieve 1 kΩ at the

interface, resulting in 5 nV/√

Hz flat line in the high frequency range, in the low noise

MEG session session the impedance preparation led to 2.3 kΩ for the subject shown here,

corresponding to a noise level of around 7.5 nV/√

Hz.

The evoked response to the somatosensory stimulation is analyzed on the basis of their

time-frequency transform. In Figure 3.8 the averaged SEP/F (N=6000) are projected on

the time-frequency plane, normalized with respect to the pre-stimulus interval. The fil-

tered time traces for the broadband N20 and N20m, sigma (450-750 Hz) and kappa (850-

1200 Hz) bands are reported sequentially below. A strict correspondence links magnetic

48

3.4. Low noise EEG/MEG recordings of kappa band components

Figure 3.7.: Spectral representation of one minute resting state for MEG (above) and EEG(below) systems from experiment I (blue) and experiment II (red).

and electric projections over the scalp, as they co-vary in terms of power across sessions

for the same subject. In the sigma range a strong relation between MEG and EEG is ob-

servable for both experimental sessions. In particular it is possible to distinguish, given

the good SNR, two distinct components, immediately preceding and following the N20

peak: this result matches the data reported in the literature, corresponding respectively

to the presynaptic contribution from ascending thalamocortical pathways, and the postsy-

naptic feature generated by intracortical population activity.

In addition to this, the noise level of the two MEG systems does not show signifi-

cant difference, meaning that for this amount of trials the improvement from 1.62 to 0.5

fT/√

Hz is not so critical for sigma range signatures detection. It appears evident, instead,

how the poor quality of the standard MEG recording (Figure 3.8a) prevents the possi-

bility to detect distinct components in the kappa range, while with the low noise MEG

system the lower noise of the device leads to the extraction of a ripple around 20 ms

after the stimulus onset (Figure 3.8c). Such components are possibly present also in the

data recorded with standard MEG, but, given the limitation of the system, they remain

hindered by the noise level. In Figure 3.9 the superimposition of evoked responses from

49

3. High frequency SEP (hf-SEP) - kappa band

Figure 3.8.: (a) Time Frequency Representation (TFR) of the MEG signal from the stan-dard MEG session, and relative time trends for the averaged SEP in the fre-quency bands of interest (broadband, sigma and kappa range); (b-d) samerespectively for the EEG from from the standard MEG session and MEGand EEG from from the low noise MEG session. The data refer to the samerepresentative subject, who underwent both experimental sessions.

50

3.4. Low noise EEG/MEG recordings of kappa band components

the kappa range averaged SEF, for the same representative subject, and the same amount

of trials is provided. The difference in the baseline (outside of the signal window, 15-30

ms) put in evidence the system performance criticality: the burst expressed by the record-

ings of the second experimental session is distinguishable from the baseline, but do not

dominate, rather correlate, with the trend obtained with a less sensitive setup.

Independently from the session we refer to, a stable scaling effect across different spec-

tral ranges can be observed. Consistently with other SEP/F studies, a correspondence

between 10 nV and 1 fT allows a direct comparison of similar and complementary elec-

tromagnetic field projections. As shown in the lower panels of Figure 3.8, a scaling

factor of 10 characterizes the transition from the N20/N20m to the sigma range ripples in

both magnetic and electric domains. Similarly, a scaling factor of 10 must be taken into

account passing from the sigma to the kappa range. This physical property offers a funda-

mental reference in the analysis of these spectral components, their putative generators,

and the available projections onto system with specific spectral and spatial sensitivity.

Since the low noise MEG performances allows to investigate the kappa band in its mag-

netic field projection, we considered the averaged of the complete dataset (N = 12000).

SNR in high-frequency bands has been calculated for the burst time window: across sub-

jects, in the sigma-band SNR was found at 12.9 ± 5.5 / 19.8 ± 12.6 for EEG / MEG,

while in the kappa-band SNR at 3.77 ± 0.8 / 4.5 ± 2.9.

In order to investigate the role of different temporal components, we analyzed the rela-

tion of pre- and post-synaptic components in the sigma and kappa spectral range, across

subjects and measurement modalities. The complete statistical outcome is reported in

Figure 3.10. Each panel describes RMS of pre- synaptic, post-synaptic, and baseline time

intervals. In the case of 600 Hz features, it appears evident how MEG and EEG show

similar projections. Even the time trends trace each other in terms of peak and power (not

shown). In addition, the stronger contribution in terms of energy comes from the later

components.

In kappa range this trend is inverted, the power is confined in the earlier part of the

burst. This effect deserves attention, because while in the EEG data pre- and-post-

synaptic RMS provides a clear statistical difference across subjects, this is not verified

in the case of the MEG detected burst, where the overall affect is similar, but this trend is

followed only by 3 over 5 subjects. This difference can be addressed partly to the system

sensitivity and partly to the neurophysiological nature of the source, and will be examined

51

3. High frequency SEP (hf-SEP) - kappa band

Figure 3.9.: Superimposition of averaged kappa band SEP/F (N = 6000) for the two ex-perimental sessions. The burst amplitude is visible in the second session (redtrace), while in the first experiment the noise level is too high to allow a cleardistinction (green trace).

in the following paragraph.

3.5. Discussion

3.5.1. On the low noise hf-EEG

The present data demonstrate the possibility to increase the sensitivity of non-invasive

EEG recordings by minimizing the various sources of noise in the high-frequency range.

To this end, we recorded median nerve SEP in an electromagnetically shielded envi-

ronment, employed a dedicated custom-made low-noise preamplifier, and optimized the

electrode-skin impedances. This combination enabled the detection and even mapping

of neuronal evoked components in the kiloHertz frequency range non-invasively. Scalp

mappings of hf-SEP reveal that the kappa component at and above 1 kHz exhibits more

than one repeating spatial pattern and can be observed also after the N20 peak; notably,

at identical latencies patterns clearly different from the sigma burst were observed sug-

gesting that the kappa-band provides additional information. All SEP parameters, such as

onset time, wavelet duration, time evolution, and voltage distribution on the scalp, con-

52

3.5. Discussion

Figure 3.10.: . Root-mean square of pre- and post-synaptic components for sigma (upperpanels) and kappa (lower panels) bands recorded in experiment II, for EEG(left panels) and MEG (right panels). The blue bars explain the mean ofthe burst RMS before (Pre) and after (Post) the N20 peak, and the baselineâNoiseâ. The black star refers to significant difference between components(Wilcoxon test, p-value < 0.05).

53

3. High frequency SEP (hf-SEP) - kappa band

sistently indicate the detectability of several kappa-generators in subcortical as well as

intracortical structures.

Subcortical pre-synaptic and intracortical post-synaptic hf-SEP components have been

implicated in numerous studies, showing different dependence on stimulus parameters

like single vs double pulses [Klostermann et al., 2000] or stimulation frequency [Kloster-

mann et al., 1999]. In particular, cortically applied kynurenic acid (blocking glutamater-

gic receptors) selectively reduced the later high-frequency (postsynaptic) cortical burst

responses in piglet primary somatosensory cortex [Ikeda et al., 2002] without affecting

the earlier (presynaptic) subcortical components. The origin of the subcortical compo-

nents is ascribed to cuneothalmic fibers (12-15 ms ) and action potential barrages in fibers

of thalamo-cortical projection cells (15-18 ms). The origin of the cortical late (post N20)

burst oscillations has been suggested as linked to GABAergic inhibitory cells [Hashimoto

et al., 1996]; notably, bursts are not affected by modified GABAergic inhibition after lo-

razepam [Restuccia et al., 2002b] or tiagabine [Restuccia et al., 2002a]. Dosage, however,

appears critical because higher lorazepam levels can delay later burst peaks [Haueisen

et al., 2000]. The acetylcholinesterase inhibitor rivastigmine increased selectively later

burst wavelets between 18-28 ms; interestingly, pyramidal chattering cells can be cholin-

ergically activated whereas ACh does not modify the firing rate of fast-spiking GABAer-

gic interneurons [Restuccia et al., 2003].

Measurements obtained in patients with Deep Brain Stimulation (DBS) intrathalamic

electrodes distinguished a thalamic very fast component at 1 kHz from a later cortical

cortical component at around 600 Hz [Klostermann et al., 1999,Klostermann et al., 2002].

This distinction is robustly supported by invasive concomitant local-field potential (LFP)

and single unit recordings [Hanajima et al., 2004] showing a neat match between neuronal

spikes and oscillations.

Moreover, such kiloHertz kappa band contributions have been detected after median

nerve stimulation in the somatosensory cortex in epilepsy patients who underwent long-

term subdural electrode monitoring (Sakura et al. 2009) and in resting state recording in

extremely localized cortical regions [Usui et al., 2010], as well as in epidural measure-

ments at lower structures such as the spinal cord and dorsal column nuclei [Insola et al.,

2008, Insola et al., 2010].

The presence of such a fast firing rate can be understood in the face of firing proper-

ties, e.g., of thalamic cells capable to discharge at high frequencies (800-1000 Hz) during

54

3.5. Discussion

spindle oscillations [Steriade et al., 1993]. This confirms the kiloHertz range as a natural

response mode, which is not artificially driven by the electric stimulation. Alternately,

even if cellular burst displayed interspike intervals of 2 to 3 ms [Hanajima et al., 2004],

they could contribute to some, but not necessarily to all peaks of a 1 kHz SEP component;

thus, multiple burst generators working at shifted phases could superimpose [Kloster-

mann, 2005]. The present non-invasive recordings cannot specify the cellular origin of

the kappa oscillations. Nevertheless, different spatial patterns before and after the N20

peak were identified for the kappa band. While inverse multi-source reconstructions are

not stable given the present signal-to-noise ratio one can detect a contrast between an

initial well defined monopolar pattern (typical for a subcortical radial source, such as

deep thalamocortical fibers) and a later circumscribed tripolar (maybe even more com-

plex) kappa-burst pattern (arguing for multiple superficial cortical sources); this matches

comparably narrow focal hf-SEP patterns as described also in invasive (epidural) record-

ings [Sakura et al., 2009].

3.5.2. On the low noise combined hf-MEG/EEG

The complementary information gained by simultaneous EEG and MEG recordings was

already shown for high-frequency SEP/F in the 600 Hz range [Curio et al., 1994]. The

sigma range in particular has been largely investigated in the last two decades with these

two modalities (for a complete review see [Curio, 2005, Ozaki and Hashimoto, 2011]).

Our interest here it is test the detectability of faster and weaker components, already

isolated with custom made low-noise EEG [Scheer et al., 2011] and characterized as

distinct from the well-known sigma burst [Fedele et al., 2012], but not yet detected with

MEG. The opportunity to measure with a custome-made low noise MEG system, and

the possibility to exploit two consecutive stage of technological development in terms of

system noise reduction, has allowed the collection of the presented data.

The choice to show two datasets, recorded in separate sessions with two different sys-

tems, allows us to describe the criticality of the high-frequency detectability in terms of

system noise level with respect to different spectral components.

The main contribution of the study is the empirical demonstration of MEG system

noise figure criticality in the detection of high-frequency components in the kHz range.

Two low-noise systems were tested, and only the second, with baseline of 0.5 fT/√

Hz,

could provide a clear burst standing out of the noise level. The first system, present-

55

3. High frequency SEP (hf-SEP) - kappa band

ing 1.62 fT/√

Hz, did not necessarily fail in the detection, but the baseline level did not

allow a distinct kappa burst isolation. In technical terms, the lower noise for MEG is

provided by an enlargement of the pick-up loop, from a diameter of 20 (system I) to

45 mm. Intuitively, the noise level at the input is given by the magnetic field B, given

by the magnetic flux divided by the loop area. Since the flux is constant and the loop

area increases, the system provides a lower noise level, which leads to a lower baseline

in the high-frequency range. This results in a theoretical increase of a factor of 5 in

terms of pick-up loop area, which accounts for the improvement from standard to low

noise MEG setup. This certainly does not explain alone the observed difference between

the datasets, as the signal flux approaching the sensor is neither uniform nor extended

over the all surface available from the coil placed in the proximity of the scalp. How-

ever, the comparison between the kappa burst amplitude and the level of the baseline

across different setup shows the central role the system noise figure: the burst detected

by the low noise MEG would still be missed with the standard MEG system. Technical

improvements and experimental evidences demonstrate the low-noise MEG system crit-

icality in the detectability of high-frequency weak components. The different sensitivity

in MEG/EEG systems co-recording must be taken into account in the interpretation of the

results. The corresponding scaling factor described across spectral ranges shows as 10 nV

in the EEG roughly corresponds to 1 fT in the MEG. This scaling effect does not concern

just peaks and bursts in the signal time window, but relates to the baseline, and the system

noise level: the low-noise EEG preamplifer presents 4.8 nV/√

Hz at the input, while the

low-noise MEG 0.5 fT/√

Hz. Therefore, the two systems offer an equivalent technical

sensitivity, allowing a direct comparison of electromagnetic projections of complemen-

tary physiological information. In this sense, we are allowed to consider SEP and SEF

variability not dependent on the recording system physical properties, but on the nature

of the neurophysiological generators.

Exploiting the two system sensitivity, we focused on the projection of pre- and post-

synaptic activity in the sigma and kappa band. As reported in Figure 3.10, we analyzed

the power of burst section immediately preceding and following the N20 peak. In the case

of the sigma band EEG and MEG carry a similar pattern in terms of peaks and time-lags,

and the statistical analysis show a dominant later component detection for both systems.

In the kappa range the situation is more complex: for the EEG the earlier component is

dominant, while for the MEG this is statistically consistent across subjects. Such discrep-

56

3.5. Discussion

ancy possibly relates to the sources position and orientation: the MEG, for its intrinsic

physical configuration, is more sensitive to tangential components, while the EEG to ra-

dial ones [Haueisen et al., 2012]. For the same reason the EEG offers a better detection of

deeper sources. These results would suggest a deeper and radial orientation for the first

part of the kappa burst, but only a combined multichannel study could provide sufficient

evidence.

In conclusion, it has been shown that low-noise MEG performance matches well with

simultaneously recorded low-noise EEG for the non-invasive detection of somatosensory

evoked activity at and above 1 kHz. The EEG and MEG system utilized in this study

are technically equivalent in terms of input noise level, and their sensitivity allows to

observe a correlated scaling effect in the spectral features of the evoked response for both

recording modalities, and to attribute entirely to the generative sources the variability

between electric and magnetic projections detected on the scalp. In particular, we report

that a difference in temporal features in the sigma and kappa burst can be observed, with

post-synaptic components dominating in the 600 Hz range, and pre-synaptic components

expressing higher power in the 1 kHz range.

57

4. Towards single trial resolution

In the previous chapter high frequency spike-like activity detection up to 1 kHz at the

level of the scalp has been demonstrated. The SEP high frequency components elicited

by electrical median nerve stimulation could be isolated because the low input noise figure

achieved with the made-in-house dedicated recording system, and oscillations as small as

tens of nV peak-to-peak could be identified after averaging across a massive number of

trials. Nevertheless, even if the phase-locked features of interest could be separated from

the diverse noise contributions, the analysis was limited to the data sensor space, relying

on scalp derived potential, without a tailored spatio-temporal sources characterization,

neither retrieval of single trial information.

In this chapter we explore the combination of low-noise technology together with ma-

chine learning techniques, and describe the biophysical constraints to be satisfied in or-

der to achieve single trial level information in critical Signal-to-Noise Ratio conditions.

Computer simulations were run to evaluate the algorithms performance with respect to

different generative models, added-energy and phase reset. Finally, further characteriza-

tion of sigma and kappa burst is presented, by applying optimal decomposition on SEP

data obtained at different stimulus rate. Part of this chapter is adapted from [Fedele et al.,

2013]. The EEG recordings described and analyzed in this Chapter are listed in A.1 as

recordings IV.

4.1. Introduction

The EEG detectability of neuronal responses to perceptual or cognitive paradigms is af-

fected by noise, which affects the distinction of the stimulus related activity. Notably,

such responses exhibit reproducible signatures in the spatio-temporal domain, and are

phase-locked to the stimulus onset. As expression of a structured biological reaction to

an external input, they are named event-related potential (ERP).

ERP are generated by different sensory, motor, or cognitive events, that trigger one or

59

4. Towards single trial resolution

more specific brain areas, giving rise to neuronal activation, and producing an electromag-

netic projection on the scalp. The direct relation between stimulus and ERP allows the

investigation of underlying cognitive processes through the analysis of temporal, spec-

tral and spatial features. Nevertheless, even if the brain source projecting to the scalp

can be identified, it remains unclear how the underlying neuronal activity is generated.

Three basic biophysical models of ERP generation have been proposed so far: added en-

ergy, phase-reset and baseline-shift. In the added-energy model, the stimulus elicits ERP

linearly superimposed on the background ongoing activity [Shah et al., 2004, Mäkinen

et al., 2005,Mazaheri and Jensen, 2006,Haufe et al., 2011]. In the phase-reset model, the

stimulus synchronizes the phase of ongoing oscillations [Makeig et al., 2002, Fell et al.,

2004]. In the baseline-shift model, ERs are produced through the amplitude modulation

of ongoing oscillations, which do not have zero mean [Nikulin et al., 2007]. The ERP,

phase-locked to the stimulus, can be distinguished from the background through averag-

ing across many trials. Since the averaged response is not informative about the underly-

ing generating mechanisms, the interest has been directed towards the selection of criteria

or parameters guiding to the distinctive underlying neural process [Sauseng et al., 2007].

If, on one side, the neurophysiological investigation focuses, at the microscopic scale, on

the properties and the interaction between single neuronal units, in terms of response to

an incoming stimulus and of their coupling, directly affecting the local spatial averaged

potential, on the other hand non-invasive techniques as EEG and MEG, can provide only

an abstract representation of the complex scheme of biochemical interactions, at a larger

macroscopic scale,. In this sense it is not possible, in the absence of strongly reliable

prior information on the nature of the neural population involved, and moreover on the

level of functional interconnection, to discern between different generative mechanisms

in neurophysiological terms [Telenczuk et al., 2010].

A sub-class of event-related potentials is represented by the somatosensory evoked

potential (SEP), elicited by electrical median nerve stimulation. The interesting aspect

of SEP is the presence of fast spike-like activity oscillating at frequencies above 400

Hz, with stimulus phase-locked ripples as tiny as tens of nV [Curio, 2005, Ozaki and

Hashimoto, 2011]. High frequency SEP (hf-SEP) represent an opportunity to non-invasively

investigate neuronal spiking activity, as well as a challenging detectability problem in

terms of Signal-to-Noise Ratio (SNR). Here we provide an analytic approach that, re-

gardless of the generative mechanisms of the detected evoked response, allows to extract

60

4.1. Introduction

neurophysiological features even in critical SNR conditions.

Traditional EEG analysis methods such as trial-averaging improve the SNR typically

considering the time course of individual channels. Modern EEG recording system pro-

vide up to 256 of sensors. This has created a need for tools capable to simultaneously

exploit this multivariate information. The activity in multiple channels is often visual-

ized as a topographic map across the scalp, and various methods are commonly used

to integrate this activity for localizing neuronal sources at a specific location within the

brain [Niedermeyer, 1996, Mosher et al., 1999, Darvas et al., 2004, Michel et al., 2004].

Source localization represents a mathematical ill-posed problem: the number of sources

assigned to different patches of cortical tissue is generally higher than the number of

available sensors. Since to each of the sources 6 variables (3 for the position, 3 for the

orientation) must be assigned, it turns out that many possible current distributions can

lead to the same observed EEG activity. To resolve this ambiguity, localization meth-

ods try to explain the spatio-temporal statistics of the observed data by constraining the

possible source distributions in space using plausible priors relating to anatomy of the

head. An alternative approach, avoiding any spatial modeling assumptions with regard

to the sources anatomy relies entirely on the statistics of the observed data. Recently,

various multivariate signal processing algorithms have been proposed, linearly combin-

ing EEG multivariate signals to generate an informative and aggregated data representa-

tion [Chapman and McCrary, 1995,Makeig et al., 1996,Ramoser et al., 2000,Parra et al.,

2002,Blankertz et al., 2008,Blankertz et al., 2011,Lemm et al., 2011]. More specifically,

denoting with x(t) the vector of multidimensional EEG data at time t, and applying to

x(t) the weighting vector w, the projection y(t) is generated as

y(t) = wT x(t) (4.1)

This linear projection y(t) can be obtained considering different strategies. Here, we

explore the potential of two criteria: the maximum power ratio and the maximum correla-

tion between two dataset. The first is described and implemented in the framework of the

Common Spatial Pattern (CSP), the last in the Canonical Correlation Analysis (CCA).

The CSP decomposition is based on the differences in power of different dataset, ex-

pressed in their covariance matrix, while the CCA relies on their correlation, expressed

as cross-covariance. Each of these assumptions on x(t) leads to the mathematical formu-

61

4. Towards single trial resolution

lation in terms of Generalized EigenValue Decomposition (GEDV), allowing to compute

optimal spatial filter w and to extract the desired projections y(t), which are called also

components or features. Here, an innovative approach is suggested, maximally exploiting

the prior-information provided from the synchronized nature of the evoked response. The

considered machine learning techniques are described and evaluated in the context of a

computational biophysical scenario. Supported by the outcome of the simulation, fea-

turing burst-like oscillatory sources, we show the opportunity to distinguish contributing

sources in the scalp derived hf-SEP even in situation of severely low SNR. The compos-

ite nature of the scalp derived SEP sigma burst (around 600 Hz) has been characterized

using a wide range of stimulation frequencies [Klostermann et al., 1999]. Lately, the

non-invasive detection of even faster components has been proved possible, at around 1

kHz, named kappa bursts [Scheer et al., 2011], before accessible only by invasive record-

ings [Klostermann et al., 2002, Hanajima et al., 2004]. Here we sample the refractory

behavior of sigma and kappa burst by stimulating the median nerve at 1 and 8 Hz, aim-

ing to disentangle diverse contributions at the level of the extracted sources. Given the

critical SNR of the kappa burst, a fair comparison, other than the topographic map tempo-

ral evolution [Fedele et al., 2012], has not been possible yet. Optimal feature extraction

techniques are applied on low-noise system recordings, in order to extract and compare

high frequency range components, and differentiate their respective contributions also on

the basis of their spatiotemporal profile. Following this introduction a description of the

computational framework, the experimental recordings, and the algorithmic approaches

is provided. The results are organized to show at first the performance of the considered

analytical strategies evaluated in a simulated biophysical scenario; afterwards decompo-

sition techniques are applied on the SEP, with particular focus on hf-SEP, in the sigma

and kappa band. Optimal source analysis serves further characterization of hf-SEP in

source space in order to differentiate features belonging to separate frequency bands, on

the basis of spatial and temporal properties, and stimulation rate. Finally, the ensemble

of presented computational and experimental results is discussed, in order to clarify the

biological constraints as well as to describe the role of optimal mathematical approaches

to non-invasively describe poor SNR ERP components.

62

4.2. Methods

4.2. Methods

4.2.1. Simulation settings

The forward model is realized using a realistic Montreal head model [Holmes et al.,

1998], with N = 2 142 dipoles arranged in a cubic grid with 1 cm side length. Since

the low-noise EEG recording setup consists of 30 monopolar derivation, in order to com-

pare the outcome of the simulation with experimental data, the forward model projects to

30 scalp EEG channels positions, with leadfield matrix calculated according to [Nolte and

Dassios, 2005]. We computed two set of simulations, one mimicking added-energy and

one the phase-reset macroscopic generative model. The implementation of two models

here aims to show the sensitivity of the decomposition techniques, and not the achieve-

ment of a distinction at a microscopic biological level. The single trial was considered

to have 100 samples, the first 50 samples belonging to the pre-stimulus and the last 50

samples corresponding to the post-stimulus time interval. The sampling frequency was

set to 5 kHz, in order to have 10 ms for pre- and post-stimulus, with t = 0 for the stimulus

onset, directly consistent with the experimental data, where a signal window of 10 ms

is defined. For the generation of noise we used N-1 uncorrelated dipoles mimicking 1/f

type of noise, filtered in the frequency range of interest (400-800 Hz). The dipoles have

random spatial location and orientation. Importantly, this type of noise produces spatial

correlations in the sensor space. The source model relates to the considered macroscopic

generative model. The pre-stimulus interval is a flat zero line in the case of added-energy

model, and a 600 Hz oscillation with random phase in the case of phase-reset model. The

post-stimulus interval presents a pure 600 Hz oscillation (5.8 cycles, phase = 0 at t = 0).

No additional sensor noise was considered, as, for the algorithm performance evaluation

in terms of source features extraction, additional uncorrelated noise would represent a

more composite but not more demanding scenario. The forward model can be written as

x(t) = Ls(t) (4.2)

where x(t) is the multivariate EEG signal matrix, [t x k], with t samples and k channels,

L is the forward model lead field matrix, [k x N], with each row expressing a specific

spatial pattern, s(t) is the signal in the source space, [N x t], with one spatio-temporal

pattern fixed for all trials, and all the other dipoles left to vary randomly. The simulation

63

4. Towards single trial resolution

were parameterized with respect to: 1) the Signal-to-Noise ratio (SNR), calculated at the

scalp level as the ratio between the root mean square (RMS) of the source projection

(SigRMS) and the mean of the RMS of the N-1 random dipoles projections (NoiseRMS):

SNR =SigRMS

NoiseRMS(4.3)

This definition is useful to compare simulation and experimental results, where the

averaged phase locked component represents the best estimation for the signal (SigRMS).

The SNR was tuned to vary between 0.01 and 1, in steps of 0.01; 2) the number of trials,

ranging between a minimum of 100 and a maximum of 5000. Each condition was iterated

20 times, for each SNR and trial number combination.

4.2.2. Experimental setup

SEP measurements were performed in four healthy subjects, inside an electrically and

magnetically shielded room (Vakuumschmelze AK3b). Thirty sintered Ag/AgCl ring

electrodes were placed using a fabric cap with holders at positions according to the 10/20

system. The reference was placed on the nasion, the ground on left cheek. The elec-

trodeâskin interfaces were carefully prepared using standard abrasive paste until for all

electrodes impedances at or below 1 kΩ were achieved; impedance stability was checked

and confirmed repeatedly during all measurements. Data were acquired using a band-

width of 0.16-2000 Hz (total gain 10000, ADC rate 5 kHz, resolution 20 bits). We utilised

a low-noise custom-made amplifier, with a white noise level of 4.8 nV/√

Hz [Scheer et al.,

2006]. During the recordings the subject was comfortably resting on a bed in a relaxed

position. A constant-current square wave electrical stimulus of 200 µs width, 5-8 mA

(1.5 motor threshold). Two stimulation repetition rates were applied transcutaneously to

the median nerve at the right wrist: one slower, at 0.99 Hz, for 6 experimental blocks of

15 minutes (block type A), and one faster, at 8.1 Hz, for two blocks of 5 minutes (block

type B). For all subjects the measurement blocks sequence was: A-A-B-A-A-B-A-A.

64

4.2. Methods

4.2.3. Data Analysis

SEP preprocessing

The data were epoched with respect to the stimulus onset. A pre-stimulus interval of 50

ms and a post-stimulus interval of 100 ms were used. Stimulus artefacts were removed in

single-trials in each channel by cubic interpolation from -5 to +5 ms around the stimulus

onset. Trials exceeding three times the average standard deviation in the frequency range

450-2000 Hz were rejected (less than 5% for all subjects). Channels with noise level

higher than expected in the high frequency range (> 450 Hz) were removed from the

analysis (channel CP4 for subject 1; channel C2 for subject 4).

Sensor space analysis

The averaged signal was projected in the time-frequency plane by S-transform. Time

average signal, Power Spectral Density and Time-Frequency representation were anal-

ysed by visual inspection of all channels, subjects and stimulation condition. Three fre-

quency band of interest were identified: N20 (50-300 Hz), sigma (450-850 Hz), kappa

(900-1600 Hz). Butterworth filters (order = 8) were applied to conduct further analysis

in each selected spectral range. For each frequency band the respective SNR has been

calculated as

SNR =SigRMS

NoiseRMS(4.4)

where SigRMS stands for RMS of the filtered averaged signal in the 15-25 ms post-

stimulus time window, while NoiseRMS refers to average RMS of filtered single trial pre-

stimulus interval. In this sense SNR is an index of detectability for the burst before

the averaging procedure, consistently with the definition used for the evaluation of the

simulations outcome.

Source/Component Space Analysis

We compare in the two set of simulations and in the real hf-EEG data the performance

of three multivariate analysis techniques: Common Spatial Pattern (CSP), Canonical Cor-

relation (CCA), Canonical Correlation Average regression (CCAr). All are based on

GEVD of covariance matrices, computed in specific temporal intervals of the analysed

65

4. Towards single trial resolution

dataset. The relation between spatial filters and pattern has been described in section

2.3.4 [Blankertz et al., 2011, Haufe et al., 2014].

Common spatial pattern (CSP)

CSP computes the simultaneous diagonalization of two covariance matrices, by maxi-

mizing their ratio. For each dataset we define two classes: the signal, relative to a prede-

fined time window in the post-stimulus interval, and baseline, relative to a pre-stimulus

baseline. For each of the two classes we compute the covariance matrices Cpost and Cpre.

Referring to the notation introduced in section 2.3.4, the spatial filters are extracted by

solving

maxwwT SdwwT Scw

(4.5)

withSd =Cpost −Cpre

Sc =Cpost +Cpre

(4.6)

obtaining w, the spatial filter maximizing the power ratio between the two classes. In

the simulations the definition is straightforward, since the pre- and post-stimulus interval

are by construction the time interval of interest. In the case of SEP recording, the pre-

processed data are filtered in the frequency band of interest, and for each trial we set the

baseline temporal window ranging from -25 to -15 relative to the stimulus onset, while

the signal window is set from 15 to 25 ms post-stimulus, as in this interval the expected

cortical contribution to hf-SEP is maximal. The obtained eigenvalues represent the power

ratio of each source across conditions, while the eigenvector are the corresponding spatial

filters.

Canonical Correlation Analysis (CCA)

CCA computes maximizing the correlation of two multivariate dataset, X and Y . It pro-

vides the optimal basis vectors wx and wy diagonalizing the cross-correlation matrix, Cxy,

such that the correlation between the projections of the variables onto these basis vectors

are mutually maximized. The mathematical problem can be formulated as follows, as

introduced in Fundamentals,

66

4.2. Methods

maxwx,wyρ =wT

x XY T wy√wT

x XXT wxwTy YY T wy

=wT

x Cxywy√wT

x CxxwxwTy Cyywy

(4.7)

Multiple orthogonal projections can be computed simultaneously by solving the gen-

eralized eigenvalue problem [Knutsson et al., 1998]:(0 Cxy

CTxy 0

)(wx

wy

)= ρ

(Cxx 0

0 Cyy

)(wx

wy

)(4.8)

An alternative efficient and stable implementation is the QR decomposition of X and

Y , followed by a singular value decomposition [Bjoerck and Golub, 1971], as imple-

mented in the Matlab canoncorr.m function. Different implementations of this mathe-

matical framework can be used, depending on the choice of the datasets X and Y which

are plugged into the algorithm.

A straightforward choice is to use the concatenation of even trials as X and the con-

catenation of odd trials as Y :

X = (X2X4...XN)

Y = (X1X3...XN−1)(4.9)

where each XN of size [t x k], is extracted for the temporal window of interest from

each of the N trials.

Canonical Correlation Analysis regression (CCAr)

Canonical Correlation Average regression (CCAr) follows the mathematical frame-

work of CCA. The difference to the previously presented approach is the dataset Y which

is plugged into the algorithm. In the previous section, the covariance matrix was estimated

as the sample covariance matrix of single-trial EEG. However, if the power contribution

of the evoked component is buried under noise CCA and CSP would fail to extract reliable

filters. Instead, since the evoked component is phase-locked, a more powerful approach

consists in finding spatial filters that maximize the correlation of single-trial data to the

trial-averaged EEG. Following the notation of the previous section this can be formulated

as:

67

4. Towards single trial resolution

X = (X1X2...XN)

Y = (〈X〉〈X〉 ...〈X〉)

with 〈X〉= 1N

ΣNi=1Xi

(4.10)

This approach was derived independently from a similar previous implementation [Rivet

et al., 2009], but analytical comparison and simulation results show their equivalence. A

similar strategy has been applied to ICA-based decomposition [Lemm et al., 2006].

Performance Evaluation

For the simulated data as well as for SEP, the performance of CSP, CCA and CCAr is

presented in terms of Similarity Index (see section 2.3.3) between the estimated spatial

pattern and the original spatial pattern (simulated data), or between estimated patterns of

different subset of data. A similarity index greater than 0.85 was chosen as threshold for

corresponding spatial filters reliability.

Source Reconstruction

Source reconstruction with current equivalent dipole modeling was performed. For

each subject and each condition. The first CCAr pattern for N20, Sigma and Kappa

range SEPs were subjected to BESA (MEGIS Software GmbH, GrÃfelfing, Germany).

A model with at least 75% of goodness-of-fit (GOF) was accepted. The head model used

in BESA was a spherical four-shell of 85-mm radius [Berg and Scherg, 1994]. Fitted

dipoles coordinates are reported in unit sphere coordinates.

Latency Analysis

Sigma and kappa burst peaks are identified in the hf-SEP burst for the two stimulation

rates. For each frequency band corresponding peaks elicited by 1 Hz and 8 Hz stimula-

tion are marked after visual inspection of the averaged strongest source. Latencies shifts

correlation is calculated, and visualized on a cartesian diagram (x-axis for the 1 Hz stimu-

lation burst latencies, and y-axis for the 8 Hz stimulation burst latencies). A statistical test

is performed to compare latencies shifts distribution to zero-mean Gaussian distribution.

68

4.3. Results

Figure 4.1.: Simulations outcome expressed as Similarity Index for (a) Added-energy and(b) Phase-reset generative mechanisms for the set range of SNR and numberof trials, for Common Spatial Pattern (CSP), Canonical Correlation Analysis(CCAr), Canonical Correlation Average regression (CCAr).

4.3. Results

4.3.1. Simulation

The simulation outcome is summarized in Figure 4.1. The comparison between the three

decomposition techniques, Common Spatial Pattern (CSP) and Canonical Correlation

Analysis (CCA) and Canonical Correlation Average regression (CCAr) is illustrated in

terms of Similarity Index between estimated pattern and known source pattern for dif-

ferent number of trials and SNR (x- and y-axis of each plot). The first row describes

the algorithms performance in the added-energy model case, while the second row refers

the phase-reset model. Looking at CSP and CCA, it appears evident that when pre- and

post-stimulus variance are similar at the level of the source (phase reset mechanism), the

power ratio is less reliable for SNR<0.2, while the correlation based optimization detects

correctly the source pattern, independently from the generative model, until SNR>0.08.

Interestingly, CCAr performs well in both situations, retrieving the correct scalp source

contribution even for SNR values as critical as 0.02.

69

4. Towards single trial resolution

4.3.2. Sensor space analysis of SEP data

In Figure 4.2 an example of channel with prominent high frequency content is provided.

The corresponding electrode is placed on the left hemisphere, in the proximity of the so-

matosensory area (position C3, 10-20 system). It is expected to observe a pronounced

N20 peak in time averaged trace (Figure 4.2 a). The interesting information regarding

high frequency energy distribution is visible in the time-frequency plane (Figure 4.2 b),

where baseline normalized power is represented in logarithmic scale. Color coded rectan-

gles highlight high-frequency contributions, in green for the sigma band, and in magenta

for the kappa band. These boundaries are depicted on the time trace as well as in the

Power Spectral Density (PSD) estimation (Figure 4.2 c). It is important to notice that

if the time-frequency plane provides a normalized version of the power distribution, the

PSD is a spectral estimation in the short time interval of interest, offering the absolute

value for the diverse spectral contributions.

Such analysis, performed on all channels, subjects and stimulation conditions has led

to the identification of three frequency bands of interest (see section 4.2.3). SNR for

N20, sigma and kappa ranges have been calculated, and corresponding values for this

channel are marked on the simulation outcome for the added-energy generative mech-

anisms, proposed in Figure 4.2d. The SNR across subjects was found at 0.38 ± 0.17

for the N 20 spectral range, 0.15 ± 0.07 for the sigma burst, and 0.06 ± 0.03 for the

kappa burst. The agreement between simulation results and algorithm performances is

then tested on hf-SEP. Moreover, the macroscopic generative mechanisms can be distin-

guished, by comparison of CSP and CCA pattern/filter extraction. This will subject of

discussion in the final section.

4.3.3. Source space analysis of SEP data

The band-pass filtered epochs covariance decomposition has been performed in three

spectral ranges by CSP, CCA and CCAr. The results for the strongest component for

one subject are depicted in Figures 4.3- 4.5. The reliability of each decomposition is

provided by the validation step Similarity Index, at the lower left of each represented

spatial pattern. Matching the prediction emerging from the simulated data in the added-

energy model, we observed no difference in the spatial and temporal pattern computed for

N20 and sigma bands ( Figure 4.3 and 4.4). Similarity indexes indicate almost identical

patterns in the case of the N20, and of the sigma burst. The single trial temporal evolution

70

4.3. Results

Figure 4.2.: Sensor space analysis. (a-c) Example of SEP for a scalp channel with promi-nent high frequency content in the hf-frequency bands of interest (channelC3, subject 4). (a) Averaged SEP (b) Time-Frequency Representation bybaseline-normalized Stockwell transform. (c) Voltage Spectral Density. Thesigma band is indicated in green (450-800 Hz), while the kappa (900-1400)in magenta. They are outlined in the time, frequency and time frequency rep-resentation, and corresponding expected performance from the algorithm ismarked on the algorithmic performances. (d) Simulation outcomes for Com-mon Spatial Pattern (CSP), Canonical Correlation Analysis (CCAr), Canon-ical Correlation Average regression (CCAr).

71

4. Towards single trial resolution

Figure 4.3.: Source space analysis for the N20 band (50-300 Hz) for a single subject. (a)Spatial patterns (first component) computed by CSP, CCA and CCAr, accom-panied by the Similarity Index. (b) Single trial source extraction, computedin the data source space, for the CSP, CCA, CCAr spatial filters. (c) Compar-ison of the sources average, color coded for CSP, CCa, CCAr. (d) Matrix ofsimilarity Index for the pattern extracted with the three different algorithmicapproaches.

of the sources allow to visualize the N20 peak alignment around 18 ms across epochs, as

well as an outstanding ripple at the same time delay in the sigma band.

For the kappa burst only CCAr provides a reliable source extraction, as predicted in

Figure 4.2d. CSP and CCA validation performance are around 0.5, while CCAr reaches

0.9. This is reflected in a strongest averaged burst, which nevertheless is not clearly vis-

ible at single trial level. The low resolution in terms of single epoch will be debated in

section 4.4. Direct comparison of the algorithm performances on simulated and exper-

imental data is provided in Figure 4.6. Each point of the graph represent the computed

similarity index between the pattern estimated given a limited number of trials (on the

x-axis), and the best possible pattern (the assigned source in the case of the simulation,

and the validated pattern obtained with all trials in the experimental case). On the left

panel outcomes for the sigma range are illustrated: the only divergence between simula-

tion and real data is given by CCAr for a number of trials lower than 1000. In the kappa

range CCAr outperforms even for a lower number of trials, resembling the simulated

72

4.3. Results

Figure 4.4.: Source space analysis for the sigma band (450-850 Hz) for a single subject.Same layout as Figure 4.3

case. Since CCAr performs better than CSP and CCA for all spectral ranges of interest,

the following analysis will rely on the CCAr derived patterns and filters. The complete

number of available trials have been used, even if a number of trials ranging between 500

and 1000 would be sufficient to achieve the following results,

4.3.4. Source reconstruction on patterns

Patterns obtained for all frequency ranges were compared across subjects. The strongest

component for all decompositions resulted in a bipolar pattern, depicted in Figure 4.7 for

the 1 Hz stimulation rate experimental condition. A further spatial characterization of the

source distribution was performed by single dipole fitting. In Figure 4.7 average dipole

localization is presented. Since the number of subjects is 4, and the setup consists of 30

channels which not homogenously sample the head surface, the statistical power of these

data is limited. Nevertheless, the localization of the dipoles outlines a closer alignment

of the hf-sources, while low and high frequency dipoles location are distinctly separated,

as reported in spherical coordinates in Table 4.1 .

73

4. Towards single trial resolution

Figure 4.5.: Source space analysis for the sigma band (450-850 Hz) for a single subject.Same layout as Figure 4.3

Figure 4.6.: Comparison of CSP, CCA and CCAr performances on simulated and exper-imental data. (a) In the sigma burst case, the three algorithm are representedwith dotted line for the simulation, and continuous line for the experimentaldata (mean ±standarderror).(b)samelayoutas(a) f orthekappaband.

74

4.3. Results

rang

eN

20Si

gma

Kap

paco

nditi

onco

ord

xy

zG

OF

xy

zG

OF

xy

zG

OF

1H

zm

ean

-0.5

60.

070.

3995

.37

-0.4

0.17

0.45

98.6

-0.3

50.

170.

3587

.32

std

0.06

0.14

0.07

3.9

0.08

0.07

0.04

0.3

0.07

0.11

0.14

7.1

8H

zm

ean

-0.6

00.

3293

.1-0

.47

0.13

0.43

97.6

-0.3

20.

320.

3188

.33

std

0.07

30.

180.

143.

40.

060.

10.

061.

70.

210.

10.

214.

9

Tabl

e4.

1.:D

ipol

efit

unit

sphe

reco

ordi

nate

san

dG

oodn

ess-

of-F

it(G

OF)

for

the

first

com

pone

nts

inN

20,

sigm

aan

dka

ppa

rang

e,fo

r1H

zan

d8

Hz

stim

ulat

ion

rate

.

75

4. Towards single trial resolution

Figure 4.7.: Source Analysis for the 1 Hz stimulation experimental condition. Spatialbipolar pattern obtained with CCAr for all subjects and frequency bands ofinterest. Mean equivalent dipole localization across subjects for each exam-ined spectral range.

4.3.5. Latency analysis

High frequency components have been compared in terms of sigma and kappa burst peak

latency for the strongest bipolar component across the two stimulation conditions. In

Figure 4.8a spatial and temporal pattern for one subject for sigma and kappa burst are

shown. Even if the source spatial distribution is very similar (Figure 4.8a), the time

pattern (Figure 4.8b) indicates a progression in the peak delays of the 8 Hz with respect

to the 1 Hz stimulated evoked response. Systematic peak detection for both conditions

has been performed, and represented in a 1 Hz delay- 8 Hz delay plane (Figure 4.8c)

for the sigma and kappa burst (left and right panels, respectively). The quadrant bisector

(blue), placed at 45 degrees, represents absence of delay. The linear fit (red) shows a

progressive divergence from the bisector (ttest, p-value <0.05). The delay distributions of

the two bursts are not statistically distinct (ttest, p-value>0.05).

76

4.3. Results

Figure 4.8.: Peak Latency Analysis. (a) Spatial and (b) temporal patterns are shown forthe sigma (upper panels) and kappa (lower panels) burst in the two experi-mental condition, at 1 and 8 Hz stimulation rate. (c) Corresponding detectedpeaks are represented according to the delay in the two stimulation conditionsfor the sigma (left panel) and the kappa (right panel) delays. In both cases theslope of the linear fit is steeper than the I Cartesian quadrant bisector (t-test,p<0.05).

77

4. Towards single trial resolution

4.3.6. Time envelope analysis

Explorative analysis of time envelopes of the averaged sources expressing similar dipolar

pattern has been conducted, for the sigma and kappa spectral range in the stimulation

rates experimental conditions. In Figure 4.9 the averaged source envelopes for the same

frequency band are compared across stimulation condition. The temporal trend is ex-

tended to the baseline in the prestimulus and in the poststimulus interval (-40 to 40 ms

with respect to the stimulus onset), in order to account for different noise levels. In the

sigma range Figure 4.9, left), the power expressed by the burst elicited by 1 Hz stimula-

tion is higher in all subjects, with a later falling time to baseline. Also the kappa burst

(Figure 4.9, right) appears to be enhanced by slower stimulation rate, with not consistent

effect on the ripple duration.

Since the spatial localization shows similar features, further analysis has been per-

formed on the temporal evolution of the sources. For this purpose amplitude envelopes

of the strongest sigma and kappa range components, normalized to their maximum, have

been superimposed along the time axis for each subject and experimental condition. The

outcome is depicted in Figure 4.10, where the envelopes are shown on the 10-40 ms

post-stimulus interval. Since the comparison should take in account the different SNR

of the bursts belonging to two bands, the RMS of each envelope is calculated from the

prestimulus interval. A dotted line represents the confidence interval of three times RMS

following the envelope evolution. In the 1 Hz stimulation subject 1 and 4 show a more

sustained sigma burst activity, while for subject 2 and 3 the trend is still different but

less pronounced. In the 8 Hz stimulation such effect is partially lost in all subjects, but

still present in 1 and 4. In particular in subject 1 is observable a very strong late contri-

bution: the normalization here penalizes the first part of the sigma burst, which would

correspond to the kappa burst evolution as in all the other subjects, due to the high SNR

of the post-synaptic sigma contribution.

4.4. Discussion

The non-invasive detection of high frequency EEG up to 1 kHz SEP has been already

proved possible by means of low-noise technology [Scheer et al., 2011](, and different

spatiotemporal contributions of sigma and kappa bursts have been characterized [Fedele

et al., 2012]. The main issue related to the high-frequency SEP ripples is the weak power

78

4.4. Discussion

Figure 4.9.: Time envelopes for sigma (left) and kappa (right) range averages of thesources expressing similar dipolar spatial patterns are compared. The timeinterval considered is extended from the prestimulus baseline to to after burstbaseline. Envelopes are color-coded, according to the experimental stimula-tion rate, of 1 Hz (blue) and 8 Hz (red). computed as three times the standarddeviation of the baseline envelope, is also shown (light blue and magentadotted lines, for sigma and kappa, respectively). Results are shown for thedifferent subjects (rows) and experimental stimulation rate (columns).

79

4. Towards single trial resolution

Figure 4.10.: Time envelopes for sigma (blue) and kappa (red) range averages of thesources expressing similar dipolar spatial patterns are compared by theirnormalized amplitude envelopes. The confidence interval, computed asthree times the standard deviation of the baseline envelope, is also shown(light blue and magenta dotted lines, for sigma and kappa, respectively).Results are shown for the different subjects (rows) and experimental stimu-lation rate (columns)

80

4.4. Discussion

expressed at the level of the scalp, which requires a high number of repetitions, in order

to preserve the phase-locked components, otherwise hindered by noise. The non-invasive

detection of hf-SEP, relatively to the N20 and sigma band spectral ranges, is extensively

covered in the literature [Curio, 2005, Ozaki and Hashimoto, 2011]. Moreover, invasive

studies relate the phase-locked burst to cortical and subcortical spiking activity [Kloster-

mann et al., 2002, Baker et al., 2003, Hanajima et al., 2004]. This scenario describes the

hf-SEP as peculiar neurophysiological scenario, optimal to serve as test case to evaluate

the potential of combined low-noise technology and machine learning approaches in the

characterization of evoked responses in critical SNR conditions.

Three algorithmic approaches have been described, tested on simulated data set, at

different levels of SNR. The CSP has been applied in neuroscience to solve classification

problems [Koles et al., 1995, Ramoser et al., 2000, Parra et al., 2005, Blankertz et al.,

2008] in the traditional EEG spectral range. It relies on the power ratio expressed by

two dataset, measured by their covariance matrices. The CCA, developed by [Hotelling,

1992], has not been yet widely used in EEG/MEG research. Thus, the phase-locked

spike-like activity of the hf-SEP is a perfect candidate to benefit from the cross-covariance

optimization.

It is still object of discussion whether it could be possible to disentangle phase-reset

from added-energy generative model from non-invasive EEG/MEG data. The investiga-

tion of ensemble of single unit activity is precluded for non-invasive recording techniques

as EEG, given the limited available spatial resolution. However it is possible, given suf-

ficient SNR and prior knowledge on the nature of the source, to characterize the source

in terms of its macroscopic neurophysiological properties. In particular, in the case of the

sigma burst, it is known that spectrally and temporally confined oscillations reflect spik-

ing activity of underlying neural substrates [Baker et al., 2003,Telenczuk et al., 2011], and

with the low-noise recording system, single trial visibility can be achieved [Waterstraat et

al.,2014, under review.]. The biophysical framework showed that power ratio optimiza-

tion would fail for the SNR of interest (SNR<0.2), in case of phase-reset macroscopic

model, while CCA, which does not relate to a baseline, performs equally in both cases,

up to a limit imposed by the SNR. Even if real data features extraction demonstrates

equivalent CSP and CCA performance for the 600 Hz SEP, favoring the added-energy

macroscopic model, it is still not possible to infer a definitive conclusion: zero-energy of

lower energy pre-stimulus baseline could be equally provided by absence of activity or

81

4. Towards single trial resolution

local spatial averaging [Telenczuk et al., 2010]. Nevertheless, this bio-computational sce-

nario, featuring the detection of one narrow-band highly synchronized evoked response, is

useful to test the sensitivity of different algorithmic approaches, and in particular the po-

tentiality of CCAr, capable to retrieve the sources of interest, independently by generative

mechanisms and in highly demanding SNR conditions. Both simulations and experimen-

tal data results systematically confirm the reliability of CCAr, mathematically equivalent

to the already published xDawn [Rivet et al., 2009]. The version implemented here ex-

ploits the phase synchronization at single trial level with respect to the best estimation

of the covariance matrix of the evoked components, computed from the averaged signal.

This forces the decomposition to converge in terms of correlation on the true available

information, which allows the extraction of correct features even in demanding SNR con-

ditions. It has been shown how CCAr provides reliable spatial filters with less than 1000

trials, for all the spectral ranges under study, matching the expectation offered by simula-

tions. The mismatch remains between synthetic and real data analysis in the sigma range

for a limited number of trials: clearly the forward model, in the implementation proposed

here, is not capturing the complexity of the scenario, possibly given the presence of more

than one source. This would result in a noisy estimation of the covariance matrix for a

limited number of available trials, which would bias the diagonalization procedure.

As a general result, N20 and sigma can be visualized at single trial level, as shown

by the stripe-like pattern from Figure 4.3b and 4.4b. The kappa burst, instead, is still

not easily distinguishable from the noise level. However, this is expected, given the poor

SNR. CCA-based approaches decrease the noise contribution taking advantage of trial-to-

trial synchronization. In this respect, as for CSP, having source and noise different spatio-

temporal configurations, the extracted filters provide clear oscillatory temporal features,

but their prominence over the baseline is dictated by the SNR level.

The opportunity to access source space opens the way to further characterization: di-

verse contribution can be analyzed separately, as has been done for the bipolar compo-

nents relative to the investigated spectral ranges. In particular, sigma and kappa sources

have been compared with respect to space, time and response variability to different stim-

ulation rate. A comparison of the spatial information, by dipole fit on the spatial pattern,

has not outlined a clear difference in localization. This could be in part due to location of

true sources, as we expect to find sources belonging to a restricted neuronal pool active

along the somatosensory pathway, and in part due to to the limited number of subjects,

82

4.4. Discussion

and the non-optimality of the setup for source reconstruction purposes, which restrict the

statistical power of the data. Response to different stimulation rates has outlined pro-

gressive peaks latency, as well as a decrease on the later components, in both sigma and

kappa burst. This last is a well-known results in the case of 600 Hz ripple [Klostermann

et al., 1999], since it could prove the presence of pre- and post-synaptic components,

with respect to the N20 peak, marking the synaptic event in area 3b of the somatosen-

sory cortex. If the pre-synaptic burst contribution comes from ascending thalamo-cortical

fibers, the later part is possibly generated by intra-cortical GABA interneurons [Ozaki

and Hashimoto, 2011], whose refractory period is affected by the stimulation rate. The

same would be expected to happen for the 1 kHz components, as the time resolved am-

plitude envelope show: amplitude envelopes demonstrated different temporal evolution

of the sigma and kappa burst, suggesting the presence of different mechanisms even for

the generation of similar spatial patterns.

Moreover, trial-to-trial variability, for single and multiple patterns can now be investi-

gated. The early SEP is an important feature of the perceptual system, given its temporal

reliability. The thalamo-cortical system is involved in the regulation of arousal and at-

tention [Portas et al., 1998], and SEP variability can be related to transitions between

different brain states, such as phases of sleep, or variations in the levels of consciousness,

attention, expectation and learning [Li et al., 1999, Rosanova and Timofeev, 2005, Fox

et al., 2006, Fontanini and Katz, 2008]. In the present study EEG recordings were per-

formed during awake state, instructing the subject to keep the eyes open, but non-invasive

experiments performed in healthy human subjects [Gobbelé et al., 2000, Halboni et al.,

2000] showed that the amplitude of scalp hf-EEG is sensitive to specific neural conditions

preceding the stimulus, such as sleep phases and fluctuating vigilance or attention. Inves-

tigation of correlated variability would then be possible combining different experimental

protocols, low-noise technology, possibly complemented by new optimization techniques

for cross-band analysis [Dähne et al., 2013].

In conclusion, a methodological framework for the source characterization of high fre-

quency somatosensory evoked responses has been presented. Our approach allows ex-

traction of single trials in source space as tested with simulated and experimental data

set. The use of low-noise EEG technology allows the detection of spike-like burst fea-

tures up and above 1 kHz, and as small as tens of nV over the scalp. The methodological

framework described, showing the improvement in the detection of weak evoked poten-

83

4. Towards single trial resolution

tials, is not limited to the sigma or kappa band range, but can also be applied to other

scenarios with low SNR phase-locked responses.

84

5. Low-noise bio-amplifier technology

Electroencephalographic potentials of the range of fractions of µV can be measured non-

invasively from the human scalp. Devices designed for this purpose go under the name

of electroencephalographs. They are typically characterized by an electronic amplifier

for bio-potential, named bio-amplifiers, connected at the input to the electrodes applied

to the scalp and at the output to an Analog to Digital Converter (ADC), which allows

the data storage in a memory card or in PC disk for further off line analysis. The bio-

physics that explains the nature, the propagation and the detectability of neural sources

have been already introduced. In this chapter the focus is on the technological aspects

developed in order to realize a low-noise, high performing portable system, for the non-

invasive detection of low and high frequency EEG potentials (from near DC to 2 kHz).

The chapter is divided in five sections: in the first section a brief historical overview of

the EEG system design, main issues and solutions regarding environmental interference

are outlined; in the second part the criticality of the low-noise input stage for the de-

tection of high frequency component is discussed; in the third part a new bio-amplifier

design for high-frequency EEG (HF-EEG) is proposed, from the analytical model to the

practical realization; in the fourth a set of performed measurements is presented, in order

to demonstrate the achievable data quality, and finally the potential of the technological

progress is discussed. Recordings reported here have been performed in PTB, Berlin

(recordings V, A.1) and in Charite, Campus Benjamin Franklin, Berlin (recordings VI,

A.1). For the electromagnetic interference theoretical model and hardware implemen-

tation I acknowledge the fundamental contribution of Hans-Juergen Scheer and Frank

Petsche from PTB, Berlin.

5.1. Introduction

Since the first EEG measurement published by Berger in 1929 [Berger, 1938], a lot has

been developed to bring the neurophysiological non-invasive scalp recording as a gold

85

5. Low-noise bio-amplifier technology

standard of the clinical practice [Goldensohn, 1998]. Few years later Fisher and Lowen-

back demonstrated the first spikes in EEG recordings due to epileptic events. The first

EEG laboratory was founded at Massachusetts general hospital in 1936, but only in 1958

a committee led by Jasper standardized the 10/20 system for electrodes spatial positions

on the scalp. In the 60 data processing became prevalent within EEG signal processing,

and with the advent of digital EEG technology, the opportunity to realize topographic

mapping made EEG popular in a variety of clinical fields. From the 90 clinical practice

and cognitive neuroscience research could already count on real time monitoring sys-

tems. Today, most EEG machines rely on the same basic principles, differing largely

only in the software produced by EEG manufactures. As seen in US Patent No. 7336991,

many, if not all current devices are comprised of a largely digital system [Yanagihara,

K., Ishikawa, T., Imajo, 2008]. The electrodes are connected to the patient, and then feed

directly into a preamplifier, implemented on the device main board (passive electrodes) or

directly in proximity of the electrode ring (active electrodes). The amplified multichan-

nel signal is converted in digital format by a fast ADC, and conveyed to software, which

manages the online visualization and processing, as well as the data storage.

5.1.1. Single Ended versus Differential Inputs

The analog input stage can be designed in different ways. For bio-amplifiers these designs

are classified in two main types: differential input and single ended input. In order to

provide an intuitive explanation, let us consider a three point measurement scheme, with

the main lead (EEG channel), the reference, and the ground applied on the body, or the

scalp. The output of the preamplifier, derived with respect to the ground potential, is sent

to the ADC.

A single ended input measures the voltage between the EEG channel and reference

electrode, which is directly connected to the amplifier internal ground (Figure 5.1a). A

differential input measures the voltage between two individual inputs (EEG channel and

reference), both referred to the amplifier ground potential (Figure 5.1b). This second

configuration requires a more complex design and larger number of components, but

has been preferred for long time, because of its intrinsic theoretical independence from

environmental sources of interference. The target of this work is to minimize the effect

of the interference as well as the noise figure of the device at the input, which clearly

depends on the chosen configuration. For reasons that will be quantitatively addressed

86

5.1. Introduction

Figure 5.1.: Amplifier input configurations. a) Single ended (SE), with reference elec-trode connected to the circuit ground. b) Differential input, where the inputsare floating with respect to circuit ground.

in the following, the single ended design has been preferred in the final version of the

HF-EEG amplifier.

5.1.2. Common Mode and Isolation Mode

The target of the preamplifier is the detection of the potential emitted by the source of

interest, contained in the difference between the two inputs (EEG channel and reference),

defined as Differential Mode. The electromagnetic coupling between human body, leads,

and the environmental electromagnetic fields generates an additional potential not belong-

ing to the source of interest, but detected by the recording system, defined as Common

Mode. When the amplifier ground is floating with respect to the earth ground, another

contribution of interference is given by the Isolation Mode, constituted by the poten-

tial between the amplifier internal ground and the earth ground. The goodness of a de-

sign [Nagel, 2000] is defined in terms of

- Common Mode Rejection Ratio (CMRR): The ratio between the amplitude of a com-

mon mode signal and the amplitude of a differential signal that would produce the same

output amplitude or as the ratio of the differential gain over the common-mode gain:

CMRR =GD

GCM(5.1)

87

5. Low-noise bio-amplifier technology

- Isolation Mode Rejection Ratio (IMRR): The ratio between the isolation voltage,

VISO, and the amplitude of the isolation signal appearing at the output of the isolation am-

plifier, or as isolation voltage divided by output voltage VOUT in the absence of differential

and common mode signal:

IMRR =VISO

VOUT(5.2)

5.1.3. Interference model

The first complete review in the field was given by [Huhta and Webster, 1973], who

analyzed power line interference in three-electrode ECG recordings with grounded am-

plifiers, with differential input stage. They isolated four main contributions on signal

path, introducing a fundamental model for future developments:

1) magnetic induction in input leads;

2) displacement currents in those leads;

3) displacement currents in the body;

4) common-mode voltage contribution because of the amplifiers limited common-

mode rejection ratio (CMRR) compounded by electrodes and common-mode input impedance

imbalance.

The complete model is explained in Figure 5.2 (adapted by [Wood et al., 1995])

And the environmental electromagnetic interferences described by equation 5.2.

Vn = 2π f SB︸ ︷︷ ︸magnetic induction

+ (Id1Ze1− Id2Ze2)︸ ︷︷ ︸displacement current in leads

+ IbZt︸︷︷︸displacement current intissue

+ IbZr︸︷︷︸CM potential

1CMRR︸ ︷︷ ︸

commonmodere jectionratio

+ZCM1

ZCM1 +Ze1− ZCM2

ZCM2 +Ze2︸ ︷︷ ︸potential divider e f f ect

(5.3)

The first two contributions can be limited respectively by twisting the electrodes lead,

reducing loops area offered to the magnetic flux, and shielding the leads, to minimize ca-

pacitive coupling. The third is luckily not an issue in EEG, since the electrodes are close

enough to neglect body displacement currents. The Common Mode reduction was partly

88

5.1. Introduction

Figure 5.2.: Biophysical model scheme of the power line interference on a physiologi-cal measurement with a standard differential input bioamplfier. Stray capaci-tances describe the coupling of: human body to power line (Cp), Earth ground(Cb); leads to power line (Cd1, Cd2); leads shields to power line(C1, C2); leadsto guarding (Cg1, Cg2, Cg3), recording system power line (Camp) and earthgreound(Ciso). Impedances represent electrical properties of the tissue (Zt),the electrodes (Ze1, Ze2, Zr), amplifier input (Z1, Z2), and model differen-tial mode (ZDM), commn mode (ZCM), and isolation mode (ZISO). Currentscoupled to the body (Ib) and to the leads (Id1, Id2) are responsible for theCommon Mode.

89

5. Low-noise bio-amplifier technology

solved by the Driven Right Leg circuit, introduced by [Winter and Webster, 1983], and

applicable to isolated and non-isolated amplifiers. They proposed to reduce the interfer-

ence by increasing the amplifiers effective CMRR or by reducing the common-mode volt-

age, for example by increasing the isolation impedance. The role of the Isolation mode

rejection was modeled by [Pallás-Areny, 1988], who analytically compared the charac-

teristics of isolated and non-isolated configuration scheme, concluding that isolated am-

plifiers, needed to ensure patient safety, only help in interference reduction if their IMRR

is high enough. The role of capacitive coupling and displacement currents coupled to

amplifier common was further investigated [Metting van Rijn et al., 1990, Wood et al.,

1995]. Nowadays the isolated configuration is part of the standard practice in commer-

cial amplifiers design, since isolation can be easily achieved in three different modalities:

transformer isolation, capacitor isolation, and optic isolation. This last one is usually

implemented in series to the ADC [MettingVanRijn et al., 1993], in a system design in-

clusive of battery power supply, and offering a potential infinite isolation with respect

to ground earth. Isolated systems development is directed towards low consumption,

portable devices, providing high performance in terms of interference rejection together

with durable monitoring capacity [Yazicioglu et al., 2008]. The possibility to use digital

integrated programmable devices allows the implementation of more precise solutions, as

a Digital Driven Right Leg feedback loop, optimized for the frequency of interest [Haber-

man and Spinelli, 2010]. In this case it is possible even to reconsider the complexity of

a differential input stage and obtain the same performances in terms of CMRR with a

single-ended input [Haberman and Spinelli, 2012].

5.1.4. Low-noise bio-amplifier: biophysical model and design

An EEG system has to provide acceptable Signal-to-Noise Ratio (SNR), defining as Sig-

nal the potential expressed on the scalp by the biological source of interest, and as Noise

any other contribution added to the signal. In the previous paragraph the environmental

interference from the power line has been discussed, together with its relation to the com-

mon mode voltage at the input. Here we assume to be immune to this contribution (as for

example in an electromagnetically shielded room), and focus on the parameters influenc-

ing the design of a low noise figure amplifier, critical for the detection of neurophysiolog-

ical features in the 300-2000 Hz spectral range. An indicative description of the physical

contribution for this spectral domain has been provided by [Scheer et al., 2006]: theoreti-

90

5.1. Introduction

Figure 5.3.: Noise model of a bio-potential measurement with bioelectrical ebio, interfaceeint and amplifier noise sources, signal of interest Sig. Bioelectrical noisecomprises all background noise generated by the investigated subject, in EEGmainly that of the cortex. Noise generated at the GND electrode appears asa common mode signal and is rejected by the differential amplifier (from[Scheer et al., 2006], with permission).

cal considerations are compared to experimental EEG recordings, performed with devices

presenting different noise levels at the input. Investigating the contributions from techni-

cal and biological sources of noise in human scalp recordings during relaxed condition,

the authors were able to disentangle three main components, as illustrated in Figure 5.3:

1. Bioelectric sources, ebio

2. Interface at the skin, eint

3. Amplifier input noise, eamp

These three contributions can be reasonably considered statistically independent and,

quadratically summed together, in order to have a first approximation of the noise level at

the input of the measurement chain [Netzer, 1981],

91

5. Low-noise bio-amplifier technology

eTOT =√

e2bio + e2

int + e2amp (5.4)

where the spectral densities are given in terms of V/√

Hz. The bioelectric sources

express the biological background contributions. They are characterized by spectral 1/f

trend, which sinks, at high frequencies, into the flat system noise level. In the case of

evoked responses, which are phase- locked to a triggered stimulus, it is possible to average

out the background noise, and extract the signal of interest with a certain number of

trials. Statistical properties of the multivariate signal can be further exploited by machine

learning based approaches [Blankertz et al., 2008, Celka et al., 2008]. Interface noise

relates to the electrode gel-skin interface. Considering the value of the impedance at the

interface, it is possible to compute the noise contribution in terms of thermal noise by the

Nyquist-Johnson relation,

eint =√

4kT Re[Z] (5.5)

where k is the Boltzmann constant and Re(z) is the real part of the impedance at the

absolute temperature T. In terms of the parameters of Figure fig:Noise model, Z equals

the inter-electrode impedance Zint = Z1 + Z2 at the body surface temperature T ≈ 310

K. In general, it is important to remember that 1 kΩ equals 4 nV/√

Hz. In the ampli-

fier, the noise is generated by thermally activated electronic fluctuations inherent to the

physical nature of the components. It can be modelled by one voltage source and two

current sources, applied to the adopted configuration. The experimental counterpart of

this theoretical biophysical framework in terms of power spectral densities is reported

in Figure 5.4, for scalp EEG recordings obtained respectively with a commercial EEG

and an in-house-made dedicated low-noise EEG; the input shortcut spectral density of

the low-noise amplifier it is depicted (Figure 5.4, upper panel) and typical values are

indicated as dotted lines on physiological recording spectral representation (Figure 5.4,

lower panel) Notice that every single point observed in the graph is characterized by the

three contributions outlined above. In the case to a commercial amplifier (noise level

30 nV/√

Hz, as indicated by the AAEM - American Association of Electro-diagnostics

Medicine), it is possible to observe the 1/f spectral trend, typical of neurophysiological

measurement [Buzsaki and Draguhn, 2004], until a few hundreds Hz, where the estimated

92

5.1. Introduction

spectral power becomes flat (dashed line). At this point the background bioelectric noise

is smaller than the system noise. Skin abrasion and lower input noise amplifier allow

the access to biological contributions at higher frequencies, typically from 400 to more

than 1000 Hz, to encounter again the noise floor of the system (notice that the spectral

estimation is above the input shortcut level because of the impedance contribution).

For the measurement of ultra-fast EEG signals the amplifier noise should be one half

of or less than the thermal noise of the inter-electrode resistance [Scheer et al., 2006]. In

this case, noise introduced by the amplifier is less than 12% of the total. Without skin

preparation the real part of inter-electrode impedance is in the order of about 25 kΩ, so

that an amplifier noise of 10 nV/√

Hz is sufficiently low. On the other hand, when skin

abrasion or puncturing is applied, resistance of around 1 kΩ can be achieved and the

amplifier noise should be below 2 nV/√

Hz. This model is fundamental to understand

the criticality of a low-noise system, and the needed improvement in terms of detection

power with respect to higher frequency contributions, otherwise hindered in the noise of

the measurement system. In the low frequency range from 1 to 100 Hz, bioelectric noise

is, by far, the dominating factor, while at higher frequencies, amplifier and thermal noise

are the primary sources of noise. The thermal noise can be reduced by careful impedance

preparation. Analog electronics components available on the market provide system noise

level 5 to 6 times lower than AAEM requirements.

5.1.5. Noise calculation from the input stage

In Figure 5.5a the scheme for the non-inverting single ended configuration is described.

The signal is represented as a voltage source Vs with its output impedance Rs , which

corresponds to the electrodes impedance in the case of physiological recordings. R1 and

R2 are electronic components determining the gain of the design ( G = 1+ R1 / R2). The

total noise is calculated considering Vs = 0, and taking into account the op-amp current and

voltage input noise, in and en, the thermal noise of the impedance R1 and R2 (Figure 5.5b).

Two computational are reported, one for a FET input operational amplifier (op-amp),

and one for a bipolar input op-amp. Data sheet values for in and en and computation

results are reported in Table5.1. The two components are comparable in terms of perfor-

mance. The first one was implemented in the amplifier used in [Scheer et al., 2006]). The

second was chosen for the new design implementation, since it is available in Surface

Montage package, more suitable for a compact multichannel design.

93

5. Low-noise bio-amplifier technology

Figure 5.4.: Typical measured voltage spectral densities related to amplifier inputs (dif-ferent FFT data are used below and above 100 Hz). Top: noise floor of thehigh and low noise amplifier with inputs connected to the ground terminal.Bottom: relaxed EEG simultaneously recorded with the same amplifiers inparallel at C3; dashed lines show the noise floor of the two amplifiers

94

5.1. Introduction

Figure 5.5.: Noise calculation model for the analog input stage. (a) Drawing for the non-inverting single ended input configuration; (b) mathematical framework forthe noise level estimation at the output of the analog input stage. k is theBoltzmann constant and T the temperature in Kelvin.

Datasheet AD743 (FET) OPA2078-Pin SO-8

Vn [nV/√

Hz] 2.9 2.2

In [fA/√

Hz] 6.7 500

Total input noise (input shortcut) 3.0 2.3[nV/

√Hz]

Total input noise (Z = 1 kΩ) 5.1 4.8[nV/

√Hz]

Table 5.1.: Commercial data sheet values for AD743 and OPA207 and resulting total in-put noise for the configuration described in Figure 5.5

95

5. Low-noise bio-amplifier technology

5.2. The new low-noise amplifier design

The main problems and proposed solutions have led to envisage two main requirements

arising from the biophysical scenario: the rejection of environmental interference, and

high sensitivity to the biological target activity, provided by low noise input stage. Here

the concept, the design and realization of a setup embodying the required specifications

is described. In general, the measurement system chain can characterized as follows: the

physiological signal enters the measurement chain at the level of the analog input stage;

the amplified signal feeds an ADC and the digital data stream is delivered by optical link

to a storage unit. The compact design and the digital communication allow to neglect

any environmental noise contribution on the later sections of the system. Therefore the

focus first on the physical framework of the interference rejection at the level analog input

stage; afterwards, the integration of the different units is described.

5.2.1. Electro-technical model for the interference estimation

A general scheme of the physiological measurement physical framework is shown in

Figure 5.6. The Power Line interference Vpl is represented as a 230 V - 50 Hz signal

generator. The Body and the Amplifier Case are coupled to the environment by four

capacitances:

- Cpb , capacitance between power line and body (typical value 100-300 pF);

- Cbe, capacitance between body and earth (typical value 1-3 pF);

- Cps, capacitance between power line and amplifier case (typical value «1 pF);

- Cse, capacitance between amplifier case and earth (typical value 20-30 pF;

- Cgs, capacitance between amplifier internal ground and case (typical value 200 pF-1

nF);

The singled ended input stage is connected to the body through Zel , the active lead

impedance, while Zbg is the impedance of the reference electrode, connected to the am-

plifier internal ground. A further lead is the ground electrode, Zbs, which is referred to

the amplifier case and separated from the amplifier ground by a capacitance Cgs. Since

presenting a low-noise design, we consider an impedance of Z ≈ 1 kΩ confined by practi-

cal limitations of the electrode preparation procedure together with a limited contribution

of Johnson-Nyquist noise contribution, to result in a noise level as low as 4 nV/√

Hz.

All the interference current flows through Zbg, as the op-amp input stage presents a high

96

5.2. The new low-noise amplifier design

Figure 5.6.: Biophysical model of Body-Amplifier coupling to electric power line inter-ference. The body is coupled to the power line active lead by Cpb and to Earthground by Cbg. The amplifier case is coupled to the power line active leadby Cps and to Earth ground by Cse. Body and amplifier internal ground areconnected through reference electrode with impedance Zbg, while Body andamplifier case are connected through ground electrode with impedance Zbs.Typical values are reported

input impedance. To measure the interference mode, the voltage drop across Zbg, with

respect to Vpl , needs to be calculated. In order to simplify the analytical derivation, the

circuit can be rearranged as indicated in Figure 5.7. In the following we consider elec-

trode impedances purely resistive, which is acceptable in the frequency range of interest

(up to 2 kHz), even if they are labelled as impedances. The Body is reduced to the electric

point A, while the amplifier case potential is electric point B.

This circuit can be modeled as a Wheatstone Bridge, with the Ratio A = Cpb/Cbe and

Ratio B = Cps/Cse balancing the voltages at points A and B, respectively, as shown in

Figure 5.6. This configuration is referred in the following as the Capacitive Bridge. If

the ratios A and B perfectly match, in consequence there would neither be current flowing

between points A to B, nor interference effects Vn. Even if this cannot be practically

realized, Ratio A and Ratio B can match close enough to reduce Vn contribution below the

sensitivity of the recording system.

97

5. Low-noise bio-amplifier technology

Figure 5.7.: Equivalent schematic of the electro-technical model of Figure 5.6. Point Aand B represent respectively body and amplifier case potential.

To minimize Vn, the current flow through Zbg has to be limited.

The circuit analysis can be simplified in two steps, starting from the schematic reported

in Figure 5.7:

Step1. Applying Thevenins theorem [Helmholtz, 1853, Thévenin, 1883] between pin

A and B, obtaining an equivalent voltage source with its output impedance (Figure 5.8):

Vth1 =Vp

(Zbe

Zbe +Zpb− Zse

Zse +Zps

)(5.6)

Zth1 =ZbeZpb

Zbe +Zpb+

ZseZps

Zse +Zps(5.7)

Step2. Applying Thevenins theorem again on Zbs, we can further simplify, as in Fig-

ure 5.9:

Vth2 =Vth1Zbs

Zth1 +Zbs(5.8)

98

5.2. The new low-noise amplifier design

Figure 5.8.: Equivalent schematic of the electro-technical model of Figure 5.6, after thefirst Thevenin simplification step.

Figure 5.9.: Equivalent schematic of the electro-technical model of Figure 5.6, after thesecond Thevenin simplification step.

99

5. Low-noise bio-amplifier technology

Zth2 =Zth1Zbs

Zth1 +Zbs(5.9)

Taken together, Vn can be expressed in terms of circuital and environmental parameters

as

Vn =Vth2Zbg

Zth2 +Zbg +Zgs

=Vth1Zbs

Zth1 +Zbs

ZbgZth1Zbs

Zth1 +Zbs+Zbg +Zgs

=Vp

(Zbe

Zbe +Zpb− Zse

Zse +Zps

)Zbs

ZbeZpb

Zbe +Zpb+

ZseZps

Zse +Zps+Zbs

ZbgZbeZpb

Zbe +ZpbZbs +

ZseZps

Zse +ZpsZbs

ZbeZpb

Zbe +Zpb+

ZseZps

Zse +Zps+Zbs

+Zbg +Zgs

(5.10)

Since Cbe » Cse, Cps,Cpb we can simplify Zth1 as follows

Zth1 =1

Cbe +Cpb+

1Cse +Cps

=Cse +Cps +Cbe +Cpb

CpsCpb +CpsCbe +CseCpb +CseCbe≈ Cbe

CbeCse(5.11)

and considering also that

Zbs <<1

Cse

Zgs >> Zbg

Zgs >>1

Zth1

(5.12)

100

5.2. The new low-noise amplifier design

In the end we can write:

Vn =Vp

(Zbe

Zbe +Zpb− Zse

Zse +Zps

)|ω|CseZbs |ω|CgsZbg (5.13)

Equation 5.13 provides a quantitative characterization of the main parameters describ-

ing model. Cgs strictly depends on the board surface and design; the optic isolation lowers

consistently Cse; Zbs and Zbg can be reduced during the preparation. The term into the

brackets refers to the capacitive bridge. Since it is practically impossible to control the

value of each single coupling, the focus is on the two ratios, the first relative to the body

and the second to the amplifier case.

5.2.2. Stray capacitances estimation

In order to estimate the influence of the coupling on the interference present at the input,

it is necessary to provide a value for the four stray capacitances composing the bridge

outlined in Figure 5.6. Methods of coupling capacitance measurement are available in the

literature, offering a sufficient sensitivity in the case of power line-to-body capacitance

estimation (few pF). Here, however, we deal with values as low as fractions of pF, as

present in small battery-powered devices, and more importantly we refer to the ratio

rather than their absolute values, increasing the criticality of a correct estimation. For

this reason, here, we implement a recently proposed approach [Haberman et al., 2011],

more suitable for low capacitance values, which has been implemented to characterize

the measurement set up described in the following.

Briefly, the electric-field power line interference is described in Figure 5.10a. The two

capacitance CP and CB represent the coupling of point A to the power line cord and to

the Earth ground, respectively. After inserting a load impedance RL in parallel to CB, the

voltage between point A and ground is given by

VL =VPL2π fPLCPRL√

1+(2π fPL(CB +CP)R2)2 (5.14)

where VPL is the power line voltage and fPL the power line frequency. Assuming RL

variable, we can calculate the limit for its extreme upper and lower values, and derive

asymptotical approximation of VL,

101

5. Low-noise bio-amplifier technology

Figure 5.10.: Stray Capacitance measurement. (a) General electrical scheme. (b) Experi-mental setup. Adapted from [Haberman et al., 2011]

f or RL→ 0

VL ∼=VPL2π fPLCPRL

(5.15)

f or RL→ ∞

VL ∼=VPL1

1+ CBCP

(5.16)

Choosing small RL values it is possible to extract the value for CP from equation 5.15.

Afterwards, implementing high load impedances, also CB can be measured, by inverting

equation 5.16. The experimental set-up for the parametric stray capacitance estimation is

presented in Figure 5.14. The input of this circuit is connected between point A and Earth

ground, and its input impedance represents our RL. Since very low and very high load

impedance need to be available, a bootstrap configuration is used. Resistors R1 and R3 are

fixed to 5.6 MΩ and 100 kΩ, respectively, while R2 values are allowed to vary between

10 kΩ and 10 MΩ, as reported in Table 5.2(left column). With the ground is connected

102

5.3. Set up implementation

R2 RL (switch low) RL (switch high)11 kΩ 11 kΩ 6.2 MΩ

56 kΩ 56 kΩ 8.8 MΩ

91 kΩ 91 kΩ 57.8 MΩ

1.7 MΩ 1.7 MΩ 108 MΩ

10 MΩ 10 MΩ 571 MΩ

20 MΩ 20 MΩ 1.16 GΩ

Table 5.2.: Input impedance of the stray capacitance measurement setup. The RL valueis controlled by two switches. The first selects among different R2 values,the second connects the input to bootstrap circuit, providing the input loadimpedance given by equation 5.17.

is on the Low pin, RL = R2. With the switch on the High pin, the corresponding circuit

input impedance is given by

RL = R1 +R2 +R1R2

R3(5.17)

allowing RL to reach values as high as 1 GΩ (Table 5.2, right column). The circuit is

implemented using the operational amplifier OPA129 of Texas Instruments, characterized

by an ultra high input impedance and low bias current, critical for loads in the 100 MΩ

â 1 GΩ range. Shielded-cable capacitances are minimized by shield driver, connected to

the negative input of the op-amp. The collected values for VL were fitted by Minimum

Least Square estimation in low and RL ranges, to compute CP and CB values.

The configuration of the noise sources can be more complex in a realistic scenario, as

introduced by equation 5.3; in the following we present a set of measurements, performed

in different capacitive bridge unbalance and the electromagnetic interference conditions.

5.3. Set up implementation

5.3.1. First single channel prototype

The new design concept was implemented in its first version in a single channel proto-

type. The analog stage presents one single-ended input with electronic noise level of 2.4

nV/√

Hz. The analog output is connected to 24-Bit-Audio-ADC (sampling frequency 48

103

5. Low-noise bio-amplifier technology

kHz) with digital optical S/PDIF output (AAD24). The digital stream is delivered to an

acquisition PC in USB format through a Creative Sound Blaster X-Fi HD. The soundcard

is equipped with a 24 bit DAC, providing an analog copy of the measured signal. The

recording setup is run by AVS Audio Editor Software (freeware version). The analog

stage and the ADC is placed in a metallic case, which is connected to the EEG ground

electrode during recordings. The case presents input connectors for the electrodes, and

accommodates guides for the optic fibers at the output. This preliminary version was

used to test the interference rejection and design performance inside and outside an elec-

tromagnetically shielded environment.

5.3.2. HF-EEG amplifier

The new low-noise high performing design was integrated in a battery powered portable

system, referred in the following as the HF-EEG amplifier. It consists of 3 main physical

components: the electrodes box, the main box, the optic-USB converter. The Block

Diagram of the complete design is provided in Figure 5.11.

The electrodes box has 10 input pins for the 8 input channels, 1 reference and 1 ground

electrodes, and is connected to a male D-connector through a gray 10 wires flat cable

(Figure 5.12).

The main box contains the analog input stage, ADC and µController, for the manage-

ment of the on board functionalities and the communication with user’s software. The

analog input consist of 8 single ended low noise preamplifiers (OPA209: 2.3 nV/√

Hz,

Gain = 100) feeding 8 single-ended buffers with differential output. The analog EEG

signal is collected by 8 sigma-delta 24-bit Analog-to-Digital converters (ADS1278) pro-

viding a digital output stream in SPI format, sampled at 10 kHz. The µController collects

the data stream, adding to it the build-in trigger level bits (5 V TTL, 4 Hz, duty cycle

= 0.032) and manages the transmission to the Optical to Digital Converter. The unit is

equipped with an impedance measurement circuit (not shown in the diagram). The ele-

ments described are implemented on a two level board. The box offers on the front panel

the D-connector for the electrodes box or for the battery charger and on the rear panel the

on/off switch button, a green light led indicating the status of the device a red light inter-

nal led, the optic cable connector (Figure 5.12). When the button is pressed a power-up

unit (Attiny 25) is activated, and power supply is delivered from the batteries to the active

elements through a series of voltage regulators.

104

5.3. Set up implementation

Figure 5.11.: HF-EEG amplifier design block diagram.

105

5. Low-noise bio-amplifier technology

The optic-USB converter has as input a data optical link (Tx/Rx), a trigger optical link

(Rx), and two digital outputs, an USB port to Laptop and a Trigger Digital signal to drive

the trigger generator (Figure 5.12). The software communication is realized by virtual

COM port. From the USB converter the data stream is finally stored in a binary file on

the laptop hard disk.

The setup can be controlled by laptop or workstation by a dedicated made-in-house

software, which was implemented in the Labview Platform in PTB. The User interface

allows data stream visualization and recordings, on-line FFT, impedance measurement,

and digital trigger toggling.

5.4. EEG Recordings

Here we present the measurements performed at different stages of the hardware develop-

ment. The focus is kept on the advantages offered by the proposed system in comparison

to an available low-noise set up; afterwards, the HF-EEG amplifier usability for neuro-

physiological data collection is demonstrated.

5.4.1. Inside/outside of the shielded room

In previous studies, the high-frequency EEG recordings have been collected by low-noise

setup explicitly designed to be operated an electromagnetically shielded environment.

Here we compare the performance between this previous system and the implementa-

tion of the new design, demonstrating the improvement in interference rejection while

maintaining a low- noise input front-end (recordings V, A.1)

Settings

Using a fabric EEG-cap with holders at positions according to the 10/20 system, sin-

tered Ag/AgCl ring electrodes were placed using a fabric cap with holders at positions

according to the 10/20 system on the scalp, with the active lead aton F3, the reference

aton C3’ and the ground aton TP7. Impedances were kept below 2 kΩ.

EEG measurements were performed with two systems: the already available PTB

low-noise setup specifically designed to be operated in an electromagnetically shielded

room [Scheer et al., 2006], and the first prototype of HF-EEG amplifier, described in

paragraph 5.3.2. Recordings were performed in relaxed and task-free condition for 5

106

5.4. EEG Recordings

Figure 5.12.: The HF-EEG amplifier components: the electrodes box, the main box, theoptic-USB converter.

107

5. Low-noise bio-amplifier technology

minutes, inside and outside the electrically and magnetically shielded room (Vakuum-

schmelze AK3b). The presence of power line interference was evaluated by Power Spec-

trum estimation, monitored online by connecting the output of the two devices to Digital

Spectrum Analyzer.

Results

A comparison of Power Spectra is provided in Figure 5.13. Recordings inside the

shielded room are equivalently free of interference, while when the electromagnetic en-

vironmental shield is absent, the new design outperforms on the previous system. This is

observed for the 50 Hz power line interference, and, more importantly for the purpose of

this work, for the high-frequency range (500-1000 Hz). Moreover, a slight improvement

of the noise level in the high-frequency range can be observeved for the new system, as

is expected by due to the single-ended configuration, which halves the electronic noise

contribution of the analog input stage.

5.4.2. Neurophysiological SEP recordings in a clinical environment

Settings

Eight sintered Ag/AgCl ring electrodes were placed on the scalp coverwith the goal of

optimal coverage of the area above the left hemispheric somatosensory cortex. Reference

and ground electrode were placed at the right and left wing of the noise. As required,

impedances were kept below 1 kΩ (recordings VI, A.1)

In order to prove the importance of the position of the amplifier main box in a real case

scenario, recordings were performed in two distinct conditions: in one case the amplifier

box is placed about 1 m apart from the subject. In the second case it is secured on the

subject chest, avoiding any contact between the amplifier shield and the skin. In this

second case the stray capacitances bridge is expected to be almost balanced.

Using the circuit and the procedure described in paragraph 5.2.2, an estimation of the

elements of the capacitive bridge was obtained: point A shown in Figure 5.10a corre-

sponds alternatively to the body and to the amplifier case. Intuitively, body and amplifier

related capacitances depend on their exposed surface, relative distance and orientation

towards environmental sources of interference. Given the difference in terms of the ex-

posed surfaces of body and amplifier case it would be highly demanding to exactly match

their values, however, the main interest here is to match the individual ratios between the

108

5.4. EEG Recordings

Figure 5.13.: Spectral comparision of different measurement systems inside ans outsidethe shielded room (SR) in relaxed and task-free EEG recordings. Full-bandPower spectrum (left). Zoom on 50 Hz and relative harmonics in the HFspectral range (right).

109

5. Low-noise bio-amplifier technology

coupling to power line and earth ground. This can empirically be achieved by putting the

case as close as possible to the body, so that both of them are similarly influenced by the

electromagnetic interference, even if at different absolute levels.

It is worth noticing that the electro-technical model described in paragraph 5.1.4 ac-

counts for stray capacitances to the body and the amplifier case. Moreover, other factor

might play a prominent role as magnetic field coupled to the wires loops and the power

line coupling to the leads. It is very difficult to theoretically estimate these factors. In this

respect we minimize their contribution by twisting the electrodes cables, and by shielding

the cables bundles, connecting the shield to the amplifier ground electrode, and therefore

to the amplifier main box case. This does not significantly influences the ground elec-

trode impedance, but could increase the capacitance between case potential and amplifier

internal ground, named in the model as Cgs. This parameter does not overcome the value

of 1 nF for the set up used. Moreover, the balanced C bridge minimizes also the conse-

quences of the power-line to lead coupling. SEP elcited by right median nerve stimulation

(4 Hz stimulation frequency, 1.5xmotor threshold = 8.5 mA, N = 6000 trials) were also

recorded in the case of amplifier adjacent to the subject’s body. The measurement were

performed in a clinical environment, particularly aggressive in terms of electromagnetic

power line interference. The subject was placed at the center of the room, keeping a

minimal distance of 0.5 meter from walls and electrical devices.

The recordings in the task-free relaxed condition were evaluated by estimating the

power spectral density (Welch’s algorithm, Matlab). SEP data were averaged and pro-

jected into the time-frequency domain by the Stockwell transform [Stockwell et al.,

1996], Bands of interest for further analysis were selected by normalizing the amplitudes

in each frequency band by a pre-stimulus baseline.

Results

In Table 5.3 the results of three sets of measurements are reported: the estimated stray

capacitances relative to the subject body, to the amplifier case put at 1 m from the body,

and to the amplifier case placed close to the subject. The values of the electrical parame-

ters from equation 5.10 are assigned as follows: Vp = 230VRMS;Zbs = 1kΩ; Zbg = 1kΩ;

Cgs = 1000pF.

Capacitance Ratios are reported in Table 5.4: the two different spatial configurations

for the amplifier case position lead to a difference by a factor of 20 in the coupling ratios.

By plugging these values in equation 5.13, it is demonstrated that a careful position-

110

5.4. EEG Recordings

Estimated Stray Capacitance [pF]Body Amplifier Case (Far) Amplifier Case (Close)

Cpb Cbe Cps Cse Cps Cse

4.89E-2 1.60E+2 3.25E-1 4.15E+1 6.06E-3 1.67E+1

Table 5.3.: Estimated stray capacitance for the subject body, and the amplifier case, placedat 1 m (far), and in proximity (close) to the subject.

Cpowerline/Cearth RatioBody Ampl. Far Ampl. Close

0.000305 0.007834 0.000362

Table 5.4.: Capacitance ratio for the subject’s body, the amplifier case placed at 1 m fromthe subject (Far), and the amplifier case placed in proximity of the subject(Close).

ing of the amplifier case lowers the interference by more than two orders of magnitude

(Table 5.5). This theoretical evaluation demonstrates how the optimized position of the

amplifier case with respect to the subject body improves the interference rejection up to

a factor of 50 dB. In order to achieve this performance in practice, apart from optimiz-

ing the amplifier case location, the impedances of the reference and ground electrodes

are important, as indicated by equation 5.13. The first must be low to minimize the ef-

fect of capacitive currents in the leads, while the second offers a discharge path to any

interference contribution, in parallel to the signal path.

Interference rejection of the HF-EEG amplifier final implementation is reported in

Figure 5.14. One representative bipolar channel derivation (FC3-CP3) is represented in

the spectral domain, showing the difference between capacitive in an bridge unbalanced

Estimated Interference at the input Vpp/Cearth

Ampl. Far Ampl. Close4.72E-08 1.64E-10

Table 5.5.: Theoretical computation of the interference at the amplifier input with caseplaced at 1 m from the subject (Far), and with the amplifier case placed inproximity of the subject (Close).

111

5. Low-noise bio-amplifier technology

(blue) and balanced condition (green). The HF-EEG provides a noise free scalp potential,

in the range of the requested noise level in the high frequency spectral domain: around

7.5 nV/√

Hz at 500 Hz, and asymptotically approaching 6 nV/√

Hz at higher frequencies,

as it is expected by a bipolar derivation with impedances below 1 kΩ.

the average of SEP recordings in the bipolar derivation FC3-CP3 was projected in the

Time-Frequency domain, and full band time trend, as well as averaged power spectral

density values are shown (Figure 5.14). Distinct spectral components extend from 100

to 1400 kHz, with three main contributions concentrated around 20 ms after the stimulus

onset. The N20 peak is visible in the broadband time trend, while sigma and kappa burst

are extracted by band pass filtering at 450-750 Hz and 800-1400 Hz, respectively.

Since averaging across N=6000 trials is sufficient to isolate the somatosensory response

features of interest, the findings of N20 and sigma burst are consistent with previous

recordings reported in the literature. The bottom panel, instead, shows the kappa burst â

for the first time measured in an un-shielded environment, with components at around 1

kHz clearly standing out the baseline level at about 20 ms post-stimulus.

5.5. Discussion

In this chapter the steps which led to the realization of HF-EEG low-noise portable mul-

tichannel recording system have been described. Some elements of electronic design

theory and a brief historical overview have been introduced, in order to welcome the

reader to the issue of combining interference rejection to low noise design.

The electromagnetic interference is one of the major problems in research-related as

well as in clinical measurements. Nowadays most commercial systems, based on differ-

ential input design, are characterized by Common Mode rejection of 70-90 dB. Unfortu-

nately, the devices available on the market are designed for EEG recordings in standard

spectral domains, ranging mostly from near DC to 100 Hz. In this frequencies the brain

signal has a SNR that allows non-invasive access to neurophysiological parameters of

interest, as scalp projections of post-synaptic potential. Since the interest of this work is

to detect also faster and, at the same time, weaker components, the necessity to improve

the sensitivity to physiological sources, screening out environmental interference, has

been encountered. In other words, power line harmonics effect needed to be minimized,

compatibly with a design providing a critically low system noise figure.

112

5.5. Discussion

The solution has been found in the design presented in paragraph 5.1.4. The essen-

tial need to have electrodes-skin impedances in the range of 1 kΩ, in order to minimize

the thermal noise contribution at the skin -gel-electrode interface, was revised as an ad-

vantage: the low impedance ground electrode, connected to the amplifier case, offers to

the stray currents an alternative path, in parallel to the one represented by the reference

electrode and the internal capacitance between the amplifier ground and the case (Fig-

ure 5.8, Zbs in parallel to Zbg and Cgs). Concurrently, a minimized current flow on the

reference electrode decreases the effect of the interference along the signal path. This

conceptual kernel, inserted in the middle of the capacitive bridge, has allowed the imple-

mentation of a single-ended input, theoretically more sensitive to interference effects. The

single-ended design halves the noise contribution from the device, limited to 2.3 nV/√

Hz,

decreasing the system influence on the noise figure as compared to the 4 nV/√

Hz con-

tributed by the termal noise at the electrode-gel-skin interface at an impedance of 1 kΩ.

Following the demonstration of the suitability of such design for an electrically un-

shielded environment (Figure 5.14), a portable set-up has been implemented, from the

analog front-end to the software user interface. The inferior performance of the previous

system is due to its different design [Scheer et al., 2006]: both active leads and refer-

ence are implemented as a single-ended input, and the ground electrode is connected to

the amplifier ground, making the stray current to entirely flow in the signal path. More-

over, since active channels and reference channel incorporate an operational amplifier,

the noise floor is higher, resulting in a measured 4.8 nV/√

Hz, comparable with a good

impedance preparation.

Placing the HF-EEG amplifier case close to the subject’s body, we empower high

quality data recordings in a clinical environment. This was indicated primarily by the

computations of the electro-technical model summarized in equation 5.10, and proved

afterwards experimentally. Nevertheless, a perfect match between model and real mea-

surement is not possible, since the presented mathematical framework describes only the

principle of the design, and not the true physical condition of the measurement. Other fac-

tors contribute to the interference at the input, as described by equation 1. With respect

to magnetic fields, we could minimize the loops surface by twisting the electrodes cables.

With respect to stray capacitance to the leads, cables shielding to case potential was exe-

cuted. These preventive measures could not be optimal in preventing stray currents into

the leads and capacitive coupling to the leads is not easy to model. Typical values for the

113

5. Low-noise bio-amplifier technology

Figure 5.14.: Comparison of Voltage spectral density of task-free EEG recordings, withamplifier at 1 m (ampl. far) and in proximity (ampl. close) to the subject.Full-band Power spectrum (left). Zoom on 50 Hz and relative harmonics inthe HF spectral range (right).

currents involved range around 100 pA [Wood et al., 1995]. For a reference impedance of

1 kΩ this would correspond to an interference of some pp at the input, which is compati-

ble with the recordings performed in the clinical environment with amplifier far from the

subject’s body. Even if it is not possible to precisely model the effect of currents coupled

to the leads, it must be observed that the amplifier case placed close to the body not only

balances the capacitive bridge with respect to the body, but minimizes the current flow

through the reference electrode, simultaneously reducing the effect of other sources of

interference.

With the implemented HF-EEG system, for the first time evoked kappa band com-

ponents as weak as 100 nVpp, elicited in the somatosensory cortex by electrical median

nerve stimulation were detected non-invasively. These encouraging results suggest to

apply this new technological achievement to record non-invasively also from patients

affected by epilepsy in whom high-frequency pathological activity has been shown inva-

sively [Bagshaw et al., 2009].

The HF-EEG amplifier setup has been certified for medical safety in accordance with

Council Directive 93/42/EEC for medical devices (MDD), and CE mark has been applied.

This allow researchers, technicians and medical doctor to use it to record healthy subjects

114

5.5. Discussion

Figure 5.15.: Neurophysiological band-passed components for channel derivation F3-C3’. Upper panel: Normalized time-frequency S-Transform. Bottom panel:SEP time curves for the broad band average, and for sigma (450-750 Hz)and kappa spectral ranges (800-1400 Hz).

115

5. Low-noise bio-amplifier technology

as well as patients. Further investigation by means of low-noise multichannel EEG must

be performed, to collect high frequency SEP recordings, and possibly investigate the non-

invasive detectability of fast components in epileptic patients.

By a technical point of view further progresses are still possible, since this system

still presents relatively long cables, exposed to displacement currents. The best design

solution remain could be implemented with active electrodes [Chimene and Pallas-Areny,

2000], and enlarging the number of available channels to provide a platform for features

detection in low-noise recordings.

116

6. Summary and Conclusion

The ensemble of previous and present contributions describe HFO as an important phys-

iological signature of the human brain. Fast rhythmic activity in the context of so-

matosensory stimulation was extensively characterized in humans and animals [Curio,

2005,Ozaki and Hashimoto, 2011], and their non-invasive detectability was demonstrated

up to 1 kHz [Scheer et al., 2011, Fedele et al., 2012]. Importantly HFO variability relates

to diverse pathological conditions, and particular attention has been dedicated, in the re-

cent years, to HFO in epilepsy. Several studies have demonstrated the presence of fast

ripples in epileptic tissues at micro and macro scales, in pre-ictal, ictal and post-ictal time

periods [Bragin et al., 2010,Jacobs et al., 2012,Zijlmans et al., 2012]. Even if events in the

ripples range (80-250 Hz) were recorded with scalp EEG, reports about HFO components

at the scalp level are missing.

A quantitative analysis of the HFO detectability must rely on the characterization of

the biophysical limiting factors. Poor SNR certainly represent a major concern, even if

the obstacles to the HFO non-invasive detection are not quantitatively clear yet. In this

review we do not claim to provide a clear-cut solution. Our primary interest is to gather

together elements belonging to biology and physics, in order to offer a detailed overview

of this complex scenario. The heterogeneous way in reporting findings has made this

reconstruction particularly difficult. In epilepsy studies in particular, systematic infor-

mations regarding detected events amplitudes/power range, location, central frequency,

occurrence, duration, and setup technical features are partially omitted, as many impor-

tant contributions cover in turn only some of these aspects. In this sense, a standardized

framework would provide the scientific community with a rich database, in order to allow

comparison based on multiple parameters.

In order to clarify the opportunity to detect such components from the scalp, demon-

strating their not artefactual origin [Benar et al., 2010, Muthukumaraswamy, 2013], si-

multaneous recordings with macroEEG and ECoG grid would be requested. The setup

should satisfy technical requirements as high spatial sampling, low noise at the interface,

117

6. Summary and Conclusion

sampling frequency of at least 2 kHz, in order to relate HFO locally arising from deep or

cortical sources to the potential available at the level of the scalp.

In this context hf-SEP represent the physiological workhorse to test progress in tech-

nology as well as in data analysis approaches. Namely, the phase-locked nature of the

hf-SEP provides a rare opportunity to rely on consistent prior information about the de-

tected spatiotemporal features: 1) source are localized in area 3b [Nakano and Hashimoto,

1999,Haueisen et al., 2000,Haueisen et al., 2001,Gobbelé et al., 2004], and confined to a

limited cortical patch, as neural population responding to somatotopic mapping by means

of median and ulnar nerve stimulation show two distinct dipolar sources localized 1 cm

apart from each other [Curio et al., 1997]; 2) even if a weak amplitude of few hundreds of

nVpp is available on the scalp, the precise phase locked spiking pattern guarantees post-

averaging visibility. Stability of the response and well controlled experimental protocol

provide an optimal testing framework for small signal detectability. Moreover, low-noise

technology opens the way to promising progress towards single trial resolution, quantified

in terms of SNR [Waterstraat et al., 2012].

The issue of the detectability power could also be resolved in principle by simultane-

ous recording in the case of hf-SEP. Sigma band laminar field HFO measured with needle

electrodes (1 mm diameter) are in the range of 40 µVpp, with corresponding scalp EEG

in monkey (around 200 nVpp) attenuated by factor of 100 [Shimazu et al., 2000]. Sub-

dural recordings following either by submotor threshold stimulation [Kojima et al., 2001]

or supramotor threshold stimulation [Baker et al., 2003] detect ripples in the 5 µVpp

range. The different spatial scale, the choice for the reference, and the focal nature of the

somatosensory response, can account for differences of one order of magnitude between

subdural probes and laminar contacts [Peterson et al., 1995]. Even if a fair comparison is

not possible, it is worth noticing how a similar dampening factor is observed for higher

bands: subdurally recorded hf-SEP above 1 kHz range around 1 µVpp [Sakura et al.,

2009] mirroring similar attenuation with respect to scalp recordings performed with low

noise bioamplifier, as ripples in the order of tens of nVpp could be resolved after massive

off-line averaging [Scheer et al., 2011, Fedele et al., 2012].

In the context of epileptic interictal events, contributions in the 80-250 Hz ripples band

at epidural level, ranging between 50-200 µV could be observed [Urrestarazu et al.,

2007, Jacobs et al., 2008, Schevon et al., 2009, Jacobs et al., 2009, Kobayashi et al.,

2010, Andrade-Valenca et al., 2012, Cho et al., 2012]. Similar features were detected

118

also from the scalp, mostly in ictal events, as large as 10-50 µV pp [Kobayashi et al.,

2004, Inoue et al., 2008, Kobayashi et al., 2010, Andrade-Valenca et al., 2011, Kobayashi

et al., 2011, Kobayashi et al., 2013].

Interestingly, some of these studies describing Ripple Band features also report the

presence of Fast Ripple band (250-500 Hz) components, with HFO in the order of 10 to

100 µV at cortical level [Jacobs et al., 2008, Schevon et al., 2009, Usui et al., 2010, Cho

et al., 2012]. Thus, evidence for a scaling factor of 2 to 10 is provided for the same

recording site.

Macroscopic measurements cannot unambiguously clarify whether this difference should

be referred to different neural population, confined to smaller (and diverse) synchronized

areas, or correspond to low-pass filtering effect described theoretically in biophysical

models [Koch, 2004, Leski et al., 2013]. Despite the generative mechanisms responsible

for difference in power, assuming the linear far field propagation, we encounter detectabil-

ity issue very similar to the hf-SEP, possibly even less critical. Macroscopic electric field

propagation allows calculation of the number of post-synaptic potentials, somatic and ax-

onal action potential needed to achieve a detectable ripple at the level of the scalp. In

the electrically homogenous matter this would require i.e. the synchronous activity of

approximately 2000 somatic cortical spikes (see 2.2), geometrically oriented in a con-

structive way, and powerful enough to overtake incoherent neighbouring background ac-

tivity. Critically, cell density varies not only across species but also brain areas [Collins

et al., 2010]. The localized nature of HFO has been shown through epidural electrode grid

recordings. Empirical results for lower frequency ranges constrain scalp detectability to a

source cortical area of 5-10 cm2 [Tao et al., 2007,Yamazaki et al., 2012], even if referred

to interictal epileptic spikes, meaning hundreds of µV in amplitude in a frequency range

severely affected by background activity [Cosandier-Rimele et al., 2012].

In this scenario, the opportunity to benefit of an optimized recording setup becomes

critical. Low-noise technology design, and biophysical model of the influencing fac-

tors [Scheer et al., 2006] proved the detection of signals as low as 300 nVpp under spe-

cific conditions (Low noise FET input, 1 kΩ impedance, 100 Hz wide band-pass) possi-

ble. A low-noise input multichannel recording device, carrying outstanding robustness to

environmental interference, and implemented as a portable system, has been described,

realized and tested 5. At the current status the HF-EEG system is certified for medical

safety, in accordance with Council Directive 93/42/EEC for medical devices (MDD), and

119

6. Summary and Conclusion

CE mark has been applied. This allows researchers, technicians and medical doctors to

use it to record healthy subjects as well as patients.

It has been demonstrated how demanding SNR signatures, as the kappa band com-

ponents, can be detected and extensively characterized from recordings realized inside

an electromagnetically shielded environment ( 3 and 4). These results could be repli-

cated by mean of the new system also in a clinical environment 5. Moreover, further

steps towards single trial visibility have been performed testing diverse machine learning

approaches in order to increase the SNR, and modeling the biophysical constrains in a

simulated theoretical framework.

Taken together, tools targeted to the non-invasive detection of high-frequency scalp

EEG have been optimized. A sensitive recording system can be now applied to test the

detectability of the HFO in other context, primarily neocortical epilepsies. The opportu-

nity to extend this sensitivity to a dense multichannel setup would open the way to higher

spatial sampling on the recording site [Yamazaki et al., 2012], complemented by multi-

variate analysis technique optimized to improve SNR [Blankertz et al., 2008,Celka et al.,

2008], pointing to the golden standard of single trial resolution.

120

A. Appendix

A.1. Experimental Sessions

The set of recordings performed during this work are listed in Table A.1. While the

details about preparation, set-up, and target of the experimental session are explained in

the chapters 3-5, here we summarize them in a synthetic outlook.

Each recording is labeled with a Latin number, assigned according to order followed

in the dissertation. The protocols considered here are:

- resting state data, in which the subject is required to relax, and remain awake keeping

eyes open,

- SEP/F (Somatosensory Evoked Potential/Field), obtained by repeated delivery of cur-

rent pulses at the level of the median nerve, in order to elicit the somatosensory evoked

response.

Recordings I-V were conducted in PTB, Berlin, while recording VI was performed in

Charite, Campus Benjamin Franklin, Berlin. The PTB represents an excellence for the

development of low-noise technology, and allowed access to EEG systems configurable

to obtain a noise figure from 2.7 to 12 nV/√

Hz, with montage ranging from one single

bipolar channel to 30 common referenced derivations, and operated in an electromag-

netically shielded environment [Scheer et al., 2006]. Moreover, it was possible to com-

bine low-noise EEG technology, together with consecutive implementations of low-noise

MEG system (recordings IIIa, IIIb). Supported by a deeply experienced environment in

dedicated electronic system design, the candidate could finally achieve the realization of

a multichannel portable system, named HF-EEG, suitable for a non electromagnetically

shielded room. This first implementation of this new design was compared to existing

low-noise devices (V): the final portable recording system, certified by CE mark and ap-

proved by Charite ethical committee, is now available in department of neurophysiology

121

A. Appendix

of Charite, Campus Benjamin Franklin, Berlin, where its sensitivity in non-invasive de-

tection of cortical population spikes was demonstrated (VI).

122

A.1. Experimental Sessions

Rec

ordi

ngPr

otoc

olD

evic

e,no

ise

leve

l,L

ocat

ion

Cha

pter

deriv

atio

nI

Res

t,SE

PE

EG

,2.7

and

12nV

/√H

zE

M-S

R3

bipo

lard

eriv

atio

nPT

B,B

erlin

IIR

est,

SEP

EE

G,4

.8nV

/√H

zE

M-S

R,

3m

ultic

hann

elsy

stem

PTB

,Ber

linII

IaR

est,

SEP/

FE

EG

,4.8

nV/√

Hz,

ME

G,1

.9fT

/√H

z,E

M-S

R,

3bi

pola

rder

ivat

ion,

sing

lech

anne

lPT

B,B

erlin

IIIb

Res

t,SE

P/F

EE

G,4

.8nV

/√H

z,M

EG

,0.5

fT/√

Hz,

EM

-SR

,3

bipo

lard

eriv

atio

n,si

ngle

chan

nel

PTB

,Ber

linIV

SEP

EE

G,4

.8nV

/√H

z,E

M-S

R,

4m

ultic

hann

elsy

stem

PTB

,Ber

linV

Res

tE

EG

,4.8

nV/√

Hz,

EE

G,2

.3nV

/√H

z,E

M-S

R,

5bi

pola

rder

ivat

ion,

bipo

lard

eriv

atio

n,PT

B,B

erlin

VI

Res

t,SE

PH

F-E

EG

,2.3

nV/√

Hz,

Neu

roph

ysic

sD

epar

tmen

t,5

bipo

lard

eriv

atio

nC

BF,

Cha

rite

,Ber

lin.

Tabl

eA

.1.:

Lis

tan

dde

scri

ptio

nof

the

reco

rdin

gse

ssio

nspe

rfor

med

,w

ithre

fere

nce

toPr

otoc

ol,

Dev

ice

setti

ngs,

and

thes

isC

hapt

erin

setu

pan

dres

ults

are

repo

rted

inde

tail.

Res

t:re

stin

gst

ate;

SEP:

Som

atos

enso

ryE

voke

dPo

tent

ial

bym

ean

ofm

edia

nne

rve

stim

ulat

ion.

EM

-SR

:Ele

ctro

mag

netic

ally

Shie

lded

Roo

m.

PTB

:Phy

sika

lisch

-Tec

hnis

che

Bun

desa

nsta

lt;C

BF:

Cam

pus

Ben

jam

inFr

ankl

in.

123

List of Equations

2.1 Recording system noise floor . . . . . . . . . . . . . . . . . . . . . . . . . 182.2 Thermal noise at the interface . . . . . . . . . . . . . . . . . . . . . . . . . 182.3 Dipole far field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.4 Quadrupole far field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.5 S-Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.6 STFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.7 Gaussian window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.8 Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.9 Mother Wavelet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.10 IIR filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.11 Forward Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.12 Forward Model with noise . . . . . . . . . . . . . . . . . . . . . . . . . . 232.13 Forward Model in Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 242.14 CSP, double diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . 252.15 CSP, GEDV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.16 CSP, difference between 2 classes . . . . . . . . . . . . . . . . . . . . . . 252.17 CSP, Power Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.18 CCA, formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.19 CCA, GEDV system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.20 CCA, GEDV simplified system . . . . . . . . . . . . . . . . . . . . . . . 262.21 Similarity Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.1 BW-normalized noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343.2 PTB ampl noise budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354.1 Spatial filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614.2 Forward Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.3 SNR definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644.4 SNR definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654.5 CSP, Power Ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664.6 CSP, difference between 2 classes . . . . . . . . . . . . . . . . . . . . . . 664.7 CCA, formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674.8 CCA, GEDV double problem . . . . . . . . . . . . . . . . . . . . . . . . . 674.9 CCA, datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

125

List of Equations

4.10 CCAr, datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685.1 CMRR, definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875.2 IMRR, definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885.3 EM Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 885.4 Noise model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 925.5 Thermal noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 925.6 Thevenin, step I, voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . 985.7 Thevenin, step I, impedance . . . . . . . . . . . . . . . . . . . . . . . . . 985.8 Thevenin, step II, voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . 1005.9 Thevenin, step II, impedance . . . . . . . . . . . . . . . . . . . . . . . . . 1005.10 Complete electrotechnical model . . . . . . . . . . . . . . . . . . . . . . 1005.11 Complete model, simplification I . . . . . . . . . . . . . . . . . . . . . . 1005.12 Complete model, simplification II . . . . . . . . . . . . . . . . . . . . . . 1005.13 Complete simplified model . . . . . . . . . . . . . . . . . . . . . . . . . 1015.14 Stray Capacitance measurement . . . . . . . . . . . . . . . . . . . . . . . 1015.15 Stray Capacitance low Load . . . . . . . . . . . . . . . . . . . . . . . . . 1025.16 Stray Capacitance high Load . . . . . . . . . . . . . . . . . . . . . . . . 1025.17 Stray Capacitance varying Load . . . . . . . . . . . . . . . . . . . . . . . 103

126

List of Figures

2.1 Amplitude Density Spectrum . . . . . . . . . . . . . . . . . . . . . . . . 172.2 SEP Experimental protocol . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.1 Simplified circuit diagram PTB amplifier . . . . . . . . . . . . . . . . . . 363.2 FFT of background noise . . . . . . . . . . . . . . . . . . . . . . . . . . 383.3 Kappa ripple detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.4 hf-SEP Time-Frequency representation . . . . . . . . . . . . . . . . . . . 433.5 hf-SEP scalpmaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.6 hf-SEP butterflies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463.7 M/EEG resting state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493.8 M/EEG TFr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503.9 M/EEG kappa band . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.10 M/EEG group stats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.1 Simulations outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694.2 Sensor space analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714.3 Source space N20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 724.4 Source space sigma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734.5 Source space kappa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744.6 Comparison simulation - real data . . . . . . . . . . . . . . . . . . . . . 744.7 Source reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764.8 Latency analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774.9 Envelope sigma and kappa . . . . . . . . . . . . . . . . . . . . . . . . . 794.10 Envelope 1 and 8 Hz . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

5.1 Amplifier Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875.2 Simulations outcome . . . . . . . . . . . . . . . . . . . . . . . . . . . . 895.3 Noise model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915.4 Measured voltage spectral densities . . . . . . . . . . . . . . . . . . . . 945.5 Noise calculation model . . . . . . . . . . . . . . . . . . . . . . . . . . . 955.6 Biophysical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975.7 Equivalent schematic I . . . . . . . . . . . . . . . . . . . . . . . . . . . 985.8 Equivalent schematic II . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

127

List of Figures

5.9 Equivalent schematic III . . . . . . . . . . . . . . . . . . . . . . . . . . 995.10 Stray Capacitance measurement . . . . . . . . . . . . . . . . . . . . . . 1025.11 HF-EEG block diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 1055.12 HF-EEG amplifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1075.13 Low noise EEG Power spectra . . . . . . . . . . . . . . . . . . . . . . . 1095.14 HF-EEG Voltage spectral density . . . . . . . . . . . . . . . . . . . . . . 1145.15 HF-EEG SEP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

128

List of Tables

2.1 Action Potentials Count . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3.1 Noise level measurement at rest . . . . . . . . . . . . . . . . . . . . . . 373.2 Filter settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.1 Source reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

5.1 Noise calculation model . . . . . . . . . . . . . . . . . . . . . . . . . . . 955.2 Stray capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1035.3 Estimated stray capacitance . . . . . . . . . . . . . . . . . . . . . . . . . 1115.4 Estimated Capacitance ratio . . . . . . . . . . . . . . . . . . . . . . . . . 1115.5 Theoretical Interference . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

A.1 List of the recording sessions . . . . . . . . . . . . . . . . . . . . . . . . 123

129

Bibliography[Akiyama et al., 2006] Akiyama, T., Otsubo, H., Ochi, A., Galicia, E. Z., Weiss, S. K., Don-

ner, E. J., Rutka, J. T., and Snead 3rd, O. C. (2006). Topographic movie of ictal high-frequency oscillations on the brain surface using subdural EEG in neocortical epilepsy. Epilep-sia, 47(11):1953–1957.

[Alegre et al., 2006] Alegre, M., Urriza, J., Valencia, M., Muruzábal, J., Iriarte, J., and Artieda,J. (2006). High-frequency oscillations in the somatosensory evoked potentials of patients withcortical myoclonus: pathophysiologic implications. Journal of clinical neurophysiology : offi-cial publication of the American Electroencephalographic Society, 23:265–272.

[Allen et al., 1992] Allen, P. J., Fish, D. R., and Smith, S. J. (1992). Very high-frequency rhyth-mic activity during SEEG suppression in frontal lobe epilepsy.

[Andrade-Valenca et al., 2012] Andrade-Valenca, L., Mari, F., Jacobs, J., Zijlmans, M., Olivier,A., Gotman, J., and Dubeau, F. (2012). Interictal high frequency oscillations (HFOs) in patientswith focal epilepsy and normal MRI. Clin Neurophysiol, 123(1):100–105.

[Andrade-Valenca et al., 2011] Andrade-Valenca, L. P., Dubeau, F., Mari, F., Zelmann, R., andGotman, J. (2011). Interictal scalp fast oscillations as a marker of the seizure onset zone.Neurology, 77:524–531.

[Bagshaw et al., 2009] Bagshaw, A. P., Jacobs, J., LeVan, P., Dubeau, F., and Gotman, J. (2009).Effect of sleep stage on interictal high-frequency oscillations recorded from depth macroelec-trodes in patients with focal epilepsy. Epilepsia, 50(4):617–628.

[Baillet et al., 2001] Baillet, S., Mosher, J., and Leahy, R. (2001). Electromagnetic brain map-ping. IEEE Signal Processing Magazine, 18.

[Baker et al., 2003] Baker, S. N., Curio, G., and Lemon, R. N. (2003). EEG oscillations at 600Hz are macroscopic markers for cortical spike bursts. The Journal of physiology, 550:529–534.

[Benar et al., 2010] Benar, C. G., Chauviere, L., Bartolomei, F., and Wendling, F. (2010). Pitfallsof high-pass filtering for detecting epileptic oscillations: a technical note on "false" ripples.Clin Neurophysiol, 121(3):301–310.

[Berg and Scherg, 1994] Berg, P. and Scherg, M. (1994). A fast method for forward computationof multiple-shell spherical head models. Electroencephalography and clinical neurophysiol-ogy, 90:58–64.

[Berger, 1938] Berger, H. (1938). Das Elektrenkephalogramm des Menschen. Nova ActaLeopoldina, 6:173–179.

[Bjoerck and Golub, 1971] Bjoerck, A. and Golub, G. H. (1971). Numerical methods for com-puting angles between linear subspaces.

131

Bibliography

[Blankertz et al., 2011] Blankertz, B., Lemm, S., Treder, M., Haufe, S., and Müller, K.-R. (2011).Single-trial analysis and classification of ERP components–a tutorial. NeuroImage, 56:814–825.

[Blankertz et al., 2008] Blankertz, B., Tomioka, R., Lemm, S., Kawanabe, M., and Muller, K.-R. (2008). Optimizing Spatial filters for Robust EEG Single-Trial Analysis. IEEE SignalProcessing Magazine, 25.

[Bragin et al., 1999] Bragin, A., Engel, J., Wilson, C. L., Fried, I., and Mathern, G. W. (1999).Hippocampal and entorhinal cortex high-frequency oscillations (100–500 Hz) in human epilep-tic brain and in kainic acid–treated rats with chronic seizures. Epilepsia, 40(2):127–37.

[Bragin et al., 2010] Bragin, A., Engel Jr., J., and Staba, R. J. (2010). High-frequency oscillationsin epileptic brain. Curr Opin Neurol, 23(2):151–156.

[Bragin et al., 2002] Bragin, A., Wilson, C. L., Staba, R. J., Reddick, M., Fried, I., and EngelJr., J. (2002). Interictal high-frequency oscillations (80-500 Hz) in the human epileptic brain:entorhinal cortex. Ann Neurol, 52(4):407–415.

[Butterworth, 1930] Butterworth, S. (1930). On the Theory of Filter Amplifiers, volume 7.

[Buzsaki and Draguhn, 2004] Buzsaki, G. and Draguhn, A. (2004). Neuronal oscillations in cor-tical networks. Science, 304(5679):1926–1929.

[Celka et al., 2008] Celka, P., Le, K. N., and Cutmore, T. R. H. (2008). Noise reduction in rhyth-mic and multitrial biosignals with applications to event-related potentials. IEEE transactionson bio-medical engineering, 55:1809–1821.

[Chapman and McCrary, 1995] Chapman, R. M. and McCrary, J. W. (1995). EP component iden-tification and measurement by principal components analysis. Brain and cognition, 27:288–310.

[Chatillon et al., 2013] Chatillon, C. E., Zelmann, R., Hall, J. A., Olivier, A., Dubeau, F., andGotman, J. (2013). Influence of contact size on the detection of HFOs in human intracerebralEEG recordings. Clin Neurophysiol, 124(8):1541–1546.

[Chimene and Pallas-Areny, 2000] Chimene, M. and Pallas-Areny, R. (2000). A comprehensivemodel for power line interference in biopotential measurements. IEEE Transactions on Instru-mentation and Measurement, 49.

[Cho et al., 2012] Cho, J. R., Joo, E. Y., Koo, D. L., Hong, S. C., and Hong, S. B. (2012). Clin-ical utility of interictal high-frequency oscillations recorded with subdural macroelectrodes inpartial epilepsy. J Clin Neurol, 8(1):22–34.

[Collins et al., 2010] Collins, C. E., Airey, D. C., Young, N. A., Leitch, D. B., and Kaas, J. H.(2010). Neuron densities vary across and within cortical areas in primates. Proc Natl Acad SciU S A, 107(36):15927–15932.

[Coppola et al., 2005] Coppola, G., Vandenheede, M., Di Clemente, L., Ambrosini, A., Fumal,A., De Pasqua, V., and Schoenen, J. (2005). Somatosensory evoked high-frequency oscillationsreflecting thalamo-cortical activity are decreased in migraine patients between attacks. Brain :a journal of neurology, 128:98–103.

132

Bibliography

[Cosandier-Rimele et al., 2012] Cosandier-Rimele, D., Bartolomei, F., Merlet, I., Chauvel, P.,and Wendling, F. (2012). Recording of fast activity at the onset of partial seizures: depth EEGvs. scalp EEG. Neuroimage, 59(4):3474–3487.

[Cracco and Cracco, 1976] Cracco, R. Q. and Cracco, J. B. (1976). Somatosensory evoked po-tential in man: far field potentials. Electroencephalogr Clin Neurophysiol, 41:460–466.

[Crepon et al., 2010] Crepon, B., Navarro, V., Hasboun, D., Clemenceau, S., Martinerie, J.,Baulac, M., Adam, C., and Le Van Quyen, M. (2010). Mapping interictal oscillations greaterthan 200 Hz recorded with intracranial macroelectrodes in human epilepsy. Brain, 133(Pt1):33–45.

[Curio, 2005] Curio, G. (2005). Ultrafast EEG activities. In: Niedermeyer E, da Silva, FL, ed-itors.Electroencephalography basic principles clinical applications and related fields. 5th ed.Philadelphia: Lippincott Williams Wilkins; p. 495â504.

[Curio et al., 1994] Curio, G., Mackert, B. M., Burghoff, M., Koetitz, R., Abraham-Fuchs, K.,and Harer, W. (1994). Localization of evoked neuromagnetic 600 Hz activity in the cerebralsomatosensory system. Electroencephalogr Clin Neurophysiol, 91(6):483–487.

[Curio et al., 1997] Curio, G., Mackert, B. M., Burghoff, M., Neumann, J., Nolte, G., Scherg, M.,and Marx, P. (1997). Somatotopic source arrangement of 600 Hz oscillatory magnetic fields atthe human primary somatosensory hand cortex. Neuroscience letters, 234:131–134.

[Dähne et al., 2013] Dähne, S., Meinecke, F. C., Haufe, S., Höhne, J., Tangermann, M., Müller,K.-R., and Nikulin, V. V. (2013). SPoC: A novel framework for relating the amplitude ofneuronal oscillations to behaviorally relevant parameters. NeuroImage.

[Darvas et al., 2004] Darvas, F., Pantazis, D., Kucukaltun-Yildirim, E., and Leahy, R. M. (2004).Mapping human brain function with MEG and EEG: methods and validation. NeuroImage, 23Suppl 1:S289–99.

[Drung et al., 2007] Drung, D., Assmann, C., Beyer, J., Kirste, A., Peters, M., Ruede, F., andSchurig, T. (2007). Highly Sensitive and Easy-to-Use SQUID Sensors. IEEE Transactions onApplied Superconductivity, 17.

[Emori et al., 1991] Emori, T., Yamada, T., Seki, Y., Yasuhara, A., Ando, K., Honda, Y., Leis,A. A., and Vachatimanont, P. (1991). Recovery functions of fast frequency potentials in theinitial negative wave of median SEP. Electroencephalography and clinical neurophysiology,78:116–123.

[Engel, 2011] Engel, J. (2011). Biomarkers in epilepsy: foreword.

[Fedele et al., 2012] Fedele, T., Scheer, H., Waterstraat, G., Telenczuk, B., Burghoff, M., andCurio, G. (2012). Towards non-invasive multi-unit spike recordings: Mapping 1kHz EEGsignals over human somatosensory cortex. Clinical Neurophysiology.

[Fedele et al., 2013] Fedele, T., Scheer, H.-J., Burghoff, M., Waterstraat, G., Nikulin, V., andCurio, G. (2013). Distinction between added-energy and phase-resetting mechanisms in non-invasively detected somatosensory evoked responses. Conference proceedings : ... AnnualInternational Conference of the IEEE Engineering in Medicine and Biology Society. IEEEEngineering in Medicine and Biology Society. Conference, 2013:1688–1691.

133

Bibliography

[Fell et al., 2004] Fell, J., Dietl, T., Grunwald, T., Kurthen, M., Klaver, P., Trautner, P., Schaller,C., Elger, C. E., and Fernández, G. (2004). Neural bases of cognitive ERPs: more than phasereset. Technical report.

[Fisher et al., 1992] Fisher, R. S., Webber, W. R., Lesser, R. P., Arroyo, S., and Uematsu, S.(1992). High-frequency EEG activity at the start of seizures. Journal of clinical neurophysiol-ogy : official publication of the American Electroencephalographic Society, 9:441–448.

[Fontanini and Katz, 2008] Fontanini, A. and Katz, D. B. (2008). Behavioral states, networkstates, and sensory response variability. Journal of neurophysiology, 100:1160–1168.

[Fox et al., 2006] Fox, M. D., Snyder, A. Z., Zacks, J. M., and Raichle, M. E. (2006). Coherentspontaneous activity accounts for trial-to-trial variability in human evoked brain responses.Technical report.

[Freeman et al., 2003] Freeman, W. J., Holmes, M. D., Burke, B. C., and Vanhatalo, S. (2003).Spatial spectra of scalp EEG and EMG from awake humans. Clinical neurophysiology : officialjournal of the International Federation of Clinical Neurophysiology, 114:1053–1068.

[Gobbelé et al., 1999] Gobbelé, R., Buchner, H., Scherg, M., and Curio, G. (1999). Stability ofhigh-frequency (600 Hz) components in human somatosensory evoked potentials under varia-tion of stimulus rate–evidence for a thalamic origin. Clinical neurophysiology : official journalof the International Federation of Clinical Neurophysiology, 110:1659–1663.

[Gobbelé et al., 2003a] Gobbelé, R., Waberski, T. D., Dieckhöfer, A., Kawohl, W., Klostermann,F., Curio, G., and Buchner, H. (2003a). Patterns of disturbed impulse propagation in multiplesclerosis identified by low and high frequency somatosensory evoked potential components.Journal of clinical neurophysiology : official publication of the American Electroencephalo-graphic Society, 20(4):283–90.

[Gobbelé et al., 2000] Gobbelé, R., Waberski, T. D., Kuelkens, S., Sturm, W., Curio, G., andBuchner, H. (2000). Thalamic and cortical high-frequency (600 Hz) somatosensory-evokedpotential (SEP) components are modulated by slight arousal changes in awake subjects. Exper-imental brain research. Experimentelle Hirnforschung. Experimentation cerebrale, 133:506–513.

[Gobbelé et al., 2004] Gobbelé, R., Waberski, T. D., Simon, H., Peters, E., Klostermann, F., Cu-rio, G., and Buchner, H. (2004). Different origins of low- and high-frequency components (600Hz) of human somatosensory evoked potentials. Clinical neurophysiology : official journal ofthe International Federation of Clinical Neurophysiology, 115:927–937.

[Gobbelé et al., 2003b] Gobbelé, R., Waberski, T. D., Thyerlei, D., Thissen, M., Darvas, F.,Klostermann, F., Curio, G., and Buchner, H. (2003b). Functional dissociation of a subcor-tical and cortical component of high-frequency oscillations in human somatosensory evokedpotentials by motor interference. Neuroscience Letters, 350(2):97–100.

[Goldensohn, 1998] Goldensohn, E. S. (1998). Animal electricity from Bologna to Boston. Elec-troencephalogr Clin Neurophysiol, 106(2):94–100.

[Gyorgy Buzsaki, 2011] Gyorgy Buzsaki (2011). Rhythms of the Brain [Paperback]. OxfordUniversity Press, USA; 1 edition.

134

Bibliography

[Haberman et al., 2011] Haberman, M., Cassino, A., and Spinelli, E. (2011). Estimation of straycoupling capacitances in biopotential measurements. Med Biol Eng Comput, 49(9):1067–1071.

[Haberman and Spinelli, 2010] Haberman, M. and Spinelli, E. (2010). A digital driven right legcircuit. Conference proceedings : ... Annual International Conference of the IEEE Engineeringin Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Confer-ence, 2010:6559–6562.

[Haberman and Spinelli, 2012] Haberman, M. A. and Spinelli, E. M. (2012). A MultichannelEEG Acquisition Scheme Based on Single Ended Amplifiers and Digital DRL.

[Haegelen et al., 2013] Haegelen, C., Perucca, P., Chatillon, C. E., Andrade-Valenca, L., Zel-mann, R., Jacobs, J., Collins, D. L., Dubeau, F., Olivier, A., and Gotman, J. (2013). High-frequency oscillations, extent of surgical resection, and surgical outcome in drug-resistant focalepilepsy. Epilepsia, 54(5):848–857.

[Halboni et al., 2000] Halboni, P., Kaminski, R., Gobbelé, R., Züchner, S., Waberski, T. D., Her-rmann, C. S., Töpper, R., and Buchner, H. (2000). Sleep stage dependant changes of the high-frequency part of the somatosensory evoked potentials at the thalamus and cortex. Clinicalneurophysiology : official journal of the International Federation of Clinical Neurophysiology,111:2277–2284.

[Hanajima et al., 2004] Hanajima, R., Chen, R., Ashby, P., Lozano, A. M., Hutchison, W. D.,Davis, K. D., and Dostrovsky, J. O. (2004). Very fast oscillations evoked by median nervestimulation in the human thalamus and subthalamic nucleus. J Neurophysiol, 92(6):3171–3182.

[Hashimoto et al., 1999] Hashimoto, I., Kimura, T., Fukushima, T., Iguchi, Y., Saito, Y., Terasaki,O., and Sakuma, K. (1999). Reciprocal modulation of somatosensory evoked N20m primaryresponse and high-frequency oscillations by interference stimulation. Clinical neurophysiology: official journal of the International Federation of Clinical Neurophysiology, 110:1445–1451.

[Hashimoto et al., 1996] Hashimoto, I., Mashiko, T., and Imada, T. (1996). Somatic evokedhigh-frequency magnetic oscillations reflect activity of inhibitory interneurons in the humansomatosensory cortex. Electroencephalography and clinical neurophysiology, 100:189–203.

[Haueisen et al., 2012] Haueisen, J., Funke, M., Güllmar, D., and Eichardt, R. (2012). Tangentialand Radial Epileptic Spike Activity.

[Haueisen et al., 2000] Haueisen, J., Heuer, T., Nowak, H., Liepert, J., Weiller, C., Okada, Y., andCurio, G. (2000). The influence of lorazepam on somatosensory-evoked fast frequency (600Hz) activity in MEG. Brain research, 874:10–14.

[Haueisen et al., 2001] Haueisen, J., Schack, B., Meier, T., Curio, G., and Okada, Y. (2001). Mul-tiplicity in the high-frequency signals during the short-latency somatosensory evoked corticalactivity in humans. Technical report.

[Haufe et al., 2014] Haufe, S., Meinecke, F., Görgen, K., Dähne, S., Haynes, J.-D., Blankertz,B., and Bieß mann, F. (2014). On the interpretation of weight vectors of linear models inmultivariate neuroimaging. NeuroImage, 87:96–110.

135

Bibliography

[Haufe et al., 2011] Haufe, S., Tomioka, R., Dickhaus, T., Sannelli, C., Blankertz, B., Nolte, G.,and Müller, K.-R. (2011). Large-scale EEG/MEG source localization with spatial flexibility.NeuroImage, 54(2):851–9.

[Helmholtz, 1853] Helmholtz, H. (1853). Ueber einige Gesetze der Vertheilung elektrischerStröme in körperlichen Leitern mit Anwendung auf die thierisch-elektrischen Versuche. Ann.Phys. Chem., 89:211–233, 353–377.

[Holmes et al., 1998] Holmes, C. J., Hoge, R., Collins, L., Woods, R., Toga, A. W., and Evans,A. C. (1998). Enhancement of MR images using registration for signal averaging. Journal ofcomputer assisted tomography, 22(2):324–33.

[Hotelling, 1992] Hotelling, H. (1992). Relations Between Two Sets of Variates BT - Break-throughs in Statistics. In Breakthroughs in Statistics, pages 162–190.

[Huhta and Webster, 1973] Huhta, J. C. and Webster, J. G. (1973). 60-HZ interference in elec-trocardiography. IEEE transactions on bio-medical engineering, 20:91–101.

[Ikeda et al., 2002] Ikeda, A., Miyamoto, S., and Shibasaki, H. (2002). Cortical motor mappingin epilepsy patients: information from subdural electrodes in presurgical evaluation. Epilepsia,43 Suppl 9:56–60.

[Inoue et al., 2001] Inoue, K., Hashimoto, I., and Nakamura, S. (2001). High-frequency oscilla-tions in human posterior tibial somatosensory evoked potentials are enhanced in patients withParkinson’s disease and multiple system atrophy. Neuroscience letters, 297:89–92.

[Inoue et al., 2008] Inoue, T., Kobayashi, K., Oka, M., Yoshinaga, H., and Ohtsuka, Y. (2008).Spectral characteristics of EEG gamma rhythms associated with epileptic spasms. Brain Dev,30(5):321–328.

[Insola et al., 2008] Insola, A., Padua, L., Mazzone, P., and Valeriani, M. (2008). Unmasking ofpresynaptic and postsynaptic high-frequency oscillations in epidural cervical somatosensoryevoked potentials during voluntary movement. Clin Neurophysiol, 119(1):237–245.

[Insola et al., 2010] Insola, A., Padua, L., Mazzone, P., and Valeriani, M. (2010). Effect of move-ment on SEPs generated by dorsal column nuclei. Clin Neurophysiol, 121(6):921–929.

[Jacobs et al., 2008] Jacobs, J., LeVan, P., Chander, R., Hall, J., Dubeau, F., and Gotman, J.(2008). Interictal high-frequency oscillations (80-500 Hz) are an indicator of seizure onsetareas independent of spikes in the human epileptic brain. Epilepsia, 49(11):1893–1907.

[Jacobs et al., 2012] Jacobs, J., Staba, R., Asano, E., Otsubo, H., Wu, J. Y., Zijlmans, M., Mo-hamed, I., Kahane, P., Dubeau, F., Navarro, V., and Gotman, J. (2012). High-frequency oscil-lations (HFOs) in clinical epilepsy. Prog Neurobiol, 98(3):302–315.

[Jacobs et al., 2009] Jacobs, J., Zelmann, R., Jirsch, J., Chander, R., Dubeau, C. E., and Gotman,J. (2009). High frequency oscillations (80-500 Hz) in the preictal period in patients with focalseizures. Epilepsia, 50(7):1780–1792.

[Jacobs et al., 2010] Jacobs, J., Zijlmans, M., Zelmann, R., Chatillon, C. E., Hall, J., Olivier,A., Dubeau, F., and Gotman, J. (2010). High-frequency electroencephalographic oscillationscorrelate with outcome of epilepsy surgery. Ann Neurol, 67(2):209–220.

136

Bibliography

[Jirsch et al., 2006] Jirsch, J. D., Urrestarazu, E., LeVan, P., Olivier, A., Dubeau, F., and Gotman,J. (2006). High-frequency oscillations during human focal seizures. Brain, 129(Pt 6):1593–1608.

[Jones et al., 2000] Jones, M. S., MacDonald, K. D., Choi, B., Dudek, F. E., and Barth, D. S.(2000). Intracellular correlates of fast (>200 Hz) electrical oscillations in rat somatosensorycortex. Journal of neurophysiology, 84:1505–18.

[Kerber et al., 2013] Kerber, K., Levan, P., Dümpelmann, M., Fauser, S., Korinthenberg, R.,Schulze-Bonhage, A., and Jacobs, J. (2013). High frequency oscillations mirror disease ac-tivity in patients with focal cortical dysplasia. Epilepsia, 54:1428–36.

[Khosravani et al., 2009] Khosravani, H., Mehrotra, N., Rigby, M., Hader, W. J., Pinnegar, C. R.,Pillay, N., Wiebe, S., and Federico, P. (2009). Spatial localization and time-dependant changesof electrographic high frequency oscillations in human temporal lobe epilepsy. Epilepsia,50(4):605–616.

[Klostermann, 2005] Klostermann, F. (2005). 500-1000 Hz responses in the somatosensory sys-tem: approaching generators and function. Clinical EEG and neuroscience : official journal ofthe EEG and Clinical Neuroscience Society (ENCS), 36:293–305.

[Klostermann et al., 2000] Klostermann, F., Funk, T., Vesper, J., Siedenberg, R., and Curio, G.(2000). Propofol narcosis dissociates human intrathalamic and cortical high-frequency (> 400hz) SEP components. Neuroreport, 11:2607–2610.

[Klostermann et al., 2002] Klostermann, F., Gobbele, R., Buchner, H., and Curio, G. (2002). In-trathalamic non-propagating generators of high-frequency (1000 Hz) somatosensory evokedpotential (SEP) bursts recorded subcortically in man. Technical report.

[Klostermann et al., 2001] Klostermann, F., Gobbele, R., Buchner, H., Siedenberg, R., and Curio,G. (2001). Differential gating of slow postsynaptic and high-frequency spike-like componentsin human somatosensory evoked potentials under isometric motor interference. Technical re-port.

[Klostermann et al., 1999] Klostermann, F., Nolte, G., and Curio, G. (1999). Multiple generatorsof 600 Hz wavelets in human SEP unmasked by varying stimulus rates. Neuroreport, 10:1625–1629.

[Klostermann et al., 1998] Klostermann, F., Nolte, G., Losch, F., and Curio, G. (1998). Differ-ential recruitment of high frequency wavelets (600 Hz) and primary cortical response (N20) inhuman median nerve somatosensory evoked potentials. Neurosci Lett, 256(2):101–104.

[Knutsson et al., 1998] Knutsson, H., Borga, M., and Landelius, T. (1998). Learning multidi-mensional signal processing. Proceedings. Fourteenth International Conference on PatternRecognition (Cat. No.98EX170), 2.

[Kobayashi et al., 2009] Kobayashi, K., Inoue, T., Watanabe, Y., Oka, M., Endoh, F., Yoshinaga,H., and Ohtsuka, Y. (2009). Spectral analysis of EEG gamma rhythms associated with tonicseizures in Lennox-Gastaut syndrome. Epilepsy Res, 86(1):15–22.

[Kobayashi et al., 2013] Kobayashi, K., Miya, K., Akiyama, T., Endoh, F., Oka, M., Yoshinaga,H., and Ohtsuka, Y. (2013). Cortical contribution to scalp EEG gamma rhythms associatedwith epileptic spasms. Brain Dev, 35(8):762–770.

137

Bibliography

[Kobayashi et al., 2004] Kobayashi, K., Oka, M., Akiyama, T., Inoue, T., Abiru, K., Ogino, T.,Yoshinaga, H., Ohtsuka, Y., and Oka, E. (2004). Very fast rhythmic activity on scalp EEGassociated with epileptic spasms. Epilepsia, 45:488–496.

[Kobayashi et al., 2010] Kobayashi, K., Watanabe, Y., Inoue, T., Oka, M., Yoshinaga, H., andOhtsuka, Y. (2010). Scalp-recorded high-frequency oscillations in childhood sleep-inducedelectrical status epilepticus. Epilepsia, 51(10):2190–2194.

[Kobayashi et al., 2011] Kobayashi, K., Yoshinaga, H., Toda, Y., Inoue, T., Oka, M., and Oht-suka, Y. (2011). High-frequency oscillations in idiopathic partial epilepsy of childhood. Epilep-sia, 52(10):1812–1819.

[Koch, 2004] Koch, C. (2004). Biophysics of Computation: Information Processing in SingleNeurons. Oxford University Press.

[Kojima et al., 2001] Kojima, Y., Uozumi, T., Akamatsu, N., Matsunaga, K., Urasaki, E., andTsuji, S. (2001). Somatosensory evoked high frequency oscillations recorded from subdu-ral electrodes. Clinical neurophysiology : official journal of the International Federation ofClinical Neurophysiology, 112:2261–2264.

[Koles et al., 1995] Koles, Z. J., Lind, J. C., and Soong, A. C. (1995). Spatio-temporal decompo-sition of the EEG: a general approach to the isolation and localization of sources. Electroen-cephalography and clinical neurophysiology, 95:219–230.

[Kubota et al., 2004] Kubota, M., Tran, T. D., Hirose, H., Kimura, I., and Sakakihara, Y. (2004).Patients with benign rolandic epilepsy have a longer duration of somatosensory evoked high-frequency oscillations. Pediatrics international : official journal of the Japan Pediatric Society,46:631–634.

[Lemm et al., 2011] Lemm, S., Blankertz, B., Dickhaus, T., and Müller, K. R. (2011). Introduc-tion to machine learning for brain imaging. NeuroImage, 56:387–399.

[Lemm et al., 2006] Lemm, S., Curio, G., Hlushchuk, Y., and Müller, K.-R. (2006). Enhancingthe signal-to-noise ratio of ICA-based extracted ERPs. Technical report.

[Leski et al., 2013] Leski, S., Linden, H., Tetzlaff, T., Pettersen, K. H., and Einevoll, G. T. (2013).Frequency dependence of signal power and spatial reach of the local field potential. PLoSComput Biol, 9(7):e1003137.

[Li et al., 1999] Li, B., Funke, K., Wörgötter, F., and Eysel, U. T. (1999). Correlated variationsin EEG pattern and visual responsiveness of cat lateral geniculate relay cells. The Journal ofphysiology, 514 ( Pt 3:857–874.

[Liepert et al., 2001] Liepert, J., Haueisen, J., Hegemann, S., and Weiller, C. (2001). Disinhibi-tion of somatosensory and motor cortex in mitochondriopathy without myoclonus.

[Lindén et al., 2010] Lindén, H., Pettersen, K. H., and Einevoll, G. T. (2010). Intrinsic dendriticfiltering gives low-pass power spectra of local field potentials. Journal of computational neu-roscience, 29:423–444.

[Logothetis et al., 2007] Logothetis, N. K., Kayser, C., and Oeltermann, A. (2007). In vivo mea-surement of cortical impedance spectrum in monkeys: implications for signal propagation.Neuron, 55(5):809–823.

138

Bibliography

[Mackert et al., 2000] Mackert, B. M., Weisenbach, S., Nolte, G., and Curio, G. (2000). Rapidrecovery (20 ms) of human 600 Hz electroencephalographic wavelets after double stimulationof sensory nerves. Neuroscience letters, 286:83–86.

[Makeig et al., 1996] Makeig, S., J. Bell., A., Jung, T.-P., and Sejnowski, T. J. (1996). Indepen-dent Component Analysis of Electroencephalographic Data. In Advances in Neural Informa-tion Processing Systems, volume 8, pages 145–151.

[Makeig et al., 2002] Makeig, S., Westerfield, M., Jung, T. P., Enghoff, S., Townsend, J., Courch-esne, E., and Sejnowski, T. J. (2002). Dynamic brain sources of visual evoked responses.Science (New York, N.Y.), 295:690–694.

[Mäkinen et al., 2005] Mäkinen, V., Tiitinen, H., and May, P. (2005). Auditory event-relatedresponses are generated independently of ongoing brain activity. Technical report.

[Mazaheri and Jensen, 2006] Mazaheri, A. and Jensen, O. (2006). Posterior alpha activity is notphase-reset by visual stimuli. Proceedings of the National Academy of Sciences of the UnitedStates of America, 103:2948–2952.

[Metting van Rijn et al., 1990] Metting van Rijn, A. C., Peper, A., and Grimbergen, C. A. (1990).High-quality recording of bioelectric events. Part 1. Interference reduction, theory and practice.Medical & biological engineering & computing, 28:389–397.

[MettingVanRijn et al., 1993] MettingVanRijn, A. C., Kuiper, A. P., Linnenbank, A. C., andGrimbergen, C. A. (1993). Patient isolation in multichannel bioelectric recordings by digi-tal transmission through a single optical fiber. IEEE transactions on bio-medical engineering,40:302–308.

[Michel et al., 2004] Michel, C. M., Murray, M. M., Lantz, G., Gonzalez, S., Spinelli, L., andGrave de Peralta, R. (2004). EEG source imaging. Clinical neurophysiology : official journalof the International Federation of Clinical Neurophysiology, 115:2195–2222.

[Mochizuki et al., 1999] Mochizuki, H., Ugawa, Y., Machii, K., Terao, Y., Hanajima, R., Fu-rubayashi, T., Uesugi, H., and Kanazawa, I. (1999). Somatosensory evoked high-frequencyoscillation in Parkinson’s disease and myoclonus epilepsy. Clinical neurophysiology : officialjournal of the International Federation of Clinical Neurophysiology, 110:185–191.

[Mosher et al., 1999] Mosher, J. C., Baillet, S., and Leahy, R. M. (1999). EEG source localizationand imaging using multiple signal classification approaches. J Clin Neurophysiol, 16:225–238.

[Murakami and Okada, 2006] Murakami, S. and Okada, Y. (2006). Contributions of principalneocortical neurons to magnetoencephalography and electroencephalography signals. J Phys-iol, 575(Pt 3):925–936.

[Murakami et al., 2008] Murakami, T., Sakuma, K., Nomura, T., Nakashima, K., and Hashimoto,I. (2008). High-frequency oscillations change in parallel with short-interval intracortical inhi-bition after theta burst magnetic stimulation. Clinical neurophysiology : official journal of theInternational Federation of Clinical Neurophysiology, 119:301–308.

[Muthukumaraswamy, 2013] Muthukumaraswamy, S. D. (2013). High-frequency brain activityand muscle artifacts in MEG/EEG: a review and recommendations. Front Hum Neurosci,7:138.

139

Bibliography

[Nagel, 2000] Nagel, J. (2000). Biopotential Amplifiers. In Bronzino, J. D. ., editor, The Biomed-ical Engineering Handbook: Second Edition, chapter Chapter 70. Crc press edition.

[Nakano and Hashimoto, 1999] Nakano, S. and Hashimoto, I. (1999). The later part of high-frequency oscillations in human somatosensory evoked potentials is enhanced in aged subjects.Neuroscience letters, 276:83–86.

[Nakano and Hashimoto, 2000] Nakano, S. and Hashimoto, I. (2000). High-frequency oscilla-tions in human somatosensory evoked potentials are enhanced in school children. Neuroscienceletters, 291:113–116.

[Nariai et al., 2011] Nariai, H., Nagasawa, T., Juhasz, C., Sood, S., Chugani, H. T., and Asano, E.(2011). Statistical mapping of ictal high-frequency oscillations in epileptic spasms. Epilepsia,52(1):63–74.

[Netzer, 1981] Netzer, Y. (1981). The design of low-noise amplifiers. Proceedings of the IEEE,69(6):728–741.

[Niedermeyer, 1996] Niedermeyer, E. (1996). Dipole theory and electroencephalography. Clini-cal EEG (electroencephalography), 27:121–131.

[Niedermeyer and Sherman, 2001] Niedermeyer, E. and Sherman, D. L. (2001). Ultrafast fre-quencies: a new challenge for electroencephalography, with remarks on ultraslow frequencies.Clinical EEG (electroencephalography), 32:119–121.

[Nikulin et al., 2007] Nikulin, V. V., Linkenkaer-Hansen, K., Nolte, G., Lemm, S., Müller, K. R.,Ilmoniemi, R. J., and Curio, G. (2007). A novel mechanism for evoked responses in the humanbrain. The European journal of neuroscience, 25:3146–3154.

[Nolte and Dassios, 2005] Nolte, G. and Dassios, G. (2005). Analytic expansion of the EEG leadfield for realistic volume conductors. Physics in medicine and biology, 50:3807–3823.

[Norra et al., 2004] Norra, C., Waberski, T. D., Kawohl, W., Kunert, H. J., Hock, D., Gobbelé, R.,Buchner, H., and Hoff, P. (2004). High-frequency somatosensory thalamocortical oscillationsand psychopathology in schizophrenia. Neuropsychobiology, 49:71–80.

[Nunez and Srinivasan, 2006] Nunez, P. L. and Srinivasan, R. (2006). Electric Fields of theBrain: The Neurophysics of EEG, volume 4. Oxford University Press.

[Okada et al., 1997] Okada, Y. C., Wu, J., and Kyuhou, S. (1997). Genesis of MEG sig-nals in a mammalian CNS structure. Electroencephalography and clinical neurophysiology,103(4):474–85.

[Olivier et al., 1994] Olivier, A., Germano, I. M., Cukiert, A., and Peters, T. (1994). Framelessstereotaxy for surgery of the epilepsies: preliminary experience. Technical note. J Neurosurg,81(4):629–633.

[Oostendorp et al., 2000] Oostendorp, T. F., Delbeke, J., and Stegeman, D. F. (2000). The con-ductivity of the human skull: results of in vivo and in vitro measurements. IEEE transactionson bio-medical engineering, 47:1487–1492.

[Oppenheim et al., 1999] Oppenheim, A. V., Schafer, R. W., and Buck, J. R. (1999). DiscreteTime Signal Processing, volume 1999.

140

Bibliography

[Ozaki and Hashimoto, 2011] Ozaki, I. and Hashimoto, I. (2011). Exploring the physiology andfunction of high-frequency oscillations (HFOs) from the somatosensory cortex. Clin Neuro-physiol, 122(10):1908–1923.

[Pallás-Areny, 1988] Pallás-Areny, R. (1988). Interference-rejection characteristics of biopoten-tial amplifiers: a comparative analysis. IEEE transactions on bio-medical engineering, 35:953–959.

[Parra et al., 2002] Parra, L., Alvino, C., Tang, A., Pearlmutter, B., Yeung, N., Osman, A., andSajda, P. (2002). Linear spatial integration for single-trial detection in encephalography. Tech-nical report.

[Parra et al., 2005] Parra, L. C., Spence, C. D., Gerson, A. D., and Sajda, P. (2005). Recipes forthe linear analysis of EEG. NeuroImage, 28:326–341.

[Peterson et al., 1995] Peterson, N. N., Schroeder, C. E., and Arezzo, J. C. (1995). Neural gen-erators of early cortical somatosensory evoked potentials in the awake monkey. Electroen-cephalography and clinical neurophysiology, 96:248–260.

[Pfurtscheller and Cooper, 1975] Pfurtscheller, G. and Cooper, R. (1975). Frequency dependenceof the transmission of the EEG from cortex to scalp. Electroencephalography and clinicalneurophysiology, 38:93–96.

[Portas et al., 1998] Portas, C. M., Rees, G., Howseman, A. M., Josephs, O., Turner, R., andFrith, C. D. (1998). A specific role for the thalamus in mediating the interaction of attentionand arousal in humans. The Journal of neuroscience : the official journal of the Society forNeuroscience, 18:8979–8989.

[Ramoser et al., 2000] Ramoser, H., Müller-Gerking, J., and Pfurtscheller, G. (2000). Optimalspatial filtering of single trial EEG during imagined hand movement. Technical report.

[Restuccia et al., 2003] Restuccia, D., Marca, G. D., Valeriani, M., Rubino, M., Paciello, N., Vol-lono, C., Capuano, A., and Tonali, P. (2003). Influence of cholinergic circuitries in generationof high-frequency somatosensory evoked potentials. Clinical Neurophysiology, 114(8):1538–1548.

[Restuccia et al., 2007] Restuccia, D., Ulivelli, M., De Capua, A., Bartalini, S., and Rossi, S.(2007). Modulation of high-frequency (600 Hz) somatosensory-evoked potentials after rTMSof the primary sensory cortex. Eur J Neurosci, 26(8):2349–2358.

[Restuccia et al., 2002a] Restuccia, D., Valeriani, M., Grassi, E., Gentili, G., Mazza, S., Tonali,P., and Mauguière, F. (2002a). Contribution of GABAergic cortical circuitry in shaping so-matosensory evoked scalp responses: specific changes after single-dose administration oftiagabine. Technical report.

[Restuccia et al., 2002b] Restuccia, D., Valeriani, M., Grassi, E., Mazza, S., and Tonali, P.(2002b). Dissociated changes of somatosensory evoked low-frequency scalp responses and600 Hz bursts after single-dose administration of lorazepam. Brain research, 946:1–11.

[Rivet et al., 2009] Rivet, B., Souloumiac, A., Attina, V., and Gibert, G. (2009). xDAWN algo-rithm to enhance evoked potentials: application to brain-computer interface. IEEE transactionson bio-medical engineering, 56:2035–2043.

141

Bibliography

[Rosanova and Timofeev, 2005] Rosanova, M. and Timofeev, I. (2005). Neuronal mechanismsmediating the variability of somatosensory evoked potentials during sleep oscillations in cats.The Journal of physiology, 562:569–582.

[Rossini et al., 1985] Rossini, P. M., Basciani, M., Di Stefano, E., Febbo, A., and Mercuri, N.(1985). Short-latency scalp somatosensory evoked potentials and central spine to scalp propa-gation characteristics during peroneal and median nerve stimulation in multiple sclerosis. Elec-troencephalography and clinical neurophysiology, 60:197–206.

[Sakuma et al., 2004] Sakuma, K., Takeshima, T., Ishizaki, K., and Nakashima, K. (2004). So-matosensory evoked high-frequency oscillations in migraine patients.

[Sakura et al., 2009] Sakura, Y., Terada, K., Usui, K., Baba, K., Usui, N., Umeoka, S., Yam-aguchi, M., Matsuda, K., Tottori, T., Mihara, T., Nakamura, F., and Inoue, Y. (2009). Very high-frequency oscillations (over 1000 Hz) of somatosensory-evoked potentials directly recordedfrom the human brain. Journal of clinical neurophysiology : official publication of the Ameri-can Electroencephalographic Society, 26:414–421.

[Sauseng et al., 2007] Sauseng, P., Klimesch, W., Gruber, W. R., Hanslmayr, S., Freunberger,R., and Doppelmayr, M. (2007). Are event-related potential components generated by phaseresetting of brain oscillations? A critical discussion. Neuroscience, 146(4):1435–1444.

[Scheer et al., 2011] Scheer, H. J., Fedele, T., Curio, G., and Burghoff, M. (2011). Extension ofnon-invasive EEG into the kHz range for evoked thalamocortical activity by means of very lownoise amplifiers. Physiol Meas, 32(12):N73–9.

[Scheer et al., 2006] Scheer, H. J., Sander, T., and Trahms, L. (2006). The influence of amplifier,interface and biological noise on signal quality in high-resolution EEG recordings. PhysiolMeas, 27(2):109–117.

[Schevon et al., 2009] Schevon, C. A., Trevelyan, A. J., Schroeder, C. E., Goodman, R. R., McK-hann Jr., G., and Emerson, R. G. (2009). Spatial characterization of interictal high frequencyoscillations in epileptic neocortex. Brain, 132(Pt 11):3047–3059.

[Shah et al., 2004] Shah, A. S., Bressler, S. L., Knuth, K. H., Ding, M., Mehta, A. D., Ulbert,I., and Schroeder, C. E. (2004). Neural dynamics and the fundamental mechanisms of event-related brain potentials. Cerebral cortex (New York, N.Y. : 1991), 14:476–483.

[Shimazu et al., 2000] Shimazu, H., Kaji, R., Tsujimoto, T., Kohara, N., Ikeda, A., Kimura, J.,and Shibasaki, H. (2000). High-frequency SEP components generated in the somatosensorycortex of the monkey. Neuroreport, 11:2821–2826.

[Speckmann and Elger, 2005] Speckmann, E.-J. and Elger, C. E. (2005). Introduction to the neu-rophysiological basis of the EEG and DC Potentials. In Electroencephalography - Basic prin-ciples, clinical applications, and related fields, pages 17–30.

[Staba et al., 2004] Staba, R. J., Wilson, C. L., Bragin, A., Jhung, D., Fried, I., and Engel Jr.,J. (2004). High-frequency oscillations recorded in human medial temporal lobe during sleep.Ann Neurol, 56(1):108–115.

[Steriade, 2001] Steriade, M. (2001). Impact of network activities on neuronal properties in cor-ticothalamic systems. Journal of neurophysiology, 86:1–39.

142

Bibliography

[Steriade et al., 1993] Steriade, M., Curró Dossi, R., and Contreras, D. (1993). Electrophysio-logical properties of intralaminar thalamocortical cells discharging rhythmic (approximately40 HZ) spike-bursts at approximately 1000 HZ during waking and rapid eye movement sleep.Neuroscience, 56:1–9.

[Stockwell et al., 1996] Stockwell, R., Mansinha, L., and Lowe, R. (1996). Localization of thecomplex spectrum: the S transform. IEEE Transactions on Signal Processing, 44(4):998–1001.

[Tanosaki et al., 2002] Tanosaki, M., Kimura, T., Takino, R., Iguchi, Y., Suzuki, A., Kurobe, Y.,Haruta, Y., Hoshi, Y., and Hashimoto, I. (2002). Movement interference attenuates somatosen-sory high-frequency oscillations: contribution of local axon collaterals of 3b pyramidal neu-rons. Technical report.

[Tao et al., 2007] Tao, J. X., Baldwin, M., Ray, A., Hawes-Ebersole, S., and Ebersole, J. S.(2007). The impact of cerebral source area and synchrony on recording scalp electroen-cephalography ictal patterns. Epilepsia, 48(11):2167–2176.

[Telenczuk et al., 2011] Telenczuk, B., Baker, S. N., Herz, A. V., and Curio, G. (2011). High-frequency EEG covaries with spike burst patterns detected in cortical neurons. J Neurophysiol,105(6):2951–2959.

[Telenczuk et al., 2010] Telenczuk, B., Nikulin, V. V., and Curio, G. (2010). Role of neuronalsynchrony in the generation of evoked EEG/MEG responses. J Neurophysiol, 104(6):3557–3567.

[Thévenin, 1883] Thévenin, L. (1883). Extension de la loi dâOhm aux circuits électromoteurscomplexes. Annales Télégraphiques, 10:222–224.

[Urasaki et al., 2002] Urasaki, E., Genmoto, T., Akamatsu, N., Wada, S.-i., and Yokota, A.(2002). The effects of stimulus rates on high frequency oscillations of median nervesomatosensory-evoked potentials–direct recording study from the human cerebral cortex.Technical report.

[Urrestarazu et al., 2007] Urrestarazu, E., Chander, R., Dubeau, F., and Gotman, J. (2007). Inter-ictal high-frequency oscillations (100-500 Hz) in the intracerebral EEG of epileptic patients.Brain, 130(Pt 9):2354–2366.

[Urrestarazu et al., 2006] Urrestarazu, E., Jirsch, J. D., LeVan, P., Hall, J., Avoli, M., Dubeau,F., and Gotman, J. (2006). High-frequency intracerebral EEG activity (100-500 Hz) followinginterictal spikes. Epilepsia, 47(9):1465–1476.

[Usui et al., 2010] Usui, N., Terada, K., Baba, K., Matsuda, K., Nakamura, F., Usui, K., Tottori,T., Umeoka, S., Fujitani, S., Mihara, T., and Inoue, Y. (2010). Very high frequency oscillations(over 1000 Hz) in human epilepsy. Clin Neurophysiol, 121(11):1825–1831.

[Vanhatalo et al., 2005] Vanhatalo, S., Voipio, J., and Kaila, K. (2005). Full-band EEG (FbEEG):an emerging standard in electroencephalography. Clinical neurophysiology : official journalof the International Federation of Clinical Neurophysiology, 116:1–8.

[Virtanen et al., 1997] Virtanen, J., Parkkonen, L., Ilmoniemi, R. J., Pekkonen, E., and Näätänen,R. (1997). Biopotential amplifier for simultaneous operation with biomagnetic instruments.Medical & biological engineering & computing, 35:402–408.

143

Bibliography

[Waberski et al., 2004] Waberski, T. D., Norra, C., Kawohl, W., Thyerlei, D., Hock, D., Kloster-mann, F., Curio, G., Buchner, H., Hoff, P., and Gobbele, R. (2004). Electrophysiologicalevidence for altered early cerebral somatosensory signal processing in schizophrenia. Psy-chophysiology, 41(3):361–366.

[Wang et al., 2013] Wang, S., Wang, I. Z., Bulacio, J. C., Mosher, J. C., Gonzalez-Martinez, J.,Alexopoulos, A. V., Najm, I. M., and So, N. K. (2013). Ripple classification helps to localizethe seizure-onset zone in neocortical epilepsy. Epilepsia, 54(2):370–376.

[Waterstraat et al., 2012] Waterstraat, G., Telenczuk, B., Burghoff, M., Fedele, T., Scheer,H. J., and Curio, G. (2012). Are high-frequency (600Hz) oscillations in human somatosen-sory evoked potentials due to phase-resetting phenomena? Clinical Neurophysiology,123(10):2064–2073.

[Winter and Webster, 1983] Winter, B. B. and Webster, J. G. (1983). Reduction of interferencedue to common mode voltage in biopotential amplifiers. IEEE transactions on bio-medicalengineering, 30:58–62.

[Wolters and Munck, 2007] Wolters, C. and Munck, J. (2007). Volume conduction.

[Wood et al., 1995] Wood, D. E., Ewins, D. J., and Balachandran, W. (1995). Comparative analy-sis of power-line interference between two- or three-electrode biopotential amplifiers. Medical& biological engineering & computing, 33:63–68.

[Worrell et al., 2008] Worrell, G. A., Gardner, A. B., Stead, S. M., Hu, S., Goerss, S., Cascino,G. J., Meyer, F. B., Marsh, R., and Litt, B. (2008). High-frequency oscillations in humantemporal lobe: simultaneous microwire and clinical macroelectrode recordings. Brain : ajournal of neurology, 131:928–937.

[Wu et al., 2008] Wu, J. Y., Koh, S., Sankar, R., and Mathern, G. W. (2008). Paroxysmal fastactivity: an interictal scalp EEG marker of epileptogenesis in children. Epilepsy Res, 82(1):99–106.

[Wu et al., 2010] Wu, J. Y., Sankar, R., Lerner, J. T., Matsumoto, J. H., Vinters, H. V., and Math-ern, G. W. (2010). Removing interictal fast ripples on electrocorticography linked with seizurefreedom in children. Neurology, 75(19):1686–1694.

[Yamada et al., 1988] Yamada, T., Kameyama, S., Fuchigami, Y., Nakazumi, Y., Dickins, Q. S.,and Kimura, J. (1988). Changes of short latency somatosensory evoked potential in sleep.Electroencephalography and clinical neurophysiology, 70:126–136.

[Yamazaki et al., 2012] Yamazaki, M., Tucker, D. M., Fujimoto, A., Yamazoe, T., Okanishi, T.,Yokota, T., Enoki, H., and Yamamoto, T. (2012). Comparison of dense array EEG with simul-taneous intracranial EEG for interictal spike detection and localization. Epilepsy Res, 98(2-3):166–173.

[Yanagihara, K., Ishikawa, T., Imajo, 2008] Yanagihara, K., Ishikawa, T., Imajo, K. (2008).Multi-channel biological signal telemetry systems.

[Yazicioglu et al., 2008] Yazicioglu, R., Merken, P., Puers, R., and Hoof, C. V. (2008). A 200eight-channel EEG acquisition ASIC for Ambulatory EEG Systems. 2008 IEEE InternationalSolid-State Circuits Conference - Digest of Technical Papers.

144

Bibliography

[Zijlmans et al., 2009] Zijlmans, M., Jacobs, J., Zelmann, R., Dubeau, F., and Gotman, J. (2009).High frequency oscillations and seizure frequency in patients with focal epilepsy. EpilepsyRes, 85(2-3):287–292.

[Zijlmans et al., 2012] Zijlmans, M., Jiruska, P., Zelmann, R., Leijten, F. S., Jefferys, J. G., andGotman, J. (2012). High-frequency oscillations as a new biomarker in epilepsy. Ann Neurol,71(2):169–178.

145