Neurophysiology of Fast Oscillations: Investigation of Physiological and Pathological States PDF Free Download

Name: Neurophysiology of Fast Oscillations: Investigation of Physiological and Pathological States PDF
Author: m.burgess

1 / 224

0 views•224 pages

Neurophysiology of Fast Oscillations: Investigation of Physiological and Pathological States PDF Free Download

Neurophysiology of Fast Oscillations: Investigation of Physiological and Pathological States PDF free Download. Think more deeply and widely.

Neurophysiology of Fast Oscillations: Investigation of

Physiological and Pathological States

Tamir Avigdor

Integrated Program in Neuroscience Department of Neurology and Neurosurgery McGill University,

Montréal, Canada


 
Table of contents 
Abstract ........................................................................................................................................................ 4 
Acknowledgments ........................................................................................................................................ 8 
Author Contributions ................................................................................................................................... 9 
Statement of Originality ............................................................................................................................. 11 
Chapter 1: Background ....................................................................................................................... 12 
1 Electrophysiology ............................................................................................................................. 12 
1.1 EEG generation .......................................................................................................................... 13 
1.2 EEG Recordings .......................................................................................................................... 14 
1.3 EEG processing & analysis ......................................................................................................... 17 
1.4 EEG as an oscillatory measurement ......................................................................................... 17 
1.5 Electrical source imaging........................................................................................................... 21 
2 Epilepsy ............................................................................................................................................. 26 
2.1 Definition and prevalence ......................................................................................................... 26 
2.2 Electrophysiology of epilepsy ................................................................................................... 27 
2.3 Drug-resistant epilepsy ............................................................................................................. 30 
2.4 Presurgical evaluation and localization .................................................................................... 30 
2.5 Invasive surgical therapies ........................................................................................................ 31 
3 Sleep ................................................................................................................................................. 31 
3.1 Physiology and function ............................................................................................................ 33 
3.2 Electrophysiology of sleep and sleep staging ........................................................................... 34 
3.3 Sleep macro- and microstructure ............................................................................................. 35 
4 Fast activity ....................................................................................................................................... 37 
4.1 Role in pathology ...................................................................................................................... 38 
4.2 Role in physiology ..................................................................................................................... 40 
Chapter 2: Manuscript #1: Fast oscillations> 40 Hz localize the epileptogenic zone: An electrical 
source imaging study using high-density electroencephalography ......................................................... 42 
1 Preface .......................................................................................................................................... 42 
2 Abstract......................................................................................................................................... 43 
3 Introduction .................................................................................................................................. 44 
4 Methods ........................................................................................................................................ 46 
5 Results........................................................................................................................................... 55 
6 Discussion ..................................................................................................................................... 62 


 
7 Supplementary material .............................................................................................................. 68 
Chapter 3: Manuscript #2: Consistency of electrical source imaging in presurgical evaluation of 
epilepsy across different vigilance states .................................................................................................. 71 
1 Preface .............................................................................................................................................. 71 
2 Abstract ............................................................................................................................................ 72 
3 Introduction ...................................................................................................................................... 73 
4 Methods ............................................................................................................................................ 75 
5 Results .............................................................................................................................................. 80 
6 Discussion ......................................................................................................................................... 90 
7 Supporting Information ................................................................................................................... 94 
Chapter 4: Manuscript #3: The Awakening Brain is Characterized by a Widespread and 
Spatiotemporally Heterogeneous Increase in High Frequencies ........................................................... 104 
1 Preface ............................................................................................................................................ 104 
2 Abstract .......................................................................................................................................... 105 
3 Introduction .................................................................................................................................... 106 
4 Results ............................................................................................................................................ 108 
5 Discussion ....................................................................................................................................... 119 
6 Materials and Methods .................................................................................................................. 127 
7 Supporting Information ................................................................................................................. 140 
Chapter 5 Manuscript #4: Spectral and network investigation reveals distinct power and 
connectivity patterns between phasic and tonic REM ........................................................................... 161 
1 Preface ............................................................................................................................................ 161 
2 Abstract .......................................................................................................................................... 162 
3 Introduction .................................................................................................................................... 163 
4 Methods .......................................................................................................................................... 165 
5 Results ............................................................................................................................................ 174 
6 Discussion ....................................................................................................................................... 182 
7 Supporting Information ................................................................................................................. 188 
Discussion ......................................................................................................................................... 200 
Reference .............................................................................................................................................. 206 
 
 

Abstract

The study of the human brain can be broadly divided into two domains: the study of the

healthy brain and that of the diseased brain. Both offer a different understanding of how the

brain works. Electrophysiology allows us to explore neuronal activity with high temporal and

spatial resolution. Electrophysiological measurements are usually characterized by their

oscillatory patterns. These different oscillatory components are suggested to underlie diverse

brain processes in different states. The goal of this thesis is to investigate patterns of transient

and broad-band fast oscillations (FOs) >30Hz in both physiological and pathological conditions,

as modeled by sleep for the physiological state and focal epilepsy for the pathological state.

Sleep was selected as it provides multiple differently defined states of physiological signals,

which allow testing hypotheses in various known conditions. Epilepsy was selected as it

provides a diverse array of pathologies in various brain regions that can be very focal and affect

a small region, thus making it the optimal pathological model to investigate brain function.

Finally, sleep and epilepsy intersect, as it has been shown that localizable electrical epileptiform

signals are often amplified during sleep, thus helping the clinical evaluation of the patient. In

this thesis I utilized electroencephalography (EEG) recordings, consisting of both non-invasive

scalp (1) high-density EEG (hdEEG), which offers a good spatial coverage of the brain, as well as

invasive (2) intracranial EEG (iEEG), which offers high focal accuracy to investigate FOs in

physiological and pathological conditions. This thesis tackles FOs from two angles: two studies

are based on pathological states using epilepsy as a model and two studies are on physiology

with sleep as a model. In chapter 2, I investigated the feasibility of non-invasively detecting FOs

>40Hz of epileptic patients using hdEEG (Avigdor et al., 2020), demonstrating that it is feasible

to detect and localize pathological FOs during non-rapid-eye-movement (NREM) sleep. I

showed that source localization of FOs corresponds to the epileptogenic zone, and might point

to a superficial generator rather than a deep generator. In chapter 3, I explored whether the

source of interictal epileptic activity is affected by the vigilance state (Avigdor et al., 2024).

Interictal markers are more frequent during NREM sleep. For FOs, this tendency is especially

pronounced, making them difficult to detect in other vigilance states. This raises the question

about the consistency of the source of FOs across different vigilance states. The FOs we

localized in Chapter 2 appeared at the times of interictal epileptiform discharges (IEDs), which,

although more frequent during NREM sleep, can also be detected in other vigilance states.

Therefore, in Chapter 3, I use IEDs as an example of an interictal marker to investigate their

spatial consistency across different states of vigilance. I observed that while properties of IED

changed in the sensor space, there was no apparent effect on the source space, suggesting that

source localization of IED is at best minimally by the state of vigilance. In chapter 4, I turned to

investigating the role of FOs during morning awakenings (Avigdor et al., 2025a) from sleep

using iEEG. The use of iEEG allows the investigation of higher frequencies, which, by using scalp

EEG, might be masked by noise. I observed that FOs play a role in waking up, specifically very

high frequencies above 80Hz seem to increase during the process. This increase seems to be

present from awakening both from REM and NREM sleep, and displays temporally and spatially

heterogeneous patterns, varying between regions and networks. Chapter 5 deals with

investigating the role of FOs in phasic and tonic REM sleep using iEEG (Avigdor et al., 2025b). I

analyzed segments of phasic and tonic REM periods, and discovered a spectral gradient where

tonic REM is strong in the low frequencies and phasic REM dominated in the high frequencies.

Moreover, high frequencies displayed a larger effect and were a distinguishing factor between

tonic and phasic REM. In summary, the results presented in this thesis support the emerging

importance of FOs in the human brain in both physiological and pathological conditions. FOs

have been shown to be a potential noninvasive biomarker for epilepsy and to have a newly

emerging role in sleep research.

Résumé

L'étude du cerveau humain peut être globalement divisée en deux domaines : l'étude du

cerveau sain et celle du cerveau malade. Les deux offrent une compréhension différente de la

façon dont le cerveau fonctionne. L'électrophysiologie nous permet d'explorer l'activité

neuronale avec une résolution temporelle et spatiale élevée. Les mesures

électrophysiologiques sont généralement caractérisées par leurs composantes oscillatoires. On

suggère que ces différentes composantes oscillatoires sous-tendent divers processus cérébraux

dans différents états. L'objectif de cette thèse est d'étudier les oscillations rapides (FOs)

supérieures à 30 Hz dans des conditions à la fois physiologiques et pathologiques, en prenant le

sommeil comme modèle pour l'état physiologique et l'épilepsie focale pour l'état pathologique.

Le sommeil a été choisi car il offre plusieurs états différents de signaux physiologiques, ce qui

permet de tester des hypothèses dans diverses conditions connues. L'épilepsie a été choisie

parce qu'elle fournit un éventail diversifié de pathologies affectant différentes régions du

cerveau, pouvant être très focales et toucher une petite zone, en faisant ainsi le modèle

pathologique optimal pour étudier le fonctionnement cérébral. Enfin, le sommeil et l'épilepsie

s’entrelancent, car il a été montré que des signaux électriques de l’épilepsie sont souvent

amplifiés pendant le sommeil, aidant ainsi à l'évaluation clinique du patient. Dans cette thèse,

j'ai utilisé des enregistrements d'électroencéphalographie (EEG), comprenant à la fois (1) un

EEG haute densité (hdEEG) non invasif, qui offre une bonne couverture spatiale du cerveau, et

(2) un EEG intracrânien (iEEG) invasif, qui offre une grande précision focale pour étudier les FOs

dans des conditions physiologiques et pathologiques. Cette thèse aborde les FOs sous deux

angles: deux études reposent sur des états pathologiques en utilisant l'épilepsie comme modèle

et deux études portent sur la physiologie en utilisant le sommeil comme modèle. Au chapitre 2,

j'ai étudié la faisabilité de détecter de manière non invasive des FOs supérieures à 40 Hz chez

des patients épileptiques à l'aide de l’EEG haute densité (Avigdor et al., 2020), démontrant qu'il

est possible de détecter et de localiser des FOs pathologiques pendant le sommeil à ondes

lentes (NREM). J'ai montré que la localisation de la source des FOs correspond à la zone

épileptogène et pourrait indiquer un générateur superficiel plutôt qu'un générateur profond.

Au chapitre 3, j'ai exploré si la source de l'activité épileptique intercritique est influencée par

l'état de vigilance (Avigdor et al., 2024). Les marqueurs intercritiques sont plus fréquents

pendant le sommeil NREM. Pour les FOs, cette tendance est particulièrement marquée, les

rendant difficiles à détecter dans d'autres états de vigilance. Cela soulève une question sur la

cohérence de la source des FOs dans différents états de vigilance. Les FOs que nous avons

localisées au chapitre 2 sont apparues au moment de décharges épileptiformes intercritiques

(IEDs), qui, bien que plus fréquentes pendant le sommeil NREM, peuvent également être

détectées dans d'autres états de vigilance. Par conséquent, au chapitre 3, j'utilise les IEDs

comme exemple de marqueur intercritique pour étudier leur cohérence à travers différents

états de vigilance. J'ai observé que, bien que le signal des IEDs change dans l'espace des

capteurs, il n'y a aucun effet apparent dans l'espace des sources, de sorte que le générateur

épileptique de l'activité épileptique intercritique n'est pas influencé par l'état de vigilance. Au

chapitre 4, je me suis tourné vers l'investigation du rôle des FOs lors des éveils matinaux

(Avigdor et al., 2025a) à partir du sommeil en utilisant l'iEEG. L'utilisation de l'iEEG permet

d'étudier des fréquences plus élevées qui, avec l'EEG de surface, pourraient être masquées par

l’activité musculaire. J'ai observé que les FOs jouent un rôle dans le réveil, en particulier les

fréquences très élevées au-dessus de 80 Hz semblent augmenter pendant le processus. Cette

augmentation semble être présente lors du réveil à partir à la fois du sommeil paradoxal (REM)

et NREM, et montre des schémas temporellement et spatialement hétérogènes, variant entre

régions et réseaux. Le chapitre 5 traite de l'investigation du rôle des FOs dans le sommeil REM

phasique et tonique en utilisant l'iEEG (Avigdor et al., 2025b). J'ai analysé des segments de

périodes REM phasiques et toniques et découvert un gradient spectral où le REM tonique est

fort dans les basses fréquences et le REM phasique domine dans les hautes fréquences. De

plus, les hautes fréquences montraient un effet plus important et constituaient un facteur

distinctif entre le REM tonique et phasique. En résumé, les résultats présentés dans cette thèse

soutiennent l'importance émergente des FOs dans le cerveau humain, à la fois dans des

conditions physiologiques et pathologiques. Il a été démontré que les FOs peuvent constituer

un biomarqueur non invasif potentiel pour l'épilepsie, et qu'un rôle nouvellement émergent se

dessine également dans la recherche sur le sommeil.

Acknowledgments

I am deeply grateful for the support and guidance of Professors Birgit Frauscher and Christophe

Grova during these years. Prof. Frauscher’s mentorship was invaluable; her dedication, late

nights, and weekends spent ensuring her students' success were truly exceptional. She

introduced me to the world of translational research, and it was transformative to see the

methods we use in the lab applied in clinical settings. I will always be thankful for being

accepted into her group; it's remarkable how one of her papers in Brain (2018) influenced my

entire career trajectory. Prof. Grova's methodological guidance was crucial, and his great

insights taught me how to perform rigors quantitate research. When I started, I had little

knowledge of source localization, but he patiently ensured that I and all his students truly

understood the concepts and methodologies behind our work. I deeply appreciate both

Professors for allowing me to grow in their labs. I had the honor to have Prof. Jean Gotman as

scientific mentor, his advice on how to conduct meaningful and rigorous research while keeping

the big picture in mind was transformative. I was also fortunate to collaborate with Prof. Laure

Peter-Derex on the 4th project of this thesis and collaboration on other papers. Prof. Peter-

Derex was consistently helpful, insightful, and collaborative. I extend my thanks to my

collaborators across various projects: Dr. Chifaou Abdallah for her indispensable assistance with

all clinical neurophysiological aspects. Dr. Nicolas von Ellenrieder provided invaluable technical

and theoretical advice for many projects. Prof. Jean-Marc Lina offered extensive guidance on

source localization in various projects. Many thanks to Jawata Afnan who helped with

methodological part for the second project. I would like to express my gratitude to Jessica

Royer and Prof. Boris Bernhardt who provided high-quality processed imaging data for project

one and two. Special thanks to Dr. Katharina Schiller for her assistance with the 4th project and

numerous external collaborations. I would also like to thank my fellow lab mates, who made

these years the best of my life: Alyssa Ho, Erica Minato, Vojtech Travnicek, Edward Delaire,

Sana Hanan, John Thomas, Mahdi Abadi, Valentina Hrtonova, Barbora Matouskova, Veronique

Latreille, Arielle Dascal, Ella Sahlas, Petr Klimes, Zhengchen Cai, Daniel De Latorre. Finally,

especially to Yingqi Wang and Kassem Jaber, working with each of you on your respective

projects brought me immense joy and fulfillment even more then working on my own projects.

Author Contributions

Manuscript #1: Fast oscillations> 40 Hz localize the epileptogenic zone: An electrical source

imaging study using high-density electroencephalography

Authors: Tamir Avigdor, Chifaou Abdallah, Nicolás von Ellenrieder , Tanguy Hedrich, Annalisa

Rubino, Giorgio Lo Russo, Boris Bernhardt, Lino Nobili, Christophe Grova, Birgit Frauscher. Clin

Neurophysiol. 2021 Feb;132(2):568-580

Tamir Avigdor conducted all the source localizations and fast oscillation detection and analysis,

with ideation originally by Birgit Frauscher and methodology development by Christophe Grova.

Chifaou Abdallah, Nicolás von Ellenrieder, Tanguy Hedrich, and Boris Bernhardt provided

advisory support in clinical and technical matters. Annalisa Rubino, Giorgio Lo Russo, , and Lino

Nobili collected the data.

Manuscript #2: Consistency of electrical source imaging in presurgical evaluation of epilepsy

across different vigilance states

Tamir Avigdor*, Chifaou Abdallah*, Jawata Afnan, Zhengchen Cai, Saba Rammal, Christophe

Grova, Birgit Frauscher. Ann Clin Transl Neurol. 2024 Feb;11(2):389-403

Tamir Avigdor conducted all the source localizations and analysis. Chifaou Abdallah conducted

clinical analysis and along with Birgit Frauscher identified interictal epileptiform discharges,

which were initially assesed by Saba Rammal. The ideation was initially suggested by Birgit

Frauscher and subsequently developed by Christophe Grova, Tamir Avigdor, and Chifaou

Abdallah.

Manuscript #3: The awakening brain is characterized by a widespread and spatiotemporally

heterogeneous increase in high frequencies

Tamir Avigdor, Guoping Ren, Chifaou Abdallah, François Dubeau, Christophe Grova, Birgit

Frauscher. Advanced Science. 2025; 2409608.

Tamir Avigdor conducted the analysis, with ideation originally suggested by Birgit Frauscher and

further developed by Tamir Avigdor and Christophe Grova. Birgit Frauscher performed all

awakening and electrophysiology assessments, with wakefulness segments selected by Chifaou

Abdallah and Guoping Ren, François Dubeau referred the patients, and all helped in editing the

manuscript.

Manuscript #4: Spectral and network investigation reveals distinct power and connectivity

patterns between phasic and tonic REM sleep

Tamir Avigdor, Laure Peter-Derex, Alyssa Ho, Katharina Schiller, Yingqi Wang, Chifaou Abdallah,

Edouard Delaire, Kassem Jaber, Vojtech Travnicek, Christophe Grova, Birgit Frauscher. Sleep.

2025 zsaf133.

Tamir Avigdor conducted the analysis, with ideation originally suggested by Birgit Frauscher and

further developed by Tamir Avigdor and Christophe Grova. Birgit Frauscher and Laure Peter-

Derex marked all the phasic and tonic periods. Katharina Schiller and Yingqi Wang prepared the

dataset. Chifaou Abdallah marked the wakefulness segments. Vojtech Travnicek, Alyssa Ho,

Edouard Delaire, Kassem Jaber assisted with creating clearly written methodology and figures.

Statement of Originality

To the best of my knowledge, the work presented in this thesis is novel and original.

In Project 1, I demonstrated for the first time that fast oscillations can be detected and localized

using high-density EEG in an adult drug-resistant epilepsy population. I showed that the

localization of these fast oscillations is not inferior to interictal epileptiform discharges and that

they might indicate a surface-generated epileptic source rather than a deep source.

In Project 2, I addressed the debate in the literature regarding source localization and vigilance

states. I showed that the source localization of interictal epileptiform discharges remains

consistent across vigilance states. While the amplitude and duration of the source space may

change, the origin of the source does not change, and the extent of the source remains the

same from a relative threshold perspective.

In Project 3, I investigated human awakening using intracranial EEG for the first time,

demonstrating that awakening is a heterogeneous local process. I also showed, for the first

time, the role of high frequencies in morning awakening.

In Project 4, I conducted the first multi-lobar investigation of phasic and tonic REM sleep using

intracranial EEG, reporting for the first time the differences of broad band >80hz activity

between phasic and tonic REM periods, and demonstrating the regional heterogeneity of REM

microstructure.

1. Chapter 1: Background

This thesis explores the function of high frequency oscillations in pathology and physiology. The

following background chapter provides the necessary context for understanding the concepts

discussed throughout the work. Each subsequent chapter delves deeply into a specific, current

problem in these fields, offering a focused background on that issue or phenomenon. The thesis

employs electrophysiological methods to investigate fast frequency oscillations in epilepsy and

sleep. Accordingly, this chapter will first survey the basic concepts of electrophysiology to

establish the methods that will be used throughout the work. Subsequently, it will introduce

concepts related to EEG spectral and source analysis, laying the foundation for the

methodological discussions in later chapters. This is followed by an introduction to epilepsy and

sleep research, which provides the essential background needed to understand the specific

questions posed in chapters 2 through 5. Finally, the chapter concludes with a brief overview of

the history and current understanding of fast oscillations in physiology and pathology, a topic

that will be explored further in each of the subsequent chapters.

1.1 Electrophysiology

I chose to explore fast oscillations using electrophysiological methods because analyzing fast

oscillations requires high temporal resolution, a capability currently achievable only with these

techniques. The brain is composed of tens of billions of neurons working together as a network.

The electrical changes in individual neurons during the transmission of information can be

considered as a direct correlate of neuronal processes. These electrical fluctuations of

individual neurons called post synaptic potentials, can be recorded when the activity is

synchronized using electrodes (Louis and Frey, 2016). This ability was first demonstrated in

monkeys by Richard Caton (Caton, 1875), who applied a Galvanometer to a monkey brain. A

few decades later the first human encephalogram (EEG) was recorded (Figure 1) by Hans

Berger (Berger, 1929), observing for the first time human alpha rhythm (8-12Hz). The discovery

of the spectral properties of these EEG waves was later shown to be the result of

synchronization of neuronal activity (Gloor, 1985).

Figure 1. First human EEG from a 15 y.o during wakefulness showing 10 seconds of EEG signal

(Berger, 1929). EEG was one of the first tools for the investigation of the human brain (Dietsch,

1932). To date, it is one of the main tools we use to explore the brain in both physiological and

pathological conditions. Permission to use this figure in the thesis was granted by Springer via

1.1.1 EEG generation

The signal captured by EEG is generated when a group of neurons experiences a postsynaptic

potential change at the same time (Beniczky and Schomer, 2020). The flow of current in the

extracellular medium during the postsynaptic potential is captured as a potential difference,

and when these synchronize between different neurons, we measure a potential change

between our electrodes (Figure 2). This summation results in the recorded potential on the

scalp in the order of microvolts. This potential can be described by the solid angle concept

(Gloor 1985) which states that an EEG electrode’s recorded potential is proportional to the

fraction of the surrounding space (solid angle) that the neural source occupies as seen from the

electrode. In addition, if we placed electrodes inside the brain using intracranial EEG (iEEG) we

are able to record from a more local area of the same phenomena. Using iEEG and

microelectrodes it is possible even to record single unit neuronal firing due to proximity to the

cell body (Im and Seo, 2016).

Figure 2. Generation of an EEG signal in pyramidal cells as the summation of excitatory post-

synaptic potentials (EPSP, in green) and inhibitory post-synaptic potentials (IPSP, in red)

(Beniczky and Schomer, 2020). Permission to use this figure in the thesis was granted by Wiley

via Copyright Clearance Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.1.2 EEG Recordings

In this thesis, chapters 2 and 3 utilize scalp EEG, while chapters 4 and 5 employ intracranial EEG

to explore aspects of epilepsy and sleep. The following section introduces briefly these

methodologies and discusses important considerations for recording signals using these

methods.

1.1.2.1 Scalp EEG

Acquisition of scalp EEG is done routinely these days, usually using the 10-20 international

system (Klem et al., 1999) proposed by Jasper in the 50’s, which uses the nasion, inion, and pre-

auricular landmarks to place 21 electrodes in a reproducible manner. For epilepsy presurgical

evaluation the recommendation is 25 electrodes (Seeck et al., 2017) or 32-64 according to the

American clinical neurophysiology society (Brenner et al., 2008). In addition, many electrode

caps and nets are now available that offer electrode arrays of between 32 to 512 electrodes

with a uniform distribution (Figure 3).

Figure 3. The 10-20 international system. The system aims to cover all lobes of the brain (Seeck

et al., 2017). C– central, F– frontal, P– parietal, T– temporal, O– occipital. Odd numbers- left

hemisphere, even numbers-right hemisphere, z- midline. Permission to use this figure in the

thesis was granted by Elsevier via Copyright Clearance Center, Inc under license to ‘Republish in

a thesis/dissertation’.

Moreover, in recent years, a 10-10 system for high-density (hdEEG) has been introduced to fit

64-83 electrodes in a similar manner (Figure 4).

Figure 4. The 10-10 international system. The system aims to cover all lobes of the brain

extensively (Seeck et al., 2017). C– central, F– frontal, P– parietal, T– temporal, O– occipital. C–

central, F– frontal, P– parietal, T– temporal, O– occipital. Odd numbers- left hemisphere, even

numbers-right hemisphere, z- midline. Permission to use this figure in the thesis was granted by

Elsevier via Copyright Clearance Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.1.2.2 Intracranial EEG

When recording intracranial EEG (iEEG), which is usually done in the context of epilepsy

presurgical evaluation, two methods are common which record from underneath the skull. One

is electrocorticography (ECoG), in which electrodes in grids or strips are placed on top of the

brain surface. The second is intracerebral depth electrodes, which are inserted stereotactically,

hence stereo-EEG (sEEG), into various locations across the brain. SEEG usually comprises an

electrode with multiple recording contacts on it (Figure 5). Overall, the signal of the iEEG is of

high quality, low noise, and can act as a local measurement of neuronal activity (Ball et al.,

2009). However, it lacks spatial coverage due to the limitation of the number of electrodes that

can be inserted, unlike scalp EEG which provides full head coverage.

Figure 5. SEEG electrode implantation (Ye et al., 2022) and recording setups. (A) Surgical

placement of electrode shafts. (B) Example of a 3D brain model with implanted electrode

shafts. (C) A patient and the complete recording system within a clinical setting. Permission to

use this figure in the thesis was granted by Frontiers via Copyright Clearance Center, Inc under

license to ‘Republish in a thesis/dissertation’.

1.1.2.3 Referencing and montages

The choice of reference in both methods is critically important, as we are recording a potential

difference between two electrodes. This means that an event occurring in the reference

electrode will be apparent in all channels referenced to it. Consequently, the selection of

montage i.e., how each recording electrode or contact is referenced is significant, and different

objectives require different types of montages. The initial recording montage is usually a

referential montage, in which all channels are recorded with respect to a single common

reference electrode. A commonly used montage for both scalp and intracranial EEG is the

bipolar montage, in which each electrode is referenced to a nearby electrode, thereby creating

a more localized view. Other montages, such as the common average or other computational

montages, involve computing an average of all channels and subtracting it from each channel,

thus creating a pseudo-monopolar channel. In all chapters, different types of common average

montage are used (see Methods for more details).

1.1.3 EEG processing & analysis

Signals acquired from scalp EEG, ECoG, or iEEG recordings are all measurements of summations

of neuronal activity at different levels of granularity. However, the processing of the data varies

depending on the montage (Hu et al., 2018), type of electrodes (Mathewson et al., 2017; Jobst

et al., 2020), and purpose. These factors all need to be considered to perform a meaningful EEG

analysis. When analyzing standard clinical 10-20 set-ups both a bipolar configuration and a

referential montage may be used. When analyzing sEEG, mainly a local bipolar montage is used,

while a referential montage can aid as an addition. The bipolar montage is often used in clinical

sEEG presurgical evaluation (Zaveri et al., 2006). While it offers a more localized measurement,

it also results in reduced amplitude and possible phase discontinuities due to polarity reversal

(Hu et al., 2007), depending on the source location. Electrical measurements such as EEG

always have an assumed level of noise that needs to be filtered out. However, the choice of

filter must align with factors such as the window of interest, frequency bands of interest, and

study design (Burgess, 2019). When analyzing hdEEG, the choice of montage is usually either a

monopolar referential or a common average montage. In this case of this thesis different types

common average montages were used for all the chapters (see section 1.5).

1.1.4 EEG as an oscillatory measurement

Since this thesis deals with fast oscillations, it is important to understand that EEG is

characterized by oscillatory patterns. The oscillatory pattern was first reported in humans EEG

by Berger (Berger, 1929), which described the alpha rhythm (8-13Hz). In the following years,

more distinct rhythms (Table 1) were observed and correlated with different behaviors and

states. Their role in sleep physiology (Frauscher and Gotman, 2019; Navarro-Lobato and Genzel,

2019; Cox and Fell, 2020) and epilepsy pathology (Sundaram et al., 1999; Frauscher et al., 2017)

remains a primary tool in both clinical practice and research.

Band

Name

Range

(Hz)

Role

Delta

0.3-4

Observed with great amplitude in most brain regions during deep sleep

(Knyazev, 2012) and is considered a marker for unconsciousness

Theta

4-8

Observed during drowsiness periods and mainly generated by the

hippocampus (GREEN and ARDUINI, 1954) and is also associated with

memory formation (Buzsáki and Moser, 2013).

Alpha

8-13

Discovered first by Berger and is associated with wakefulness (with

closed eyes) in occipital regions (Berger, 1929; Foster et al., 2017).

Beta

13-30

Present in frontal regions during wakefulness and associated with

motor functions (Jasper and Penfield, 1949; Khanna and Carmena,

2015) as well as working memory (Schmidt et al., 2019).

Gamma

30-70

Recorded in the mid-part of the century in the visual system (Doty et

al., 1964) was in recent decades linked to consciousness and cognition

(Mably and Colgin, 2018) as well as its relation to pathological

conditions (Fell et al., 2010).

Ripples

80-

250

High-frequency oscillations which were discovered at the end of the

last century (Fisher et al., 1992), and were shown to be a seizure-

independent biomarker of epilepsy (Frauscher et al., 2017), as well as

to have a potential role in memory (Buzsáki and Silva, 2012).

Fast

Ripples

250-

500

Fast Ripples were discovered (Bragin et al., 1999) [Bragin et al., 1999]

and shown to be able of pointing to the EZ.

Very high-

frequency

oscillations

500-

1000

Very high-frequency oscillations were recently discovered in the

context of epilepsy (Usui et al., 2015), and proposed to be an even

more specific biomarker for the EZ (Brázdil et al., 2017).

Table 1. Presented is information on the traditional and more recently introduced frequency

bands. The frequency range and a short descriptive history and function for each band are

provided. These characteristics are seen both in iEEG and hdEEG for the traditional bands, yet

for FO bands currently, most evidence corroborating these characteristics were gathered using

iEEG as they are more difficult to assess in the scalp given the signal-to-noise ratio as well as

confound with muscle artifacts.

In Chapters 3 and 4, I explore sleep physiology using intracranial EEG. Analysis of the signal in

the sensor space is the primer analysis technique. Here I will give a brief introduction into

spectral and connectivity analysis, and in the method sections of Chapters 3 and 4 the specific

adaptation of these methods will be expanded on.

1.1.4.1 Spectral analysis

The patterns discovered in different frequency bands made spectral analysis of immediate

importance to the field. Using the discrete Fourier transform, we can assess the contribution of

each frequency to the signal (Al-Fahoum and Al-Fraihat, 2014). This results in the ability to

analyze the different components of the signal, which are a mixture of activation of different

neuronal populations. This analysis was first performed manually for EEG (Dietsch, 1932). The

computation time of the discrete Fourier transform is large (complexity of O(N²)). Luckily, in

1965, the fast Fourier transform (FFT) was invented by Cooley & Tukey and quickly adopted for

EEG analysis (Brigham, 1974). Since then, computational power has risen significantly, and

today any regular PC can perform EEG spectral analysis in a time-efficient manner. Other

methods offer better spectral estimates, such as Welch's method and multitaper (Babadi and

Brown, 2014), and can also be used to estimate the spectrum over time resulting in a

spectrogram (Figure. 6).

Figure 6. A depiction of an estimation of the spectrum over time using multitaper (Avigdor et

al., 2021) (A) an average of each band (B) a spectrogram (C-E) raw EEG in different time points.

Permission to use this figure in the thesis was granted by Elsevier. via Copyright Clearance

Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.1.4.2 Connectivity analysis

While the brain has regions with unique functions(Lazar, 2010; Fan et al., 2016), it also

functions as a network (Park and Friston, 2013; Bazinet et al., 2023), and connections between

regions are the bedrock of neuroscience from the cellular level to interhemispheric

communication. Connectivity analysis aims to study how different regions behave together over

time, and how groups of regions function in different states. In the context of EEG, the aim is to

assess if the signals from two different sources (different channels) behave in a time-locked

manner, i.e., if over time activity in region A can inform us about the activity in region B. This

helps determine if these two regions are connected and to what degree. Two broad categories

exist in connectivity analysis: one focuses on the time domain, i.e., the raw signal, and the other

on the frequency domain i.e., after some transformation such as Fourier or Hilbert transform.

When assessing connectivity based on the raw or filtered amplitude, the aim is to assess if the

amplitudes of both signals rise and fall between the two signals in a predictable manner.

Measurements of this connectivity can vary from simple correlations to complex autoregressive

models (Wendling et al., 2009; Cao et al., 2022). The latter category also tries to assess similar

connections but uses the transformed data. The first type of frequency domain connectivity

measure suggested was coherence (Guevara and Corsi-Cabrera, 1996), which is analogous to

correlation in the frequency domain. Later-developed connectivity measures (Figure 7), such as

the phase locking value (PLV) (Lachaux et al., 1999) or phase locking index (PLI) (Stam et al.,

2007a), build on this idea and attempt to assess connectivity purely based on the relation

between two signals In this thesis, I will utilize PLV, as intracranial EEG is less contaminated by

volume conduction (see the full description and justification in chapters 3 and 4). The

advantage of this approach is that it is not sensitive to amplitude and can measure exact

temporal dynamics; however, it is more susceptible to common noise.

Figure 7. Illustration of the phase locking value calculation process (Lachaux et al., 1999),

highlighting how the clustering of phase angles is utilized to estimate the consistency of the

phase lag between two signals. It’s achieved by averaging these phase differences over trials,

we obtain a complex value u for each latency t, whose magnitude (abs(u)) represents the

phase-locking value. Permission to use this figure in the thesis was granted by Wiley via

1.1.5 Electrical source imaging

In chapters 2 and 3, I explore the source localization of interictal biomarkers. To estimate the

potential source, I utilized electrical source imaging (ESI), a method that allows for estimating

the underlying brain topography from scalp EEG (Michel and He, 2019). To estimate the

topography of potentials that will give rise to our measurements, we first need to understand

the influence each location in the brain would have on each electrode. This is known as the

forward problem, which is well-posed and can be solved using the quasi-static approximation of

Maxwell’s equations. Solving the forward problem provides us with a gain matrix that informs

us how each potential generator at each point in the brain will affect each of our electrodes. In

addition, we need to have a model of the geometry of the brain, called the head model, this can

be a simple sphere, or a complex realistic model based on imaging. There are two common

approaches: Boundary Element Models (Hamalainen and Sarvas, 1989) and Finite Element

Models (Marino et al., 1993). The Boundary Element Model, which I use in this thesis, uses

surface meshes from Magnetic Resonance Imaging (MRI) segmentations, assuming

homogeneous and isotropic conductivity for the brain, skull, and scalp. The different head

tissues have different connectivity which is particularly important for EEG. In this thesis, I

utilized the following values for the outer skull, and scalp, assigning values of 0.33, 0.0165, and

0.33 S/m respectively (Ferree et al., 2000; Hoekema et al., 2003; Lai et al., 2005). A common

tool for computing the head model is The OpenMEEG BEM solver (Gramfort et al., 2011). After

the creation of the head model co-registration of the scalp electrodes to the scalp surface is

required (Figure 8). Different methods exist of electrode digitization, some manual such as the

Polhemus FastTrack system, and some optical such as the Geodesic Photogrammetry System.

The former is more accurate, and the latter is useful when over 200 channel caps are used. We

then use this gain matrix when attempting to solve the inverse problem. The inverse problem

aims to determine the most reasonable distribution of sources that would produce our

observed measurements. When recording scalp EEG, we measure a summation of activity from

various nearby regions, making the problem non-trivial as we have only a limited external

sample of the actual activity. This problem has been shown to be an ill-posed problem known

as Helmholtz theorem, which requires us to regularize to arrive at a solution. This means we

need to make assumptions about our signal and expected topography and incorporate these

assumptions into the regularizations. Over the years, various methods have been proposed to

achieve the most accurate estimation of the probable topography that would give rise to our

external measurements (Grova et al., 2006; Chowdhury et al., 2013; Michel and Brunet, 2019).

Figure 8. Pipeline of Modern EEG (Michel and He, 2019) Source Imaging: hdEEG (e.g., 256

channels) is recorded with high temporal resolution. Electrode positions are precisely

determined, for instance, using a photogrammetry system (e.g., EGI Inc.). From these

recordings, scalp potential maps are reconstructed for each time instant. Structural MRI of the

subject is acquired, and electrodes are coregistered with the head surface. The brain is

segmented, and solution points are distributed across the gray matter. Finally, a distributed

inverse solution is applied to the EEG map, utilizing an individualized volume conductor model

derived from the MRI data. Permission to use this figure in the thesis was granted by Elsevier

via Copyright Clearance Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.1.5.1 Approaches : Dipole modeling Vs. distributed methods

Generally, the types of approaches to perform ESI and solve the inverse problem has two broad

categories. Dipole modeling (Rose and Ebersole, 2009), which uses either one or a few sources

and tries to only model the location, amplitude and origination of this dipole(s) to arrive with

the best fit that explains the most variation in the signal. Another approach is using distributed

methods (Michel et al., 1999) which assume that all the brain has varied levels of activation at

each time, and that all the sources contribute to the creation of the signal. Dipole modeling can

be accurate when assessing a single or a few very focal sources, while distributed methods offer

a spatiotemporal estimation of the potential source and its spatial extent. However, the latter

can be more computationally demanding.

1.1.5.2 Maximum entropy on the mean

In this thesis, I will utilize only the following framework for ESI; chapters 2 and 3 provide a full

description of the methods encompassed within this framework. The following is an

introduction to the framework: within the category of distrusted methods several approaches

are available (Michel and He, 2019). They are all trying to approach the regularization of the

inverse problem in different ways (Grech et al., 2008). One of these approaches is the

maximum entropy on the mean (MEM) framework (Amblard et al., 2004), which attempts

restate the problem in a Bayesian probabilistic manner. MEM aims to estimate the distribution

that maximizes the uncertainty about missing information in the data relative to the prior

reference distribution. MEM uses a spatial prior, which assumes that brain activity is organized

into cortical parcels which share the same order of magnitude as the number of sensors. When

a parcel is active, a Gaussian distribution is used as its activity prior, while an inactive parcel is

represented by a Dirac distribution that effectively shuts off activity in that parcel (Figure 9).

This allows MEM to activate or deactivate parcels during the localization process, while

allowing for local contrast within the active parcels on the cortical surface. In recent years MEM

was shown to be more sensitive to the spatial extent comparing to other methods such as

minimum norm estimate (Grova et al., 2006; Pellegrino et al., 2020). MEM was further

developed into various techniques, each concentrating on different components of the signal.

This enables more targeted localization depending on the specific type of event being analyzed.

All the source localizations in this thesis were conducted using the MEM framework.

Figure 9. The maximum entropy on the mean framework (a) MEM employs a reference

distribution (Grova et al., 2008), to address the inverse problem by maximizing relative entropy

(or equivalently, minimizing the Kullback-Leibler divergence) between the source distribution

and the reference distribution, while ensuring the data is accurately explained.(b) Reference

distribution is defined as the resting brain activity into each cortical parcels. Permission to use

this figure in the thesis was granted by Elsevier via Copyright Clearance Center, Inc under

license to ‘Republish in a thesis/dissertation’.

1.1.5.2.1 Coherent MEM

Chapter 3 of the thesis utilizes the Coherent MEM (cMEM) method which was first introduced by

Chowdhury et al. in 2013, it incorporates a data-driven spatial prior that remains fixed over time, making

it "coherent." This means that the parcellation of the brain is not recalibrated with each time point, but

instead remains stable across the entire duration of the analysis (Chowdhury et al., 2013; Chowdhury et

al., 2016). In the case of cMEM, the assumption is that brain activity is organized into stable cortical

parcels that persist throughout the experiment or clinical observation period. The effectiveness of

cMEM has been demonstrated for the localization of interictal epileptic activity, where it has shown

considerable sensitivity to the source and the spatial extent of interictal epileptic activity (Grova et al.,

2006).

1.1.5.2.2 Wavelet MEM

Chapter 4 of the thesis utilizes the wavelet MEM (wMEM) approach which utilizes a discrete

wavelet transform (Daubechies wavelets) to capture oscillatory patterns in the data before

employing the MEM solver. wMEM is designed to localize brain oscillatory patterns. By applying

a wavelet transformation wMEM effectively characterizes oscillatory patterns (Lina et al.,

2014). In this method, the time-domain representation of the data is replaced with a time-scale

representation. This approach was shown to be useful to localize fast oscillations using

magnetoencephalography (von Ellenrieder et al., 2016). In this thesis all oscillatory activities

were localized using wMEM. For a full description of cMEM and wMEM please refer to the

method and supplementary sections of chapters 2 and 3.

1.2 Epilepsy

In this thesis, the investigation of fast oscillations in a pathological state employs epilepsy as the

model. Epilepsy was chosen not only because of the well-established connection between fast

oscillations and the disorder but also because electrophysiology is a cornerstone of epilepsy

care. The following section provides a brief introduction to epilepsy as a disorder and to the

role of electrophysiology in its diagnosis. Chapters 2 and 3 will offer a more detailed

background, focusing specifically on the challenges of source localizing epilepsy biomarkers.

Epilepsy is a disorder characterized by recurrent seizures (Scheffer et al., 2017). It is an umbrella

term that describes a symptom rather than a cause, as epilepsy can have various causes,

ranging from traumatic brain injury to tumors to being (Sanchez-Carpintero Abad et al., 2007;

Chen et al., 2018a; Fordington and Manford, 2020).

1.2.1 Definition and prevalence

Epilepsy is one of the most common neurological disorders affecting up to 1% of the population

(Beghi, 2020). Epilepsy can occur at all ages and does not depend on sex (Fiest et al., 2017).

Although seizures have been described past (Asadi-Pooya and Rostami, 2017), modern clinical

practice requires ‘At least two unprovoked (or reflex) seizures occurring more than 24 hours

apart or One unprovoked (or reflex) seizure and a probability of further seizures similar to the

general recurrence risk (at least 60%) after two unprovoked seizures, occurring over the next 10

years or Diagnosis of an epilepsy syndrome’ according to the international league against

epilepsy. The diagnosis of epilepsy will almost always require a scalp EEG to attempt to capture

epileptic activity to confirm the diagnosis.

1.2.2 Electrophysiology of epilepsy

EEG was an early adopted tool in the field, with one of the first tests conducted around 1935 by

Wilder Penfield at the Montreal Neurological Institute. It was observed that during a seizure,

the scalp EEG shows a fast, large-amplitude desynchronized signal. This observation was

important in understanding the origin of the seizure, as the region where the first changes were

apparent is more likely to be the region of cause. This area was later named the ictal or seizure

onset zone (SOZ), which serves as an approximation of the epileptogenic zone (EZ). The EZ is

the theoretical area of the cortex essential for generating epileptic seizures. It includes the

region that produces seizures before surgery and any adjacent tissue that might cause seizures

if not removed (Rosenow and Lüders, 2001). As the field developed, more electrical biomarkers

were discovered, such as interictal epileptiform discharges (IED) or spikes, which define the

irritative zone (IZ) which is a region that generates IEDs (Rosenow and Lüders, 2001). These

different regions overlap, but they are not identical (Figure 10).

Figure 10. Overlaps of the different epileptic zones (Loddenkemper, 2010). Depicted are a

hypothetical example of different types of epileptic zones and their overlap with the

epileptogenic zones. the Permission to use this figure in the thesis was granted by Springer via

IEDs are brief (<200ms) bursts of abnormal brain activity seen on EEGs which occur between

seizures (Figure 11). IED are abnormal electrical events arising from the synchronous firing of

hyperexcitable neurons and are associated with epilepsy (Ward, 1960). One hypothesis is that

increased neural excitability promotes these spikes, which may accumulate to a critical spatial

or temporal density and trigger a seizure (Jensen and Yaari, 1988). Another hypothesis suggests

a causal relationship between interictal spikes and ictogenesis (Avoli et al., 2002).

Figure 11. One of the first reported interictal epileptiform discharges (Ward, 1960). Depicted is

a spontaneous interictal activity of a single cell in a monkey (dash line =10 msec) Permission to

use this figure in the thesis was granted by Wiley via Copyright Clearance Center, Inc under

license to ‘Republish in a thesis/dissertation’.

In recent years IEDs with fast oscillations as a component have been identified(Ren et al., 2015)

(Thomas et al., 2023; Shi et al., 2024) as additional biomarkers for the EZ (Figure 12). However,

most studies on IEDs with fast oscillations come from iEEG, where they are easier to detect. In

this thesis, I attempted to noninvasively localize the fast oscillatory component of these signals.

For further details, please refer to the Fast Oscillation section and Chapter 2.

Figure 12. Depiction of spike gamma (Thomas et al., 2023) (A) The morphology of a spike is

illustrated by identifying peak N1, with P1 and P2 as the preceding and succeeding troughs, and

N2 as the succeeding crest. Spike features such as rise/fall amplitude, slope, and duration were

calculated. Spikes were extracted as 300-ms segments (−75 to +225 ms relative to N1) for

Teager energy evaluation. (B) Preceding gamma activity is shown in two patients with a 600-ms

SEEG segment centered on the spike and its scalogram. Gamma activity, observed just before

the spike, was detected using a 30–100 Hz bandpass filter and Morse wavelet decomposition.

Permission to use this figure in the thesis was granted by Wiley via Copyright Clearance Center,

Inc under license to ‘Republish in a thesis/dissertation’.

1.2.3 Drug-resistant epilepsy

Following an epilepsy diagnosis, the first attempt to control seizures is with antiseizure

medications (Kanner and Bicchi, 2022). If the seizures are not well controlled, the patient might

be ‘drug-resistant’ defined as “Drug resistant epilepsy may be defined as failure of adequate

trials of two tolerated and appropriately chosen and used antiepileptic drugs schedules

(whether as monotherapies or in combination) to achieve sustained seizure freedom.”

according to the international league against epilepsy (Kwan et al., 2010). This definition of is

supported by evidence indicating that medical drug resistance can often be recognized after

two unsuccessful trials of antiseizure medications. With each failed medications trial, the

probability of achieving seizure control through medication alone decreases. Ultimately, up to

30% of patients will remain inadequately controlled by medication (Dalic and Cook, 2016). This

proportion seems to persist despite new anti-epileptic drugs becoming available (Kwan and

Palmini, 2017). These patients will then be defined as having refractory epilepsy and will be

considered for surgical options such as resection, minimally invasive ablation or

neuromodulation (Rugg-Gunn et al., 2020).

1.2.4 Presurgical evaluation and localization

When an epilepsy patient is diagnosed with drug-resistant epilepsy by an epileptologist the

patient may start a presurgical evaluation (Rosenow and Lüders, 2001) to determine if they will

benefit from an invasive procedure. Presurgical evaluation begins with a phase 1 assessment

that includes high-resolution MRI, video scalp EEG, and a detailed neuropsychological

evaluation (Baumgartner, 2012). If these tests clearly localize the epileptogenic zone and

adequately assess postoperative risks, the patient may proceed directly to surgery. However, if

the results are ambiguous, additional invasive iEEG test might be offered in phase 2 using

defending if there is a presumed focality (Astner-Rohracher et al., 2022). Phase 2 may also

include auxiliary test such as hdEEG, ESI, Positron Emission Tomography, Single Photon

Emission, Functional Magnetic Resonance Imaging (fMRI), Computed Tomography, EEG-fMRI.

All these tests are conducted with the aim of determining if a single or multiple sources may be

generating the seizures. As well as for assessing with which degree of confidence a potential

source for the EZ can be delineated and what is the extent. If deemed positive, the patient may

be offered a surgical option aimed at treating the potential EZ with the goal of achieving seizure

freedom or, as a palliative option, to reduce seizure burden and improve the patient’s quality of

life.

1.2.5 Invasive surgical therapies

Invasive procedures can include resection (Rugg-Gunn et al., 2020), where the hypothesized

epileptogenic area is surgically removed, or minimally invasive ablative surgery such as

thermoregulation (Pollandt and Bleck, 2018), and laser ablation (Grewal and Tatum, 2019)

these methods are used to ablate a small target tissue, in contrast to a large resection.

Additionally, neuromodulation techniques such as deep brain stimulation (Fisher, 2023),

responsive neurostimulation (Skarpaas et al., 2019), or Vagus nerve stimulation (Gonzalez et al.,

2019) involve implanting devices to deliver electrical current, thus modulating or disrupting the

activity in targeted regions in order to reduce seizure frequency. However, in most cases only a

resection or ablation of the complete epileptogenic area will result in seizure-freedom. In

Chapter 2, I will utilize patients who have undergone resective surgeries and were rendered

seizure-free (Engel 1A) to determine if the localization of fast oscillation can adequately assess

the epileptogenic zone, as defined by the resection cavity of the surgery, as the ground truth.

1.3 Sleep

Sleep plays a central role throughout this work. Chapter 2 examines fast oscillations during

sleep, and Chapter 3 investigates the effects of sleep and wakefulness on interictal markers,

Chapter 4 explores the termination of the sleep process (i.e., awakening), and Chapter 5 delves

into subtle aspects of fast activity during sleep. The introduction to human sleep research

provides an in-depth overview of current perspectives on sleep and its interactions with

epilepsy, as further detailed in the introductions to Chapters 3–5. Sleep is a ubiquitous

phenomenon across all species (Miyazaki et al., 2017), yet its role and purpose are still debated.

Human sleep has been a subject of electrographic investigation since 1935, when Loomis

started his laboratory and discovered the macro- and microstructure of sleep (Figure 13),

specifically non-rapid eye movement sleep (NREM) (Loomis et al., 1935), followed by the

discovery of REM sleep in 1953 (Aserinsky and Kleitman, 1953) (Figure 14).

Figure 13. First recording NREM sleep (Loomis et al., 1935). (A) the subject is asleep, and

marked trains of brain rhythms are observed in response to the stimuli. (B), when the subject is

awake, no such rhythms are present despite identical sound stimulation. Time in seconds is

indicated by dots along the top. Permission to use this figure in the thesis was granted by

Science|AAAS via Copyright Clearance Center, Inc under license to ‘Republish in a

thesis/dissertation’.

Since then, human sleep has been the subject of investigation (Carley and Farabi, 2016) and

found to play an important role in memory (Rasch and Born, 2013), attention (Hudson et al.,

2020), and behavior (Zhu et al., 2024).

Figure 14. First recording of REM sleep (Aserinsky and Kleitman, 1953). Sample recording

showing rapid eye movements in a sleeping subject. RV represents vertical leads on the right

eye, RH indicates horizontal leads on the right eye, and RF corresponds to the right frontal EEG

lead. Permission to use this figure in the thesis was granted by Science|AAAS via Copyright

Clearance Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.3.1 Physiology and function

Sleep is a structured process occurring in regular intervals, which vary between species and are

governed by the circadian rhythm of the species (Miyazaki et al., 2017). In mammals, sleep is

controlled by a system located in the brainstem called the Ascending Reticular Activating

System (ARAS), discovered by Moruzzi and Magoun in the 1940s (Moruzzi and Magoun, 1949).

This system starts with the reticular formation in the brainstem (Figure 15) and projects dorsally

through the midbrain and relays via the thalamus to the cortex. They observed that animals in a

deeply sedated or sleeping state would become alert when a stimulus would be applied to this

region. They further demonstrated that if these ascending connections were disrupted, the

animal lapsed into something akin to a permanent sleep-like state.

Figure 15. A schematic depicts the key components of the ascending arousal system

(Eggermont, 2014). It highlights major structures, including the brainstem, hypothalamic

regions, and basal forebrain areas. Permission to use this figure in the thesis was granted by

Wiley via Copyright Clearance Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.3.2 Electrophysiology of sleep and sleep staging

During sleep the brain changes its electrical signature. The signature changes from a low-

voltage desynchronized activity during wakefulness to a slow-wave synchronized activity along

with an additional low-voltage desynchronized sleep (Navarro-Lobato and Genzel, 2019). In

humans, sleep is divided into two main stages (Figure 16): NREM sleep and REM sleep

(Academy of Sleep Medicine, 2020). NREM sleep is further categorized into three stages: N1 is

the ‘lightest’ stage, marking the transition from wakefulness to sleep. It is characterized by

theta waves and low voltage. N2 is a deeper stage where heart rate slows, body temperature

drops, and patterns such as sleep spindles and K complexes occur. N3, also known as deep

sleep or slow-wave sleep, is characterized by high amplitude delta activity and large slow

waves. In contrast, REM sleep characterized by wake-like EEG with accompanied periods of

rapid eye movements and atonia. REM sleep can be subdivided into two microstates: (1) phasic

REM, characterized by rapid eye movements, and (2) tonic REM, which does not have these eye

movements

Figure 16. Sleep stages and a full night hypnogram (Fröhlich, 2016). EEG

(electroencephalogram) assesses brain activity and neural oscillations, EOG (electrooculogram)

distinguishes REM from non-REM sleep, and EMG (electromyogram) measures muscle tone.

Permission to use this figure in the thesis was granted by Elsevier via Copyright Clearance

Center, Inc under license to ‘Republish in a thesis/dissertation’.

1.3.3 Sleep macro- and microstructure

Sleep progresses cyclically, with each cycle lasting approximately 90 minutes. This is often

referred to as the sleep architecture or macrostructure. A typical cycle begins with NREM sleep,

moving sequentially from N1 to N3, before transitioning to REM sleep. The cycle then repeats,

with the proportion of REM sleep increasing in subsequent cycles as the night progresses. Sleep

also has a distinct microstructure (Academy of Sleep Medicine, 2020), which are the transient

events within its broader macrostructure. The key microstructures are presented in Table 2 and

Figure 17. Throughout this work, micro- and macrostructural information was used to score

sleep data. Chapter 5 focuses on the distinction between phasic and tonic REM sleep, using

REMs as indicators, forming the basis of the study. This chapter also provides a detailed review

of current research on the differences between phasic and tonic REM sleep.

ATTRIBUTE

DESCRIPTION

Sleep

Spindles

Bursts of rhythmic brain activity occurring predominantly during N2 sleep. They

last 0.5–3 seconds and have a frequency of 10–16 Hz.

K-

Complexes

Large slow waves that appear during N2 sleep, often in response to external

stimuli or spontaneously. Characterized by a sharp upward deflection followed by

a slower downward wave.

Delta

Waves

High-amplitude, low-frequency waves (0.5–4 Hz) that dominate during N3 sleep.

Rapid Eye

Movements

The defining feature of REM sleep, characterized by quick, simultaneous

movements of both eyes. Associated with vivid dreaming and increased brain

activity. EEG patterns during REM resemble wakefulness with a mix of alpha, beta,

and theta waves.

Sawtooth

Waves

A hallmark of REM sleep, appearing as low-amplitude, triangular-shaped waves on

an EEG.

Arousal

Events

Brief episodes of increased faster activity, often triggered by internal or external

stimuli.

Cyclic

Alternating

Pattern

(CAP)

A periodic fluctuation between phases of stable and unstable sleep within NREM

sleep.

Table 2. Description of sleep microstructure elements

Figure 17. Sleep microstructure events and their pace in the macrostructure (Pan et al., 2021).

On the left, the EEG characteristics associated with each stage are detailed. Abbreviations: REM

(rapid eye movement sleep), N1 (non-REM sleep stage 1), N2 (non-REM sleep stage 2), and N3

(non-REM sleep stage 3). Permission for the use of this figure was given by MDPI under the

journal’s open access policy.

1.4 Fast activity

This thesis investigates the role of fast oscillations in both pathology and physiology. Chapters 2

and 3 deal with the pathological aspects of fast oscillations and interictal markers, while

Chapters 4 and 5 explore the role of fast oscillations during sleep. The following section

provides an introduction to research on fast frequencies in both physiology and pathology, and

each subsequent chapter offers an in-depth overview of the current state of research on fast

oscillations as it pertains to its specific focus. Generally, fast activity falls into two categories:

transitory fast oscillations, i.e., brief events lasting on the order up to tens of milliseconds

(Frauscher et al., 2017), also known as high-frequency oscillations , which I will refer to as fast

oscillations from now on, and broadband activity above 30 Hz (Herrmann et al., 2010), spanning

the gamma band and upward, which I will refer to as high frequency activity. The original band

was discovered and described in 1938 by Jasper and Andrews (Jasper and Andrews, 1938) in the

human cortex, as well as in the visual cortex of monkeys in 1964 (Hughes, 1964). The

generation of high frequency activity was linked to the notion that individual neurons can

oscillate in a narrow band across various frequencies, potentially aiding information coding

(Buzsaki and Draguhn, 2004). However, progress in understanding fast oscillations was slow

due to their difficulty in being observed using non-invasive methods (Muthukumaraswamy,

2013). Fast escalation have been shown, especially in the gamma band, to be correlated with

FMRI activity (Foucher et al., 2003; Nir et al., 2007) and associated interneuronal firing. In

recent years, the role of fast oscillations in physiology as an important contributor to memory

(Kucewicz et al., 2014), attention (Ray et al., 2008), cognition (Kaiser and Lutzenberger, 2005),

and consciousness (Kahn et al., 1997) has emerged. Simultaneously, the role of fast oscillations

in pathology has started to be elucidated, especially in the field of epilepsy (Frauscher et al.,

2017).

1.4.1 Role in pathology

First used in the context of epilepsy in 1992 fast oscillations were linked to the seizure onset

zone (Allen et al., 1992). Later (Bragin et al., 1999), fast oscillation events up to 500 Hz in

mesiotemporal lobe epilepsy were observed in patients using microelectrode . They were

classified as ripples (80–250 Hz) and fast ripples (>250 Hz). Gotman’s group later demonstrated

that fast oscillations up to 500 Hz could be recorded with macroelectrodes (Jirsch et al., 2006).

Fast oscillations are commonly observed alongside interictal epileptiform discharges (Figure

18).

Figure 18. Fast oscillations (Frauscher et al., 2018a) from a 34-year-old female undergoing

presurgical evaluation with stereo EEG at the Montreal Neurological Institute and Hospital.

Permission to use this figure in the thesis was granted by Wiley via Copyright Clearance Center,

Inc under license to ‘Republish in a thesis/dissertation’.

Three patterns were identified (Urrestarazu et al., 2007): (1) 64% occurred with spikes and were

visible as riding on the spike in unfiltered signals, (2) 17% occurred with spikes but were

invisible in unfiltered signals, and (3) 19% occurred independently of spikes in timing and

location. These fast oscillations were claimed to be highly specific to the SOZ (Jacobs et al.,

2012; Wang et al., 2013), demonstrating a 95% specificity Fast ripples identified 52% of SOZs,

ripples 38%, and spikes 33%. Ripples co-occurring with spikes were even more SOZ-specific than

ripples alone. Resection of regions with high rates of fast oscillations was shown to be a positive

prognostic for good outcome after surgery (van Klink et al., 2014), although these claims were

recently challenged (Zweiphenning et al., 2022). Most information regarding fast oscillations in

epilepsy comes from invasive iEEG studies. Chapter 2 of this thesis focuses on translating these

findings from iEEG into non-invasive hdEEG (Avigdor et al., 2020). For a more detailed

description of non-invasive attempts to localize fast oscillations, refer to the Introduction

Chapter 2. In addition, fast oscillations have an interaction with sleep similarly to other

interictal biomarkers. They were shown to be most frequent during NREM sleep (Staba et al.,

2004), and are least frequent during REM sleep and wakefulness. Lastly, fast oscillations have

been demonstrated to have potential diagnostic value in as Alzheimer’s disease (Lisgaras and

Scharfman, 2023) , schizophrenia (Goda et al., 2015), bipolar disorder (Garcia-Rill et al., 2019)

multiple sclerosis (Gobbelé et al., 2003), and Parkinson’s disease (Johnson et al., 2021).

1.4.2 Role in physiology

The study of the relation between physiology and high frequencies mainly focused on the

gamma band, yet even fast oscillations can be observed in healthy tissue (Figure 19).

Figure 19. Ripple rates cortical map (Frauscher et al., 2018a). The top panel presents the 95th

percentile ripple rate for each brain region, while the bottom panel displays individual channel

rates. Each dot represents a channel, with its size and color indicating the ripple rate. Views are

provided for the lateral (left) and medial (right) cortex. Permission to use this figure in the

thesis was granted by Wiley via Copyright Clearance Center, Inc under license to ‘Republish in a

thesis/dissertation’.

It was shown to be an evoked response for many sensory domains, such as the auditory

(Pantev, 1995), visual (Chen and Herrmann, 2001; Hoogenboom et al., 2006), somatosensory

(Chen and Herrmann, 2001), and olfactory system (Eeckman and Freeman, 1990). The

ubiquitous presence of high frequency activity, associated with nearly all human cognitive

functions, suggests that it may be fundamentally important to brain function. After further

exploration it was discovered that high frequency activity was also linked to memory (Kucewicz

et al., 2014; Kucewicz et al., 2024), language (Bastiaansen and Hagoort, 2006), cognition (Mably

and Colgin, 2018), and visual awareness (Schurger et al., 2006). Most recently it was suggested

that even consciousness (Rusalova, 2006; Ferrari-Marinho et al., 2020; Juan et al., 2023) might

be linked to high frequency activity. The latter was always in dialog with the phenomena of

sleep, which demonstrate a natural state of altered consciousness. Not surprisingly, sleep itself

was linked to high frequency activity, yet counterintuitive considering the basic intuition about

sleep as a “slower” state. It was shown to be able to distinguish between NREM, REM and

wakefulness (Gross and Gotman, 1999), to have local variability in power and connectivity

(Cantero et al., 2004), and was suggested to even play a role in REM microstates (Gross and

Gotman, 1999; Nishida et al., 2005; Simor et al., 2020) . Taken together, high frequency activity

appears to be involved in nearly every human function

In summary, the correlation between fast oscillations and high-frequency broadband activity

with both pathology and physiology, particularly in the contexts of epilepsy and sleep, is well

established. Epilepsy and sleep not only exhibit a bidirectional relationship but also share

common methodologies as disciplines. These findings underscore the importance of studying

high-frequency activity in both pathological and physiological contexts. Consequently, this

thesis will investigate transient fast oscillations in epilepsy as a model for pathology, and

broadband high-frequency activity in sleep as a model for physiology.

2. Chapter 2: Manuscript #1: Fast oscillations> 40 Hz localize

the epileptogenic zone: An electrical source imaging study

using high-density electroencephalography

Published as: Tamir Avigdor, Chifaou Abdallah, Nicolás von Ellenrieder , Tanguy Hedrich,

Annalisa Rubino, Giorgio Lo Russo, Boris Bernhardt, Lino Nobili, Christophe Grova, Birgit

Frauscher. Clin Neurophysiol. 2021 Feb;132(2):568-580

2.1 Preface

I started by examining the pathological aspects of high frequencies. It has been shown that fast

oscillations might act as biomarkers for the epileptogenic zone. However, most evidence comes from

sEEG, which is capable of detecting fast oscillations due to the relatively low noise recording enabled by

the intracranial nature of sEEG. The goal was to determine if fast oscillations could be detected in a non-

invasive manner and, if so, whether they could be used to source localize the epileptogenic zone. For

this purpose, I used a cohort of patients who underwent successful surgery with seizure-free outcome

(Engel 1A). Using high-density EEG, I analyzed interictal epileptiform discharges during NREM sleep and

attempted to identify fast oscillations at the time of these interictal discharges. Once identified, I applied

a source localization method sensitive to spatial extent to assess the ability of these fast oscillations to

indicate the epileptogenic zone. I observed that fast oscillations are detectable on scalp EEG and that

source localization of these events points to the epileptogenic zone. Moreover, the ability to detect and

localize these events was feasible only for superficial sources rather than deep ones, suggesting that the

presence of these events on scalp EEG might indicate a superficial generator. In summary, fast

oscillations are observable using scalp EEG, and their source localization using HD-EEG is feasible. This

study demonstrated the possibility of non-invasively using fast oscillations as part of epilepsy surgical

evaluations.

2.2 Abstract

Objective

Fast Oscillations (FO) >40 Hz are a promising biomarker of the epileptogenic zone (EZ). Evidence

using scalp electroencephalography (EEG) remains scarce. We assessed if electrical source

imaging of FO using 256-channel high-density EEG (HD-EEG) is useful for EZ identification.

Methods

We analyzed HD-EEG recordings of 10 focal drug-resistant epilepsy patients with seizure-free

postsurgical outcome. We marked FO candidate events at the time of epileptic spikes and

verified them by screening for an isolated peak in the time-frequency plot. We performed

electrical source imaging of spikes and FO within the Maximum Entropy of the Mean

framework. Source localization maps were validated against the surgical cavity.

Results

We identified FO in five out of 10 patients who had a superficial or intermediate deep

generator. The maximum of the FO maps was localized inside the cavity in all patients (100%).

Analysis with a reduced electrode coverage using the 10–10 and 10–20 system showed a

decreased localization accuracy of 60% and 40% respectively.

Conclusions

FO recorded with HD-EEG localize the EZ. HD-EEG is better suited to detect and localize FO than

conventional EEG approaches.

Significance

This study acts as proof-of-concept that FO localization using 256-channel HD-EEG is a viable

marker of the EZ.

2.3 Introduction

Epilepsy is a chronic condition characterized by recurrent seizures accompanied by negative

impact on quality of life (Hinnell et al., 2010). A significant number of 30% of patients with focal

epilepsy are drug-resistant, and these numbers did not change over the past 30 years despite

the development of multiple new antiepileptic drugs (Chen et al., 2018b). The therapy of choice

for focal drug-resistant epilepsy is epilepsy surgery (Vakharia et al., 2018). The aim of epilepsy

surgery is to remove the epileptogenic zone (EZ) defined as the area of cortex that needs to be

removed to achieve seizure freedom (Rosenow and Lüders, 2001). In current practice, the

seizure-onset zone (SOZ) is used as the main proxy marker for the EZ. A sustained seizure-free

condition is currently, however, obtained in only 50% of carefully selected patients (Krucoff et

al., 2017; Mohan et al., 2018; West et al., 2019). This is likely due to inaccurate localization of

the EZ or a network involvement larger than initially expected (Englot, 2018). This underlines

the need to develop new markers and localization techniques for better identification of the EZ

and hence improved surgical outcomes. Recently, Fast Oscillations (FO) have been identified as

promising novel interictal marker for the EZ (Frauscher et al., 2017). Most evidence comes from

intracranial electroencephalography (iEEG) suggesting that resection of areas with high FO rates

is associated with good surgical outcome, and that presence of FO after resection is predictive

of postsurgical seizure relapse (Höller et al., 2015; Frauscher et al., 2017; van 't Klooster et al.,

2017; Jacobs et al., 2018; Tamilia et al., 2018). In contrast, evidence from non-invasive methods

is rather scarce (Thomschewski et al., 2019), and combination of FO detection with source

localization was mainly obtained from magnetoencephalography (MEG). MEG studies from

different centers suggest that it is possible to detect and then localize isolated FO events > 40

Hz from interictal MEG recordings with satisfactory data quality (Papadelis et al., 2016; van

Klink et al., 2016a; von Ellenrieder et al., 2016). Four studies validated FO localization against

the postsurgical cavity in good outcome patients as gold standard for the EZ; they pointed

towards usefulness of FO for delineating the EZ (van Klink et al., 2017; Velmurugan et al., 2019;

Yin et al., 2019; Tamilia et al., 2020). In contrast to MEG, high-density EEG (HD-EEG) would have

the great advantage that it is easy to maintain and more affordable given the high cost of

operating a MEG device. From a technical point of view, however FO recording and localization

using HDEEG has even more challenges to overcome. The main hurdle is a more complex

solution of the forward problem needed for source localization given the necessity of assessing

accurately the conductivity of the head tissues, especially the skull (Goldenholz et al., 2009; de

Munck et al., 2012; Harri and Puce, 2017; Ilmoniemi and Sarvas, 2019). There is first evidence

that FO can be detected also in the scalp EEG (Andrade-Valenca et al., 2011; Kobayashi et al.,

2011; von Ellenrieder et al., 2012b; Zelmann et al., 2014; Pizzo et al., 2016; van Klink et al.,

2016b). Some of these studies showed concordance with the area of the seizure-onset zone

(SOZ) as determined by intracranial EEG or the resection cavity in patients with good seizure

outcome (Kuhnke et al., 2018; Kuhnke et al., 2019; Tamilia et al., 2020). Moreover, it was

shown that presence of scalp FO is associated with the depth of the epileptic generator, with

FO being present in case of a generator located in more superficial cortical regions (Cuello-

Oderiz et al., 2017). Finally, it was suggested that scalp EEG might be superior for the detection

of FO due to higher rates of FO detected in scalp EEG compared to MEG (van Klink et al., 2019;

Tamilia et al., 2020). Given the small size of the generators of FO (von Ellenrieder et al., 2014),

HD-EEG might then be superior to standard EEG. So far, there is only one study reporting scalp

FO using HD-EEG with 128 electrodes in the sensor space (Kuhnke et al., 2018). In this proof-of

concept study, we assessed the feasibility of detection and localization of FO in epileptic

patients using a HDEEG array of 256 electrodes. We examined whether localization of FO > 40

Hz using HD-EEG is capable of delineating the EZ (Rosenow and Lüders, 2001). We then

compared results to standard spike source localization and to conventional EEG approaches

using the 10–10 or 10–20 EEG system. For validation, we opted to use the surgical cavity in

postsurgically seizure-free patients (Engel Ia) with a follow-up duration of > 2 years as best

approximation of the EZ. We opted to not use the intracranially identified SOZ as alternative

validation standard, given that a considerable number of patients do not become seizure-free

after removal of the SOZ (Krucoff et al., 2017).

2.4 Methods

Subject selection

A total of ten patients with seizure-free outcome > 2 years (Engel 1A (Engel et al., 1993)) after

epilepsy surgery procedure were selected from the HD-EEG database of the Claudio Munari

Epilepsy Center in Milan between 2015–2016. Selection criteria were presence of > 10 epileptic

spikes during HD-EEG recording, presence of non-rapid eye movement (NREM) sleep in order to

have the least artifacts possible for FO detection (Zijlmans et al., 2017), absence of previous

surgery, availability of electrode positions and magnetic resonance imaging (MRI) co-

registration and satisfactory data quality. Fig. 1 presents the flowchart for patient selection.

Note that data quality was a negligible reason for exclusion. Patient demographic and clinical

information is provided in Table 1. The study conforms to the Declaration of Helsinki and was

approved by Niguarda Hospital in Milan, Italy. A written informed consent was signed by all

patients prior to study participation.

HD-EEG data acquisition and preprocessing

HD-EEG was recorded using a 256-electrode EEG system (Electrical Geodesic Inc., EGI, now

Magstim EGI, Eden Prairie, MN, USA) with a sampling rate of 500 Hz and hardware filter settings

of 0.3 Hz for the high pass and 200 Hz for the low pass filter. The recordings were performed

with long-term EEG nets using gel and lasted approximately 1.5 h; this duration was chosen to

enable the patient to fall asleep. The impedance of the selected electrodes was kept under 1

kΩ. Cz was the recording reference for this study. For analysis, we created an average

referential montage. HD-EEG processing was performed with the Brainstorm software package

(Tadel et al., 2011) For interictal spike detection and analysis, preprocessing included 0.3–70 Hz

band-pass filtering and direct current correction (baseline window from 1000 ms to 500 ms

before the marked spikes).

Fig. 1. Flowchart describing the selection process from the complete high-density

electroencephalogram database of surgical epilepsy patients comprising 87 patients to the final

10 patients analyzed in this study.

No filter was applied for FO events. The EEG sensor positions were estimated using

digitalization with a SofTaxicOptic system (EMS s.r.l., Bologna, Italy). A linear coregistration with

a pre-implant MRI (Achieva 1.5 T, Philips Healthcare) was performed. The digitized positions of

the electrodes were then coregistered to the scalp surface segmented from the anatomical MRI

of each patient, using a surface matching algorithm within the Brainstorm software. The

accuracy of the resulting registration was then verified visually. The HD-EEG electrodes located

on the face and on the neck (~40 channels) were excluded for further analysis in order to avoid

artifacts caused by muscle or poor-contact electrodes (Hedrich et al., 2017).

Interictal event marking

Spikes were marked at their peak by an epileptologist (CA). All spike events occurring during

N1, N2 or N3 sleep were screened for FO candidate events by a second epileptologist (BF). FO

were defined as at least four oscillations clearly standing out of the background EEG in the

gamma (40–80 Hz) and ripple (80–160 Hz) frequency band in the same electrodes as the spike

that was marked at that time. FO candidate events were verified using a time–frequency

representation screened for an isolated peak within the FO band (gamma or ripple) with no

other visible peaks in the time frequency (TF) plot in the same threshold within 0.5 s before and

0.5 s after (Fig. 2). TF representations were computed using the Morlet wavelet with a central

frequency of 1 Hz and a window duration of 3 s and plotted as a power scale. We used this two-

step approach in order to avoid misclassification of the filtering effect as ‘‘true” FO. Given the

low signal to noise ratio (SNR) of FO, we attempted source localization only in patients with a

minimum number of 5 FO in either the 40–80 or 80–160 Hz window.

Table 1

Demographic and clinical information of the study sample

Sex

Age

Epilepsy

type

Epilepsy

onset

(years)

EEG ictal

EEG

interictal

MRI

Surgery

Size of

the

cavity

Pathology

Patients with Fast Oscillations

TPO

R T

R PT

R hemisphere atrophy with suprasylvian

maximum

TPO disconnection

212.6

cm2

gliosis

L F

L F polar

L anterior frontal

resection

59.81

cm2

FCD IIa

R T

R T ganglioglioma

RT resection and

lesionectomy

59.98

cm2

ganglioglioma

R F

dorsolateral

R F

dorsolateral

FCD R precentral sulcus

R F resection

48.39

cm2

gliosis

FTI

R FT

bilateral T, R

> L

L T pole cyst

R FT insular resection

83.58

cm2

n.a.

Patients without Fast Oscillations

R T

R FT

R T resection

52.26

cm2

FCD IIa

L FCP

L T pole hypoplasia, L HS, L hemispheric

atrophy

L T lobectomy plus

insula

47.64

cm2

gliosis

Cingulate

L FCT

L F

dorsolateral

L central cingulate lesion

L central cingulate

lesionectomy

4.69

cm2

tuberous

sclerosis

L F

L T

L F perisylvian gliosis

L orbitofrontal

lesionectomy

64.11

cm2

FCD IIb

L FCT

L F polymicrogyria + increase volume of L

amygdala

L T resection

47.5

cm2

Legend. C, central; EEG, electroencephalography; MRI, magnetic resonance imaging; F, frontal;

FCD, focal cortical dysplasia; FO, fast oscillation; HS, hippocampal sclerosis; I, insula; L, left; M,

man; na, not available; O, occipital; P, parietal; R, right; T, temporal; W, woman. The size of the

surgical cavity was estimated as area (cm2 ) along the cortical surface.

Electrical source imaging

The 1.5 T anatomical MRI was segmented, and the cortical surface was obtained using the

FreeSurfer software package (v = 6.0, http://surfer.nmr.mgh.harvard.edu). The EEG forward

problem was solved using the boundary element method (BEM) (Kybic et al., 2006). The gain

matrix was calculated using a 3-layer BEM model for brain, skull and scalp (conductivity of 0.33,

0.0165, 0.33 S/m respectively) using OpenMEEG (Gramfort et al., 2010) implemented in the

Brainstorm software. The inverse problem was solved using the Maximum Entropy on the

Mean (MEM) method (Amblard et al., 2004) MEM is a nonlinear distributed inverse problem

method, for which the prior model is built using a data-driven parcellation (DDP) technique in

order to cluster the cortical surface into K parcels (Lapalme et al., 2006). To do so, we used the

multivariate source pre-localization (MSP) method (Mattout et al., 2005) which is a projection

method that estimates a coefficient, which characterizes the possible contribution of each

dipolar source to the data. We make use of it in the MEM reference model (or prior model), in

which a hidden variable is associated to each parcel in order to model the probability of the

parcel to be active or to be switched off. This method will be referred to as the coherent MEM

(cMEM), which was carefully evaluated for its ability to recover the spatial extent of the

underlying generators (Chowdhury et al., 2013; Chowdhury et al., 2016; Hedrich et al., 2017;

Pellegrino et al., 2018; Pellegrino et al., 2020). The wavelet based MEM extension (wMEM) of

the MEM framework (Lina et al., 2014) decomposes the signal in a discrete wavelet basis before

performing MEM source localization on each TF box. Therefore, wMEM is particularly suited to

localize oscillatory patterns, as evaluated with realistic simulations, but also when localizing

FO(von Ellenrieder et al., 2016), oscillatory patterns at seizure onset(Pellegrino et al., 2016) or

resting state ongoing oscillations (Aydin et al., 2020). These two different variations of MEM

were used in this study: wavelet MEM (wMEM) was adopted to localize FO, whereas cMEM

(cMEM) was adopted to localize spikes. We used the standard settings of the wMEM and cMEM

as provided in Brainstorm(Chowdhury et al., 2013; Lina et al., 2014), except that for wMEM we

set the amount of vanishing moments for the Daubechies wavelet to 8. The use of 8 vanishing

moments instead of the default 4 was done in order to capture the complexity of FO. The

baseline to model the diagonal noise covariance matrix for both wMEM and cMEM was chosen

for each patient visually during an artifact- and spike-free 2-second period. wMEM was

performed on the marked duration of the FO by selecting the only TF box exhibiting the largest

amount of energy as recommended (von Ellenrieder et al., 2016). To select which TF boxes

were considered as localizable using the wMEM framework, a global-spatial wavelet power was

first computed for each time–frequency box, by summing up the energy of each wavelet

coefficient over all the sensors. Total power was then estimated by considering all frequencies

(scales) and time samples of a specific segment of the duration of the FO. TF boxes that were

not contributing to 99% of the total cumulative power, within the selected window, were

considered as too low SNR and therefore non localizable. FOs, for which no localizable TF box

was identified within the frequency band of interest, were labelled as non localizable FO

events. The spike map was computed for 50 to + 50 ms from the spike peak using cMEM. Spikes

were filtered using a 0.3–70 Hz finite impulse response filter prior to cMEM localization; for FO

no filter was applied as wMEM acts a filter which is chosen by a specific frequency band, i.e.

40–80 Hz or 80–160 Hz.

Fig. 2. Detection of fast oscillations (FO) using scalp high-density electroencephalogram (HD-

EEG). This figure provides examples of FO for each patient. (A) unfiltered signal showing the

epileptic spike with the FO underlined in black (B) filtered signal (40–80 Hz) demonstrating the

FO event over the background (C) time frequency plot showing the isolated peak in the 40–80

Hz frequency band atop of the spike. Time scale is set to the peak of the event and 125 ms

before and after are given.

Epileptic spike and FO consensus maps

In order to take into account reliability and reproducibility of FO or spikes, we applied the

concept of consensus maps of source localization(Chowdhury et al., 2018). In order to estimate

these consensus maps for every patient, we first applied cMEM or wMEM to generate source

maps of every single event, FO or spikes. In a second step, we estimated a similarity index

between all single event source maps (FO or spikes), based on spatio-temporal correlation

around the peak of the event The source maps were then clustered using a hierarchical

clustering approach (Ward’s hierarchical clustering) followed by thresholding of the

dendrogram to obtain the optimal number of clusters. The cluster exhibiting most events was

selected and the consensus map was finally obtained by averaging all the single event maps of

this cluster. This consensus approach was considered instead of simple averaging of the event

followed by one source localization, in order to enhance reliability between source localization

results and reducing the influence of more noisy maps. We previously carefully demonstrated

the robustness of this approach when considering either EEG source imaging, MEG source

imaging or EEG-MEG fusion source imaging(Chowdhury et al., 2018).

Comparing HD-EEG against conventional EEG approaches

Our main results used the 256-electrode montage (after removal of ~ 40 face and neck

artifactual electrodes per patient). We then compared our results to the standard 10–10 system

(73 electrodes) and the 10–20 system (25 electrodes). The 10–10 and 10–20 system was

approximated from the full montage as per the recommendations of EGI geodesic systems (see

Fig. 3 for selected electrodes). All FO were localized using these three montages and results

were compared using the same metrics (see below).

Fig. 3. Illustration of the distribution of electrodes in the 256 channel high-density

electroencephalography (HD-EEG) system versus the standard 10–10 system (73 electrodes)

and the 10–20 system (25 electrodes). Note that the 10–10 and 10–20 systems were

approximated from the full HD-EEG montage as per the recommendations of the EGI geodesic

systems manual.

Generator depth

The depth of the generator was assessed on the basis of all available clinical information. The

generator depth was divided into 3 categories: superficial, intermediate and deep as done in

previous work(Cuello-Oderiz et al., 2017). Superficial was considered as involving the neocortex

adjacent to the skull including the bottom of the sulcus; deep was considered as involving the

medial aspects of the frontal, parietal, occipital, and temporal lobes and temporo-occipital

basal regions; and intermediate as generators found in regions not fitting the above two

categories.

Evaluation metrics & statistics

All final spikes and FO consensus maps were tresholded at 50% of the maximum reconstructed

intensity for both visualization and statistical purposes. We previously demonstrated that MEM

source localization results provide maps with high contrast, therefore results on the spatial

extent are relatively stable within a large range of detection threshold, as opposed to other

standard source localization techniques (Chowdhury et al., 2018; Pellegrino et al., 2018;

Pellegrino et al., 2020). In order to assess the success or failure of the localization, we

considered the surgical cavity as a reference. Since all selected patients are > 2 year

postsurgically seizurefree (Engel 1a), we are sure that the presumed EZ was indeed localized

within the cavity. The surgical cavity was fitted as a surface based region of interest (scout) on

the presurgical cortical surface considered as our brain source model (Supplementary Fig. 1).

This was done visually using the post-surgical MRI co-registered to the intact cortical surface of

the pre-surgical MRI, using the Brainstorm software. The coregistration between presurgical

and postsurgical MRI was obtained with a 6 parameter rigid body coregistration using the MINC

toolkit (https://bic-mni.github.io/). All evaluation metrics were then estimated using this scout

of the surgical cavity as our clinical reference. We assessed the following validation metrics: the

minimum Distance Localization Error (Dmin), Spatial Dispersion (SD), the Spatial Map

Intersection (SMI), and the SNR. For each of the validation metrics described above, we

considered only the consensus localization map as defined as the average localization obtained

from the cluster exhibiting the largest number of single events for both spikes and FO.

- Dmin: the minimum distance localization error was computed as the Euclidean

distance in mm from the maximum of the map to the closest vertex belonging to the

cavity. Whenever this maximum was located inside the cavity, Dmin was set to 0 mm.

- Spatial Dispersion (SD): the SD metric measured the spatial spread (in mm) of the

localization around the Ground Truth considered here as the surgical cavity. It was

computed as the root mean square of the distance from the cavity weighted by the

energy of the source localization map on each vertex.

󰇛󰇛

󰇜



 





Where Θ denoted the surgical cavity, and  is the amplitude results of the MEM solver

(Cmem or wMEM) for each dipolar source , while 󰇛

󰇜 function provides the

minimum Euclidean distance between the dipolar source i to the closest vertex within

the surgical cavity Θ.

- Spatial Map Intersection (SMI): SMI was defined as the percentage of the thresholded

localization map that was falling within the surgical cavity region of interest, such that if

all the thresholded localization was inside the scout then it equals 100% intersection.

SMI󰆓



󰆓



 Where 󰆒 refers to a binarized version of  after 50% threshold.

- Signal to noise ratio (SNR): the SNR was calculated using the mean amplitude at the

time of the event divided by the standard deviation of a baseline period of 10 ms long

for 200 ms prior to the event, the calculation was done on the filtered signal for each

type of event (0.3–70 Hz for spikes and 40–80 Hz or 80–160 Hz for FO) in accordance to

the localization filters. For spikes the SNR calculation was made on the average spike.

For FO, SNR was calculated for each event and then averaged over all events of each

patient, followed by averaging over all patients to reach the final SNR.

We used the Welch’s t-test to compare the different metrics. The Fisher exact test was used to

assess the distribution of generator depth categories between patients with and without FO.

2.5 Results

Detection of FO

FO in the gamma band (40–80 Hz) were found in five out of ten patients (Table 2). FO in the

ripple band (>80 Hz) did not reach the minimum of 5 requested events per patient. The mean

number of FO > 40 Hz was 12.2 ± 3.9 (range, 10–20) with a mean SNR of 3.4 ± 1.4. Fig. 2 shows

representative examples of FO of the five patients in the unfiltered signal, the filtered signal, as

well as the corresponding time frequency plot. FO preceded spikes in 78.8% of cases with an

onset of 0.1 ± 0.03 s before the peak of the spike. Spikes were present in all ten patients. The

median number of spikes was 41 (range, 12–303) with a mean SNR of 9.3 ± 2.7. As expected,

the SNR was significantly lower in FO than in spikes (p < 0.05).

Electrical source imaging

FO sources were localized in 5 out of 5 patients inside the surgical cavity (Fig. 3) resulting in a

minimum distance Dmin of 0 mm. FO source maps were localized within the surgical cavity with

an SD of 9.4 ± 3.2 mm and a SMI of 62.0 ± 15.0%. Spike sources localized inside the surgical

cavity in 9 out of 10 patients (Fig. 4, Supplementary Fig. 2) with a Dmin of 6.5 ± 20.4 mm.

Source maps were localized within the surgical cavity with an SD of 6.2 ± 10.9 mm and a SMI of

76.0 ± 30.0%. The spread outside the cavity as well as the activation intersection with the

surgical cavity did not differ significantly between FO and spikes (SD: p = 0.17; SPI: p = 0.09). For

further details see Fig. 5.

Table 2

Depth of the epileptic generator, epileptic spikes, and fast oscillations (FO).

Depth of

generator

# Spikes

localized

# FO

localized

# of Spikes in

the

consensus

map

# of FO in

the

consensus

map

Localizable

FO main

frequency

(Hz) range

Patients with Fast Oscillations

superficial

303

164

42-65

superficial

258

101

41-73

intermediate

155

65-80

superficial

40-55

superficial

42-80

Patients without Fast Oscillations

deep

intermediate

deep

For each patient the number of localizable epileptic spikes and fast oscillations (FO) are

presented. The depth of the generator is given. It is based on the available clinical information

and rated by an epileptologist. Note that FOs > 80 Hz were found in 8/10 patients. As their

number did not reach five or more FOs > 80 Hz, they were not included in the analysis.

Usefulness of the consensus map approach for FO source localization

The consensus map approach for FO sources resulted in less disperse maps compared to the

one obtained by simply averaging of all FO maps (SD difference between cluster versus

averaged map: 10.3 ± 4.8 mm versus 9.4 ± 3.2 mm; p = 0.4). This difference in SD became

significant when no thresholding was considered prior to the estimation of SD, resulting in a SD

of 18.1 ± 3.1 mm for the averaged map compared to 16.3 ± 3.2 mm for the consensus map (p =

0.007). Our clustering approach was used instead of traditional averaging in order to have a

more statistically robust localization. In the case of a small number of events as present for FO

the clustering acts as a denoising mechanism as even one ‘‘bad” FO can misinform the whole

average map. This demonstrated that the hierarchical clustering approach considered for the

consensus map helped to remove the impact of more noisy FO events, resulting in a more

reliable localization. This is illustrated in Fig. 6, which shows only the activations outside the

cavity for averaged versus consensus FO source maps, when no threshold was applied.

FO localization performance in HD-EEG vs the conventional 10–10 & 10–20 EEG arrays

We examined the localization accuracy of FO using the 10–10 system with 73 electrodes and

the 10–20 system with 25 electrodes (Figs. 3, 7, Supplementary Table 1). When considering the

10–10 system, among the FO detected using the HD-EEG array, we could identify FO in 4 out of

5 patients in the 10–10 electrodes. For the 10–20 system, we could identify FO in 3 out of 5

patients. Among the FO detected, all of them exhibited sufficient energy in the time–frequency

representation and were therefore localizable using our proposed strategy. Results showed

that using the full HD-EEG array we obtained a maximum source accurately localized in the

cavity (Dmin = 0 mm) for all cases (Fig. 7), whereas with the 10–10 and 10–20 system, one and

two FO localization exhibited a non-zero Dmin in each case (respectively) . This means that the

10–10 system managed to correctly identify the EZ using FO in 60% (3 out of 5 patients) and the

10–20 system in 40% (2 out of 5 patients) of patients when detection and FO localization was

considered. SD which measures the off-target spatial spread around the surgical cavity was

shown to be lower with HD-EEG compared to that of the 10–20 system in all cases and to that

of the 10–10 system in 2/4 of cases. SMI which represents the amount of spatial overlap with

the surgical cavity, was higher with HD-EEG compared to the 10–10 system and 10–20 system

in 2 of 4 cases and 2 of 3 cases respectively. SD and SMI values for each patient are provided in

Supplementary Table 1.

Fig. 4. Results of electrical source imaging for fast oscillations (FO) and spikes. FO and spikes

were source localized using the Maximum Entropy on the Mean (MEM) method (wavelet MEM

for FO and coherent MEM for spikes) in patients who showed > 5 FO. The surgical cavity was

fitted on the brain model and was marked as grey area. (A) mean scalp topography of the

consensus map (defined as cluster exhibiting most single events) of FO for each patient. Power

topography maps were averaged together for the peak of the event time course over the peak

frequency as visualized by the time frequency plot (B) consensus map of electrical source

imaging (wMEM) of selected FO for each patient (C) mean scalp topography of the consensus

map of spikes for each patient (D) consensus map of electrical source imaging (cMEM) of

selected spikes for each patient. Please note that scalp topographies (column A and C) are

presented following the same orientation as the corresponding source maps. All source

localization results are presented using a color map scaled to the maximum reconstructed

intensity of the corresponding map and thresholded at 50% of their maximum value. The

current amplitude of sources of FO was as expected several orders of magnitude lower than

that of spike sources. Patient 10 s large cavity is due to a disconnecting surgery.

Generator depth

We found that all patients with localizable FO actually had a surface close generator, with 4

patients being classified as having a superficial generator and 1 patient being classified as

having an intermediate generator (Table 2). In the 5 patients not showing sufficient numbers of

localizable FO, the generator proved to be deep in 4 cases and intermediate in 1 (p = 0.02). In

the patient in whom the spike map localized outside the surgical cavity, the generator was

actually localized in the cingulate gyrus, which is a deep region usually not detectable from

scalp EEG.

Fig. 5. Validation metrics of the results of electrical source imaging of fast oscillations (FO) and

spikes. Different measures were used to evaluate the success rate of the detectability and

localization of FO compared to spikes (A) Distance Localization Error (DLE) (Dmin) as measured

from the maximum of the localization map to the closest point of the cavity (Dmin = 0 mm

means within cavity; the larger the number, the further is the localization peak outside the

cavity) (B) Spatial dispersion (SD) maps (50% threshold) measuring the amount of activity

outside the cavity in relation to the distance from the closest cavity point (C) Spatial Map

Intersection (SMI) between the activation maps (50% threshold) with the cavity such that 100%

means that the localization map is completely localized inside the cavity.

Fig. 6. Comparison between electrical source imaging performed with the average vs. the unthresholded

consensus map for fast oscillations (FO). FO were clustered using hierarchical clustering. The cluster

exhibiting the most events was chosen as a consensus map. The cavity is marked in gray and the

spurious activity outside the cavity is then depicted. Note that the maximum is inside the obscured

cavity (as shown in Fig. 3) and only the spread is shown; thus the maximum is obscured by the cavity.

Fig. 7. Results of electrical source imaging for fast oscillations (FO) for the 10–10 and the 10–20

electroencephalogram system. FO were source localized using the wavelet Maximum Entropy on the

Mean method. The cavity is superimposed in light gray color. Final maps were either based on the best

cluster or averaged in the case of < 8 FO events.

2.6 Discussion

This study presents a proof-of-concept that electrical source imaging of FO, if present, is able to

identify the EZ in patients with drug-resistant focal epilepsy using HD-EEG. We demonstrated

that (i) FO can be recorded in scalp high-density EEG during short recording sessions lasting 90

minutes, (ii) the presence or absence of sufficient numbers of FO is dependent on the depth of

the epileptic generator, and (iii) electrical source imaging of FO using HD-EEG is able to

correctly localize the EZ being superior to conventional EEG approaches using the 10–10 or 10–

20 EEG systems.

Use of FO for identification of the EZ

FO were present in 50% of the investigated patients despite only short segments with light

sleep used for detection of FO. Interestingly all patients exhibiting FO and spikes had a surface

close generator, whereas the patients in whom we did not identify FO but only spikes had an

either intermediate or deep generator. This finding is in keeping with a previous study

investigating highfrequency oscillations in the scalp demonstrating that scalp FO are correlated

with the depth of the epileptic generator (Cuello-Oderiz et al., 2017). We found that the

maxima of the electrical source localization maps of FO correctly localized the EZ as

approximated by the surgical cavity in all patients. The ability of FO to non-invasively localize

the EZ was recently demonstrated using MEG (Nissen et al., 2016; von Ellenrieder et al., 2016;

Velmurugan et al., 2019; Yin et al., 2019; Tamilia et al., 2020). Velmurugan and colleagues

(Velmurugan et al., 2019) have recently shown the utility of high-frequency oscillations in

correctly localizing the EZ in MEG in 52 patients thus demonstrating its potential as a scalp

biomarker. Yet, while MEG has advantages such as better SNR and less complicated head

modeling impacting spatial resolution when compared to EEG (Ilmoniemi and Sarvas, 2019), the

lack of availability in most centers and its high maintenance cost makes it less ideal as a tool for

everyday use in presurgical evaluation. In addition, it is worth noting that while EEG has been

shown to detect higher numbers of FO (van Klink et al., 2019; Tamilia et al., 2020), spatial

coverage is key due to the low SNR and a high-density array is preferable. Thus, this proof of

principle study using HD-EEG with 256 electrodes together with the already existing evidence in

MEG and EEG (Thomschewski et al., 2019; van Klink et al., 2019; Tamilia et al., 2020) points to

HD-EEG being a feasible candidate for FO localization of the EZ during presurgical evaluation.

Comparison of FO to spike sources

When comparing FO to spike source localization, we found that FO source localization and spike

source localization were concordant with the latter resulting in localizations in all 10 study

subjects. Note is made that one of the spike localizations was outside and far from the epileptic

generator. The patient was identified as having a deep generator in the posterior cingulate

gyrus determined by a focal MRI lesion and positron emission tomography (PET)

hypometabolism. The spikes localized to the anterior temporal lobe corresponded hence to

propagated activity. In this patient, FO were absent suggesting that source localization results in

presence of FO are very likely pointing to the ‘‘true” onset generator (Cuello-Oderiz et al.,

2017). Future research will assess the clinical validity of these results. In order to assess the

feasibility of using scalp FO, and their eventual occurrence when superficial generators are

involved, as a marker of the spike onset zone, a larger cohort and longer recording durations

are needed. Our finding might hence be able to address a longstanding problem inherent to

non-invasive source localization. However, careful evaluation of propagation of source

localization along the peak of the spike vs. FO localization was out of the scope of this study and

will be considered in future investigations. A recent study showed that electrical source imaging

seems to be more accurate when performed at the time of spike onset instead of the peak of

the spike(Plummer et al., 2019).

Different measures for source localization quality

In clinical practice most studies assessed the concordance of the source with the assumed EZ at

the lobar(Duez et al., 2019; Rampp et al., 2019) or sub-lobar level (Zijlmans et al., 2017). We

evaluated our findings using different quantitative validation metrics, several of them being

similar to the ones proposed in Pellegrino et al. (Pellegrino et al., 2018; Pellegrino et al., 2020)

and Chowdhury et al. (Chowdhury et al., 2018) in order to assess the quality of our results as

objectively as possible regarding the source map maxima, its spatial extent, and the spatial

intersection with the resection cavity. The originality of our proposed approach was also to

consider an accurate delineation of the resection cavity as our reference for the evaluation of

source localization results, whereas such a comparison was more qualitative (Abdallah et al.,

2017) or semi-qualitative in other studies Pellegrino et al (Pellegrino et al., 2018; Pellegrino et

al., 2020). The accuracy of FO localization was high with the maxima localized to the surgical

cavity in all patients. Albeit being not significant, the extent of the source seemed to be more

widespread in FO compared to spikes. However, whereas the estimation of the spatial extent of

spike maps with cMEM has been carefully evaluated by our group (Chowdhury et al., 2016;

Pellegrino et al., 2016; Pellegrino et al., 2020), the evaluation of the spatial extent of wMEM for

FO would require further careful investigation. Moreover, it is difficult to disentangle whether

the spatial extent of FO maps was true extent or only resulted from the localization of low SNR

events. This might be potentially explained by the small number of FO given the short recording

time of 90 minutes, as well as by the lack of consolidated sleep in the present study; longer

recording durations preferably overnight are likely to be more favorable regarding FO quantity

and SNR. Future research performing prolonged overnight sleep recordings is awaited for

confirmation.

FO localization performance with different EEG montage

We confirmed evidence of previous work that a higher and denser electrode coverage is

preferable to detect FO given their small generators(Kuhnke et al., 2018). Moreover, we

demonstrated that this also impacts FO source localization. We showed that the correct

localization of the EZ declines from 100% with HD-EEG to 60% with the 10–10 system and 40%

with the 10–20 system. The use of HD-EEG seems therefore to be even more important for FO

than spikes due to a lower SNR(Song et al., 2015). Further research with a larger cohort is

needed in order to clinically validate our method for epilepsy patients undergoing surgery and

see how our localization algorithm will perform.

Importance of the consensus map approach for FO source maps

In this study we applied for the first time for spike and FO in HD-EEG the consensus map

approach we recently proposed as a more robust approach than event averaging to provide

reliable source localization, while taking into account the reproducibility of single discharges

source maps(Chowdhury et al., 2018). In this study, this was our first attempt to consider

consensus maps from wMEM results for FO localization. We therefore first performed a single

event FO source localization, using only the TF box exhibiting the largest amount of energy

along the FO, following the exact same methodology proposed in MEG by our group (von

Ellenrieder et al., 2016; Chowdhury et al., 2018). We then compared every single FO source

map using a hierarchical clustering in order to separate the data from events that were not in

agreement with the majority of events and which would therefore add noise to the maps, as

done in our previous work for spikes (Chowdhury et al., 2018). This consensus map approach

seems to be particularly useful for FO, which in case of a discordant FO event tend to create a

noisier map. Fig. 4 shows a significant improvement of source localization of FO using this

consensus map approach, when compared to standard averaging of all FO maps, resulting in a

reduction of the SD of the source map outside the presumed EZ as approximated by the surgical

cavity. Whether the fact of a lower accuracy of 0.87 in a recent combined EEG/MEG study

(Tamilia et al., 2020) can be explained by the advantage that a consensus map could offer

awaits further confirmation.

cMEM and wMEM for source localization of epileptic discharges

Most previous work in source localization of FO in MEG used the Beamformer technique, as it is

assumed to be able to detect distributed and deep sources (Hu et al., 2017). Yet Beamformer is

not a source localization method in its proper sense, but corresponds rather to a statistical

dipole scanning approach, iteratively assessing how likely it would be to fit an equivalent

current dipole at a specific position in a 3D grid covering the brain (Hillebrand et al., 2005). One

important feature of the Beamforming technique is its inherent denoising properties

(Hillebrand et al., 2005; Cheyne et al., 2007), as a spatial filtering approach, which is probably

the main reason why several groups have considered this localization approach for FO

(Belardinelli et al., 2012; van Klink et al., 2018). In this study, we chose the MEM framework, as

one of the only distributed approaches that has been proposed to carefully recover the

generators of epileptic discharges together with their spatial extent (Birot et al., 2011;

Chowdhury et al., 2016). We also considered the time–frequency based extension of MEM

(wMEM), providing us with the unique property of localization in the frequency domain(Lina et

al., 2014), thus making it a good tool for the localization of oscillatory events, as previously

demonstrated for FO in MEG and low-density EEG (von Ellenrieder et al., 2016; Tamilia et al.,

2020), localization of oscillatory patterns at the seizure onset in EEG and MEG (Pellegrino et al.,

2016; Pellegrino et al., 2018) and ongoing MEG resting state fluctuations (Aydin et al., 2020).

For future investigations, it might be interesting to use a combined Beamformer-wMEM

approach in order to increase the chance of FO detection with Beamformer applied as a

denoiser in the source space followed by localization of the FO from denoised scalp data using

wMEM.

SNR as a technical challenge

The SNR of EEGs can vary widely with different conditions as it is very susceptible to muscle

artifact notably as well as interferences with outside noise (Islam et al., 2016). This latter

challenge might be potentially overcome by the use of a low-noise amplifier (Fedele et al.,

2015), which could make it possible to obtain low noise recordings in the future. We attribute

the larger spread of FO maps compared to spike maps to the difference in the respective SNR

which was shown to create a less focused localization for oscillatory activity with no spatial

smoothing(Lina et al., 2014). The low SNR might also explain the fact that we detected mainly

FO in the gamma frequency range (40–80 Hz) and not FO > 80 Hz. Prolonged clean recordings

can be extremely valuable for FO identification, as data have the least artifacts in the high-

frequency domain during sleep (Zijlmans et al., 2017). Yet in the case of prolonged recordings it

might not be feasible to visually mark and validate each FO and a computational approach

might be needed (Höller et al., 2018). Another interesting aspect that might influence the

results is the type of electrode being used. Recent research demonstrated that tripolar EEG

electrodes might yield some benefit with higher signal quality (Toole et al., 2019).

Limitations

First, we would like to acknowledge that our sample size is small. However, all 10 patients

selected for this study were postsurgically seizure-free with an at least 2-year postsurgical

follow-up, which allowed us to have the best estimate possible for approximation of the EZ in

this proof-of-principle study. An alternative would have been to validate against the

intracranially identified SOZ (Papadelis et al., 2016; Kuhnke et al., 2018; Dirodi et al., 2019) .

Given however, that up to ~ 50 % of patients in whom the SOZ has been surgically removed, do

not become seizure-free after surgery (Krucoff et al., 2017), this seemed to us to not be the

best option for the validation of our approach, even though a large resection diminishes the

sensitivity to the real EZ (see patient 1 for an example). Second, we only had short recordings of

approximately 1.5 h, during which patients achieved mostly only light sleep. Prolonged

recordings of overnight sleep as planned in future research as well as use of denoising

techniques as possible with Beamformer (Hillebrand et al., 2005; Cheyne et al., 2007) could

result in higher patient numbers in whom FO can be identified and results regarding the extent

of the source might be improved given higher event numbers and therefore better SNR. Third,

we marked FO at the time of spikes as suggested by other authors (Nissen et al., 2016;

Velmurugan et al., 2019), as FO at the time of spikes have a greater chance of a correct

localization than FO occurring independent of spikes (Dirodi et al., 2019). To avoid a confound

with a filtering effect of spikes (Bénar et al., 2010), we followed a two-step procedure to

exclude false positives based on both visual signal inspection and presence of an isolated peak

in the time–frequency representation in order to avoid misclassification of FO. This study

demonstrated the ability of 256-channel HD-EEG to correctly identify the EZ using source

localization of FO. Presence or absence of FO was shown to be dependent on the presence of a

surface close generator. This points to an added value of FO source localization, as presence of

concordant spike and FO sources could confirm correct localization of the EZ, whereas lack of

FO sources might point to the fact that the identified spike source could be the correlate of

rather propagated activity and not the primary source, a problem inherent to non-invasive

source imaging. We further confirmed the added value of HD-EEG for both FO detection and

source localization, when comparing our findings to conventional approaches using the 10–10

and 10–20 EEG system. Further research is required to assess the clinical validity of these

results. In order to assess the feasibility of using scalp FO in clinical practice, a larger cohort

which includes successful and unsuccessful surgery outcomes is required to address the

sensitivity and specificity of scalp FO in localization of the EZ and to compare their accuracy

with that of spikes as traditional biomarker of epilepsy.

2.7 Supplementary material

Supplementary Figure 1. Superimposing the scout of the cavity on the postsurgical MRI. The

scouts were later used to determine the localization. The postsurgical MRI was co-registered

into the space of the presurgical MRI using the MINC toolkit (https://bic-mni.github.io/). A 3D

representation with the moveable surfaces was superimposed over the brain

segmentation model. The brain was made semi-transparent in order to better visualize the

superimposed post-surgical MRI. Once the resection cavity was identified, a scout was carefully

fitted on the brain and grown until it filled the whole cavity. The scout was visually inspected

and compared to the postsurgical MRI by a board certified neurologist.

Supplementary Figure 2. Results of electrical source imaging of spikes for the 5 patients who

did not have FO (see Fig. 3). Topography maps are provided on the left panel, source maps are

provided on the right panel. The surgical cavity is superimposed on the brain model of the

individual patients in dark grey color. Note that the topographical maps are oriented to the

same direction as the source localization map results.

Supplementary Table 1. Fast oscillations (FO) localization metrices. The table shows Spatial

Dispersion (SD) and Spatial Map Intersection (SMI) for the 256 electrodes high-density

electroencephalogram (HD-EEG), 10-10 system, and 10-20 system. HD-EEG had a superior FO

detection capability compared to the 10-10 system and the 10-20 system. Spatial dispersion

(SD) which measures the off target spread was shown to be lower with HD-EEG compared to

that of the 10-20 system in all cases and that of the 10-10 system in 50% of cases. Spatial Map

Intersection (SMI) which represents the amount of overlap with the surgical cavity, was higher

with HD-EEG in 67% of cases as assessed with the 10-20 system, and 50% of cases as assessed

with the 10-10 system.

# Patient

SD HD- EEG

SD 10-10

SD 10-20

SMI HD- EEG

SMI 10-10

SMI 10-20

6.57

0.77

10.61

1.09

12.87

0.67

0.95

0.37

13.26

0.58

4.78

9.88

10.05

0.74

0.53

0.91

11.63

16.96

15.83

0.35

0.3

0.05

3. Chapter 3: Manuscript #2: Consistency of electrical

source imaging in presurgical evaluation of epilepsy

across different vigilance states

Published as: Tamir Avigdor*, Chifaou Abdallah*, Jawata Afnan, Zhengchen Cai, Saba Rammal,

Christophe Grova, Birgit Frauscher. Ann Clin Transl Neurol. 2024 Feb;11(2):389-403

3.1 Preface

The previous study demonstrated that fast oscillations on top of interictal epileptiform

discharges are localizable and might point to the epileptogenic zone. Given that these fast

oscillations are predominantly found during NREM sleep, as most interictal epileptiform

discharges also occur during this stage, it is important to investigate the possible influence of

the state of vigilance on it. There is an ongoing debate in the literature regarding the

consistency of source localization for interictal epileptiform discharges across the different

states of vigilance. If the results of source localization for interictal epileptiform discharges vary

between sleep stages, it would mean we cannot fully trust the localization results of fast

oscillations during NREM sleep. This is particularly problematic because fast oscillations are not

present in large enough quantities and with high enough signal-to-noise ratio during

wakefulness or REM sleep for meaningful comparison. To address whether source localization

of interictal epileptiform discharges is consistent across different vigilance stages and by

extension, whether the localization of fast oscillations during NREM sleep is representative, I

used HD-EEG to source localize interictal epileptiform discharges across all states of vigilance.

The goal was to determine whether the results of source localization are spatially consistent

across these states. I found that, while there was an effect of the stage of vigilance on the signal

in the sensor space (e.g., amplitude), there was no effect on the spatial results of source

localization for interictal epileptiform discharges. This demonstrates that the source localization

of these discharges is spatially consistent across vigilance stages. By extension, this validates

the findings of the study in Chapter 2, confirming that the source localization of fast oscillations

during NREM sleep can be trusted as representative. In summary, source localization of

interictal epileptiform discharges is invariant to the state of vigilance. Current clinical practice,

which predominantly records during wakefulness, remains valid as long as a sufficient number

of interictal events are recorded and the signal-to-noise ratio is adequate. Additionally, the

source localization of fast oscillations during NREM sleep might be representative of the true

source, rather than being influenced by the vigilance stage.

3.2 Abstract

Objective: The use of electrical source imaging (ESI) in assessing the source of interictal

epileptic discharges (IEDs) is gaining increasing popularity in presurgical work-up of patients

with drug-resistant focal epilepsy. While vigilance affects the ability to locate IEDs and identify

the epileptogenic zone, we know little about its impact on ESI.

Methods: We studied overnight high-density electroencephalography recordings in focal drug-

resistant epilepsy. IEDs were marked visually in each vigilance state, and examined in the

sensor and source space. ESIs were calculated and compared between all vigilance states and

the clinical ground truth. Two conditions were considered within each vigilance state, an

unequalized and an equalized number of IEDs

Results: The number, amplitude and duration of IEDs were affected by the vigilance state, with

N3 sleep presenting the highest number, amplitude and duration for both conditions (p<0.001),

while signal-to-noise ratio only differed in the unequalized condition (p<0.001). The vigilance

state did not affect channel involvement (p>0.05). ESI maps, showed no differences in distance,

quality, extent, or maxima distances compared to the clinical ground truth for both conditions

(p>0.05). Only when an absolute reference (wakefulness) was used, the channel involvement

(p<0.05) and ESI source extent (p<0.01) were impacted during Rapid-Eye-Movement (REM)

sleep. Clustering of amplitude-sensitive and -insensitive ESI maps pointed to amplitude rather

than the spatial profile as the driver (p<0.05)

Interpretation: IED ESI results are stable across vigilance states, including REM sleep, if

controlled for amplitude and IED number. ESI is thus stable and invariant to the vigilance state.

3.3 Introduction

The treatment of choice for patients with drug-resistant epilepsy is epilepsy surgery with the

objective of resecting the epileptogenic zone (EZ) (Jehi, 2018). Currently the standard diagnostic

work-up for presurgical epilepsy evaluation consists of video-electroencephalography (EEG),

structural magnetic resonance imaging, and neuropsychology. In recent years there is

increasing evidence of the value of electrical source imaging (ESI) in the presurgical evaluation

of epilepsy (Brodbeck et al., 2011; Megevand and Seeck, 2020). ESI provides a way to detect the

source of scalp EEG potentials in the brain. ESI is performed by solving an inverse problem

(Grech et al., 2008), which aims to estimate the most likely source of the observed scalp EEG

potentials. There are two main approaches. The first is dipole modeling (Ebersole and Hawes-

Ebersole, 2007), which attempts to identify a single point source or a few point sources in the

brain that explain the distribution of scalp EEG potentials. The second approach involves

distributed methods (Plummer et al., 2010), which place dipoles throughout the cortical surface

in alignment with the brain's structure. This approach then attempts to determine each dipole’s

amplitude contribution in order to explain the distribution of scalp EEG potentials. The

advantage of distributed methods is its ability to assess the source extent, i.e., the boundaries

of the active region. In the context of epilepsy, when ESI uses distributed methods and is

performed correctly (Michel and Brunet, 2019), it can accurately pinpoint the source of

epileptic activity and delineate its extent. To achieve accurate localizations, low noise and high

spatial sampling with high-density EEG (HD-EEG) (Song et al., 2015) are preferred.

Increasing evidence from intracranial EEG revealed the relevance of the effect of different

states of vigilance on epileptic activity (Herman et al., 2001; Frauscher and Gotman, 2019).

Interictal epileptic discharges (IEDs) were observed to increase in frequency during non-rapid

eye movement sleep (NREM), and to be suppressed during REM sleep (Malow et al., 1997; Ng

and Pavlova, 2013). The state of vigilance was also shown to influence IED morphology and

topography, with more circumscribed IEDs occurring during REM as opposed to NREM sleep

(Sammaritano et al., 1991; Fouad et al., 2022). In addition, the distribution and type of seizures

(Ho et al., 2023), but not their spatio-temporal dynamics (Hannan et al., 2023) have been shown

to be influenced by the vigilance state. Finally, there is some evidence from ESI that REM sleep

may reveal topographically new foci that are not present in other vigilance states (Sammaritano

et al., 1991; McLeod et al., 2020), although evidence on the latter remains sparse.

A recent review on the topography of IEDs across the different states of vigilance stated that

REM IEDs were in better agreement with the clinical SOZ than NREM IEDs (82 vs. 60%) (McLeod

et al., 2020). However, only two studies used ESI, one in a small sample of 6 patients using HD-

EEG (Kang et al., 2020) and the other one in 16 patients using low-resolution 25-electrode ESI

(McLeod et al., 2022). Both studies suggested REM ESIs offer unique spatial coordinates and

extent when compared to other vigilance states. However, the accuracy of ESI and estimation

of the underlying source extent of the generators can vary when considering a source imaging

method not designed to recover the extent of the underlying generators of IEDs. In addition,

other attributes such as a low-density array or low numbers of IEDs would also affect ESI

accuracy. Therefore, more work is warranted to clarify these findings in larger sample sizes with

higher numbers of IEDs during REM sleep and HD-EEG. This is particularly relevant, as currently

ESI is mostly performed with IEDs recorded during the awake state in short-term recordings. To

substantiate evidence that REM may achieve more accurate localization results would have

direct impact on our clinical practice.

Source localization is not trivial and a good spatial sampling from scalp recordings is crucial for

obtaining a reliable source map (Song et al., 2015). In addition, source localization methods are

known to be influenced by the signal-to-noise ratio (SNR) (Bast et al., 2006), and assessing the

underlying source extent remains a challenging task (Chowdhury et al., 2013; Chowdhury et al.,

2016; Sohrabpour et al., 2020). To address these challenges, one can use HD-EEG for spatial

coverage and a distributed ESI methodology sensitive to source localization and extent. One of

these methods is the coherent Maximum Entropy on the Mean (cMEM) method (Amblard et

al., 2004; Lapalme et al., 2006). Our group carefully evaluated the ability of cMEM to recover

the source extent of the generators, when compared to other standard distributed source

modelling approaches (Grova et al., 2006; Chowdhury et al., 2013; Chowdhury et al., 2016;

Hedrich et al., 2017; Pellegrino et al., 2020). Applying cMEM localization on manually marked

IEDs, recorded using a HD-EEG array, we were able to produce accurate results in identifying

the EZ (Avigdor et al., 2020).

In this study, we compared the IED source localization maps across different vigilance states

when applied to HD-EEG data and using cMEM as an ESI method sensitive to the source extent.

We report here that, contrary to our primary hypothesis, IED source localization during REM

sleep does not seem to offer a superior localization of IEDs when compared to any other state

of vigilance.

3.4 Methods

Subjects

Twenty-six consecutive patients (7 females; mean age, 33.56 ±10.91 years) with a diagnosis of

drug-resistant focal epilepsy who underwent between 24-48hr long recordings of combined

polysomnography (PSG) and HD-EEG (83 electrodes) at the Montreal Neurological Institute and

Hospital between January 2019 and July 2022 were screened. Patients were consecutively

recruited based on the referral of the clinical team for the indication of epilepsy presurgical

evaluation, availability of a technician to place all electrodes necessary for the HD-EEG

recording, and absence of prior brain surgery. The presence of IEDs in prior recordings was not

a prerequisite to perform HD-EEG. We excluded 8 patients that did not have a minimum of 5

IEDs in REM sleep (this was the first criterion), one whose recording contained too many

artifacts to reliably score sleep, and one patient, whose brain imaging could not be properly

segmented due to an extensive encephalomalacic lesion. A total of sixteen patients (7 females;

mean age, 33.56 ± 10.91 years) were included. Our cohort consisted of 10 temporal (62.5%) and

6 extratemporal patients. A subset of 11 patients underwent surgery, out of which 8 ( 72%)

were classified as seizure-free (Engel IA) after 23.5 [14-44] months of follow-up (Table 2 ). All

study participants provided written informed consent in agreement with the Research Ethics

Board at the Montreal Neurological Institute (REB00010120). All data were analyzed in an

average reference montage.

HD-EEG recordings and signal analysis

HD-EEG recordings were collected during the patients’ hospitalization for presurgical evaluation

in the epilepsy monitoring unit. Recordings were performed with the Nihon Koden system

(Tokyo, Japan) using collodion glued 83 electrodes placed according to the 10–10 EEG system

and sampled at 1000 Hz. Electrodes positions together with additional headshape points were

digitized on the scalp of the patients using a Polhemus localizer device (Vermont, USA). Then,

the electrodes were aligned on the head model (see SI methods) using reference points (nasion,

right and left ear). Coregistration was refined using a surface fitting approach (iterative closest

point) using Brainstorm. Finally, electrode positions were projected on the scalp surface.

Marking of IEDs and definition of the clinical ground truth

IEDs were marked by visual inspection of a neurophysiologist in all vigilance states. Not all IEDs

were marked, but we marked rather a sufficient amount of IEDs providing good SNR with a

minimum of 5 IEDs in each state of vigilance of the predominant IED type as done in our

previous work (Horrillo-Maysonnial et al., 2023). This was done in order to propose an analysis

similar to the one considered for clinical purposes, in which not all available IEDs are marked

but only the ones that are deemed to have high quality. This is important as ESI is affected by

the quality, the amount, and SNR of the final average IED which is being localized. IEDs during

wakefulness were marked prior to the night sleep recording. The IEDs were averaged ±5 s

around the negative peak of the IED, then the midpoint of the average IED, defined as half way

of the ascending slope of the IED, was marked by visual inspection, and confirmed by a

neurophysiologist. This midpoint was considered when evaluating ESI maps. There is debate in

the literature about the point of the averaged IED that should be used when calculating the ESI,

with some authors advocating for the point at 50% of the rising phase or the take-off (Brodbeck

et al., 2011), while others show no differences between this point and the peak of the IED

(Brodbeck et al., 2011; Plummer et al., 2019; Abdallah et al., 2022).We chose the midpoint as a

comprise between a good SNR and early IED propagations.

The clinical ground truth was defined by epileptologists (C.A., B.F.) based on phase 1 or 2

presurgical evaluation with long-term video EEG monitoring, anatomical 3T MRI, positron

emission tomography and neuropsychological evaluation, or stereo-electroencephalography

where applicable. This ground truth was drawn as a region of interest (ROI) along the cortical

surface on each patient’s head model. This ROI was then used as the ground truth for all

comparisons.

Electrical source imaging & Evaluation of different vigilance states

For each patient, the averaged IED was assessed for the 4 vigilance states (N2, N3, REM, Wake)

as well as compared to the IEDs chosen for ESI done in clinical routine (clinical IEDs) which were

selected on the bases of good visual SNR. This assessment was performed both in the sensor

space (see SI methods) at the peak of the averaged IED as well as in the source space at the

midpoint of the averaged IED (Fig. 1). The sensor space refers to the EEG representations

observed on a standard tracing at the scalp level, while the source space pertains to the

outcomes of the ESI, mapping the brain's sources of these scalp tracings. The peak was chosen

as the best representation of the sensor space as viewed by an electrophysiologist, while the

midpoint of the IED was chosen for the source space, in order to represent a compromise

between SNR and propagation.

Figure. 1. Study workflow. The study used all night high-density electroencephalography (HD-

EEG) recordings to assess the sensor (at IED peak) and source space (at midpoint) influence of

the different vigilance states, and compared them to the clinically selected IEDs as well as to

the clinical ground truth in the source space. Dmin: distance minimum; ESI: electrical source

imaging; IED, interictal epileptic discharge; SD: spatial dispersion SNR: signal-to-noise ratio.

The ESI maps were calculated using a depth-weighted version of cMEM (see SI methods) (Afnan

and Gotman, 2023). The inverse problem was solved using the maximum entropy of the mean

(MEM) framework (Amblard et al., 2004). The coherent MEM’s (cMEM) prior distribution uses a

hidden variable associated to each parcel which is tuning the probability of the parcel to be

active or to be switched off. Using such a flexible prior distribution, the cMEM method is able to

recover accurately the source extent of the underlying generators (Grova et al., 2006;

Chowdhury et al., 2016). The ESI was done between ±50 ms around the peak of the interictal

epileptic discharge, but only the ESI results at the mid-point of the averaged IEDs were

reported. We proposed two-level assessments of our evaluation metrics: first by considering all

marked IEDs in each state (unequalized condition), second by randomly selecting IEDs in equal

proportion to the number of IEDs in the state with the least available amount (equalized

condition). The latter was performed in order to address the fact that some vigilance states,

such as N3 sleep state, usually have higher IED rates, while minimizing the SNR effect on signal

averaging when assessing source localization results. Our proposed approach allows us to

investigate both the “real world” clinical situation in which this discrepancy indeed exists, as

well as to test whether the final results are simply driven by SNR discrepancies. We evaluated

the sensor level using the following metrics: (1) Peak amplitude (2) IED duration (3) SNR (4)

Channel involvement. We then evaluated the source level and assessed (1) Minimum distance

(Dmin) (2) Spatial dispersion (3) D-maxima (4) Source extent (5) Spatial profile. All metrics are

provided and defined in Table 1 and in the SI methods section. Finally, we also performed a

sub-lobar clinical analysis of the ESI maps to determine the sub-lobar concordance of the maps

in the various vigilance states.

Definition

Interpretation

Space

Peak amplitude

(mv)

The peak negative amplitude of the

averaged IED's most negative channel

The peak level at which the IED appears on the EEG trace. A

higher number indicates a stronger electrical potential

associated with the IED

Sensor

IED Duration

(ms)

The duration from the start to the end of

the averaged IED

A higher number indicates a wider IED

Sensor

Signal-to-noise

ratio

The difference between the IED's peak

and its background

A higher number signifies that the IED stands out more from

its surroundings. A value of one indicates the noise level

Sensor

Channel

involvement

The number of channels exceeding a

specific amplitude threshold during the

peak of the IED

The higher the number of involved channels, the larger is the

field spread on the scalp EEG

Sensor

Minimum

distance (mm)

The minimum Euclidian distance

between the maximum point of the ESI

map to the ground truth ROI

Reflects how good the ESI is in relation to the ground truth,

such that if the maximum of the ESI is within the ground truth

the value is zero

Source

Spatial

dispersion (mm)

The quantity and strength of activated

areas in the ESI that lie outside the

ground truth's ROI

A smaller number indicates better localization quality relative

to the ground truth. A value of zero signifies that all activity

was confined within the ground truth's ROI

Source

D-maxima (mm)

The distance between the peak points of

the ESI maps from two different vigilance

states

A lower number indicates that the peaks of the ESI maps are

closer together in space

Source

Source extent

The proportion of vertices in the ESI map

that are marked as active beyond a

certain amplitude threshold

A higher value indicates a more widespread ESI map

Source

Spatial profile

The spatial distribution of amplitudes in

different areas of the brain

Examining the clustering of the spatial profiles of ESI maps

helps to evaluate the overall similarity between maps. A value

of one indicates identical maps

Source

Table.1. Metrics and definitions employed in the paper to define terms related to both sensor

and source space analysis. EEG –electroencephalogram, ESI - electrical source imaging, IED-

interictal epileptiform discharge, ROI – region of interest.

Statistical analysis

We assessed evaluation metrics with ANOVA and Tukey or Friedman and Nemenyi tests,

depending on distribution. The Shapiro-Wilk test checked distribution normality. P-values were

corrected for multiple comparisons using the false discovery rate with alpha set to 0.05. Effect

sizes were calculated with Cohen’s d, Eta squared, Cliff’s d, and Kendall's W. Metrics followed

either normal (mean, standard deviation) or non-normal (median, interquartile ranges)

distributions. Results are reported for both conditions (unequalized, equalized IEDs). We also

compared two hierarchical clustering approaches using a Chi square test to check NREM, REM,

and wakefulness label distribution. A paired t-test assessed χ2 differences between clusters

with/without amplitude information.

3.5 Results

IED characteristics vary across the different vigilance states

We analyzed a total of 3449 IEDs in 16 patients (Table 2). We analyzed the average IED in each

vigilance state as well as the clinically selected average IED, first in the sensor space and later in

the source space (Table 3).

Age at HD-

EEG

Sex

Age at

seizure-

onset

MRI

SEEG

Ground truth

(sub-lobar level)

Surgery/Eng

el *

Normal

Yes

(IEDs: L amygdala,

hippocampus, entorhinal

cortex, anterior fusiform;

SOZ: L amygdala,

hippocampus, entorhinal

cortex)

L MTLE

yes/IIIb

(2 y)

R temporal

neocortex atrophy,

atrophy/agenesis of

the R piriform

Yes

(IEDs: R posterior insula,

posterior temporal gyrus;

SOZ: R posterior temporo-

insular junction)

R posterior

temporo-insular

yes/Ia (>1y)

Normal

Yes

(IEDs: R superior, middle,

inferior occipital gyri; SOZ: R

superior and middle gyri)

R latero-occipital

yes/Ib

(>1 y)

Right hippocampal

sclerosis

R MTLE

Right hippocampal

sclerosis

R MTLE

yes/Ia

(10 m)

Normal

Yes

(IEDs: L hippocampus,

amygdala, and entorhinal

cortex; SOZ: L hippocampus,

amygdala, and entorhinal

cortex)

L MTLE

yes/Ia

(>4 y)

Right hippocampal

sclerosis

R MTLE

yes/Ia

(>3 y)

R hippocampal

atrophy

R MTLE

yes/Ia

(>3 y)

R posterior insula

FCD

R parieto-

operculo-insular

junction

Possible R

hippocampal

sclerosis

R MTLE

yes/Ia

(>1 y)

R lingual

encephalomalacic

lesion

R mesio-occipital

yes/IV

(> 1 y)

R mesial premotor

FCD

R mesial premotor

Bilateral

mesiotemporal

abnormalities (L

FCD, R atrophy)

L MTLE

Normal

Yes

(IEDs: R posterior fusiform,

lingual gyrys; SOZ: no

spontaneous seizure

recorded)

R mesial temporo-

occipital

yes/Ia

(> 2 y)

R fronto-polar FCD

R fronto-polar

Early

childhood

L hippocampal

sclerosis

L MTLE

yes/III

(> 1 y)

Table. 2 Patient’s demographics *follow up >1 year after surgery., L - left, R - right, MTLE -

Mesio-temporal lobe epilepsy, FCD - Focal cortical dysplasia.

# IEDs

Anatomical location (Sub-lobar source localization)

L temporal pole

114

R posterior &

lateral

temporal

R posterior

temporo-insular

R posterior

temporo-insular

R posterior &

lateral temporal

R posterior

temporo-insular

223

R posterior &

lateral

temporal

R posterior &

lateral temporal

R posterior &

lateral temporal

R posterior &

lateral temporal

R posterior &

lateral temporal

R temporal

pole

R temporal pole

R anterior &

lateral

temporal

R anterior & lateral

temporal

R temporal pole

R anterior &

lateral temporal

R anterior &

lateral temporal

L mesio-

temporal

L mesio-temporal

L mesio-

temporal

L mesio-

temporal

L mesio-

temporal

R anterior &

lateral

temporal

R anterior & lateral

temporal

R anterior &

basal temporal

R anterior &

lateral temporal

R anterior &

lateral temporal

R temporal

pole

R temporal pole

R anterior &

lateral temporal

R temporal pole

128

R central

R lateral parietal

R posterior &

basal temporal

R anterior &

basal temporal

226

R temporal

pole

R temporal pole

R mesio-

temporal

R temporal pole

R temporal

pole

R anterior & lateral

temporal

R temporal pole

R anterior &

basal temporal

R temporal pole

113

R mesial

premotor

R mesial premotor

R central

L anterior &

lateral

temporal

L temporal pole

L anterior &

basal temporal

102

187

177

R mesial

temporo-

occipital

R mesial temporo-

occipital

R mesial

temporo-

occipital

R mesial

temporo-

occipital

R mesial

temporo-

occipital

R dorsolateral

prefrontal

R dorsolateral

prefrontal

R dorsolateral

prefrontal

R dorsolateral

prefrontal

R dorsolateral

prefrontal

L temporal pole

Table. 3 Interictal epileptic discharges count for every vigilance state, and the corresponding

sub lobar results of the electrical source imaging map.

In the sensor space we found that the number of IEDs, reflecting their discoverability on the

scalp EEG, was affected by the vigilance state (F(4,16)=23.89, p<0.001) with N3 exhibiting the

highest numbers. Furthermore, we observed an effect of the vigilance state on the IED

amplitude for both the unequalized and the equalized conditions (F(4,16) =21.1, 19.65; p

<0.001, <0.001; w =0.33, 0.31 respectively). The average unequalized and equalized IED peak

amplitude during N3 was overall larger when compared to REM (p=0.004, 0.002; d=0.84, 0.78)

and wakefulness conditions (p=0.005, 0.002; d=0.82, 0.80) (Fig. 2A). A similar observation was

found between N2 and REM (p=0.007, 0.005; d=0.76, 0.72), as well as between the clinically

selected IEDs and REM (p=0.007, 0.002; d=0.75, 0.72). The IED duration from take-off to take-

down was also affected by the state of vigilance in both conditions (F(15,60) =13.22, 10.17;

p<0.001, <0.001; ŋ=0.72, 0.68). The duration in N3 was longer (Fig. 2B) when compared to REM

(p=0.07, 0.03; d=0.65, 0.74), and wakefulness (p=0.02, 0.03; d=0.9, 0.8). The SNR of the

unequalized condition was similarly affected by the vigilance state (F(4,16) =23.05, p <0.001, w

=0.36). N3 and the clinically selected average IED were exhibiting a higher SNR (Fig. 2C) when

compared to REM (N3-R: p<0.001, d=0.65; Clinical-R: p<0.001, d=0.81) and wakefulness (N3-W:

p<0.001, d=0.56; Clinical-W: d=0.85, p<0.001). When considering the same number of IEDs in

each condition, no significant differences in SNR were observed between any states (p>0.05).

Figure. 2 Interictal epileptic discharge (IED) characteristics in different states of vigilance.

Distributions of average IED sensor space evaluation metrics for each vigilance state N2, N3,

REM, Wake, and clinically selected IEDs, represented as violin plots. (A) amplitude of the IED

peak in V (B) IED duration in ms (C) signal-to-noise ratio measured on the peak channel, for

unequalized (top row) and equalized conditions (bottom row). The lines connecting the dots are

linking each patient through all the vigilance states. Significance level were notes as *<0.05,

**<0.01, ***<0.001.

IED sensor involvement varies between different vigilance states

Here we examined whether the number of channels that are involved during an IED varies

between states. To do so, we counted the number of channels in which the amplitude during

the IED peak reached at least 20% of the amplitude of the highest amplitude channel (relative).

We found that there were no statistical differences between the different vigilance states (Fig.

3A, p>0.05). We then assessed whether there is an absolute difference rather than a relative

one between the states. To test this, we used an absolute threshold (20% of the maximum

amplitude of the average IED during wakefulness) for comparisons between states. With this

approach, we found that there was a significant effect for both equalized and unequalized

conditions (F(3,16) =8.74, 10.0; p =0.01, 0.032; w =0.21, 0.18), however, during the posthoc

analysis no statistical difference was found after correction for multiple comparisons or

correction for the unequalized condition. The equalized condition displayed a lower count

channels involved during IEDs in REM (Fig. 3B) when compared to IEDs in N2 (p=0.04; d=0.58),

N3 (p=0.04; d=0.63).

Figure. 3 Sensor space variation between vigilance states. Averaged IED channel involvement in

different states of vigilance as measured by an amplitude threshold (A) relative channel

involvement defined as channels with amplitude >50% of the maximum peak within each state

(B) absolute channel involvement defined as channels with amplitude >50of the maximum peak

of the wakefulness average IED The lines connecting the dots are linking each patient through

all the vigilance states. Significance level were notes as *<0.05, **<0.01, ***<0.001.

Localization accuracy of IEDs does not vary across different vigilance states

To assess the localization accuracy across different states, in relation to its ability to correctly

point and delineate the ROI of the ground truth, we used the ESI source maps estimated at the

midpoint of the average IEDs in each state (Fig. 4). We found no significant difference between

the vigilance states in either condition (i.e. equalized or unequalized number of IEDs averaged),

when assessing the minimum distance from the maximum ESI map activation to the clinical ROI

(Fig. 5A, all p>0.05). In addition, no significant difference was observed for accuracy (Fig. 5B) of

the localization as reflected in the spatial dispersion, measuring the spatial spread of the ESI

map around the clinical ROI (all p>0.05). Finally, we found no significant differences when

comparing distances between ESI maps maxima in different vigilance states ( Fig. 5C), yet a

tendency of N2-N3 localizations to be closer was observed in both states (all p>0.05). The

localization classification on a sublobar level (SI Table 1) also demonstrated the maps’

similarities between different ESI maps (see Fig. 4), where sublobar concordance between the

clinical ground truth and the ESI map exhibiting the largest activity was found for 15/16 patients

in both conditions in all vigilance states.

Figure. 4. Patient example #3. Source localization maps estimated and voltage maps at the

midpoint of average IEDs for all vigilance states (N2, N3, REM, Wake) and for the clinically

chosen IEDs, when considering: (A) all available IEDs, (B) equalized number of IEDs, in this case,

27 which was the number of IEDs marked during N3 sleep. Note that the scales differ between

source maps.

Figure. 5. Accuracy of ESI maps in different states. The source localization maps were compared

to a clinical ROI marking the presumed epileptogenic zone. (A) Distance minimum (Dmin):

measures the distance between the maximum vertex to the scout (B) spatial dispersion:

measures the spatial spread (in mm) of the localization around the ground truth (C) D-maxima:

measures the distance between each pair of maps maximum vertex in all states of vigilance.

Effect of the vigilance state on the ESI source extent

In this section, we investigated the impact of the vigilance state on the source extent of the IED

generators in the source space after applying ESI using cMEM. We found that the source extent

as expressed by the percentage of activated vertices did not differ between the states when we

considered a relative threshold to estimate the extent (Fig. 6A) for both unequalized and

equalized conditions (All p>0.05). Similarly, to prior investigation at the sensor level (Fig. 3), we

also investigated the source extent when considering the same absolute threshold for all maps,

therefore mimicking the clinical investigation. We considered the absolute threshold as 20% of

the wakefulness ESI source map. In this condition, we found a significant effect of the vigilance

state on the source extent (Fig. 6B) in both the unequalized and equalized conditions (F(3,16)

=12.32, 14.73; p=0.001, 0.004; w =0.26, 0.3). The ESI map during REM was exhibiting a lower

percentage of active vertices, i.e. a smaller source extent when compared to N3 (p=0.02, 0.008;

d=0.67, 0.82) and to the clinically proposed solution (p=0.02, 0.03; d=0.64, 0.75), therefore

suggesting a more focal involvement of the IED generators during REM.

Figure. 6 Source space variation between vigilance states. Source extent of average IEDs source

maps in different states of vigilance after applying an amplitude threshold in the source space:

(A) Relative threshold: percentage of activated vertices > 20% of the maximum vertex within

each specific state (B) Absolute threshold: percentage of activated vertices >20% of the

maximum vertex in the wakefulness average map, considered as reference.

Finally, we investigated whether the difference in the source extent observed during REM sleep

was mainly driven by amplitude differences, as the discrepancy between relative and absolute

thresholds results suggested (Fig. 7A-C). To examine these spatial profiles, we performed a

clustering analysis after applying cMEM on all available single IEDs in each vigilance state. We

proposed two data-driven clustering approaches of all those ESI maps, one including and one

not including ESI amplitude information (see methods). Each clustering resulted in 3 classes,

and such a classification was compared with the true vigilance state labeling (Awake, NREM,

REM) using a χ2 metric. We found that amplitude-sensitive clustering (χ2=15.96 ± 9.42) was

exhibiting a more skewed distribution (Fig. 7D) when compared to the amplitude-invariant one

which only uses the shape of ESI distribution while omitting all information related to amplitude

(χ2=13.54 ± 9.47; p=0.03; d=0.24). These results are supporting the idea that for most patients

the amplitude of the ESI maps aided in distinguishing between the maps in different vigilance

states, rather than changes in the spatial distribution of the sources themselves.

Figure. 7 The effect of relative and absolute threshold in the sensor and source space (A)

Spatial distribution examples of IEDs during NREM, REM and Wake from patient #5, who is a

39-year-old female with drug-resistant right mesiotemporal lobe epilepsy. (A) (i) The interictal

EEG showed IEDs predominantly over the right anterior to mid-temporal region (maximum

amplitude over F8, FT8, FT10, T4, Zy2 (ii) the corresponding voltage map (iii) the corresponding

source maps of the IEDs (C) 2-D view of the high-density EEG with the selected electrodes in red

which were selected in panel A. (D) Clustering differences with and without amplitude

information. Differences in χ2 results for all patients, with positive values indicating that the

clustering improved differentiation between the different states of vigilance when including

amplitude information in the clustering.

3.6 Discussion

In recent years the effect of the vigilance state on the ability to correctly localize the EZ

received increasing attention (Frauscher and Gotman, 2019; McLeod et al., 2020), with some

authors suggesting that the state of vigilance can alter the results of ESI’s spatial profile.

Confirmation of this hypothesis has the potential to change how we perform ESI in clinical

practice. Here, we aimed to understand the influence of the vigilance state on the ability to

accurately identify the sources of IEDs. We compared the localizability of IEDs during different

vigilance states against two gold standards, the set of clinically selected IEDs (i.e. when not

taking into account the vigilance state) and the clinical ground truth defined by clinical

judgment based on all presurgical information. Our overall findings indicate that: (i) NREM is

associated with higher IED numbers, higher amplitude peaks, and longer duration, but that (ii)

localizability of IEDs is not significantly impacted by the vigilance state; and that (iii) the effect

of the state of vigilance on IED localizability is primarily driven by amplitude and IED numbers

rather than the state of vigilance itself. Indeed, the spatial profile of the IED sources was not

significantly influenced by the vigilance state when controlled for IED amplitude. Overall, in

contrary to our primary hypothesis, our results suggest that the state of vigilance may hence

not be a crucial factor for IED source localization.

The state of vigilance does not significantly impact results of ESI

ESI is becoming a useful tool in presurgical evaluation of epilepsy patients (Megevand and

Seeck, 2020). However, if an interaction between the vigilance state and ESI results exists, it has

the potential to impact patient care by either hindering or improving the accuracy of ESI results.

We indeed found a high consistency between the various vigilance states and the clinical

ground truth (Fig. 5). ESI studies in the field of sleep and IEDs are scarce. To date, only two

studies have been conducted, one using a low-density array with 25 channels (McLeod et al.,

2022), and the other using a high-density array with 256 channels (Kang et al., 2020). Our study

supports the findings of the first study of McLeod et al. (McLeod et al., 2022), which found no

difference in source extent between different vigilance states except when comparing the

extent during REM sleep between unifocal and multifocal patients. However, since our study

only included unifocal patients, we cannot confirm this observation. Furthermore, this study

reported that REM had the most discordance with other vigilance states. However, in our

analysis, we did not find any significant differences when using the maximum localization

distances between different vigilance states. This discrepancy may be due to differences in the

methodological approaches. The second study by Kang et al. (Kang et al., 2020) reported that

ESI localizations were exhibiting fewer vertices involved during REM when compared to NREM.

Our results also support this claim, but only when we considered an absolute threshold defined

using the awake state for reference (as shown in Figures 3B and 6B). Indeed, in our study, we

carefully demonstrated that the little effect on ESI accuracy in different vigilance states was

mainly driven by ESI amplitude and not by the spatial profile.

Source localization as a measure of source extent

ESI can be challenging as it can vary depending on factors such as the density of the array (Lantz

et al., 2003), the SNR (Bast et al., 2006), or the methodology (Hedrich et al., 2017) used.

Assessing the source extent of the sources can be even more difficult, as different methods can

result in different findings (Samuelsson et al., 2021), and most standard linear approaches are

indeed not sensitive to the underlying source extent (Chowdhury et al., 2013; Chowdhury et

al., 2016; Pellegrino et al., 2020). This is particularly problematic in cases of deep epileptic

sources, as deep sources are traditionally difficult to detect using ESI (Krishnaswamy et al.,

2017). Given these limitations and our interest in assessing accurately the underlying source

extent of the generators, we considered an updated version of cMEM that has been shown to

accurately address this matter (Chowdhury et al., 2016; Hedrich et al., 2017). While relying on

cMEM’s ability to recover the source extent of the generators, here we adapted a depth-

weighting implementation of cMEM (Cai et al., 2022), allowing deeper sources to be amplified

and assessed more accurately. We used this method to allow inclusion of mesiotemporal lobe

epilepsy cases in our cohort. For the same reason, we also included a surface segmentation of

both hippocampi in the source space, these were modeled as distributed dipolar sources

perpendicular to this hippocampus surface (Attal and Schwartz, 2013). A previous HD-EEG ESI

study compared MNE and cMEM in different vigilance states (Kang et al., 2020). Interestingly,

the authors found that the effect of the vigilance state on the extent was reduced when cMEM

was used reaching only borderline significance (p=0.04). This discrepancy may explain our

results, as we analyzed more patients (16 vs. 6) and higher IED numbers using cMEM which

would result in more spatially reliable source maps.

Clinical quality of ESI is more impacted by amplitude and number of IEDs than the state of

vigilance

The quality and spatial accuracy of ESI in relation to the clinical ground truth, as shown by

spatial dispersion, assessing how much the ESI map was spreading around the clinical ROI, was

more sensitive to the number of IEDs than to the vigilance state. An increase in spatial

dispersion was noted in the equalized condition, when lower number of IEDs were averaged

(Fig. 5B), with no difference observed between vigilance states. Additionally, while no

significant variation was found in the maxima distances between the different vigilance states,

the overall D-maxima distance ranges also increased in the equalized condition, i.e. when a

lower number of IEDs was averaged (Fig. 5C). Interestingly, the distance of the maximum of the

map from the ground truth Dmin appeared unaffected (Fig. 5A). This highlights the crucial role

of the number of IEDs, in determining the reliability of the ESI extent. This important issue

might have impacted previous studies which explored the extent of the ESI using a low number

of IEDs (Kang et al., 2020; McLeod et al., 2022), and more likely lower SNR data. Of the initial

cohort in our study, 8 patients (30%) did not reach the minimum of 5 IEDs during REM sleep,

which we chose as an arbitrary criterion to achieve a reliable source localization. Our findings

also align with previous research (McLeod et al., 2020) indicating that these characteristics are

typically different between the various stages of vigilance with the largest differences being

observed between NREM and REM sleep. However, there is some heterogeneity of the

definition of the clinical ground truth at the individual level. Eleven patients out of 16

underwent surgery (5 had SEEG prior to surgery). Eight are Engel I and 3 are Engel II-IV. For the

5 remaining patients, 1 had hippocampal sclerosis and 4 had focal cortical dysplasia type II

lesions visible on the MRI. lt is important to mention that despite this heterogeneity of the

definition of the ground truth, a great concordance between the MRI lesion, the semiology, the

video-EEG monitoring as well as the high-density EEG, and the neuropsychological findings was

observed in all 5 patients who did not yet have surgery.

Choice of the threshold matters for IED and source extent

Previous studies have suggested that the channel involvement (Kang et al., 2020) and source

extent (Kang et al., 2020; McLeod et al., 2022) may be more focal during REM sleep. However,

our analysis using relative thresholds found no significant difference. Yet, when using an

absolute threshold defined considering data in the awake state, we did observe an effect and

found that REM sleep scalp data and localizations were exhibiting fewer involved channels and

vertices (Figures 3B,6B), aligning with these of the previous investigations. The reoccurrence of

the same discrepancy in the extent between relative and absolute threshold in both sensor (Fig.

3) and source space (Fig. 6) raises the question of what drives the source extent in those

results. Our results showed that incorporating amplitude information improved the clustering

results (Fig. 7), indicating that the amplitude itself contains valuable information, rather than

just the source distribution, to classify ESI maps to each specific vigilance state. This could also

explain the discrepancies observed in previous studies which did not consider this factor (Kang

et al., 2020; McLeod et al., 2022).

Limitations

This study utilized a moderate-sized epilepsy cohort, as we needed to exclude 30% of patients

due to stringent criteria of minimum 5 IEDs per vigilance state. Despite achieving source

assessment with Engel Class 1 outcomes only in patients undergoing surgery during

recruitment, we leveraged the clinical ground truth to mitigate this limitation. To ensure clinical

relevance beyond statistical significance, we calculated effect sizes and compared sublobar

level results. Finally, not all ground truths are equivalent in our study. A perfect ground truth of

an Engel 1A outcome was only available in 8 patients.

Conclusion

The results of this study did not support the hypothesis that REM sleep results in ESI maps that

are more focal or topographically distinct. This attests to the reliability of ESI across the

different vigilance states. Our findings also demonstrated that amplitude and the number of

IEDs is indeed more important than the state of vigilance.

3.7 Supporting Information

Methods

Anatomical MRI acquisition and head model

All patients underwent MRI scanning on a 3 Tesla Siemens Magnetom Prisma-Fit equipped with

a 64-channels head coil. The acquisition included high-resolution T1-weighted magnetic

resonance imaging (MRI) using a 3D magnetization - prepared rapid gradient-echo sequence

(MPRAGE; 0.8 mm isovoxels, TR = 2300 ms, TE = 3.14 ms, TI = 900 ms, flip angle = 9°, FOV =

256x256 mm2). The 3 T anatomical MRI was segmented, and the cortical surface was obtained

using the FreeSurfer software (http://surfer.nmr.mgh.harvard.edu). The forward problem was

solved using a boundary element method (Kybic et al., 2006) with 3-layer for brain, skull and

scalp (Lai et al., 2005) (conductivity of 0.33, 0.0165, 0.33 S/m) using OpenMEEG(Gramfort et al.,

2011) implemented in Brainstorm. The mid surface defined as the middle layer between gray

matter/pial and gray/white matter interfaces (Fischl et al., 2002), together with a surface

segmentation of both hippocampi, were considered as our source space for electrical source

imaging (ESI).

Electrical source imaging

For the purpose of this study, we implemented a new depth-weighted extension of the cMEM

framework (Chowdhury et al., 2016), in order to better address deep seated epileptic sources.

In this appendix, we will first review the main methodological concepts of the MEM framework

before introducing the depth weighting implementation that was carefully evaluated by Cai et

al. (Cai et al., 2022) in Near Infra-Red Spectroscopy 3D reconstruction and adapted here for ESI.

For this study, we considered a depth-weighted version of cMEM (Afnan and Gotman, 2023),

this version being able to localize more accurately deeper sources. This new improvement is

paramount for this paper, as a large portion of our cohort was composed of mesiotemporal

epilepsy cases (n=9). To do so, we applied the depth weighted extension of cMEM that we first

proposed in Cai et al. (Cai et al., 2022) and we also added surfaces of bilateral hippocampi in

the source space model. The activity estimates of each vertex calculated by electrical source

imaging (ESI) are theoretically subject to uncertainty, which is modelled by the variance

parameter in the source covariance matrix of the inverse problem. Deeper sources tend to have

greater uncertainty in ESI, resulting in higher variance values in the covariance matrix when

compared to superficial sources. Therefore, an a-priori source covariance matrix should

appropriately account for the variance differences across vertices. To do so, the diagonal of the

source covariance was weighted by the forward model of each particular vertex, quantifying

the influence of source depth. The weighting was done a-priori by setting the weighting

hyperparameter  =½ (see section on depth weighting below) as a midway compromise (0

would represent no depth weighting, whereas 1 would mean prioritizing deep sources) (Lin et

al., 2006). The noise covariance matrix estimate was done using a clean 2-second segment from

the same recording, this covariance was shared by all ESIs for a given patient for all vigilance

states.

MEM framework

Within the MEM framework, the amplitude of current density  estimated for J, i.e. amplitude

of J at each location in the source space (i.e., cortical and subcortical surfaces) at each time

sample, is considered as a random variable, described by the following probability

distribution󰇛󰇜󰇛󰇜. The Kullback-Leibler divergence or -entropy of 󰇛󰇜 relative to a

prior distribution 󰇛󰇜 (Eq.1)

(1)

󰇛󰇜󰇡󰇛󰇜

󰇛󰇜󰇢󰇛󰇜

󰇛󰇜󰇛󰇛󰇜󰇜󰇛󰇜



where 󰇛󰇜 is the -density of 󰇛󰇜 defined as 󰇛󰇜󰇛󰇜󰇛󰇜. Following a Bayesian

approach to introduce the data fit, we denote  as the set of probability distributions on j

that explain the mean of data (Eq. 2):

(2)

󰇣󰇟󰇠

 󰇤 

where  represents the measured amplitude changes, 󰇟󰇠represents the statistical

expectation of j under the probability distribution  , and  is an identity matrix of 󰇛󰇜

dimension (Eq. 3) where q is the EEG channel . Therefore, within the MEM framework, a

unique solution of 󰇛󰇜 could be obtained,

(3)



󰇛󰇜

The MEM solution (Eq. 4) would find a distribution of sources that maximizes the negative

entropy and, thus, maximizes the missing information as previously described (Amblard et al.,

2004)

(4)

󰇛󰇜 󰇟󰇛󰇜󰇛󰇜󰇛󰇜󰇠



  

Each cortical parcel k is characterized by an activation state, defined by the hidden variable ,

describing if the parcel is active or not. Therefore, we denote  as the probability of  parcel

to be active, i.e., 󰇛󰇜 is a Dirac function that allows to “switch off” the parcel

when considered as inactive (i.e., ). 󰇛󰇜 is a Gaussian distribution, describing the

distribution of absorptions changes within the  parcel, when the parcel is considered as

active (). This prior model, which is specific to our MEM inference, offers a unique

opportunity to switch off some parcels of the model, resulting in accurate spatial

reconstructions of the underlying activity patterns with their spatial extent.

Initialization of the reference distribution (prior)

The spatial clustering of the cortical surface into  non-overlapping parcel was obtained using

a data driven parcellization (DDP) technique. DDP consisted in first applying a projection

method, the multivariate source prelocalization (MSP) (Mattout et al., 2005), estimating a

probability like coefficient (MSP score) between 0 and 1 for each vertex of the cortical mesh,

characterizing its contribution to the data. DDP is then obtained by using a region growing

algorithm, along the tessellated cortical surface, starting from local MSP maximum (Lapalme et

al., 2006). Once the parcellation is done, the prior distribution 󰇛󰇜 is then a joint distribution

expressed as the multiplication of individual distribution of each parcel in Eq.5 assuming

statistical independence between parcels, we initialize each parcel with a Gaussian distribution

of the active state to be a zero mean.

To initialize the prior in Eq.4, the mean of the Gaussian distribution 󰇛󰇜 , was set to

zero.  at each time point, i.e. 󰇛󰇜 was defined as follows

(5)

󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜

󰆹󰇛󰇜



Where 󰇛󰇜 is a spatial smoothness matrix, defined by Friston et al. (Friston et al., 2008),

which controls the local spatial smoothness within the parcel according to the geodesic surface

neighborhood order 󰇛󰇜 was defined as  of the averaged energy of MNE solution within

each parcel  at time t.

Depth weighting

Depth weighted MNE

Minimum Norm Estimate (MNE) (Hamalainen and Ilmoniemi, 1994) tries to assess the

distribution of sources by minimizing the L2-norm. This is usually done using The Tikhonov or

Wiener regularization and SVD regularization procedures.

To achieve depth weighting we need to initialize the source covariance matrix as 

(Eq. 6)

(6)

󰆹󰇡󰇛󰇜



󰇢󰇛󰇛󰇜󰇜

󰇛󰇜 

󰇛󰇜

Depth weighted MNE (Lin et al., 2006) uses the forward model G for each vertex in the source

model to weight the covariance matrix. Therefore, it appropriately represents the distribution

of ESI uncertainty of each vertex across different brain regions.  is a weighting strength

parameter adjusting the amount of depth compensation. The larger is , the more depth

compensation is considered.  would therefore refer to no depth compensation and an

identity source covariance model.

cMEM

We then implemented depth weighting into the MEM framework, following the strategy we

proposed in Cai et al. (Cai et al., 2022) for Near InfraRed Spectroscopy 3D reconstruction, where

we carefully evaluated the performance of the method using realistic simulations. Depth-

weighted in cMEM was applied at two levels, respectively characterized by two depth weighting

parameters, was applied to solve the depth weighted MNE, as described in Eq.6, before

using those prior to initialize the source covariance model within each parcel of the MEM

model. Therefore, we used the depth weighted version of MNE solution described by Eq. 6,

avoiding biasing the initialization of the source covariance with a standard MNE solution. Then

 was used to apply depth weighting on the source covariance matrix  of each parcel  in

Eq.7. Consequently, the depth weighted version of 󰇛󰇜 is now defined as follows:

(7)

󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜󰇛󰇜

󰆹󰇛󰇜



where  (󰇛󰇜 

󰇛󰇜) is the depth weighting matrix for each

parcel k󰆹



󰇛󰇛󰇜󰇜

Metrics

Sensor level evaluation metrics

The sensor space data were assessed using the following metrics applied on the averaged IEDs

signals:

1. The absolute amplitude maximum in µVolt of the most negative channel at IED negative peak

of the averaged IED.

2. IED duration from take-off to offset, marked visually on the averaged IED on all channels.

3. Signal-to-noise (SNR) was assessed on the channel which displayed the highest negativity at

the peak of the average IED, and defined as the mean absolute amplitude ±100ms around the

IED peak divided by the standard deviation of a 200ms baseline window selected 5 seconds

before the IED peak (Eq. 8)

(8)

SNR= 

 





 󰇛

 



󰇜



 

Where, for the EEG electrode exhibiting the highest negativity,  denotes the IED signal at time

sample j within a ±100ms window around the peak (N = 200 samples) and  denotes for the

same channel at time sample I, the baseline signal within a 200ms window (M = 200 samples).

4. Channel involvement was defined as the count of channels which exceeded a negative

threshold at the time of the IED peak, which was determined by choosing the amplitude of the

most negative channel at the time of the peak. The relative channel involvement was defined

by using the relative threshold set as 50% of the averaged IED peak. The absolute threshold was

defined by using the threshold set to 50% of the averaged IED peak considering wakefulness

data as our reference for such comparison.

Source level evaluation metrics

100

After localizing the averaged IEDs in each vigilance state, ESI maps were assessed at the

midpoint of the rising phase of the IED, using the following metrics:

1. Dmin: the minimum distance localization error was computed as the Euclidean distance from

the maximum of the map to the closest vertex belonging to the clinical ROI (i.e. ground truth).

Whenever this maximum was located inside the clinical ROI, Dmin was set to 0 mm.

2. Spatial dispersion: this metric measures the spatial spread (in mm) of the localization around

the ground truth considered here as the clinical ROI. It was computed as the root mean square

of the distance from the ROI weighted by the energy of the source localization map on each

vertex (Eq. 9).

(9)

󰇛󰇛

󰇜



 





Where Θ denotes the set of vertices belonging to the clinical ROI, and  is the amplitude results

of the cMEM solver for the dipolar source on vertex  at the midpoint of the rising phase of the

averaged IED. 󰇛

󰇜 function provides the minimum Euclidean distance between the

dipolar source  to the closest vertex within the clinical ROI Θ. SD results are provided in mm.

3. D-maxima: when comparing ESI obtained for different states of vigilance, D-maxima was

measured as the Euclidean distance between the vertices exhibiting the maximum energy in

both ESI maps to be compared.

4. Spatial extent of the underlying source: the source spatial extent was assessed as the

percentage of vertices identified as activated above a specific threshold. Following our previous

study showing that cMEM spatial extent results are stable over a large range of thresholds

(Pellegrino et al., 2020), benefitting from the excellent contrast of cMEM maps, we decided to

apply a threshold set at 20% from maximum absolute intensity of the ESI map. We are

proposing two spatial extent metrics, one relative map by setting the threshold at 20% of the

101

source map maximum itself, and one absolute threshold set as 20% of the wakefulness source

map maximum.

5. Clustering: To further assess the effect of the vigilance state on relative versus absolute

effect of spatial extent, we examined the contribution of the amplitude variations to the ESI

maps. We proposed to apply ESI using cMEM on every individual IED available, followed by a

hierarchical clustering of all individual ESI maps, in a similar manner to the methodology

proposed by Chowdhury et al. (Chowdhury et al., 2018). To do so, we computed ESI using

cMEM for each IED for all available IEDs in each vigilance state. For every IED, we only

considered ESI results at one time sample, the exact peak of every IED. Every ESI map, a column

vector  estimated at the exact peak of every IED, was first normalized, by subtracting its

spatial mean (󰇛󰇜󰇜and dividing by its norm. We then clustered all the normalized

maps into 3 clusters, and compared the resulting clustering with the true classification in

NREM, REM and Wake states. We considered a hierarchical clustering approach, involving a

similarity matrix consisting as the dot product between each normalized ESI map  to every

other normalized ESI map . This proposed similarity metric was chosen to remove the

influence of ESI source amplitude on the clustering, and was shown to separate IEDs based on

the topography only(Chowdhury et al., 2018).

(10)



󰇛󰇜

, 





Where  denotes the transposition of the column vector 

.

We then we repeated this clustering process but this time we added an additional term to the

similarity matrix representing the amplitude of the ESI maps. This amplitude was the squared

difference of the norms between each two maps. This addition was normalized to have a

maximum of 1 in each row, such that the two most different maps in term of amplitude will

have a value of one. This term was then multiplied by a parameter equal to λ = 0.1 to reflect a

102

small contribution of the amplitude differences when compared to the similarity in spatial

profile (Eq. 10, 11).

(11)

󰇛󰇜



󰇛󰇜





Where λ is a fixed parameter. The λ parameter was tuned on a subset of 5 patients. Using this

new metric two ESI maps were considered closed together if they were exhibiting similar spatial

features and similar amplitudes. This allowed us to check whether the clustering changed when

an amplitude information was introduced. If the amplitude bares no effect on the ESI, then the

two clustering methods should be similar.

6. Sub-lobar clinical analysis: the ESI maps were reviewed by an epileptologist and determined

on a sub-lobar level if the peak of the map is concordant or discordant with the clinical ground

truth.

Unequalized condition

REM

Wake

Clinical

# interictal epileptiform

discharges

24 [6-102]

26 [17-226]

16 [6-55]

14.5 [5-77]

81 [18-223]

IED peak amplitude (µV)

58.99[33.45-

76.49]

67.39 [37.58-

98.78]

27.82 [18.86-

43.08]

36.61 [21.81-

82.60]

49.26 [32.54-

89.37]

Average IED duration (ms)

71.56±31.8

83.25±30.4

71.8±28.42

62.19±21.79

73.0±26.02

Signal-to-noise ratio

8.56 [4.85-13.39]

11.93 [6.52-

27.84]

5.74 [3.43-9.05]

4.69 [3.98-7.75]

13.46 [9.18-

22.67]

# involved sensor (relative

threshold)

13.5 [9.75-15]

12.5 [ 9-14.75]

13.5 [9.5-18]

14 [11-15.25]

13 [11-14]

# involved sensor (absolute

threshold)

15.5 [11-18.75]

16.5 [13-21.5]

11.5 [3-18.5]

Reference

18.5 [10.5-22.5]

Minimum distance (mm)

7.18±13.27

16.82±11.67

12.69±12.61

16.11±13.00

16.32±10.68

Spatial dispersion (mm)

18.69±6.4

18.4±6.3

19.14±7.88

18.78± 6.72

19.88±6.94

% of activated vertices (relative

threshold)

1.38±0.54

1.46±0.63

1.59±0.54

1.52±0.7

1.71±0.8

% of activated vertices (absolute

threshold)

1.57 [0.29-2.76]

1.76 [1.105-2.96]

0.26 [0-1.105]

Reference

2.38 [1.37-3.20]

103

Equalized condition

REM

Wake

Clinical

IED peak amplitude (µV)

51.3 [30.97-

72.09]

55.39 [37.02-

95.51]

28.35 [21.44-

41.86]

37.56 [24.67-

74.07]

60.27 [38.37-

75.01]

Average IED duration (ms)

84.12±36.93

83±30.34

71.06±24.36

59. ±25.43

77.12±40.82

Signal-to-noise ratio

5.1 [ 3.73-9.69]

5.98 [3.77-8.41]

4.23 [3.06-6.07]

4.46 [2.72-7.41]

7.12 [2.97-10.09]

# involved sensor (relative

threshold)

13 [11.25-15]

12.5 [9.75-14]

14.5 [9.5-19.25]

12.5 [8.75-14.25]

12.5 [11-14]

# involved sensor (absolute

threshold)

15 [10.75-21.5]

16.5 [11.75-18.5]

11 [3-14.25]

Reference

14.5 [11-14]

Minimum distance (mm)

15.85±10.72

15.07±11.75

15.73±12.86

14.81±12.47

15.18±10.01

Spatial dispersion (mm)

18.93±6.0

18.05±5.72

19.98±6.66

17.98±6.01

16.78±5.7

% of activated vertices (relative

threshold)

1.46±0.74

1.94±0.91

1.65±0.67

1.65±0.82

1.72±0.86

% of activated vertices (absolute

threshold)

2.16 [1.22-3.86]

3.07 [1.86-5.04]

0.82 [0-1.9375 ]

Reference

2.31 [0-9.15]

Table 1. All the values measured for each metric.

104

4. Chapter 4: Manuscript #3: The Awakening Brain is

Characterized by a Widespread and Spatiotemporally

Heterogeneous Increase in High Frequencies

Tamir Avigdor, Guoping Ren, Chifaou Abdalla, François Dubeau, Christophe Grova, Birgit

Frauscher. Advanced Science 2025. 2409608.

4.1 Preface

The second part of the thesis centered around the investigation of high frequencies broad-band

activity in physiology. For that purpose, I started by investigating high frequencies during the

transition from sleep back to wakefulness, i.e., awakening. I analyzed the progression of

spectral power and phase connectivity of high frequencies during the awakening process and

compared it to traditional frequency bands. I found that high frequencies increased in a

heterogeneous and widespread manner during awakening from both NREM and REM sleep.

However, the timing of when these high frequencies returned to wakefulness baselines, or

transitioned beyond their prior sleep boundaries, varied across different locations and

networks. I suspect that the association between high-frequency increases and awakening has

not been noticed in previous studies because they primarily utilized scalp EEG, which has

limited sensitivity for detecting high-frequency contributions. This finding is particularly

interesting, as the relationship between high frequencies and cognition is an emerging area of

research. This study demonstrates an association between high frequencies and the process of

regaining full consciousness during morning awakening, suggesting that high frequencies might

play a critical role in this process. In summary, increases in high frequencies are associated with

the transition from sleep to full wakefulness. These increases exhibit both spatial and temporal

variability, highlighting another aspect of the localized nature of sleep and wakefulness

transitions.

105

4.2 Abstract

Morning awakening is part of everyday life. Surprisingly, information remains scarce on its

underlying neurophysiological correlates. Here we used simultaneous polysomnography and

stereo-electroencephalography recordings from 18 patients to assess the spectral and

connectivity content of the process of awakening at a local level 15 min before and after the

awakening. Awakenings from non-rapid eye movement sleep were accompanied by a

widespread increase in ripple (>80Hz) power in the fronto-temporal and parieto-insular regions,

with connectivity showing an almost exclusive increase in the ripple band in the somatomotor,

default, dorsal attention, and frontoparietal networks. Awakenings from rapid eye movement

sleep were characterized by a widespread and almost exclusive increase in the ripple band in all

available brain lobes, and connectivity increases mainly in the low ripple band in the limbic

system as well as the default, dorsal attention, somatomotor, and frontoparietal networks.

106

4.3 Introduction

Sleep and wakefulness were classically thought of as a binary phenomenon, i.e. you are either

awake or asleep, until sleep stages were discovered. (Loomis et al., 1937) Awakening in the

morning can be viewed as the termination of the sleep process from either rapid-eye-

movement (REM) or non-REM (NREM) sleep. Behaviorally, we sometimes feel that it can take

some time to be “fully awake”. This is supported by studies that suggest that awakening is a

more gradual process (Horne and Moseley, 2011) which can take up to 30 min to return to the

wakefulness baseline. (Ferrara et al., 2000) This has been linked to the incomplete clearance of

adenosine at the time of the awakening. (Porkka-Heiskanen and Kalinchuk, 2011) Evidence

from scalp electroencephalography (EEG) showed that awakening, compared to pre-sleep

wakefulness, was characterized by an increase in delta (0.3-4 Hz) and theta (4-8 Hz) activity in

parieto-occipital regions, while increase in activity in the alpha band (8-12 Hz) was only

significant in the eyes closed condition compared to pre-sleep wakefulness. (Ferrara et al.,

2006) In addition, a decrease in beta (13-30 Hz) activity in the occipital lobe was observed

compared to wakefulness. (Marzano et al., 2011) The sleep-wake transition was also shown to

have a dependency on the preceding sleep stage. (Gorgoni et al., 2015) However, scalp EEG is

spectrally limited by muscle-related artifacts especially in the high frequencies over >30 Hz, and

more importantly >80 Hz. In addition, scalp EEG has a limited spatial resolution and only

provides a global view, making it difficult to perform a careful examination of the local variance

in neuronal activity. However, high frequencies may be particularly important in the awakening

process as they have been linked to cognition (Bosman et al., 2014) and consciousness (Rieder

et al., 2011) which are regained during this process. Intracranial EEG, a method performed in

humans in the context of epilepsy pre-surgical work-up (Frauscher and Gotman, 2019) allows

recording from multiple regions in the brain and provides a high quality signal with minimal

muscle artifacts, a necessity for the study of high frequency activity. More specifically, we used

stereo-EEG (SEEG), a type of intracranial EEG that allows recordings from both superficial and

deep structures by inserting depth electrodes into the brain with a higher spatial resolution as

opposed to scalp EEG at the location of the implantation probe. The use of sEEG recordings also

enables the assessment of high frequencies, providing a movement and muscle-artefact free

107

signal with high temporal resolution. The use of SEEG recordings in the study of human sleep

physiology has enabled the local variability of sleep to be demonstrated between different

brain regions and bands. (Nir et al., 2011; von Ellenrieder et al., 2020b) Interestingly, studies

using EEG-functional magnetic resonance imaging (fMRI) which tried to address the spatial

sampling gap have shown that awakening involves a deactivation process that begins in the

thalamus and spreads to the cortex. (Setzer et al., 2022; Song et al., 2022) Additionally, fMRI

revealed a reduction in the typical anti-correlation between regions of networks that are

usually anti-correlated during stabilized wakefulness, specifically between the task-negative

network and the task-positive networks. (Vallat et al., 2019) This reduction also varied

depending on the sleep stage. (Vallat et al., 2019) Using transcranial Doppler ultrasonography, a

reduced cerebral blood flow was observed during NREM compared to REM sleep. (Hajak et al.,

1994; Kuboyama et al., 1997; Vallat et al., 2019)

Due to these findings, we hypothesize that awakenings will demonstrate heterogeneous

spectral and connectivity signatures involving different regions and networks in different bands

depending on the prior sleep stage i.e. awakening from NREM or REM sleep. This is of particular

interest as the EEG signature of NREM(Davis et al., 1937) differs from REM (Dement and

Kleitman, 1957) sleep, (Lee et al., 2019) which suggests a different neural process; (Brown et al.,

2012; Abel et al., 2013) thus awakening from REM compared to NREM might require a different

process. Specifically, we are interested in the involvement of high frequency activity, which has

been postulated to be important for consciousness for many years, (Kahn et al., 1997; Rusalova,

2006; Siclari et al., 2017; Siclari et al., 2018; Ferrari-Marinho et al., 2020; Stephan et al., 2021)

Interestingly however, in patients with epilepsy, loss of consciousness was reported to display

the opposite trend. (Juan et al., 2023) Additionally, we propose that the behavioral effects of

morning sleep inertia— the grogginess upon waking, that temporarily impairs cognitive

performance but fades with time awake (Tassi and Muzet, 2000)—may reflect variability in the

timing and progression of the awakening process. Thus, we were interested in examining the

long temporal dynamics both before and after the awakening. Given that sleep and

wakefulness can exhibit local changes, it is important to investigate the period leading up to the

awakening. Additionally, considering that sleep inertia may imply lingering changes in the

108

minutes following awakening, exploring this time frame is essential for a comprehensive

understanding of the whole awakening process. Here we investigated in detail the

characteristic of the neural correlates of the process of awakening. We provide the temporal

dynamics of local cortical spectral property attributes and network connectivity changes in the

high frequency bands with respect to the preceding sleep stage. More specifically, we (i)

analyzed the amplitude-sensitive high frequency spectral content and the amplitude-insensitive

phase connectivity content of the final morning awakening from 15 min prior to and 15 min

after the awakening, (ii) compared these changes to normal wakefulness and stable preceding

sleep, and (iii) compared these high frequency changes to those of traditional bands in order to

obtain a complete picture on how the human brain awakens in the morning.

4.4 Results

We analyzed the morning awakenings of 18 patients with drug-resistant focal epilepsy who

underwent simultaneous polysomnography (PSG) and SEEG electrode implantations as part of

their pre-surgical epilepsy evaluation. We assessed the local changes in power and connectivity

that occur during the awakening process in various brain regions and networks. We examined

all SEEG channels deemed healthy by an epileptologist, as well as those determined to be in the

gray matter, totaling 695 channels (38.6±33.4 per patient). We focus on high-frequency bands,

but test all bands to evaluate the specificity of the findings, for every region or for every

network (Figure 1; Supp. Figures S1, S2; and Supp. Table S1) with sufficient coverage (≥3

patients and 5 channels). We report data from 7 patient awakenings from NREM sleep and 11

from REM sleep. Finally, we report results for 6 regions and 5 networks for NREM awakenings,

and 14 regions and 7 networks for REM awakenings (Figure 1; Supp. Figures S1, S2, and Supp.

Table S1).

109

Figure 1. Available coverage in the regional and network atlases. (A) All available healthy

channels from all patients. (B) Regional coverage in the MICCAI 17 anatomical atlas. (C)

Coverage of the networks in the Yeo 7 atlas. The color gradient indicates the number of

channels across all patients in each region.

All analysis is performed at the channel level to mitigate any confounding factors between

patients or regional differences. First, the time of scalp awakening (SA) is determined visually by

a board-certified neurophysiologist for each patient using scalp EEG. In a second step, the signal

is analyzed 15 minutes prior to and 15 minutes after that time (Figure 2). We assessed this

period and compared it to reference distributions (RD) of full wakefulness (wRD) and prior sleep

(sRD). We are interested in determining when each region or network “wakes up” i.e., if there is

local sleep or wakefulness that might be different from the global state and if so when does it

change. For this purpose, we assess when the signal diverges out of the bounds of sleep and

when it converges back to the bounds of wakefulness (see Methods). Thus, we aim to identify a

time point, which we named intracranial awakening (IA) times, for each region or network from

which we can safely say that the signal is no longer within the bounds of sleep, and a time point

where it will be back within the bounds of wakefulness. After having identified these time

points, we assess segments prior to the convergence back to wakefulness and segments after

110

the divergence from sleep. We assess them to further validate the statistical differences using

Wilcoxon rank sum test with a Benjamini-Hochberg correction, effect sizes using Cliff’s Delta,

and overall magnitude differences between these periods compared to sleep or wakefulness

using a relative deviation from the RD (see Methods). This is done to not only give an estimated

time of local awakening but also to assess the level of change during this process. The final

morning awakening spectral content was estimated using Welch’s method and phase

connectivity properties using the phase locking value (PLV). This was done for each patient,

spanning from 15 min before to 15 min after the awakening (Figure 2). Results reported below

are for bands, regions and networks which displayed significant changes (Tables 1-4, for a full

description of the results for all the bands, region and networks please refer to the Supporting

information (Supp. Tables S2-5). IA times in relation to the wakefulness baseline remained

largely (90%) identical when the wRD was constructed using segments from the late afternoon

or the morning, and were all between 5-10 seconds (Supp. Tables S2-5). Spurious spikes

detections accounted for <1% of signal, and the results did not change with or without the

removal of spurious spikes.

111

Figure 2. Overall scheme of the analysis process. (A) Preprocessing involves several steps. (i)

The awakening signals from 15 min before to 15 min after the scalp awakening (SA) were taken,

as well as baseline segments from the previous sleep from -15 to -10 min before the SA, and

wakefulness from the prior day. These baseline segments were then furthered bootstrapped.

(ii) Analysis was performed in each frequency band to assess the spectral properties using

Morlet wavelet, this was done for each contact, then all the channels grouped using the MICCAI

112

17 atlas. In addition, connectivity was estimated using phase locking value (PLV) for each

available channel pair within the same network, or between two different networks available,

in the Yeo 7 networks atlas. A similar analysis (power and PLV) was also applied to every

bootstrapped baseline segment, therefore creating two normative reference distributions (RD),

one from the previous sleep (sRD) and one from wakefulness (wRD). (iii) Then, in addition to

the original power/PLV, we also estimated their corresponding percentiles in comparison with

the sRD and the wRD. (iv) The original power/PLV and the percentile signals were combined

using a weighted median for all channels and patients within a given region or network. (v) The

final signals for analysis are presented. The original power/PLV signal and two categorized

percentiles signals one with respect to the sRD and one with respect to the wRD. The

categorized percentiles labelled as above 95% (red) within 5-95% (green) and below the 5%

(blue). (B) Description of the awakening process of the (i) power or PLV every 5 min for each

available region/network, as well as an example of the continuous temporal changes in one

region or network (in this example the frontal lobe) with a shaded interquartile in the region.

(ii) The categorized percentile with respect to the sRD and the (iii) categorized percentile with

respect to the wRD.

An example of the corresponding continuous changes in the representative lobe is also shown.

The points where the percentile diverges from the sRD and converges to the wRD are marked

as the intracranial awakening (IA) points. (C) Evaluation of the awakening was done from (1) the

time before the power/PLV converges to wRD and (2) after it diverges from the sRD, marked by

the IA time points i.e. the wakefulness IA and the sleep IA. A channel level analysis was

conducted using a Wilcoxon rank sum test with a Benjamini-Hochberg false discovery rate

correction. Then, we calculated the relative deviation from the RD as the weighted median

(WM) of the percentage difference at the channel level between the median awakening signal

to the median RDs.

Awakenings from NREM sleep are associated with widespread power changes

Awakening from NREM sleep was associated with spectral changes in multiple bands and

regions, with heterogeneous IA times compared to the SA (Table 1). Awakening from NREM

displayed a power increase in the high ripple band in the superior, middle, and orbital frontal

gyri and anterior part of inferior frontal gyrus, the insula, the central operculum, opercular part

of inferior frontal gyrus, the superior parietal lobule, the middle and inferior temporal gyrus,

temporal pole, planum polare, and the inferior parietal lobule (Figure 3). The relative deviation

from sRD ranged between 14-44% (d=.02-.66; p<0.01) and started between 50 s to 235 s after

113

the SA. The relative deviation from wRD ranged between 21-90% (d=.05-.74; p<0.01) in the high

ripple band and terminated between 235 s to 40 s before the SA.

As expected, lower frequencies in the delta band (Supp. Figure S3) displayed a widespread

decrease during the awakening process, involving all regions. The relative deviation from sRD

ranged between 44-270% (d=.16-.80; p<0.001), and started between 20 s to 215 s after the SA.

The relative deviation from wRD ranged between 30-81% (d=.35-.86; p<0.001) terminating

between 30 s to 65 s before the SA. In addition, a theta and alpha decrease, and a beta increase

were observed mainly in frontal regions during the awakening process (Table 1). The awakening

process showed a correlation between the power of scalp EEG at positions Fz or Cz and

intracranial EEG, primarily in the frontal neocortex at low frequencies, during awakening from

NREM sleep. As expected, there was no correlation with remote regions or structures situated

deep in the brain (Supp. Tables S7-8).

Figure 3. Spectral power in the high ripple band increases upon awakening from NREM sleep.

(A) Power from every available region across the brain is shown from 15 min before to 15 min

after the scalp awakening (SA), along with a regional example of the power with a quartile

range as the shaded area, and the respective intracranial awakening times for the sleep

114

reference distribution (sRD) and the wakefulness RD (wRD). (B) The overall percentile power in

relation to the sRD is presented for every region along with the same regional example as a bar.

same regional example as a bar. Note: Power represents the weighted median power in each

region, normalized across the entire brain. The reference distribution percentiles are based on

the weighted median in each region, which are then categorized as >95%, within 5-95%, or <

5%. Continuous power examples and comparisons to wakefulness and sleep reference

distributions are illustrated for the superior, middle, and inferior frontal regions. Deviations

<5% or >95% percentiles are deemed significant. Regions included in the example are the

superior, middle, and orbital frontal gyri, the anterior part of inferior frontal gyrus, the insula,

the central operculum, opercular part of inferior frontal gyrus, the superior parietal lobule, the

middle and inferior temporal gyrus, temporal pole, planum polare, and the inferior parietal

lobule. Data presented for 6 patients from a total of 118 channels, ranging between 15-32 per

region (Full description in Table S1 A)

High frequency phase connectivity increased in the multiple networks during the awakening

from NREM sleep

We then further explored and assessed the phase connectivity content of the awakening from

NREM sleep within and between all available networks of the Yeo 7 network atlas.

Connectivity-wise, awakening from NREM showed maximal changes occurring in only one or

two frequency bands, mostly within the ripple bands (Table 2). Connectivity increased during

the awakening from NREM in the high ripple band within the somatomotor, default, dorsal

attention, and frontoparietal networks (Figure 4). The relative deviation from sRD ranged

between 20-61% (d=.15-.79; p<0.001) and started between 45 s to 230 s after the SA. The

relative deviation from wRD within the somatomotor and dorsal attention networks ranged

between 29-124% (d=.79-.85; p<0.001), and terminated 20 s to 345 s prior to the SA.

Furthermore, connectivity increased in the ripple bands between the following networks:

somatomotor and ventral attention network, frontoparietal and default mode network, dorsal

attention and both the frontoparietal network and default mode network, and both the ventral

attention and frontoparietal networks. The relative deviation from sRD ranged between 12-59%

(d=0.13-.78; p<0.01) and started between 45 s to 390 s after the SA. The relative deviation from

wRD ranged between 125-288% (d=.10 -.72; p<0.001) and terminated 20 s to 40 s prior to the

SA, but only between the somatomotor and ventral attention network, somatomotor network

and default mode network, as well as between the default mode network and ventral attention

115

network. Lastly, the connectivity of the beta bands showed a minor decrease between the

dorsal attention network and both the frontoparietal and default mode network, and between

the default mode network and frontoparietal network. The relative deviation from sRD ranged

between 3-27% (d=.07-.33; p<0.001) and started between 95 s prior to the SA to 110 s after the

SA. Unexpectedly, no divergence from the wRD was observed during the awakening process.

Interestingly no significant changes in the PLV connectivity were observed in any other

frequencies.

Figure 4. Phase connectivity within networks in the high ripple band increases with

awakening from NREM. (A) Phase locking value (PLV) within each available network in the

brain is shown from 15 min before to 15 min after the scalp awakening (SA), along with a

network example from every available network. Quartile ranges are represented by the shaded

area. The respective intracranial awakening times for sleep reference distribution (sRD) and

wakefulness RD (wRD). (B) The overall PLV percentile in relation to the sRD within every

network complemented by a network example in the bar. (C) The overall PLV percentile in

relation to the wRD is presented by the bar, with an accompanying network example above.

Note: PLV represents the weighted median power within each network. The reference

distribution percentiles are based on the weighted median within each network, which are then

categorized as >95%, within 5-95%, or <5%. Continuous PLV examples and comparisons to

wakefulness and sleep reference distributions are illustrated for the somatomotor network.

116

Deviations below the 5% or above the 95% percentiles are deemed significant. Included

networks are: somatomotor, dorsal attention, frontoparietal and the default mode network.

Data presented for 6 patients from a total of 178 channels, ranging between 5-43 per network

(Full description in Table S1 B).

Awakening from REM sleep is associated with a widespread power increase exclusively in

high frequencies

The observation that high frequencies are associated with the awakening from NREM sleep

raised the question of whether they may also be involved in awakenings from REM sleep. REM

sleep, unlike NREM sleep, is classically characterized with a “wake-like” EEG signal. However,

high frequencies are difficult to detect on the scalp. In contrast, SEEG offers a unique

opportunity to investigate local changes in high frequencies. Therefore, we also explored the

process of awakening from REM sleep.

Awakening from REM sleep was associated with a widespread power increase almost

exclusively in the low and high ripple bands (Table 3). These increases were noted to occur

widespread over the brain (Figure 5, full list in Table 3). The relative deviation from sRD ranged

between 5-39% (d=.03-.44; p<0.01) and started between 50 s prior to the SA to 260 s after the

SA. The relative deviation from wRD ranged between 36-88% (d=.02-.57; p<0.01) and

terminated 45 s before to 45 s after the SA, with some regions failing to converge back.

Interestingly, changes in lower frequencies (delta, alpha and beta) were minimal in most

regions, except in the inferior parietal lobule, medial and basal temporal region, middle and

inferior temporal gyrus, temporal pole, planum polare, superior, middle, and orbital frontal

gyri, anterior part of inferior frontal gyrus, and the superior temporal gyrus. This manifested as

a relative deviation from sRD ranging between 11- 87% (d=.11-.73; p<0.01) starting between

480 s before to 50 s after the SA, but never diverging from the wRD.

117

Figure 5. Spectral power in the high ripple band increases upon awakening from REM sleep.

(A) Power from every available region across the brain is shown from 15 min before to 15 min

after the scalp awakening (SA), along with a regional example of the power with a quartile

range as the shaded area, and the respective intracranial awakening times for sleep reference

distribution (sRD) and wakefulness RD (wRD). (B) The overall percentile power in relation to the

sRD is presented for every brain region, complemented by the same regional example as a bar.

the same regional example as a bar. Note: Power represents the weighted median power in

each region, normalized across the entire brain. The reference distribution percentiles are

based on the weighted median in each region, which are then categorized as over 95%, within

5-95%, or below 5%. Continuous power examples and comparisons to wakefulness and sleep

reference distributions are illustrated for the middle and inferior temporal gyrus, temporal

pole, and planum polare regions. Deviations below the 5% or above the 95% percentiles are

deemed significant. Included regions are: medial occipital lobe, medial and basal temporal

region, medial parietal lobe, superior parietal lobule, superior temporal gyrus, medial frontal

cortex (including medial segment of superior frontal gyrus), inferior parietal lobule, superior,

middle, and orbital frontal gyri and anterior part of inferior frontal gyrus, insula, middle and

inferior temporal gyrus, temporal pole, and planum polare, anterior and middle cingulate gyrus,

frontal operculum, medial and basal temporal region, transverse temporal gyrus and planum

temporale. Data presented for 12 patients from a total of 525 channels, ranging between 8-127

per region (Full description in Table S1 A).

118

Connectivity increases in the ripple bands within the default mode, dorsal attention,

somatomotor, and frontoparietal networks during the awakening from REM sleep

We further investigated the phase connectivity content of the awakening from REM sleep

within and between all available networks in the Yeo 7 network atlas. This investigation

revealed that awakenings from REM were associated almost exclusively with increases in the

low ripple band across multiple networks (Table 4).

When awakening from REM sleep, a widespread modest increase in the low ripple band PLV

was observed within the limbic network and to a lesser degree, in the default mode, dorsal

attention, somatomotor, and frontoparietal networks (Figure 6). This modest change in the

relative deviation from sRD ranged between 9-39% (d=.20-.60; p<0.001) and started late,

diverging between 45 s to 380 s after the SA. However, only the limbic network differed from

the wRD, displaying a relative deviation of 36-40% (d=.30-.38; p<0.001) terminating 40 s prior to

the SA, while other networks were not affected. We also observed that between networks,

there was very widespread PLV increase in the low ripple band in most, but not all network

pairs. This modest change in the relative deviation from sRD ranged between 2-56% (d=.07-.78;

p<0.01) and started late, diverging between 45 s to 430 s after the SA. However, keeping in line

with the known similarities between wakefulness and REM sleep, the phase connectivity in

most networks did not differ from the wRD during the awakening process. Notable exceptions

were the high ripple band, which showed a decrease in the PLV between the limbic,

somatomotor, visual and ventral attention networks during the awakening process. This

decrease was mainly apparent in relation to the wRD having a relative deviation ranging

between 14-81% (d=.21-.95; p<0.01). This change terminated between 75 s to 455 s prior to the

SA, however, some connections never converged back to the wRD. In addition, the PLV within

the visual network increased during the process in relation to the wRD, only showing a relative

deviation ranging between 20-40% (d=.59-.61 p<0.01). This change terminated 70 s prior to the

SA.

119

Figure 6. Phase connectivity within networks in the high ripple band increases with

awakening from REM. (A) Phase locking value (PLV) within each available network in the brain

is shown from 15 min before to 15 min after the scalp awakening (SA), along with a network

example from every available network. Quartile ranges are represented by the shaded area.

The respective intracranial awakening times for sleep reference distribution (sRD) and

wakefulness RD (wRD) (B) The overall PLV percentile in relation to the sRD within every network

complemented by a network example in the bar. (C) The overall PLV percentile in relation to

the wRD is presented by the bar, with an accompanying network example above. Note: PLV

represents the weighted median power within each network. The reference distribution

percentiles are based on the weighted median within each network, which are then categorized

as >95%, within 5-95%, or <5%. Continuous PLV examples and comparisons to wakefulness and

sleep reference distributions are illustrated for the limbic network. Deviations below the 5% or

above the 95% percentiles are deemed significant Networks include: visual, limbic, dorsal

attention, ventral attention, somatomotor and default mode. Data presented for 12 patients

from a total of 513 channels, ranging between 5-43 per network (Full description in Table S1 B).

4.5 Discussion

Awakening is a complex heterogeneous process which terminates the sleep period. Here we

examined the phenomenon of the awakening process from a local perspective using SEEG from

multiple brain regions. Using depth electrodes, we acquired not only high spatial resolution

recordings, but also high-quality recordings while avoiding contamination by muscle artifacts as

often encountered in scalp EEG recordings during body movements, as typically present during

120

awakenings. We leveraged these inherent advantages of SEEG to examine the awakening

process, which for the first time is being evaluated for its high frequency content. This is of

particular interest given the tight link between high frequencies and cognitive processes. (Fell

et al., 2010) Our results demonstrate: (i) the association of fast frequencies in the process of

awakening, (ii) some regional differences might be influenced by the prior sleep stage, and (iii)

the spatio-temporal heterogeneity of awakening from both a spectral and a phase connectivity

point of view.

Our results show that high frequencies in the ripple range >80Hz are associated with the

process of awakening (Figures 3-6). These results are in line with the known association

between fast frequencies and cognition, (Bosman et al., 2014) as well as consciousness, (Rieder

et al., 2011) which forms the essence of the transition from sleep to wakefulness. Assessing

high frequencies poses challenges without high quality, low noise, and high sampling rate

recordings. We suspect that this is the reason that to date, the investigation of awakenings

have not been able to capture the contribution of high frequency activity. Nevertheless, when

an attempt was made to investigate the sleep-wake cycle and cognition using SEEG, it revealed

the involvement of the low gamma band. (Gross and Gotman, 1999) We observed that both the

spectral content (Figures 3, 5) as well as the phase connectivity of high frequencies (Figures 4,

6) increased during the awakening process. Our spectral results correspond to a recently

described high frequency index which was able to discriminate between NREM, REM and

wakefulness in rats. (Silva-Perez et al., 2021) In our case, the spectral content of high

frequencies gradually increased during the awakening process. In addition, the connectivity

increase during the awakening process aligns with previous findings that showed a higher

connectivity in wakefulness compared to NREM in the low ripple band using another phase-

based connectivity metric (weighted phase lag index). (Banks et al., 2020) Our results, which

demonstrate an association between increased higher frequencies and awakening, align with

accumulating evidence suggesting that higher frequencies play a critical role in consciousness.

(Liu et al., 2022; McCafferty et al., 2023; Khan et al., 2024) These findings might support the

idea that changes in high-frequency activity are key during the transition from sleep to

wakefulness. When NREM parasomnias were investigated using SEEG, it was observed that the

121

interplay between high frequencies (>20 Hz) and delta activity appears to be associated with

the level of conscious awareness. (Gibbs et al., 2016) Furthermore, another recent study which

used SEEG to investigate the connectivity of transient high frequencies oscillations in

wakefulness and NREM sleep demonstrated a similar widespread synchronization across

different regions. (Dickey et al., 2022b) We opted to use ripple band activity rather than

individual ripple events in order to assess the progressive change in the signal in all the

frequency bands with a consistent methodology. In addition, it would be difficult to assess just

using ripple rates due to their high variability and sparsity of about 3±3 per min-1. (Dickey et al.,

2022b) Our findings of increases in high frequencies might be in line with past investigation into

dreaming which also reported such associations. (Siclari et al., 2017) suggest shared neural

dynamics between awakening and dream-related processes. In addition, due to the various

timing of changes in high frequencies, some of which occur before the scalp awakening when

the patient still is behaviorally asleep, and due to the lack of large-scale cross-band activity

(specifically the absence of changes in the beta band), we believe that our results reflect a real

process taking place during the awakening process and are not solely the result of movement

artifacts.

Awakenings from NREM and REM sleep display different patterns, with the awakening process

being influenced by the prior sleep stage (NREM or REM), both in terms of the spectral and

connectivity content. Spectrally, awakenings from NREM required changes in multiple bands,

with lower bands decreasing and high frequency bands increasing. Awakenings from REM

consisted mainly of an increase in the gamma-ripple bands. Interestingly, awakenings from

REM, which are classically hard to distinguish from wakefulness using scalp EEG alone, (Scarpelli

et al., 2015) displayed a widespread increase in the ripple bands and only a few changes in

lower frequencies (Table 3).

Previous investigations using scalp EEG have shown an increase in delta activity in the first few

minutes after awakening from NREM and REM compared to the pre-sleep wakefulness. (Ferrara

et al., 2006; Marzano et al., 2011) In contrast, our results showed a gradual decrease in delta

power which converged to wakefulness tens of seconds prior to SA. We suggest that this

discrepancy may be explained by differences in definitions of when divergence/convergence

122

are determined. In our study, we employed a stricter definition requiring the signal to be within

the 5-95% percentile to be considered as divergence/convergence, as opposed to accepting any

statistically significant change. Spatially, we observed changes in the delta band in the frontal,

partial, and medial regions. While these results concur with that of previous studies, (Ferrara et

al., 2006; Marzano et al., 2011) we did not have enough sampling of occipital channels to assess

the changes in delta previously reported. (Marzano et al., 2011) It was also reported that beta

activity post-awakening from REM, and to a lesser extent from NREM, (Hilditch et al., 2023) was

lower compared to pre-sleep wakefulness. We observed these beta power changes in during

awakening from NREM in the superior, middle, and orbital frontal gyri, and the anterior part of

the inferior frontal gyrus region, as well as during awakenings from REM in temporal parietal

regions. In addition, the observation of significant differences in high frequency activity

between REM sleep and wakefulness at the moment of awakening aligns with earlier findings

from high-density EEG studies. These studies reported more extensive variations in beta power

(12-20 Hz) compared to delta power (1-4 Hz, primarily localized to sensory cortices) during

consistent periods of REM sleep versus wakefulness. (Baird et al., 2018) We can speculate that

the reason more spectrally wide changes occur in awakenings from NREM might be because

NREM is more impaired during NREM (Rechtschaffen et al., 1966; Siclari et al., 2013) thus

required more changes to return to full consciousness. The observed increase in high frequency

activity during awakening may be linked to cognitive recovery and sleep inertia. Disruptions in

this activity could underlie the cognitive impairments of sleep inertia. The availability of

multiple nights with a clear morning awakening of the same patient using sEEG was not

available. We are unable to compare directly NREM and REM due to the lack of availably of

awakenings from both stages within the same patient. Thus, future research is needed to assess

the full effect of the differences between awakening from NREM and REM.

Recently, more evidence has accumulated to support the role of connectivity within the

different brain networks in facilitating transitions between sleep and wakefulness.

(Adamantidis et al., 2019; Sulaman et al., 2023) We wanted to explore these differences during

the awakening process. From a phase connectivity perspective, we found that awakening from

NREM showed a focal change almost exclusively in the low and high ripple bands (Table 2). This

123

was most apparent within network phase connectivity in the somatomotor, default mode

network, and dorsal attention networks (Figure 4), as well as in the connectivity between these

networks. However, not all networks participated, with some networks never displaying any

significant changes during the process. When awakening from REM sleep, phase connectivity

changes were mainly observed in the low ripple band (Table 4) and were prominent mainly in

the limbic network (Figure 6), as well as other networks such as the default mode network. We

opted to show the most prominent results for each analysis. In figure 6, we present the results

from the low ripple band. Our phase connectivity results align with EEG-fMRI findings, showing

that awakenings result in a different correlation connectivity pattern depending on the prior

sleep stage, (Hajak et al., 1994; Kuboyama et al., 1997; Vallat et al., 2019) particularly within the

default mode network. (Stevner et al., 2019; Vallat et al., 2019) In addition, we expanded these

findings by exploring higher frequencies which are not accessible using fMRI and were able to

demonstrate their role in the waking process. We observed a similar upward trend in both

power and phase-based connectivity in the ripple band; however, this similarity did not extend

to other bands. Further research is needed to elucidate this connection, which has also been

observed for transient high frequency oscillations (Dickey et al., 2022b). On top of that, we

focused on morning awakenings rather than awakenings from a short nap as they are more

representative of the full phenomena of human sleep. Two recent seminal studies (Setzer et al.,

2022; Song et al., 2022) on awakening using EEG-fMRI discovered that the thalamocortical

connection preceded the scalp awakening. However, we were unable to confirm these finding

due to a lack of thalamic recordings in our cohort. The more spectrally restricted signature of

REM awakening, focused on increases in high frequencies, may suggest that different states of

consciousness require distinct alterations to transition back to regular wakefulness. Future

research is needed to further clarify the potential effects of circadian rhythm on spectral and

connectivity dynamics, particularly given its demonstrated influence on sleep inertia. (Scheer et

al., 2008; Silva and Duffy, 2008; Burns et al., 2024)

Awakening is a heterogeneous process. We observed a high degree of spatial and temporal

variability during awakenings (Tables 1-4). This spatial heterogeneity adds information on the

regional variability of wakefulness(Frauscher et al., 2018b) and sleep (von Ellenrieder et al.,

124

2020b). However, the temporal heterogeneity that we observed demonstrates that the global

phenomenon of the transition between sleep to wakefulness, as reflected using scalp EEG, is

not uniform, and does not follow a clear gradient as seen, for example, for NREM to REM sleep

transitions. (Peter-Derex et al., 2023a)

The boundaries between sleep and wakefulness appeared fluid. When evaluating the

awakening process, there are two ways to examine it: one as a temporal change from sleep i.e.

comparing it to the immediate time prior to wakening, and second, as its difference to full

wakefulness i.e. comparing it to resting state from wakefulness. In order to address both of

these questions, we performed a continuous comparison at each time point to both the sRD

from -15 to -10 min prior to the SA, and to the wRD taken from the previous day. This allowed

us to assess both the signal transition from sleep to wakefulness, as well as to determine when

we returned to the bounds of normal wakefulness. Throughout the results we noticed a

discrepancy between the IAs of the wRD and the sRD. This may be explained by an expected

overlap in the boundaries between all ranges of wakefulness activity and the boundaries of

sleep. We postulate that these edges, might reflect the overlapping boundaries of the different

states of consciousness. This might be related to the notion of local regulation of

consciousness, in which different states such as sleep and wakefulness can co-occur in different

brain regions at the same time. (Ferrara and De Gennaro, 2011; Nir et al., 2011; Andrillon et al.,

2015) The later can also explain our observation that some brain region never diverged from

wakefulness or never converged back to it. We suggest that higher frequencies, although not

exclusively, might play a role in facilitating the transition between sleep and wakefulness

without ever exceeding any boundaries. In addition, awakening from REM did not converge

within the 30 min to the wRD in one area (medial and basal temporal region), perhaps as more

time might be needed for it to return back to normal wakefulness, potentially contributing to

the behavioral phenomena of sleep inertia. In addition, wakefulness (Brodbeck et al., 2012) and

arousal states(Peter-Derex et al., 2015) are heterogeneous process with many different

patterns of activity occurring within it, as reflected in the EEG. This variability may cause the

wRD to have a wide range of possible activities since the segments used were from daily life

rather than a controlled environment. This, in turn, suggests that it is more difficult to diverge

125

from our wRD, especially with a strict 5-95% threshold. This may explain why some regions

showed changes in relation to the sRD but did not show changes in relation to the wRD. Future

research is needed to better control for different sensory inputs, times, and environments of

the wakefulness distribution in order to further parse which activities diverge and which do not.

Variability between patients was visible in different regions and networks at specific time

points. Future research with a larger sample size is needed to untangle these differences with

respect to visual and auditory inputs.

Our study has the following limitations. One inherent issue of SEEG is its limited sampling of the

brain, which means that we do not have a complete brain coverage for each sleep stage leading

up to awakenings. To address this limitation, we aggregated the results across all available

channels obtained from all patients, and then grouped them using an atlas. However, this still

limits our ability to directly compare NREM and REM awakenings as we have different coverage

with their respective patients, which may not share all regions and networks. We were also

unable to control directly for the circadian rhythm when selecting for the control segment

however we choose them from the same part of the day. Moreover, this approach is limited for

subcortical-cortical connections such as thalamo-cortical, which are not usually implanted. In

addition, this aggregation may lead to under-detection of small effects, as they are more likely

to be overshadowed by inter-patient variability. Obtaining human SEEG recordings is restricted

to epileptic patients undergoing SEEG during pre-surgical evaluations in a clinical setting. Also,

patients are usually on antiseizure medication that can impact sleep architecture. Finally,

epilepsy itself causes changes to the EEG beyond epileptic spikes, and can go along with

comorbid sleep disorders. To address this concern, we carefully deselected all channels with

abnormal EEG activity following a strict protocol to address this issue as effectively as possible.

To the best of our knowledge, all of the awakenings are spontaneous awakenings. That being

said, data were recorded in a clinical epilepsy monitoring unit setting and not a sleep laboratory

so that external noise resulting in some of the awakenings cannot be totally excluded. However,

we believe that finding reproducible effects speaks against the fact that different types of

awakenings were included. Direct comparisons between NREM and REM within individual

patients were not feasible due to the study's limited duration, which allowed for only a single

126

night of sleep analysis. Comparisons across patients were also constrained by the inherent

sampling limitations of SEEG, as each implantation is individually planned to meet the specific

clinical needs of each patient. This variability in electrode coverage prevents consistent cross-

patient comparisons. As a result, the absence of findings should not be interpreted as negative

results but rather a call for future studies to gather more data on the awakening process in

these missing regions. Finally, future work might be needed with multiple nights of recordings

in the same patients, with both awakenings from NREM and REM, to directly compare them

within the same patients and region. Lastly an avenue that we could not explore here is the

influence of dream content on the awakening process; both REM and NREM contain dreams

even if their frequency of recall varies, (Siclari et al., 2018) yet their dream content differs.

(McNamara et al., 2010) While the primary limitation of SEEG lies in its restricted coverage, as it

allows access to only a limited portion of the brain in each patient. However, recent pioneering

work has demonstrated the feasibility of combining SEEG with high-density EE, (Samogin et al.,

2019; Seeber et al., 2019; Mikulan et al., 2020; Parmigiani et al., 2022; Marino and Mantini,

2024) achieving comprehensive brain coverage. Future research stands to gain significantly

from adopting this innovative yet challenging methodology to investigate neural dynamics

during awakening. By leveraging this combined approach, researchers can better assess long-

range changes in spectral power and connectivity.

In summary, the process of awakening is complex and can differ in both timing and location.

Nonetheless, high frequency bands appear to be associated with the process of awakening

from both NREM and REM sleep across multiple brain regions. Our findings indicate that, while

scalp recordings can provide a global impression of the awakening, local recordings reveal

interactions between regions that can occur several minutes before or after the scalp

awakening. Finally, we demonstrated that the prior sleep stage can influence the local spectral

and connectivity activity patterns during the awakening.

127

4.6 Materials and Methods

Patient selection

Eighteen consecutive patients with drug-resistant focal epilepsy (7 female; age=36.9±11.5

years, Table. S6) who underwent combined SEEG and PSG recordings as part of their presurgical

investigation at the Montreal Neurologic Institute and Hospital between January 2013 and

June 2022 fulfilled selection criteria (see flowchart of patient selection in Supp. chart. S1).

Inclusion criteria were clear awakenings without falling back to sleep, age >15 years, recordings

composed of at least 5 healthy non-epileptic intracranial channels outside the seizure-onset-

zone (SOZ), recordings with a minimum duration of 10 min after the morning awakening,

availability of a post-implantation imaging for accurate anatomical localization of the individual

channels, and absence of seizures in the sleep cycle prior to the morning awakening. All

awakenings corresponded to spontaneous awakenings. We did not study forced awakenings as

we were interested in the natural process of awakening. We excluded patients without a clear

final morning awakening, defined as no N1 in the 15 min after awakening, as well as patients

who did not have stable sleep, defined as ≥3 N1 epochs in the 5 min prior to awakening. The

study was approved by the Montreal Neurological Institute and Hospital Review Ethics Board

(2014-183).

Intracranial EEG and scalp EEG recordings

An average of 11.5±4.3 depth MNI SEEG electrodes (9 contacts of 0.5-1 mm, distance between

contacts 5 mm; 4 patients) or DIXI SEEG electrodes (10-15 contacts of 2 mm, distance between

contacts 1.5 mm; 14 patients) were stereotactically implanted in every patient. Scalp EEG was

obtained simultaneously with subdermal thin wire electrodes at positions Fz, Cz, Pz, F3, C3, P3,

F4, C4, P4, and additional electrodes for electrooculogram and chin electromyogram were

applied during the night of the sleep recording. The sleep recording was chosen at least 72 h

after implantation to avoid effects of anesthesia as frequently seen in the first days after

implantation. (Frauscher et al., 2015)

128

EEGs were recorded using the Harmonie EEG system (Stellate) from 2013 to 2017 and the

Neuroworkbench EEG system (Nihon Kohden) after 2017. Recordings were acquired with a

common reference (epidural electrode fixed in the bone, far from the epileptic field). Hardware

filter settings were 0.1 Hz for the high pass filter and 500 (Stellate) or 600 Hz (Nihon Kohden)

for the low pass filter. All recordings were sampled at 2 kHz.

Channel selection and classification

Each SEEG contact was clinically assessed by an epileptologist (B.F.), and only normal healthy

channels were selected. Healthy channels were defined as channels situated within healthy

tissue, as determined by MRI scans, positioned away from the area where seizures originate,

and consistently free from interictal epileptic discharges and significant slow-wave anomalies

throughout the circadian cycle. This assessment was based on the comprehensive implantation

report and a detailed examination of one night's sleep by a board-certified electrophysiologist.

Channels in the white matter were further excluded, (Frauscher et al., 2018b) resulting in 695

channels (38.6±33.4 per patient) for analysis (Figure 1).

Registration of the electrodes was done in a similar manner to our previous work (Frauscher et

al., 2018b) where Minctools and the IBIS framework (Drouin et al., 2017) were used. This

involved first linearly aligning the CT/MRI images taken after electrode implantation, which

display the locations of the electrodes, with the pre-implantation MRI scans (preMRI) for each

patient. Following this, the preMRI images were non-linearly aligned with the ICBM152 non-

linear symmetric brain model. (Fonov et al., 2011) Through this combined transformation

process, it was possible to accurately estimate the electrode positions within a unified

stereotaxic space. Channels were then classified into anatomical regions using the MICCAI atlas

(B. Landman, 2012) reduced to 17 regions (Frauscher et al., 2018a) combining hemispheres that

offer a sufficient anatomical coverage and granularity, and into brain networks using the Yeo 7

networks atlas. (Buckner et al., 2011) The Yeo 7 network atlas has recently been applied in

sleep research investigating sleep dynamics using fMRI, (Haimovici et al., 2017; Cross et al.,

2021) and offers a functional way to assess connectivity results. Registration of the electrode

129

positions was done independently for the MICCAI atlas and the Yeo 7 networks using a 5mm

radius growing region. Finally, patients were grouped using the last sleep stage prior to their

awakening, either from REM or NREM sleep. Only regions from ≥3 patients and ≥5 channels and

networks that had ≥3 patients and ≥5 channels and at least two different structures were

analyzed for spectral and connectivity content (Figure 1, Supp. Figures S1, S2, and Supp. Table

S1).

Sleep scoring and segment selection

Sleep scoring was done by a board-certified neurophysiologist (B.F.) according to the American

Academy of Sleep Medicine criteria. (Berry et al., 2017) The exact time of the SA was marked

independent of the SEEG. Final morning awakenings (6:30-9:00 AM) were selected for each

patient based on the SA time. The awakening time was determined as “final” if the following

wake period did not contain any N1 epochs after the SA. This time point was then used as the

reference awakening throughout the paper when examining changes in SEEG channels. We

then extracted 15 min before and after the SA time in order to assess the full extent of the

awakening as a complete process. 15 minutes were chosen as in the clinical setting post-

awakening the patients were disconnected shortly after.

Baseline segments were selected as references to compare each time point during the

awakening process. We used two reference distributions by: (i) using 10 min wakefulness

segments taken from the evening before the sleep recording (>2 h before sleep onset), (ii)

segments from the sleep period between 10 to 15 min prior to the SA. Segments were visually

inspected and selected by a board-certified neurophysiologist (C.A.). We selected channels

without epileptic spikes and segments that were >90 min away from seizures. Channels without

spikes were defined as non-epileptic by a board-certified epileptologist. In addition, we

quantitatively detected spikes and selected a cut-off of 3 spikes per 10 min, which corresponds

to the false positive rate of the automatic detector. (Ho et al., 2023) This sleep segment

overlapped intentionally with the -15 min to +15min from the SA in order to evaluate the

relative evolution of the awakening itself. Finally, to avoid any spurious spiking activity, all

channels were checked automatically for epileptic spikes detection (Janca et al., 2014) and

130

corresponding detected signals were discarded using a window of -100 ms to 100 ms around

the peak of each spike for both the awakening and control segments. These selections

accounted for less than 1% of the total signal. To definitively rule out any effects of removing

these signal periods, we also performed the identical analysis not rejecting the quantitatively

identified spurious spiking activity.

Preprocessing and analysis

The SEEG signals were analyzed using a common average montage, which incorporated all non-

SOZ and visually detected artifact-free channels for both power and connectivity analysis.

Signals were filtered with a Butterworth band-pass filter ranging from 0.3 to 300 Hz. We

analyzed the awakening process starting 15 min before the awakening, as detected on the scalp

EEG, and continuing until 15 min after. The analysis was done over 5-s segments, without

overlap, throughout the 30-min signal. 5-s windows were chosen as a window in order to

guarantee reliable spectral (Dumermuth and Molinari, 1987) and connectivity (Basti et al.,

2022) estimates in both high as well as low frequencies. We also constructed RD to extract

normative distributions when characterizing the awakening process, using wakefulness control

segments from between 4-8PM of the evening before the sleep recording and a sleep segment

from -15 to -10 min before the awakening (Figure 2A i). To enhance the reliability of these

baseline segments, we bootstrapped the segments using a block design. To do so, we sampled

10,000 times with replacement of raw signals in 5-s blocks from the 10 min of wakefulness or

the 5 min of sleep (Figure 2Ai). To assess the potential impact of circadian rhythms on the

wakefulness baseline, we included an additional baseline curated by an independent

epileptologist (G.R.) not involved in the original data selection. This baseline, selected from

quiet wakefulness periods between 8AM-12PM on the day of the sleep SEEG recording, was

used to compare IA times in relation to wakefulness. We investigated the spectral and

connectivity properties of the awakening process during the selected 30-min period of the

awakening process (Figure 2A ii). The analysis was conducted for the following high frequency

bands: low gamma (30-50 Hz), high gamma (50-80 Hz), low ripple (80-140 Hz), and high ripple

(140-200 Hz). This distribution was chosen to ensure equal representation in the gamma and

ripple bands. Furthermore, we also analyzed traditional frequency bands for comparison

131

purposes: delta (0.5-4 Hz), theta (4-8 Hz), alpha (8-13 Hz), and beta (13-30 Hz). Artifacts were

visually rejected as well as rejected using automatic amplitude base artifact (>5 standard

deviations). The spectral analysis was executed using a Morlet wavelet transform on each 5-s

segment without overlap. We opted for a Morlet transform featuring a 1 Hz central frequency

and a duration of 3 s to strike a balanced trade-off between temporal and frequency resolution.

(Cole and Voytek, 2017) The outputs from the Morlet wavelet transform were treated

separately for each band. This spectral analysis was carried out for every available normal

channel for each patient. For visualization, only the power was normalized between 0 and 1

across the brain, facilitating improved visualization of both temporal and regional differences.

Connectivity analysis followed, where PLV was estimated for each pair of available normal

channels within each patient. (Lachaux et al., 1999) This was done for each pair within each

available network and for each pair that were paced in two different networks. This was done

to assess the connection strength within every network as well as the strength of connection

between different networks. The PLV was used in order to capture the phase connectivity

which is insensitive to the amplitude due to the variability of amplitudes which can occur in

SEEG. (Gloor, 1985) The PLV was computed in each frequency band separately as 



󰇻󰇛󰇛󰇜󰇛󰇜



 󰇻 where  are channels, f is the frequency band of interest, T is the

signal length and  is the instantaneous phase estimated as the angle of the analytical signal

from the Hilbert transforms of both signals. Subsequently, the power or PLV signals estimated

every time point of interest during awakening process were compared to both the sRD and the

wRD in terms of percentiles relative to that RD (Figure 2A iii). This process generated three 30-

min time series: (1) the original power or PLV signals, (2) the signal represented as a percentile

when compared to the sRD, and (3) the signal represented as a percentile compared to the

wRD. These time series were then pooled for all patients, for each channel or channel pair

(Figure 2A iv). We grouped these channels or channel pairs using a weighted median over all

the channels in the region/network. The weights, which add up to one, are defined for

each power/PLV as:  

 where w = weight of the specific channel (or channel pair for

PLV),  = number of patients contributing to a region/network, and  = number of channels

contributing to the region/network for a specific patient. The power/PLV values and the

132

weights were sorted separately in an ascending order and then a cumulative sum of the

weights was calculated until it reached 0.5. The median weighed power/PLV was defined as the

matching power/PLV value of the corresponding index at 0.5. This approach was adopted to

account for the relative contribution of each patient. Finally, these time series were compared

throughout the awakening process, considering both the original power and PLV as well as a

categorized normalized time series in relation to the sRD and wRD (Figure 2A v). The latter was

achieved by categorizing the percentile time series into three bins with respect to the

corresponding RD: (1) significantly larger, percentile >95% (red), (2) within 5-95% (green), and

(3) significantly smaller, percentile <5% (blue). We provide a continuous description of the

awakening process from different perspectives (Figure 2B). First, the original power/PLV values

are presented every 5 min for all available regions or networks, along with an example of one

region/network (Figure 2B i). Second, we present a description every 5 min of the categorized

percentiles compared to the sRD (Figure 2B ii) and the wRD (Figure 2B iii) for all available

regions or networks, as well as a continuous example from a selected region/network. We then

estimated two times for IA; the first one being the divergence from sleep (when compared to

sRD) and the second one being the convergence to wakefulness (when compared to wRD).

These IA times reflect the moment the signal either diverges from the sRD or converges to the

wRD. The IA times were identified as the initial time instance after more than three continuous

epochs (90 s) during which the percentile was either falling outside (for sRD) or inside (for wRD)

the 5-95% boundaries of the corresponding RD. We choose a strict 5-95 percentile thresholds in

order to ensure that only stable non-spurious changes are considered. Therefore, the IA times

represent the first stable instance of convergence to wakefulness or divergence from sleep.

Note that the binarized data was only used for the purpose of determining the IA time. The

subsequent analysis for the statistical assessment of the differences, effect sizes, and

magnitude of change was conducted using the original power and PLV. All the analysis was

done using MATLAB 2023a, and signal processing was done using the Brainstorm toolbox.

(Tadel et al., 2011)

133

Statistical analysis

We analyzed six patients awakening from NREM sleep and twelve patients awakening from

REM sleep. We evaluated each time point in the awakening process in relation to the sRD and

wRD, by performing a continuous statistical comparison (Figure 2C). We analyzed 90 s segments

every 5 min before the wakefulness IA and after the sleep IA. This analysis was conducted on

the channel level for all the available channels in the region or network. We compared between

the median power/PLV of the awakening signal to the median RD using a two-side paired

Wilcoxon signed-rank test, with a false discovery rate (FDR) correction. The effect size was

estimated using paired Cliff's Delta. No further transformations to the data were applied, and

no datapoints were removed. This was done once for all the time points prior to the

wakefulness IA and then for all the time points after the sleep IA. Data is presented as median

and range as distributions were not always normal. If the difference achieved significance at

p<0.05 we then assessed the magnitude of the change by defining the "relative deviation". This

was calculated as the weighted median of the percentage difference on all channels between

the awakening signal and the median RD. In other words, for each channel, we assessed the

extent to which it differed from the RD and then aggregated it using a weighted median. We

provide comprehensive tables detailing the IA times for the wakefulness IA and the IA, for all

significant results i.e. when a band or an area are not present it means no significant results

were observed (full results are provided in Supp. Tables S2-5). For each one we provide the

ranges of Cliff’s Delta and the corresponding relative deviations from both the sRD and the wRD

for the spectral and connectivity analyses for each available region or network. Note that NREM

and REM sleep are not directly compared as we only selected one type of awakening per

patient. We also compared the awakening process observed on the scalp to each available

region. For this, we selected a single scalp channel (either Cz or Fz) based on visual assessment

to ensure minimal muscle-related artifacts. This channel's data was then compared to the

median power of each region at each time point using Pearson correlation. We then report the

median coefficient and the FDR corrected p values (Supp. Tables S7-8).

134

Author Contributions

T.A., and B.F. conceptualized the project and experimental design. T.A. performed the analysis.

G.R, C.A, F.D and B.F. contributed to the curation of data. T.A, B.F. and C.G. participated in the

discussion of the methodology, mathematical formulation, and results of the study.

Acknowledgments

We wish to express our gratitude to Nicolas von Ellenrieder for his support, guidance, and insights, to

Alyssa Ho for her useful edits, and to the staff and technicians at the EEG Department of the Montreal

Neurological Institute and Hospital, particularly Erica Minato, Lorraine Allard, Nicole Drouin and Chantal

Lessard.

Data Availability Statement

Processed data will be available upon request, however the raw SEEG data are not accessible due to

data privacy regulations.

Ethical Statement

The study was approved by the Montreal Neurological Institute and Hospital Review Ethics Board (2014-

183). As this research was retrospective, informed consent was waived by the MNI Director of

Professional Services.

Funding: This work was funded by project grants from the Canadian Institutes of Health Research (PJT-

175056, B.F.; PJT 159948, C.G).

Conflicts of interest: None of the authors has any conﬂict of interest to disclose. Outside of this work, BF

received speaker / advisory board honoraria from Eisai, Natus, Paladin labs, UCB, and UNEEG.

135

Region

NREM

Relative deviation from wRD

Relative deviation sRD

Inferior parietal lobule

delta

-30s (↑ 77-81%; d=.77-86; p<0.001)

delta

45s (↓95-270%; d=.59-.80;p<0.001)

theta

350s (↓ 77-83Y%; d=.23-.33; p<0.01)

low gamma

50s (↑ 24-33%; d=.19-.31;p<0.01)

theta

-40s (↑ 55-61%; d=.47-.49; p<0.001)

low ripple

200s (↑ 35-47%; d=.25-.35; p<0.01)

high ripple

150s (↑ 23-44%; d=.26-.66; p<0.001)

Central operculum and

opercular part of

inferior frontal gyrus

delta

-65s (↓ 30-63%; d=.35-.45; p<0.001)

delta

215s (↑ 124-165%; d=.46-.53; p<0.01)

high ripple

-40s (↓ 55-60%; d=.05-.08;p<0.01)

low ripple

230s (↑ 13-22%; d=.03-.26; p<0.001)

high ripple

235s (↑ 14-34%; d=.12-.16; p<0.01)

Superior, middle, and

orbital frontal gyri and

anterior part of inferior

frontal gyrus

delta

-30s (↓ 40-46%; d=.39-.48; p<0.001)

delta

45s (↓ 118-165%; d=.38-.55; p<0.001)

Beta

30s (↑ 11-25%; d=0.03-0.13; p<0.01)

beta

-45s (↓ 39-59%; d=.16-24; p<0.001)

low gamma

65s (↑ 9-13%; d=.04-.09;p<0.01)

high gamma

230s (↑ 9-19%; d=.08-.09; p<0.001)

high ripple

-40s (↓ 62-69%; d=.38-.52; p<0.001)

low ripple

190s (↑ 15-39%; d=.14-.23; p<0.01)

high ripple

50s (↑ 17-30%; d=.38-.49; p<0.001)

Insula

high ripple

-40s (↓ 80-90%; d=.11-.13;p<0.01)

delta

40s (↓ 77-105%; d=.16-.24; p<0.01)

high ripple

230s (↑ 27-33%; d=.02-.13; p<0.01)

Superior parietal lobule

high ripple

-215s (↑ 21-51%; d=.34-.74; p<0.001)

delta

45s (↓ 56-161%; d=.53-.55; p<0.001)

high ripples

50s (↑ 23-44%; d=.09-.66; p<0.001)

Middle and inferior

temporal gyrus,

temporal pole, and

planum polare

high ripple

-235s (↓ 21-24%; d=.24-.48; p<0.001)

delta

20s (↓ 44-69%; d=.42-.53; p<0.001)

low gamma

50s (↑ 6-14%; d=.14-.18;p<0.001)

low ripple

50s (↑ 18-28%; d=.11-.26;p<0.01)

high ripple

50s (↑ 19-24%; d=.06-.23;p<0.01)

Table 1. Spectral density results of awakening from NREM sleep. Reported here are spectral results for all bands which

displayed a significant difference between the awakening process to the wakefulness or sleep reference distributions (wRD and

sRD respectively). The results are presented as the time in seconds from the intracranial awakening (IA) when compared to the

RDs. Times are given for the convergence to the wRD and the divergence from the sRD. The difference between the awakening

process and the RD prior to convergence or after the divergence was tested on the channel level for all channels in the region

using a paired Wilcoxon test and assessed with Cliff’s d. The magnitude of the difference between the awakening process to the

RDs is reported as relative deviation, which is the change in percentage compared to the RDs. The direction of change is

represented as ↑↓ if the time after the awakening corresponded to an increase ↑ or decrease ↓ when compared to the

reference distribution. Note: only significant bands and regions are reported (p<0.05 after correction for performing of multiple

comparisons).

136

Network

NREM

Relative deviation from wRD

Relative deviation from sRD

Somatomotor

high ripple

-20s (↓ 119-124%; d=.83-.85; p<0.001)

high ripple

230s (↑ 20-29%; d=.15-.78; p<0.001)

Somatomotor -Ventral

attention

high ripple

-40s (↓ 153-188%; d=.65-.68; p<0.001)

high ripple

255s (↑ 20-28%; d=.25-.82; p<0.001)

Somatomotor -

Frontoparietal

high ripple

370s (↑ 20-44%; d=.20-.66; p<0.001)

Somatomotor - Default

high ripple

-20s (↓ 229-288%; d=.70-.72; p<0.001)

high ripple

390s (↑ 16-35%; d=.35-.76; p<0.001)

Dorsal attention

high ripple

-345s (↓ 29-56%; d=.79-.83; p<0.001)

low ripple

50s (↑ 18-32%; d=.13-.32; p<0.001)

high ripple

95s (↑ 48-61%; d=.34-.61; p<0.001)

Dorsal attention -

Frontoparietal

beta

110s (↓ 9-11%; d=.17-.30; p<0.001)

low ripple

80s (↑ 25-29%; d=.41-.46; p<0.001)

high ripple

45s (↑ 49-51%; d=.29-.56; p<0.001)

Dorsal attention -

Default

beta

-95s (↓ 8-27%; d=.13-.33; p<0.001)

low ripple

50s (↑ 18-23%; d=.26-.39; p<0.01)

Default

low ripple

220s (↑ 12-17%; d=.15-.17; p<0.001)

high ripple

200s (↑ 50-59%; d=.23-.35; p<0.001)

Default - Ventral

attention

high ripple

-40s (↓ 125-147%; d=.10-.22; p<0.01)

low ripple

350s (↑ 12-24%; d=.20-.22; p<0.01)

high ripple

240s (↑ 18-22%; d=.11-.16; p<0.001)

Default -

Frontoparietal

beta

55s (↓ 3-20%; d=.07-.13; p<0.001)

high ripple

380s (↑ 13-18%; d=.53-.54; p<0.001)

Frontoparietal

high ripple

45s (↑ 34-40%; d=.77-.79; p<0.001)

Table 2. Phase connectivity results of awakening from NREM. Reported here are phase locking value (PLV) results for all bands

which displayed a significant difference between the awakening process to the wakefulness or sleep reference distributions

(wRD and sRD respectively). The results are presented as the time in seconds from the intracranial awakening (IA) when

compared to the RDs. Times are given for the convergence to the wRD and the divergence from the sRD. The difference

between the awakening process and the RD prior to convergence or after the divergence was tested on the channel level for all

channels-pairs within a network or between two different networks, using a paired Wilcoxon test and assessed with Cliff’s d.

The magnitude of the difference between the awakening process to the RDs is reported as relative deviation, which is the

change in percentage compared to the RDs. The direction of change is represented as ↑↓ if the time after the awakening

corresponded to an increase ↑ or decrease ↓ when compared to the reference distribution. ND- never diverged from the RD

in any frequency band. Note: only significant bands and regions are reported (p<0.05 after correction for performing of multiple

comparisons).

137

Region

REM

wRD

sRD

Anterior and

middle cingulate

gyrus

low ripple

-45s (↓ 41-47%; d=.31-.33; p<0.01)

low ripple

260s (↑ 5-21%; d=.21-.44; p<.01)

high ripple

-40s (↓ 36-41%; d=.38-.42; p<0.01)

high ripple

45s (↑ 5-22%; d=.22-.29; p<0.01)

Central

operculum and

opercular part

of inferior

frontal gyrus

high ripple

-35s (↓ 59-62%; d=.16-.22; p<0.01)

Inferior parietal

lobule

low ripple

45s (↓ 42-47%; d=.11-.27; p<0.001)

delta

-350s (↓ 13-87%; d=.11-.73; p<0.001)

high ripple

NC (↓ 40-65%; d=.02-.28; p<0.05)

low ripple

165s (↑ 12-15%; d=.03-.15; p<0.01)

Insula

high ripple

305s (↑ 5-17%; d=.17-.25; p<0.05)

Medial and

basal temporal

region

high gamma

NC (↓ 52-92%; d=.31-.34; p<0.001)

delta

-480s (↓ 13-72%; d=.30-.42; p<0.001)

low ripple

NC (↓ 41-85%; d=.15-.52; p<0.001)

low ripple

55s (↑ 13-19%; d=.30-.31; p<0.001)

high ripple

NC (↓ 49-88%; d=.35-.57; p<0.01)

Medial frontal

cortex

low ripple

45s (↓ 37-40%; d=.18-.19; p<0.001)

low ripple

35s (↑ 6-25%; d=.18-.30; p<0.01)

high ripple

-40s (↓ 36-43%; d=.18-.19; p<0.001)

high ripple

45s (↑ 12-26%; d=.09-.12; p<0.001)

Middle and

inferior

temporal gyrus,

temporal pole,

and planum

polare

high gamma

-5s (↓ 46-50%; d=.22-.23; p<0.001)

delta

-405s (↓ 19-58%; d=.23-.43; p<0.001)

beta

25s (↑ 15-19%; d=.15-.20; p<0.001)

low ripple

10s (↓ 81-86%; d=.47-.48; p<0.001)

high gamma

75s (↑ 9-13%; d=.04-.15; p<.001)

high ripple

-10s (↓ 54-77%; d=.35-.36; p<0.001)

low ripple

-50s (↑ 18-24%; d=.20-.44; p<0.001)

high ripple

-25s (↑ 22-27%; d=.24-.34; p<0.001)

Superior,

middle, and

orbital frontal

gyri and anterior

part of inferior

frontal gyrus

low ripple

0s (↓ 39-44%; d=.16-.19; p<0.001)

delta

-480s (↓ 37-87%; d=.16-.57; p<0.01)

high ripple

0s (↓ 39-42%; d=.12-.14; p<0.001)

low ripple

55s (↑ 10-29%; d=.11-.36; p<0.001)

high ripple

50s (↑ 14-31%; d=.08-.20; p<0.001)

Superior

temporal gyrus

low ripple

-30s (↓ 43-47%; d=.25-.27; p<0.001)

alpha

50s (↑ 23-40%; d=.14-.42; p<0.001)

beta

-40s (↑ 11-24%; d=.11-.24; p<0.001)

high ripple

low ripple

55s (↑ 19-36%; d=.21-.44; p<0.01)

high ripple

-20s (↑ 12-39%; d=.09-.27; p<0.001)

Table 3. Spectral density results of awakening from REM sleep. Reported here are spectral results for all bands which

displayed a significant difference between the awakening process to the wakefulness or sleep reference distributions (wRD and

sRD respectively). The results are presented as the time in seconds from the intracranial awakening (IA) when compared to the

RDs. Times are given for the convergence to the wRD and the divergence from the sRD. The difference between the awakening

process and the RD prior to convergence or after the divergence was tested on the channel level for all channels in the region

using a paired Wilcoxon test and assessed with Cliff’s d. The magnitude of the difference between the awakening process to the

RDs is reported as relative deviation, which is the change in percentage compared to the RDs. The direction of change is

represented as ↑↓ if the time after the awakening corresponded to an increase ↑ or decrease ↓ when compared to the

reference distribution. ND- never diverged from the RD in any frequency band, NC- never converged back the RD. Note that

while the medial and basal temporal region never significantly converged to the wRD they did show an upward trend. Note:

only significant bands and regions are reported (p<0.05 after correction for performing of multiple comparisons).

138

Network

REM

wRD

sRD

Limbic

low ripples

-40s (↓ 36-40%; d=.30-.38; p<0.001)

low ripples

50s (↑ 28-39%; d=.30-.35; p<0.001)

Limbic - Ventral attention

low ripples

50s (↑ 21-24%; d=.34-.44; p<0.001)

Limbic - Visual

high ripples

NC (↑ 21-24%; d=.21-.35; p<0.001)

high ripples

230s (↓ 6-7%; d=.20-.23; p<0.01)

Limbic - Somatomotor

high ripples

-455s (↑ 15-16%; d=.51-54; p<0.001)

low ripples

45s (↑ 22-23%; d=.31-.41; p<0.001)

Somatomotor - Ventral attention

high ripples

-75s (↑ 14-42%; d=.62-.65; p<0.001)

low ripples

45s (↑ 21-22%; d=.14-.38; p<0.01)

Limbic - Visual

high ripples

NC (↑ 21-41%; d=.21-.55; p<0.001)

low ripples

95s (↑ 25-26%; d=41-47; p<0.001)

Limbic - Default

low ripples

45s (↑ 20-24%; d=.35-.41; p<0.001)

high ripples

390s (↑ 14-19%; d=.02-.29; p<0.001)

Limbic - Frontoparietal

low ripples

45s (↑ 28-39%; d=.30-.38; p<0.001)

Default

low ripples

45s (↑ 18-34%; d=.27-.51; p<0.001)

high ripples

440s (↑ 11-14%; d=.05-.21; p<0.001)

Default- Dorsal attention

low ripples

50s (↑ 9-16%; d=.37-.67; p<0.01)

Default - Ventral attention

low ripples

50s (↑ 14-19%; d=.27-.41; p<0.01)

Default - Somatomotor

low ripples

50s (↑ 18-28%; d=.30-.45; p<0.001)

high ripples

80s (↓ 8-17%; d=.09-.21; p<0.001)

Default - Visual

low ripples

430s (↑ 19-31%; d=.51-.78; p<0.01)

Default - Frontoparietal

low ripples

275s (↑ 13-23%; d=.27-.39; p<0.001)

Dorsal attention

low ripples

280s (↑ 22-25%; d=.38-.60; p<0.001)

Dorsal attention - Limbic

low ripples

80s (↑ 31-56%; d=.35-37; p<0.001)

Dorsal attention - Ventral attention

low ripples

345s (↑ 10-23%; d=.23-53; p<0.001)

Dorsal attention - Frontoparietal

low ripples

45s (↑ 14-49%; d=.27-51; p<0.001)

Visual

high ripples

-70s (↓ 20-40%; d=.59-.61; p<0.001)

Visual - Somatomotor

high ripples

NC (↑ 75-81%; d=.92-.95; p<0.001)

low ripples

225s (↑ 24-39%; d=.31-.51; p<0.01)

Somatomotor

low ripples

50s (↑ 14-21%; d=.20-.47; p<0.001)

Somatomotor- Dorsal attention

low ripples

275s (↑ 23-41%; d=.44-.77; p<0.001)

Somatomotor - Frontoparietal

low ripples

50s (↑ 18-35%; d=.26-.49; p<0.001)

139

high ripples

485s (↓ 2-14%; d=.07-.13; p<0.01)

Ventral attention - Frontoparietal

low ripples

245s (↑ 4-15%; d=.15-.33; p<0.001)

Frontoparietal

low ripples

380s (↑ 9-27%; d=.27-.39; p<0.001)

Table 4. Phase connectivity results of awakening from REM. Reported here are phase locking value (PLV) results for all bands

which displayed a significant difference between the awakening process to the wakefulness or sleep reference distributions

(wRD and sRD respectively). The results are presented as the time in seconds from the intracranial awakening (IA) when

compared to the RDs. Times are given for the convergence to the wRD and the divergence from the sRD. The difference

between the awakening process and the RD prior to convergence or after the divergence was tested, on the channel level for all

channels-pairs within a network or between two different networks, using a paired Wilcoxon test and assessed with Cliff’s d.

The magnitude of the difference between the awakening process to the RDs is reported as relative deviation, which is the

change in percentage compared to the RDs. The direction of change is represented as ↑↓ if the time after the awakening

corresponded to an increase ↑ or decrease ↓ when compared to the reference distribution. ND- never diverged from the RD

in any frequency band. NC- never converged back to the RD. Note: only significant bands and regions are reported (p<0.05 after

correction for performing of multiple comparisons).

140

4.7 Supporting Information

Chart. S1

Chart. S1 Flowchart of the patient selection process.

141

Table S1

MICCAI 17 Regions

REM

NREM

Medial occipital lobe

14 (2)

Lateral occipital lobe

4 (1)

Medial and basal temporal region

32 (7)

2 (1)

Medial parietal lobe

20 (5)

10 (2)

Superior parietal lobule

22 (5)

17 (3)

Superior temporal gyrus

24 (5)

9 (1)

Supplementary motor cortex

1 (1)

Medial frontal cortex (including medial segment of superior frontal gyrus)

16 (4)

3 (1)

Inferior parietal lobule

58 (7)

19 (4)

Central operculum and opercular part of inferior frontal gyrus

21 (3)

15 (4)

Superior, middle, and orbital frontal gyri and anterior part of inferior frontal gyrus

118 (7)

32 (6)

Insula

18 (3)

17 (4)

Middle and inferior temporal gyrus, temporal pole, and planum polare

127 (9)

18 (3)

Anterior and middle cingulate gyrus

20 (4)

12(2)

Frontal operculum

9 (4)

8 (2)

Medial and basal temporal region

32 (7)

Transverse temporal gyrus and planum temporale

8 (4)

5 (1)

Pre- and postcentral gyri

11 (2)

14 (2)

142

Figure.1

Figure S1. MICCAI17 and YEO atlases illustration. * MICCAI regions (medial occipital lobe: blue,

lateral occipital lobe: royal blue, medial and basal temporal region: light blue, transverse

temporal gyrus and planum temporale: cyan, pre- and postcentral gyri: dark green, medial

parietal lobe: green, superior parietal lobule: light green, superior temporal gyrus: lime,

supplementary motor cortex: indigo, medial frontal cortex (including medial segment of

superior frontal gyrus): lavender, inferior parietal lobule: yellow, central operculum and

opercular part of inferior frontal gyrus: orange, superior, middle and orbital frontal gyri and

anterior part of inferior frontal gyrus: dark orange, insula: pink, middle and inferior temporal

gyrus: dark red, temporal pole, and planum polare, anterior and middle cingulate gyrus: middle

light green, frontal operculum: light blue, medial and basal temporal region: red, transverse

temporal gyrus and planum temporale: crimson) **Yeo 7 networks (visual: blue, somatomotor:

light blue, dorsal attention: light green, ventral attention: lime, limbic: orange, frontoparietal:

red, default mode: red).

143

Figure S2

Figure S2. Available coverage in the regional and network atlases. (A) Regional coverage in the

MICCAI 17 anatomical regions atlas by the number of patients available in each region (B)

Regional coverage in networks in the Yeo 7 atlas by the number of patients available in each

network.

144

Figure S3

Figure S3. Spectral power in the delta band increases upon awakening from NREM sleep. (A)

Power from every available region across the brain is shown from 15 min before to 15 min after

the scalp awakening (SA), along with a regional example of the power with a quartile range as

the shaded area, and the respected intracranial awakening times for sleep reference

distribution (sRD) and wakefulness RD (wRD) (B) The power's percentile, in relation to the sRD,

is presented for every brain region, complemented by the same regional example. (C) The

power's percentile, in relation to the wRD, is detailed for each brain region, with the same

regional example as a bar. Note: Power represents the weighted median power in each region,

normalized across the entire brain. The reference distribution percentiles are based on the

weighted median in each region, which are then categorized as over 95%, within 5-95%, or

below 5%. Continuous power examples and comparisons to wakefulness and sleep reference

distributions are illustrated for the superior, middle, and inferior frontal regions. Deviations

below the 5% or above the 95% percentiles are deemed significant. Included regions are:

superior, middle, and orbital frontal gyri and anterior part of inferior frontal gyrus, the insula,

the central operculum, opercular part of inferior frontal gyrus, the superior parietal lobule, the

middle and inferior temporal gyrus, temporal pole, planum polare, and the Inferior parietal

lobule. Data presented for 6 patients from a total of 118 channels, ranging between 15-32 per

region (Full description in Table S1 A).

145

Table S2

Region

NREM

Relative deviation from wRD

Relative deviation sRD

Inferior parietal lobule

delta

-30s (↑ 77-81%; d=.77-86; p<.001)

delta

45s (↓95-270%; d=.59-.80;p<.001)

theta

-40s (↑ 55-61%; d=.47-.49; p<.001)

theta

350s (↓ 77-83Y%; d=.23-.33; p<.01)

alpha

(d=.11; p=.7)

alpha

(d=.03-.11;p=.14-.56)

beta

(d=.14; p=.09)

beta

(d=.04; p=.44)

low

gamma

(d=.2; p=.09)

low gamma

50s (↑ 24-33%; d=.19-.31;p<.01)

high

gamma

(d=.13; p=.25)

high gamma

(d=.1; p=.13)

low ripple

(d=.22; p=.04)

low ripple

200s (↑ 35-47%; d=.25-.35; p<.01)

high ripple

(d=.17; p=.25)

high ripple

150s (↑ 23-44%; d=.26-.66; p<.001)

Central operculum and

opercular part of

inferior frontal gyrus

delta

-65s (↓ 30-63%; d=.35-.45; p<.001)

delta

215s (↑ 124-165%; d=.46-.53; p<.01)

theta

(d=.16; p=.35)

theta

(d=.14; p=.35)

alpha

ND (d=.02-.09;p=.22-1)

alpha

ND (d=.05-.15;p=.33-.68)

beta

-45s (↓ 39-59%; d=.16-24; p<.001)

beta

(d=;p=)

low

gamma

(d=.09; p=.27)

low gamma

(d=.04; p=.84)

high

gamma

(d=.08; p=.5)

high gamma

(d=.19; p=.47)

low ripple

(d=.15; p=.02)

low ripple

230s (↑ 13-22%; d=.03-.26; p<.001)

high ripple

-40s (↓ 55-60%; d=.05-.08; p<0.01)

high ripple

235s (↑ 14-34%; d=.12-.16; p<.01)

Superior, middle, and

orbital frontal gyri and

anterior part of inferior

frontal gyrus

delta

-30s (↓ 40-46%; d=.39-.48; p<.001)

delta

45s (↓ 118-165%; d=.38-.55; p<.001)

Theta

(d=.13; p=.77)

theta

(d=.07; p=.52)

alpha

ND (d=.01-.16;p=.44-1)

alpha

ND (d=.01-.25;p=.23-.85)

beta

-45s (↓ 39-59%; d=.16-24; p<.001)

beta

30s (↑ 11-25%; d=.03-.13; p<.01)

low

gamma

(d=.08; p=.13)

low gamma

65s (↑ 9-13%; d=.04-.09;p<.01)

high

gamma

(d=.08; p=.10)

high gamma

230s (↑ 9-19%; d=.08-.09; p<.001)

low ripple

(d=.17; p=.05)

low ripple

190s (↑ 15-39%; d=.14-.23; p<.01)

high ripple

-40s (↓ 62-69%; d=.38-.52; p<.001)

high ripple

50s (↑ 17-30%; d=.38-.49; p<.001)

Insula

delta

ND (d=.02-.24;p=.1-.45)

delta

40s (↓ 77-105%; d=.16-.24; p<.01)

theta

(d=.13; p=.32)

theta

(d=.14; p=.32)

alpha

ND (d=.03-.10;p=.32-/69)

alpha

ND (d=.02-.14;p=.24-.92)

beta

(d=.04; p=.25)

beta

(d=.15; p=.05)

low

gamma

(d=.09; p=.16)

low gamma

(d=.04; p=.6)

high

gamma

(d=.03; p=.35)

high gamma

(d=.05; p=.3)

low ripple

(d=.18; p=.08)

low ripple

(d=.12; p=.23)

high ripple

-40, -35s (↓ 80-93%; d=.11-.17;p<.01)

high ripple

230s (↑ 27-33%; d=.02-.13; p<.01)

Superior parietal lobule

delta

(d=.55; p=.01)

delta

45s (↓ 56-161%; d=.53-.55; p<.001)

theta

(d=.09; p=.05)

theta

(d=.14; p=.03)

alpha

ND (d=.03-.10;p=.65-1)

alpha

ND (d=.05-.17; p=.53-/79)

beta

(d=.33; p=.02)

beta

(d=.26; p=.05)

low

gamma

(d=.17; p=.01)

low gamma

(d=.19; p=.11)

high

gamma

(d=.03; p=.5)

high gamma

(d=.09; p=.18)

low ripple

(d=.17; p=.05)

low ripple

(d=.13; p=.07)

high ripple

-215s (↑ 21-51%; d=.34-.74; p<.001)

high ripple

50s (↑ 23-44%; d=.09-.66; p<.001)

delta

ND (d=.05-.34;p=.13-.68)

delta

20s (↓ 44-69%; d=.42-.53; p<.001)

theta

(d=.07; p=.35)

theta

(d=.15; p=.21)

146

Middle and inferior

temporal gyrus,

temporal pole, and

planum polare

alpha

ND (d=.03-.14;p=.36-.87)

alpha

ND (d=.05-.12;p=.12-1)

beta

(d=.27; p=.08)

beta

(d=.24; p=.10)

low

gamma

(d=.16; p=.10)

low gamma

50s (↑ 6-14%; d=.14-.18;p<.001

high

gamma

(d=.05; p=.47)

high gamma

(d=.13; p=.25)

low ripple

(d=.16; p=.06)

low ripple

50s (↑ 18-28%; d=.11-.26;p<.01)

high ripple

-235s (↓ 21-24%; d=.24-.48; p<.001)

high ripple

50s (↑ 19-24%; d=.06-.23;p<.01)

Frontal operculum

Insufficient coverage

Medial and basal

temporal region

Insufficient coverage

Medial parietal lobe

Insufficient coverage

Superior temporal gyrus

Insufficient coverage

Supplementary motor

cortex

Insufficient coverage

Medial and basal

temporal region

Insufficient coverage

Medial parietal lobe

Insufficient coverage

Superior temporal gyrus

Insufficient coverage

Supplementary motor

cortex

Insufficient coverage

Medial frontal cortex

(including medial

segment of superior

frontal gyrus)

Insufficient coverage

Anterior and middle

cingulate gyrus

Insufficient coverage

Medial and basal

temporal region

Insufficient coverage

Transverse temporal

gyrus and planum

temporale

Insufficient coverage

Pre- and postcentral

gyri

Insufficient coverage

Table S2. Spectral density results of awakening from NREM sleep. Reported here are spectral results for all bands which

displayed a significant difference between the awakening process to the wakefulness or sleep reference distributions (wRD and

sRD respectively). The results are presented as the time in seconds from the intracranial awakening (IA) when compared to the

RDs. Times are given for the convergence to the wRD and the divergence from the sRD. The difference between the awakening

process and the RD prior to convergence or after the divergence was tested on the channel level for all channels in the region

using a paired Wilcoxon test and assessed with Cliff’s d. The magnitude of the difference between the awakening process to the

RDs is reported as relative deviation, which is the change in percentage compared to the RDs. The direction of change is

represented as ↑↓ if the time after the awakening corresponded to an increase ↑ or decrease ↓ when compared to the

reference distribution. Note: non significant results are reported with the median effect size and p value throughout the

duration. Two wakefulness baselines were utilized: one from the prior evening and one from the prior morning. Any differences

between these baselines, presented in this order, are detailed in the Table.

147

Table S3

Network

NREM

Relative deviation from wRD

Relative deviation from sRD

Somatomotor

delta

(d=.28; p=.12)

delta

(d=.31; p=.1)

theta

(d=.07; p=.22)

theta

(d=.04; p=.25)

alpha

(d=.02; p=.62)

alpha

(d=.02; p=.68)

beta

(d=.23; p=.1)

beta

(d=.22; p=.12)

low gamma

(d=.12; p=.15)

low gamma

(d=.08; p=.05)

high gamma

(d=.03; p=.43)

high gamma

(d=.51; p=.08)

low ripple

(d=.29; p=.23)

low ripple

(d=.16; p=.05)

high ripple

-20s (↓ 119-124%; d=.83-.85; p<.001)

high ripple

230s (↑ 20-29%; d=.15-.78; p<.001)

Somatomotor -Ventral

attention

delta

(d=.05; p=.46)

delta

(d=.33; p=.08)

theta

(d=.07; p=.22)

theta

(d=.05; p=.53)

alpha

(d=.53; p=.15)

alpha

(d=.03; p=.72)

beta

(d=.05; p=.44)

beta

(d=.12; p=.02)

low gamma

(d=.04; p=.62)

low gamma

(d=.06; p=.29)

high gamma

(d=.19; p=.32)

high gamma

(d=.62; p=.13)

low ripple

(d=.05; p=.12)

low ripple

(d=.12; p=.06)

high ripple

-40s (↓ 153-188%; d=.65-.68; p<.001)

high ripple

255s (↑ 20-28%; d=.25-.82; p<.001)

Somatomotor -

Frontoparietal

delta

(d=.33; p=.06)

delta

(d=.51; p=.1)

theta

(d=.19; p=.07)

theta

(d=.1; p=.41)

alpha

(d=.11; p=.33)

alpha

(d=.1; p=.42)

beta

(d=.46; p=.1)

beta

(d=.25; p=.01)

low gamma

(d=.06; p=.72)

low gamma

(d=.08; p=.59)

high gamma

(d=.11; p=.34)

high gamma

(d=.16; p=.09)

low ripple

(d=.23; p=.05)

low ripple

(d=.14; p=.17)

high ripple

(d=.09; p=.48)

high ripple

370s (↑ 20-44%; d=.20-.66; p<.001)

Somatomotor - Default

delta

(d=.32; p=.15)

delta

(d=.09; p=.48)

theta

(d=.12; p=.39)

theta

(d=.43; p=.09)

alpha

(d=.12; p=.44)

alpha

(d=.02; p=.94)

beta

(d=.71; p=.1)

beta

(d=.16; p=.25)

low gamma

(d=.07; p=.74)

low gamma

(d=.25; p=.03)

high gamma

(d=.06; p=.76)

high gamma

(d=.06; p=.78)

low ripple

(d=.03; p=.93)

low ripple

(d=.55; p=.1)

high ripple

-20, -25s (↓ 229-288%; d=.59-.72;

p<.001)

high ripple

390s (↑ 16-35%; d=.35-.76; p<.001)

Dorsal attention

delta

(d=.36; p=.17)

delta

(d=.17; p=.12)

theta

(d=.05; p=.69)

theta

(d=.13; p=.08)

alpha

(d=.05; p=.62)

alpha

(d=.08; p=.35)

beta

(d=.22; p=.1)

beta

(d=.19; p=.10)

low gamma

(d=.14; p=.06)

low gamma

(d=.1; p=.19)

high gamma

(d=.03; p=.83)

high gamma

(d=.23; p=.25)

low ripple

(d=.04; p=.17)

low ripple

50s (↑ 18-32%; d=.13-.32; p<.001)

high ripple

-345s (↓ 29-56%; d=.79-.83; p<.001)

high ripple

95s (↑ 48-61%; d=.34-.61; p<.001)

Dorsal attention -

Frontoparietal

delta

(d=.1; p=.4)

delta

(d=.21; p=.12)

theta

(d=.05; p=.74)

theta

(d=.15; p=.15)

alpha

(d=.06; p=.67)

alpha

(d=.07; p=.61)

beta

(d=.13; p=.12)

beta

110s (↓ 9-11%; d=.17-.30; p<.001)

low gamma

(d=.04; p=.83)

low gamma

(d=.02; p=.93)

high gamma

(d=.13; p=.23)

high gamma

(d=.14; p=.17)

low ripple

(d=.21; p=.06)

low ripple

80s (↑ 25-29%; d=.41-.46; p<.001)

high ripple

(d=.27; p=.05)

high ripple

45s (↑ 49-51%; d=.29-.56; p<.001)

Dorsal attention -

Default

delta

(d=.22; p=.14)

delta

(d=.31; p=.02)

theta

(d=.05; p=.86)

theta

(d=.11; p=.62)

alpha

(d=.07; p=.78)

alpha

(d=.05; p=.86)

beta

(d=.53; p=.06)

beta

-95s (↓ 8-27%; d=.13-.33; p<.001)

low gamma

(d=.1; p=.65)

low gamma

(d=.1; p=.63)

high gamma

(d=.06; p=.84)

high gamma

(d=.29; p=.06)

low ripple

(d=.17; p=.31)

low ripple

50s (↑ 18-23%; d=.26-.39; p<.01)

high ripple

(d=.05; p=.86)

high ripple

(d=.39; p=.11)

148

Default

delta

(d=.02; p=.97)

delta

(d=.06; p=.91)

theta

(d=.02; p=.97)

theta

(d=.04; p=.94)

alpha

(d=.04; p=.94)

alpha

(d=.01; p=.99)

beta

(d=.1; p=.83)

beta

(d=.07; p=.88)

low gamma

(d=.12; p=.75)

low gamma

(d=.11; p=.8)

high gamma

(d=.09; p=.84)

high gamma

(d=.04; p=.94)

low ripple

(d=.07; p=.29)

low ripple

220s (↑ 12-17%; d=.15-.17; p<.001)

high ripple

(d=.17; p=.19)

high ripple

200s (↑ 50-59%; d=.23-.35; p<.001)

Default - Ventral

attention

delta

(d=.15; p=.38)

delta

(d=.22; p=.16)

theta

(d=.05; p=.86)

theta

(d=.06; p=.83)

alpha

(d=.08; p=.74)

alpha

(d=.14; p=.47)

beta

(d=.35; p=.10)

beta

(d=.18; p=.27)

low gamma

(d=.04; p=.9)

low gamma

(d=.1; p=.64)

high gamma

(d=.02; p=.95)

high gamma

(d=.3; p=.5

low ripple

(d=.12; p=.55)

low ripple

350s (↑ 12-24%; d=.20-.22; p<.01)

high ripple

-40s (↓ 125-147%; d=.10-.22; p<.01)

high ripple

240s (↑ 18-22%; d=.11-.16; p<.001)

Default - Frontoparietal

delta

(d=.02; p=.93)

delta

(d=.05; p=.81)

theta

(d=.02; p=.94)

theta

(d=.07; p=.74)

alpha

(d=.08; p=.67)

alpha

(d=.03; p=.9)

beta

(d=.09; p=.56)

beta

55s (↓ 3-20%; d=.07-.13; p<.001)

low gamma

(d=.03; p=.9)

low gamma

(d=.07; p=.74)

high gamma

(d=.09; p=.55)

high gamma

(d=.06; p=.84)

low ripple

(d=.13; p=.35)

low ripple

(d=.13; p=.33)

high ripple

(d=.21; p=.06)

high ripple

380s (↑ 13-18%; d=.53-.54; p<.001)

Frontoparietal

delta

(d=.04; p=.78)

delta

(d=.02; p=.91)

theta

(d=.02; p=.91)

theta

(d=.08; p=.35)

alpha

(d=.09; p=.25)

alpha

(d=.07; p=.43)

beta

(d=.04; p=.77)

beta

(d=.14; p=.06)

low gamma

(d=.04; p=.79)

low gamma

(d=.06; p=.59)

high gamma

(d=.14; p=.06)

high gamma

(d=.06; p=.53)

low ripple

(d=.38; p=.06)

low ripple

(d=.27; p=.05)

high ripple

(d=.44; p=.05)

high ripple

45s (↑ 34-40%; d=.77-.79; p<.001)

Ventral Attention -

Frontoparietal

delta

(d=.12; p=.31)

delta

(d=.2; p=.05)

theta

(d=.07; p=.64)

theta

(d=.09; p=.53)

alpha

(d=.03; p=.91)

alpha

(d=.08; p=.61)

beta

(d=.21; p=.03)

beta

(d=.08; p=.55)

low gamma

(d=.03; p=.87)

low gamma

(d=.08; p=.6)

high gamma

(d=.03; p=.89)

high gamma

(d=.02; p=.93)

low ripple

(d=.07; p=.63)

low ripple

(d=.18; p=.07)

high ripple

(d=.04; p=.86)

high ripple

(d=.04; p=.84)

Visual

Insufficient coverage

Visual - Somatomotor

Insufficient coverage

Visual - Dorsal Attention

Insufficient coverage

Visual - Ventral

Attention

Insufficient coverage

Visual - Limbic

Insufficient coverage

Visual - Frontoparietal

Insufficient coverage

Visual - Default mode

Insufficient coverage

Somatomotor - Dorsal

Attention

Insufficient coverage

Somatomotor - Limbic

Insufficient coverage

Dorsal Attention -

Ventral Attention

Insufficient coverage

Dorsal Attention -

Limbic

Insufficient coverage

149

Ventral Attention -

Limbic

Insufficient coverage

Limbic - Frontoparietal

Insufficient coverage

Limbic - Frontoparietal

Insufficient coverage

Limbic - Default mode

Insufficient coverage

Table S3. Phase connectivity results of awakening from NREM. Reported here are phase locking value (PLV) results for all

bands which displayed a significant difference between the awakening process to the wakefulness or sleep reference

distributions (wRD and sRD respectively). The results are presented as the time in seconds from the intracranial awakening (IA)

when compared to the RDs. Times are given for the convergence to the wRD and the divergence from the sRD. The difference

between the awakening process and the RD prior to convergence or after the divergence was tested on the channel level for all

channels-pairs within a network or between two different networks, using a paired Wilcoxon test and assessed with Cliff’s d.

The magnitude of the difference between the awakening process to the RDs is reported as relative deviation, which is the

change in percentage compared to the RDs. The direction of change is represented as ↑↓ if the time after the awakening

corresponded to an increase ↑ or decrease ↓ when compared to the reference distribution. ND- never diverged from the RD

in any frequency band. Note: non significant results are reported with the median effect size and p value throughout the

duration. Two wakefulness baselines were utilized: one from the prior evening and one from the prior morning. Any differences

between these baselines, presented in this order, are detailed in the Table.

150

Table S4

Region

REM

wRD

sRD

Anterior and

middle cingulate

gyrus

delta

(d=.33; p=.06)

delta

(d=.18; p=.05)

theta

(d=.17; p=.08)

theta

(d=.07; p=.56)

alpha

(d=.21; p=.08)

alpha

(d=.09; p=.31)

beta

(d=.09; p=.41)

beta

(d=.41; p=.04)

low gamma

(d=.08; p=.28)

low gamma

(d=.08; p=.68)

high gamma

(d=.07; p=.41)

high

gamma

(d=.17; p=.12)

low ripple

-45, -55s (↓ 41-47%; d=.31-.33; p<.01)

low ripple

260s (↑ 5-21%; d=.21-.44; p<.01)

high ripple

-40s (↓ 36-41%; d=.38-.42; p<.01)

high ripple

45s (↑ 5-22%; d=.22-.29; p<.01)

Central

operculum and

opercular part

of inferior

frontal gyrus

delta

(d=.2; p=.06)

delta

-350s (↓ 13-87%; d=.11-.73; p<.001)

theta

(d=.37; p=.10)

theta

(d=.41; p=.12)

alpha

(d=.48; p=.12)

alpha

(d=.35; p=.06)

beta

(d=.27; p=.10)

beta

(d=.41; p=.1)

low gamma

(d=.24; p=.07)

low gamma

(d=.16; p=.1)

high gamma

(d=.07; p=.14)

high

gamma

(d=.14; p=.08)

low ripple

(d=.06; p=.33)

low ripple

(d=.14; p=.05)

high ripple

-35s (↓ 59-62%; d=.16-.22; p<.01)

high ripple

Inferior parietal

lobule

delta

(d=.17; p=.07)

delta

(d=.06; p=.6)

theta

(d=.09; p=.06)

theta

(d=.06; p=.07)

alpha

(d=.07; p=.11)

alpha

(d=.07; p=.13)

beta

(d=.04; p=.05)

beta

(d=.02; p=.22)

low gamma

(d=.05; p=.1)

low gamma

(d=.04; p=.05)

high gamma

(d=.04; p=.11)

high

gamma

(d=.07; p=.05)

low ripple

45s (↓ 42-47%; d=.11-.27; p<.001)

low ripple

165s (↑ 12-15%; d=.03-.15; p<.01)

high ripple

NC (↓ 40-65%; d=.02-.28; p<.05)

high ripple

Insula

delta

(d=.23; p=.06)

delta

(d=.04; p=.36)

theta

(d=.1; p=.05)

theta

(d=.13; p=.13)

alpha

(d=.15; p=.13)

alpha

(d=.09; p=.24)

beta

(d=.04; p=.33)

beta

(d=.03; p=.45)

low gamma

(d=.08; p=.01)

low gamma

(d=.1; p=.03)

high gamma

(d=.05; p=.15)

high

gamma

(d=.04; p=.18)

low ripple

(d=.03; p=.16)

low ripple

(d=.19; p=.06)

high ripple

(d=.06; p=.07)

high ripple

305s (↑ 5-17%; d=.17-.25; p<.05)

Medial and

basal temporal

region

delta

(d=.36; p=.06)

delta

-480s (↓ 13-72%; d=.30-.42; p<.001)

theta

(d=.14; p=.87)

theta

(d=.1; p=.4)

alpha

(d=.1; p=.56)

alpha

(d=.08; p=.46)

beta

(d=.04; p=.15)

beta

(d=.12; p=.11)

low gamma

(d=.16; p=.09)

low gamma

(d=.04; p=.08)

high gamma

NC (↓ 52-92%; d=.31-.34; p<.001)

high

gamma

(d=.35; p=.06)

low ripple

NC (↓ 41-85%; d=.15-.52; p<.001)

low ripple

55s (↑ 13-19%; d=.30-.31; p<.001)

high ripple

NC (↓ 49-88%; d=.35-.57; p<.01)

high ripple

Medial frontal

cortex

delta

(d=.3; p=.07)

delta

(d=.29; p=.06)

theta

(d=.16; p=.06)

theta

(d=.2; p=.05)

alpha

(d=.27; p=.1)

alpha

(d=.24; p=.11)

beta

(d=.13; p=.12)

beta

(d=.26; p=.08)

low gamma

(d=.07; p=.33)

low gamma

(d=.09; p=.43)

high gamma

(d=.05; p=.18)

high

gamma

(d=.04; p=.33)

low ripple

45s (↓ 37-40%; d=.18-.19; p<.001)

low ripple

35s (↑ 6-25%; d=.18-.30; p<.01)

high ripple

-40, -35s (↓ 36-43%; d=.18-.19; p<.001)

high ripple

45s (↑ 12-26%; d=.09-.12; p<.001)

delta

(d=.23; p=.06)

delta

-405s (↓ 19-58%; d=.23-.43; p<.001)

151

Middle and

inferior

temporal gyrus,

temporal pole,

and planum

polare

theta

(d=.06; p=.09)

theta

(d=.05; p=.15)

alpha

(d=.11; p=.1)

alpha

(d=.16; p=.08)

beta

(d=.13; p=.6)

beta

25s (↑ 15-19%; d=.15-.20; p<.001)

low gamma

(d=.07; p=.05)

low gamma

(d=.2; p=.05)

high gamma

-5s (↓ 46-50%; d=.22-.23; p<.001)

high

gamma

75s (↑ 9-13%; d=.04-.15; p<.001)

low ripple

10s (↓ 81-86%; d=.47-.48; p<.001)

low ripple

-50s (↑ 18-24%; d=.20-.44; p<.001)

high ripple

-10s (↓ 54-77%; d=.35-.36; p<.001)

high ripple

-25s (↑ 22-27%; d=.24-.34; p<.001)

Superior,

middle, and

orbital frontal

gyri and anterior

part of inferior

frontal gyrus

delta

(d=.2; p=.16)

delta

-480s (↓ 37-87%; d=.16-.57; p<.01)

theta

(d=.03; p=.13)

theta

(d=.11; p=.08)

alpha

(d=.04; p=.11)

alpha

(d=.07; p=.1)

beta

(d=.14; p=.19)

beta

(d=.05; p=.26)

low gamma

(d=.1; p=.12)

low gamma

(d=.09; p=.08)

high gamma

(d=.06; p=.06)

high

gamma

(d=.1; p=.06)

low ripple

0s (↓ 39-44%; d=.16-.19; p<.001)

low ripple

55s (↑ 10-29%; d=.11-.36; p<.001)

high ripple

0s (↓ 39-42%; d=.12-.14; p<.001)

high ripple

50s (↑ 14-31%; d=.08-.2; p<.001)

Superior

temporal gyrus

delta

(d=.16; p=.10)

delta

(d=.09; p=.18)

theta

(d=.17; p=.12)

theta

(d=.15; p=.05)

alpha

(d=.11; p=.08)

alpha

50s (↑ 23-40%; d=.14-.42; p<.001)

beta

(d=.1; p=.06)

beta

-40s (↑ 11-24%; d=.11-.24; p<.001)

low gamma

(d=.1; p=.05)

low gamma

(d=.11; p=.15)

high gamma

(d=.02; p=.33)

high

gamma

(d=.05; p=.15)

low ripple

-30s (↓ 43-47%; d=.25-.27; p<.001)

low ripple

55s (↑ 19-36%; d=.21-.44; p<.01)

high ripple

(d=.08; p=.09)

high ripple

-20s (↑ 12-39%; d=.09-.27; p<.001)

Medial parietal

lobe

delta

(d=.22; p=.09)

delta

(d=.13; p=.21)

theta

(d=.17; p=.1)

theta

(d=.09; p=.12)

alpha

(d=.16; p=.06)

alpha

(d=.09; p=.07)

beta

(d=.08; p=.12)

beta

(d=.08; p=.25)

low gamma

(d=.08; p=.05)

low gamma

(d=.13; p=.25)

high gamma

(d=.06; p=.15)

high

gamma

(d=.09; p=.5)

low ripple

(d=.12; p=.07)

low ripple

(d=.12; p=.15)

high ripple

(d=.1; p=.25)

high ripple

(d=.08; p=.64)

Superior

parietal lobule

delta

(d=.38; p=.07)

delta

(d=.38; p=.06)

theta

(d=.36; p=.05)

theta

(d=.36; p=.07)

alpha

(d=.31; p=.05)

alpha

(d=.19; p=.06)

beta

(d=.16; p=.08)

beta

(d=.16; p=.06)

low gamma

(d=.11; p=.09)

low gamma

(d=.21; p=.08)

high gamma

(d=.14; p=.13)

high

gamma

(d=.14; p=.08)

low ripple

(d=.11; p=.1)

low ripple

(d=.04; p=.53)

high ripple

(d=.11; p=.13)

high ripple

(d=.06; p=.6)

Medial frontal

cortex (including

medial segment

of superior

frontal gyrus)

delta

(d=.3; p=.06)

delta

(d=.29; p=.10)

theta

(d=.16; p=.08)

theta

(d=.2; p=.10)

alpha

(d=.27; p=.1)

alpha

(d=.24; p=.12)

beta

(d=.13; p=.12)

beta

(d=.26; p=.08)

low gamma

(d=.07; p=.33)

low gamma

(d=.09; p=.43)

high gamma

(d=.05; p=.18)

high

gamma

(d=.04; p=.33)

low ripple

(d=.21; p=.06)

low ripple

(d=.26; p=.06)

high ripple

(d=.15; p=.33)

high ripple

(d=.22; p=.05)

Frontal

operculum

delta

(d=.07; p=.53)

delta

(d=.21; p=.81)

theta

(d=.05; p=.53)

theta

(d=.11; p=.81)

alpha

(d=.06; p=.6)

alpha

(d=.28; p=.94)

beta

(d=.11; p=.6)

beta

(d=.23; p=.08)

low gamma

(d=.1; p=.6)

low gamma

(d=.16; p=.2)

high gamma

(d=.07; p=.29)

high

gamma

(d=.16; p=.08)

low ripple

(d=.11; p=.09)

low ripple

(d=.31; p=.05)

high ripple

(d=.16; p=.06)

high ripple

(d=.46; p=.06)

delta

(d=.21; p=.16)

delta

(d=.11; p=.67)

152

Transverse

temporal gyrus

and planum

temporale

theta

(d=.11; p=.24)

theta

(d=.19; p=.29)

alpha

(d=.06; p=.53)

alpha

(d=.09; p=.46)

beta

(d=.07; p=.35)

beta

(d=.09; p=.53)

low gamma

(d=.04; p=.74)

low gamma

(d=.11; p=.24)

high gamma

(d=.06; p=.67)

high

gamma

(d=.21; p=.08)

low ripple

(d=.02; p=.88)

low ripple

(d=.21; p=.05)

high ripple

(d=.09; p=.65)

high ripple

(d=.23; p=.08)

Medial occipital

lobe

Insufficient coverage

Lateral occipital

lobe

Insufficient coverage

Pre- and

postcentral gyri

Insufficient coverage

Supplementary

motor cortex

Insufficient coverage

Table S4. Spectral density results of awakening from REM sleep. Reported here are spectral results for all bands which