Develop a whole brain model personalized for patients with epilepsy

Mas e Final P ojec
Mas e in Neu oenginee ing and Rehabili a ion
De elop a whole b ain model pe sonalized o pa ien s
wi h epilepsy
Au ho : Jimena Mon ealeg e Sánchez
Ad iso s: Edmundo López-Sola, Joan F ancesc Alonso López
Neu oelec ics
May 2025
I would like o exp ess my hea el g a i ude o my hesis supe iso , Edmundo López, o his
in aluable guidance, insigh ul ad ice, and o sha ing his passion o neu oscience. I am
especially hank ul o he oppo uni y o ca y ou his esea ch a Neu oelec ics.
My since e hanks also go o he en i e B ain Modeling depa men , wi h special ecogni ion o
Rose , he esea ch di ec o .
I am deeply g a e ul o Joan F ancesc o se ing as my hesis u o and p o iding his in aluable
insigh s h oughou his jou ney.
Finally, I ex end my wa mes hanks o my iends and amily o hei unwa e ing pa ience and
suppo .
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
1
ABSTRACT
Epilepsy, a neu ological diso de a ec ing millions wo ldwide, p esen s signi ican challenges in
i s d ug- esis an o m, whe e s anda d ea men s o en ail. The GALVANI p ojec aims o
add ess hese challenges by de eloping pe sonalized whole-b ain models o op imize he apeu ic
s a egies such as ansc anial cu en s imula ion and su ge y. These models in eg a e s uc u al,
unc ional, and clinical da a o simula e seizu e dynamics and p edic ea men ou comes.
This hesis ocuses on ad ancing he pipeline o whole-b ain modeling by ansi ioning om
in asi e s e eo-elec oencephalog aphy (SEEG) o non-in asi e scalp elec oencephalog aphy
(EEG). Using neu al mass models (NMMs) and s uc u al connec i i y de i ed om di usion MRI,
he s udy de elops a amewo k o pe sonalize whole-b ain models using EEG da a. Key s eps
include gene a ing syn he ic EEG ia o wa d modeling, compa ing i o empi ical EEG, and
op imizing model pa ame e s such as exci abili y.
Valida ion esul s highligh he easibili y o cap u ing seizu e p opaga ion pa e ns using EEG-
based models. In pa icula , opog aphic ampli ude pa e ns ( opog aphic maps) show a s ong
spa ial co espondence be ween syn he ic and eal EEG, especially in egions iden i ied clinically
as seizu e onse zones. Al hough unc ional connec i i y compa isons show some quali a i e
simila i ies, u he wo k is needed o imp o e hei quan i a i e alignmen .
F equency-speci ic analyses unde sco e he ele ance o ailo ing he model o epilep ically
meaning ul bands, such as he a and gamma, which appea mos sensi i e o ic al ac i i y.
By educing he in asi eness o da a acquisi ion while main aining biologically plausible seizu e
p opaga ion dynamics, his wo k con ibu es o he de elopmen o clinically applicable, pa ien -
speci ic simula ions o epilepsy ea men . Fu u e di ec ions include in eg a ing mul imodal da a
and ex ending he amewo k o o he neu ological diso de s.
Keywo ds
Epilepsy, Neu al Mass Model, EEG, Fo wa d Me hod
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
2
RESUMEN
La epilepsia, un as o no neu ológico que a ec a a millones de pe sonas en odo el mundo,
p esen a impo an es desa íos en su o ma a maco esis en e, en la que los a amien os
es ánda a menudo allan. El p oyec o GALVANI busca abo da es os e os median e el desa ollo
de modelos ce eb ales comple os pe sonalizados que op imicen es a egias e apéu icas como
la es imulación ansc aneal po co ien e y la ci ugía. Es os modelos in eg an da os
es uc u ales, uncionales y clínicos pa a simula la dinámica de las c isis y p edeci los esul ados
del a amien o.
Es e abajo se cen a en a anza en la me odología de modelado ce eb al comple o, pasando
de la es e eoence alog a ía in asi a (SEEG) a la elec oence alog a ía no in asi a de cue o
cabelludo (EEG). U ilizando modelos de masas neu onales (NMMs) y conec i idad es uc u al
de i ada de esonancia magné ica po di usión (dMRI), se desa olla un ma co pa a pe sonaliza
modelos ce eb ales comple os usando da os de EEG. Las e apas cla e incluyen la gene ación
de EEG sin é ico median e modelado di ec o ( o wa d modeling), su compa ación con el EEG
empí ico y la op imización de pa áme os como la exci abilidad.
Los esul ados de alidación mues an la iabilidad de cap u a los pa ones de p opagación de
las c isis median e modelos basados en EEG. En pa icula , los mapas opog á icos de ampli ud
mues an una ue e co espondencia espacial en e el EEG sin é ico y el eal, especialmen e en
las egiones iden i icadas clínicamen e como zonas de inicio de la c isis. Aunque las
compa aciones de conec i idad uncional p esen an cie as simili udes cuali a i as, se equie e
abajo adicional pa a mejo a su alineación cuan i a i a.
Los análisis especí icos po banda de ecuencia sub ayan la ele ancia de adap a el modelo a
bandas clínicamen e signi ica i as como he a y gamma, que son especialmen e sensibles a la
ac i idad ic al.
Al educi la in asi idad en la adquisición de da os man eniendo una dinámica de c isis
biológicamen e plausible, es e abajo con ibuye al desa ollo de simulaciones clínicas
pe sonalizadas pa a el a amien o de la epilepsia. Las líneas u u as incluyen la in eg ación de
da os mul imodales y la ex ensión del ma co a o os as o nos neu ológicos.
Palab as cla e
Epilepsia, modelo de masas neu onales, EEG, modelado di ec o

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
3
RESUM
L’epilèpsia, un as o n neu ològic que a ec a milions de pe sones a eu del món, p esen a g ans
ep es en la se a o ma a maco esis en , en què els ac amen s es ànda d so in no són
e ec ius. El p ojec e GALVANI é com a objec iu a on a aques s ep es mi jançan el
desen olupamen de models pe sonali za s de ce ell sence pe op imi za es a ègies
e apèu iques com l’es imulació ansc anial pe co en i la ci u gia. Aques s models in eg en
dades es uc u als, uncionals i clíniques pe simula la dinàmica de les c isis i p edi els esul a s
del ac amen .
Aques a esi se cen a en a ança la me odologia de modela ge ce eb al comple en la ansició
de la es è eoence alog a ia in asi a (SEEG) a l’elec oence alog a ia (EEG) no in asi a de cui
cabellu . U ili zan models de masses neu onals (NMMs) i connec i i a es uc u al de i ada
d’ima ges de essonància magnè ica pe di usió (dMRI), s’ha desen olupa un ma c pe
pe sonali za models de ce ell sence amb dades d’EEG. Els passos clau inclouen la gene ació
d’EEG sin è ic mi jançan modela ge di ec e ( o wa d modeling), la se a compa ació amb EEG
empí ic i l’op imi zació de pa àme es com l’exci abili a .
Els esul a s de alidació demos en la iabili a de cap u a pa ons de p opagació de les c isis
mi jançan models basa s en EEG. En pa icula , els mapes opog à ics d’ampli ud mos en una
o a co espondència espacial en e l’EEG sin è ic i el eal, especialmen en les egions
clínicamen iden i icades com a zones d’inici de c isi. To i que les compa acions de connec i i a
uncional mos en ce es simili uds quali a i es, calen millo es pe aconsegui una millo alineació
quan i a i a.
Les anàlisis especí iques pe bandes de eqüència e o cen la impo ància d’adap a el model a
bandes signi ica i es des del pun de is a epilèp ic, com les bandes he a i gamma, especialmen
sensibles a l’ac i i a ic al.
Aques eball con ibueix al desen olupamen de simulacions clíniques pe sonali zades pe al
ac amen de l’epilèpsia, eduin la in asi i a en l’adquisició de dades i man enin una dinàmica
de p opagació ealis a. Les línies u u es inclouen la in eg ació de dades mul imodals i l’ampliació
del ma c a al es as o ns neu ològics.
Pa aules clau
Epilèpsia, model de masses neu onals, EEG, modela ge di ec e
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
4
CONTENTS
ABSTRACT ................................................................................................................................. 1
GLOSSARY ................................................................................................................................ 6
LIST OF FIGURES ..................................................................................................................... 8
LIST OF TABLES ...................................................................................................................... 10
1.PREFACE .............................................................................................................................. 11
1.1. O igin o he p ojec ........................................................................................................ 11
1.2. Mo i a ion ....................................................................................................................... 11
1.3. Objec i es ....................................................................................................................... 11
1.4. Requi emen s ................................................................................................................ 12
2. INTRODUCTION .................................................................................................................. 12
2.1. Backg ound .................................................................................................................... 12
2.1.1. The GALVANI P ojec ............................................................................................... 12
2.1.2. Epilepsy ................................................................................................................... 13
2.1.3. Simula ion o Pe sonalized B ain Models ................................................................. 16
2.2. Li e a u e Re iew ............................................................................................................ 17
3. METHODOLOGY .................................................................................................................. 21
4. RESULTS ............................................................................................................................. 41
4.1. Da a P ep ocessing and P ocessing ............................................................................... 41
4.2. E alua ion o Signal Fea u es and Connec i i y Me ics .................................................. 47
4.2.1. Ampli ude En elope and Low Pass ............................................................................. 47
4.2.2. Connec i i y Me ics E alua ion ............................................................................... 48
4.2.3. Topog aphic maps ................................................................................................... 50
4.2.4. F equency Band Selec ion and Spec al Analysis .................................................... 50
4.3 Compa a i e Analysis be ween Syn he ic and Empi ical EEG ......................................... 52
4.4 Ou comes o Pa ame e Op imiza ion .............................................................................. 56
4.5 Summa y o Limi a ions and Ongoing E o s ................................................................... 60
5. DISCUSSION ........................................................................................................................ 60
6. CONCLUSSIONS ................................................................................................................. 62
7. PLANNING ........................................................................................................................... 64
8. ECONOMIC ASSESSMENT ...................................................................................................... 65
9. ENVIRONMENTAL ASSESSMENT ............................................................................................. 66
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
5
10. SOCIAL AND GENDER EQUALITY ASSESSMENT ..................................................................... 68
BIBLIOGRAPHY ....................................................................................................................... 69
APPENDICES ........................................................................................................................... 72
Appendix A. SEEG Elec ode Implan a ion O e iew ......................................................... 72
Appendix B. Ana omical Localiza ion o he Epilep ogenic Zone ........................................ 73
Appendix C. EEG Elec ode Names and Layou ................................................................ 73
Appendix D. Clinical and SEEG Da a o he Second Subjec Used o Compa a i e Analysis
.......................................................................................................................................... 74
Appendix E: O e iew o he Pe sonalized Whole-B ain Modeling Pipeline (Lopez-Sola e
al., 2025) ............................................................................................................................ 74
Appendix F. Da a P ep ocessing and P ocessing .............................................................. 76
Appendix G. Gene ic Algo i hm Pa ame e s ......................................................................... 78
Appendix H: Gene ic Algo i hm S a egies o Co ical Pa ame e Op imiza ion ..................... 79
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
6
GLOSSARY
This glossa y de ines he key echnical e ms and concep s used h oughou his Mas e 's Thesis
P ojec .
o Ampli ude En elope: The ins an aneous magni ude o a signal, ob ained by applying he
Hilbe ans o m. I is used o s udy slow ampli ude luc ua ions in EEG signals.
o ARC File: A ile ha de ines he pe sonalized b ain model a chi ec u e, speci ying he
exci abili y pa ame e s o he nodes and he s uc u al connec i i y be ween egions.
o CSD (Cu en Sou ce Densi y): An es ima e o he densi y o elec ical cu en gene a ed
by neu onal sou ces, used as an in e media e s ep o simula ing EEG signals.
o dMRI (Di usion Magne ic Resonance Imaging): A neu oimaging echnique ha cap u es
he di usion o wa e molecules in b ain issue, allowing he econs uc ion o whi e ma e
ibe pa hways and s uc u al connec omes.
o EEG (Elec oencephalog aphy): A echnique o eco ding he b ain’s elec ical ac i i y ia
elec odes placed on he scalp.
o Elec ode Space: The measu emen space whe e EEG signals a e eco ded a he scalp
elec odes, wi hou econs uc ing he unde lying sou ce ac i i y.
o FEM (Fini e Elemen Me hod): A nume ical me hod used o sol e he EEG o wa d
p oblem, modeling he conduc ion o elec ical signals h ough di e en head issues
(scalp, skull, CSF, b ain).
o Fo wa d Model: The ma hema ical model ha ela es neu al sou ces inside he b ain o
he po en ials measu ed on he scalp elec odes, usually compu ed using me hods such
as FEM.
o Func ional Connec i i y (FC): A measu e o he synch oniza ion o unc ional ela ionships
be ween b ain egions, o en e alua ed using EEG o SEEG signals.
o Gene ic Algo i hm (GA): An op imiza ion me hod inspi ed by biological e olu ion, used o
adjus b ain model pa ame e s by compa ing syn he ic and empi ical EEG da a.
o Hilbe T ans o m: A ma hema ical ans o m used o compu e he ampli ude en elope o
a signal, c ucial o EEG signal ea u e ex ac ion.
o In e se Me hod: A compu a ional echnique used o es ima e he co ical sou ce ac i i y
om EEG signals eco ded a he scalp (mo ing om elec ode space o sou ce space).
o MRI (Magne ic Resonance Imaging): A neu oimaging echnique p o iding high- esolu ion
ana omical images o he b ain.
o Neu al Mass Model (NMM): A ma hema ical model simula ing he a e age ac i i y o
popula ions o neu ons wi hin a b ain egion.
o PCC (Pea son Co ela ion Coe icien ): A s a is ical measu e o linea co ela ion be ween
wo da ase s. I is used o compa e empi ical and syn he ic signals in his wo k.
o SEEG (S e eo-Elec oencephalog aphy): An in asi e me hod o eco ding elec ical b ain
ac i i y using elec odes implan ed wi hin b ain issue.
o Sou ce Space: The econs uc ed space ha ep esen s he es ima ed co ical o igins o
EEG signals.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
13
iii. Pe sonalizing hese p o ocols using hyb id b ain models (HBMs)
based on clinical da a.
i . Clinically alida ing he pe sonalized ea men s h ough pa ien
ials
Con ibu ion o GALVANI
This wo k con ibu es o he GALVANI p ojec by de eloping a pipeline ha
u ilizes non-in asi e EEG da a o pe sonalize whole-b ain models. The
con ibu ion in ol es:
o P ocessing empi ical EEG ic al da a om GALVANI pa ien s.
o Valida ing he pe sonalized models by compa ing syn he ic and eal
EEG da a.
o Replacing o complemen ing SEEG da a wi h EEG o he
pe sonaliza ion o models.
o Adjus ing key model pa ame e s, such as exci abili y and global
coupling, using op imiza ion algo i hms.
The aim is o demons a e ha EEG-based pe sonaliza ion o e s a non-
in asi e and p ecise solu ion o simula ing seizu e dynamics and e alua ing
pe sonalized ea men s in pa ien s wi h d ug- esis an ocal epilepsy.
2.1.2. Epilepsy
De ini ion and Global Impac
Epilepsy is a ch onic neu ological diso de cha ac e ized by ecu en and
unp o oked seizu es, esul ing om abno mal elec ical ac i i y in he b ain. I
a ec s app oxima ely 50 million people wo ldwide, making i one o he mos
common neu ological condi ions globally, acco ding o he Wo ld Heal h
O ganiza ion (WHO). The p e alence o epilepsy is es ima ed a 4 o 10 cases pe
1,000 people, wi h highe a es obse ed in low- and middle-income coun ies due
o limi ed access o heal hca e and inc eased exposu e o isk ac o s such as head
inju ies and in ec ions.
Pha maco esis an Epilepsy
App oxima ely 30% o pa ien s wi h epilepsy a e conside ed pha maco esis an ,
meaning hei seizu es do no espond o a leas wo app op ia ely chosen and
adminis e ed an iseizu e medica ions (ASMs). This condi ion, known as d ug-
esis an epilepsy (DRE), signi ican ly educes he quali y o li e and inc eases he
isk o como bidi ies, including dep ession, anxie y, and cogni i e impai men .

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
14
Cu en T ea men s
a. An iseizu e Medica ions (ASMs): ASMs a e he i s -line ea men o
epilepsy, aiming o supp ess seizu es by modula ing neu onal exci abili y.
Common d ugs include ca bamazepine, alp oa e, and le e i ace am. While
e ec i e o many, long- e m use can lead o side e ec s, including a igue,
dizziness, and mood changes.
b. Su gical In e en ions: Fo pa ien s wi h ocal epilepsy, su gical esec ion o
he epilep ogenic zone (EZ) may o e a cu a i e app oach. This me hod is
mos e ec i e in well-localized epilep ogenic a eas bu ca ies isks such as
neu ological de ici s.
c. Neu os imula ion The apies: Neu os imula ion me hods include:
• Vagus Ne e S imula ion (VNS): Implan able de ices ha
deli e elec ical pulses o he agus ne e, modula ing b ain
ac i i y.
• Deep B ain S imula ion (DBS): Elec odes implan ed in
speci ic b ain egions o egula e seizu e ac i i y.
• Responsi e Neu os imula ion (RNS): Closed-loop sys ems
ha de ec and supp ess abno mal b ain ac i i y in eal- ime.
1. T ansc anial Elec ical S imula ion ( ES): T ansc anial elec ical
s imula ion is a non-in asi e neu omodula ion echnique ha uses
weak elec ical cu en s applied o he scalp o modula e co ical
exci abili y. I s sub ypes include:
▪ T ansc anial Di ec Cu en S imula ion ( DCS): Applies a
cons an cu en o ei he inc ease (anodal DCS) o dec ease
(ca hodal DCS) co ical exci abili y.
▪ T ansc anial Al e na ing Cu en S imula ion ( ACS): Uses
sinusoidal cu en s o en ain b ain oscilla ions a speci ic
equencies.
▪ T ansc anial Random Noise S imula ion ( RNS): Applies
andomly a ying cu en s o in luence neu onal plas ici y.
The Role o ES in Epilepsy
T ansc anial elec ical s imula ion ( ES) has eme ged as a p omising
neu omodula o y echnique o he ea men o d ug- esis an epilepsy (DRE).
Unlike in asi e me hods such as deep b ain s imula ion (DBS) o agus ne e
s imula ion (VNS), ES is non-in asi e, sa e, and well- ole a ed. I in ol es he
applica ion o low-in ensi y elec ical cu en s (<2 mA) o modula e co ical
exci abili y h ough elec odes placed on he scalp.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
15
The wo p ima y o ms o ES a e:
1. T ansc anial Di ec Cu en S imula ion ( DCS):
• Uses a cons an cu en o modula e neu al ac i i y.
• Ca hodal s imula ion o e he epilep ogenic zone inhibi s co ical
exci abili y, while anodal s imula ion is exci a o y.
• Clinical ials ha e shown educ ions in seizu e equency, wi h e ec s
las ing o weeks a e epea ed sessions.
2. T ansc anial Al e na ing Cu en S imula ion ( ACS):
• Applies sinusoidal cu en s o en ain b ain oscilla ions.
• Al hough less s udied, p elimina y indings sugges po en ial o
supp essing epilep i o m discha ges.
Mechanisms o Ac ion
• Hype pola iza ion and depola iza ion: Depending on cu en
di ec ion, neu ons become less o mo e exci able.
• Ne wo k e ec s: Beyond local s imula ion, ES in luences la ge-
scale b ain ne wo ks, educing abno mal synch ony wi hin
epilep ogenic ne wo ks.
Clinical and Compu a ional Insigh s
Clinical s udies ha e demons a ed he e icacy o ca hodal DCS in educing
in e ic al epilep i o m discha ges and seizu e equency. Compu a ional models
complemen hese indings by simula ing elec ic ield dis ibu ions and op imizing
s imula ion pa ame e s o maximize he apeu ic e ec s while minimizing undesi ed
ac i a ion.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
16
Figu e 1 E ec s o ES on Neu al Ac i i y and Ne wo ks [3].
Illus a ion o he mechanisms o ansc anial elec ical s imula ion, including DCS and ACS. The
igu e shows elec ode placemen , cu en low h ough he b ain, and he esul ing e ec s on neu onal
exci abili y and ne wo k synch oniza ion.
Fu u e Challenges
Despi e ad ancemen s, se e al challenges emain:
- Iden i ying op imal s imula ion p o ocols and pa ame e s o indi idual
pa ien s.
- Unde s anding long- e m e ec s and po en ial neu oplas ic changes
induced by ES.
- In eg a ing ES in o s anda d clinical wo k lows as a complemen o
al e na i e o cu en ea men s.
2.1.3. Simula ion o Pe sonalized B ain Models
The use o compu a ional models o simula e b ain ac i i y has become an
inc easingly powe ul app oach in neu oscience and clinical esea ch.
These models aim o ep oduce la ge-scale b ain dynamics by
ma hema ically ep esen ing he in e ac ions be ween di e en b ain
egions. One o he mos p omising de elopmen s in his ield is he
eme gence o pe sonalized b ain models, which inco po a e subjec -
speci ic ana omical and physiological da a o simula e how an indi idual
b ain beha es unde heal hy o pa hological condi ions.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
17
In he con ex o neu ological diso de s, such as epilepsy, pe sonalized
simula ions o e a no el way o unde s and disease mechanisms and
suppo clinical decision-making. Unlike adi ional diagnos ic ools, b ain
models p o ide a dynamic and pa ien -speci ic iew o how seizu es
o igina e and p opaga e ac oss he b ain's ne wo k. They also make i
possible o es in e en ions i ually, o e ing a sa e and con olled
en i onmen o explo e he po en ial ou comes o di e en he apeu ic
s a egies.
These simula ions a e buil using a combina ion o neu oimaging da a (e.g.,
MRI, dMRI) and elec ophysiological eco dings (e.g., EEG o SEEG),
allowing esea che s o econs uc bo h he s uc u e and unc ion o he
b ain. Once he model is cons uc ed, compu a ional me hods can be used
o simula e b ain ac i i y and compa e i wi h eal da a. Th ough his i e a i e
p ocess, he model pa ame e s can be adjus ed o align wi h he pa ien ’s
condi ion, leading o highly indi idualized insigh s.
As he ield e ol es, b ain modeling is becoming an essen ial componen in
he mo emen owa d p ecision medicine, enabling he design o ea men s
ha a e ailo ed no only o he ype o pa hology, bu o he unique neu al
a chi ec u e o each pa ien .
2.2. Li e a u e Re iew
In ecen yea s, compu a ional modeling has become an inc easingly aluable ool
in he s udy and ea men o epilepsy, pa icula ly o unde s anding seizu e
dynamics and op imizing p esu gical planning. These app oaches ypically ely on
whole-b ain models composed o neu al mass models (NMMs) pe sonalized using
pa ien -speci ic da a. Mos s udies ha e u ilized in asi e eco dings, such as
s e eo elec oencephalog aphy (SEEG), due o hei high spa ial p ecision.
Howe e , ecen e o s ha e sough o adap hese pipelines o scalp EEG, which
p esen s challenges in e ms o signal quali y, spa ial esolu ion, and sou ce
localiza ion. This e iew explo es he cu en s a e o he a in in e se and o wa d
modeling, b ain connec i i y analysis, neu al mass modeling, op imiza ion
algo i hms, and he ole o equency bands, aiming o p o ide a comp ehensi e
con ex o he me hodological con ibu ions o his wo k.
In e se and Fo wa d Modeling
A cen al s ep in EEG-based b ain modeling is he es ima ion o co ical sou ce
ac i i y. This is add essed h ough in e se modeling, which econs uc s b ain
ac i i y om he scalp- eco ded signals. In [4], he au ho s use eLORETA, a linea
in e se me hod ha compu es app oxima ely 8,004 sou ces o gene a e a de ailed
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
18
h ee-dimensional model o co ical dynamics. Due o compu a ional cons ain s,
he s udy p oposes a educ ion o he sou ce space using he Desikan-Killiany
a las, esul ing in 15 ana omically ele an egions o in e es (ROIs). While his
educ ion imp o es e iciency, he model e ains enough spa ial de ail o simula e
la ge-scale b ain dynamics.
Beam o ming is ano he in e se me hod ha o e s high spa ial esolu ion,
pa icula ly when used wi h high-densi y EEG (HD-EEG). By cons uc ing adap i e
spa ial il e s, beam o ming enhances he accu acy o sou ce localiza ion and,
when combined wi h Independen Componen Analysis (ICA), imp o es he
sepa a ion o o e lapping neu al sou ces. These me hods a e explo ed in [5],
which emphasizes hei capaci y o inc ease he accu acy o unc ional mapping in
epilepsy.
Fo wa d modeling, on he o he hand, deals wi h p ojec ing simula ed co ical
ac i i y on o he scalp. Se e al o wa d models a e desc ibed in he li e a u e. The
Fini e Elemen Me hod (FEM), as desc ibed in [6], accoun s o he complex
geome y and aniso opic conduc i i y o b ain issues, p o iding highly accu a e
p ojec ions. The Bounda y Elemen Me hod (BEM), used in [4] [7], simpli ies he
head model by ocusing on bounda y su aces, o e ing a good balance be ween
compu a ional load and ana omical p ecision. O he app oaches, such as
ecip ocal me hods and concen ic sphe ical app oxima ions, a e discussed in [4]
[8], as as bu less ana omically accu a e al e na i es, app op ia e o explo a o y
o esou ce-cons ained applica ions.
Each me hod has i s own applica ion niche. FEM is pa icula ly sui ed o pa ien -
speci ic SEEG models ha equi e ine-g ained accu acy, while BEM and sphe ical
me hods a e mo e compa ible wi h non-in asi e EEG applica ions whe e
compu a ional e iciency is p io i ized.
High-densi y EEG (HD-EEG) plays a i al ole in enhancing he spa ial esolu ion
o in e se modeling echniques. As de ailed in [9], inc easing he numbe o
elec odes imp o es sampling o he scalp po en ial and enhances he abili y o
esol e closely spaced sou ces. This makes HD-EEG a aluable b idge be ween
adi ional EEG and in asi e me hods like SEEG. When used alongside ICA and
beam o ming, HD-EEG signi ican ly imp o es he ideli y o sou ce space
es ima ions, which is c ucial o eliable modeling.
S uc u al and Func ional Connec i i y
Once he neu al ac i i y is p ojec ed in o he sou ce o elec ode space, he nex
s ep in ol es cons uc ing connec i i y ma ices ha desc ibe in e ac ions be ween
di e en b ain egions. S uc u al connec i i y is de ined by he numbe and
s eng h o whi e ma e ac s linking di e en pa cels and is ypically ex ac ed
om di usion MRI (dMRI).
Func ional connec i i y, in con as , is based on he s a is ical dependencies
be ween neu al signals. Va ious me ics a e used o es ima e hese dependencies.

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
19
Phase-based me ics such as Phase Locking Value (PLV) [4], Phase Lag Index
(PLI) [5], and cohe ence [9] cap u e synch onous oscilla ions be ween egions,
while ampli ude-based me ics like Ampli ude En elope Co ela ion (AEC) [10] and
Pea son Co ela ion Coe icien (PCC) [8] assess co- luc ua ions in signal in ensi y.
Time-lagged in e ac ions can also be measu ed using c oss-co ela ion, as
desc ibed in [7].
To ensu e he obus ness o hese connec i i y es ima es, s udies such as [4]
ecommend he use o su oga e da a o es ablish s a is ical signi icance
h esholds. This app oach il e s spu ious connec ions ha may a ise due o noise
o olume conduc ion, imp o ing he in e p e abili y o unc ional ne wo ks.
Ad anced analy ical echniques u he ex end unc ional connec i i y analysis.
Di ec ed connec i i y me ics such as G ange Causali y, Di ec ed T ans e
Func ion (DTF), and Pa ial Di ec ed Cohe ence (PDC) a e used o in e he
di ec ionali y o in o ma ion low be ween b ain egions [5]. G aph heo e ical
app oaches also play an impo an ole, o e ing ne wo k-le el desc ip o s like
be weenness cen ali y o iden i y nodes ha a e c i ical o seizu e p opaga ion
and ne wo k synch oniza ion.
Neu al Mass Models (NMMs)
Whole-b ain compu a ional models a e buil by assigning neu al mass models o
he egions de ined in he pa cella ion. These models simula e he collec i e
beha io o neu onal popula ions and a e key o ep oducing obse ed b ain
dynamics. Va ious ypes o NMMs ha e been p oposed in he li e a u e. The
Jansen-Ri model [11] simula es alpha hy hms and ep esen s in e ac ions
be ween py amidal cells and inhibi o y in e neu ons. The Wendling model [8], an
ex ension o Jansen-Ri , adds addi ional inhibi o y eedback loops o cap u e he
dynamics o epilep ic discha ges. The The a model [4] concep ualizes each node
as a phase oscilla o , making i sui able o analyzing ic ogenici y.
The Lamina Neu al Mass Model (LaNMM) [12] [10] inco po a es a laye ed co ical
s uc u e o simula e bo h alpha and gamma band ac i i y. Finally, he López-Sola
model [12] [8] in oduces chlo ide accumula ion mechanisms o simula e
ic ogenesis and seizu e p opaga ion, o e ing a biophysically g ounded al e na i e
o classical NMMs.
Op imiza ion Algo i hms
Pe sonaliza ion o he whole-b ain model is achie ed h ough pa ame e
op imiza ion. Se e al op imiza ion s a egies ha e been adop ed h ough li e a u e.
Gene ic Algo i hms (GA) use e olu iona y echniques including c osso e ,
mu a ion, and selec ion o iden i y op imal pa ame e se s. Pa icle Swa m
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
20
Op imiza ion (PSO) [5] mimics he beha io o swa ming animals and excels a
explo ing high-dimensional spaces.
O he me hods include Simula ed Annealing [8], which uses p obabilis ic sampling
o a oid local minimum, and g adien descen echniques, which p o ide e icien
ine- uning. Bayesian Op imiza ion is also gaining a en ion o i s abili y o
inco po a e p io knowledge and unce ain y in he op imiza ion p ocess. These
echniques ha e been used o minimize loss unc ions based on signal simila i y,
unc ional connec i i y dis ance, o opog aphic co ela ion.
F equency Bands in Epilepsy Modeling
EEG signals a e ypically analyzed ac oss canonical equency bands, each
associa ed wi h di e en physiological and pa hological p ocesses. In [9], del a
(0.5–4 Hz) and he a (4–8 Hz) ac i i y a e shown o domina e du ing he ea ly
s ages o seizu es, while alpha (8–12 Hz) hy hms end o be supp essed. Be a
(13–30 Hz) and gamma (>30 Hz) bands a e o en associa ed wi h hype exci abili y
and synch oniza ion nea he seizu e onse zone.
Func ional connec i i y dynamics also a y ac oss equency bands. This has led
o he de elopmen o mul i equency analysis app oaches, whe e unc ional
connec i i y is compu ed independen ly o each band using me ics such as PLV,
cohe ence, and AEC. These band-speci ic ma ices a e hen used o guide
op imiza ion and alida e he models ac oss di e en dynamical egimes. The
in eg a ion o mul i equency ea u es imp o es model gene alizabili y and
enhances he iden i ica ion o epilep ic ne wo ks.
Whole-B ain Models in Epilepsy
Recen ad ances in epilepsy modeling emphasize pe sonalized whole-b ain
app oaches. The Vi ual Epilep ic Pa ien (VEP) ep esen s a p ominen
amewo k, p o iding pa ien -speci ic b ain a lases and pe sonalized la ge-scale
models ha simula e seizu e ini ia ion and p opaga ion [13] [14] [15].The VEP
u ilizes s uc u al connec i i y de i ed om dMRI and in eg a es clinical SEEG da a
o pe sonalize model pa ame e s. I has demons a ed u ili y in p edic ing epilepsy
su ge y ou comes, iden i ying seizu e onse zones, and guiding he apeu ic
in e en ions such as a ge ed neu os imula ion [16].
Ji sa e al. (2017) ini ially in oduced he VEP, es ablishing ounda ional me hods
o cons uc ing indi idualized b ain ne wo ks capable o simula ing pa ien -speci ic
seizu e dynamics [14]. Subsequen e inemen s in eg a ed p obabilis ic
amewo ks (Bayesian Vi ual Epilep ic Pa ien ) o be e cap u e unce ain y in
epilep ogenic zone localiza ion and imp o e p edic i e accu acy [13]. Addi ional
s udies employing simila pe sonalized whole-b ain models ha e examined
s uc u al ne wo k cha ac e is ics, connec i i y dis up ions, and speci ic
hypo heses ela ed o epilepsy su ge y planning.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
21
These amewo ks ha e p ima ily elied on in asi e SEEG due o supe io spa ial
p ecision. Howe e , ecen a emp s inc easingly seek o adap whole-b ain
modeling o scalp EEG, add essing p ac ical limi a ions o in asi e app oaches.
Such EEG-based modeling has employed in e se me hods, ad anced signal
p ocessing echniques, and op imiza ion algo i hms o o e come inhe en
challenges ela ed o limi ed spa ial esolu ion and sou ce localiza ion unce ain y.
Summa y and Mo i a ion
In summa y, li e a u e p o ides a solid ounda ion o building whole-b ain models
o epilepsy using s uc u al and unc ional da a. Howe e , mos exis ing
app oaches ely on sou ce- econs uc ed EEG o in asi e SEEG, bo h o which
in oduce assump ions and clinical limi a ions. This hesis p oposes an al e na i e
pipeline ha pe o ms model i ing di ec ly in he elec ode space using s anda d
scalp EEG eco dings. By a oiding he in e se modeling s ep and combining signal
p ocessing echniques wi h whole-b ain simula ions, his app oach aims o enable
a mo e accessible and non-in asi e amewo k o pe sonalized modeling in
clinical se ings.
3. METHODOLOGY
This sec ion desc ibes he me hodological amewo k de eloped o pe sonalize whole-b ain
compu a ional models using non-in asi e EEG da a, wi h he objec i e o eplica ing seizu e
p opaga ion in pa ien s diagnosed wi h d ug- esis an epilepsy. The app oach builds upon exis ing
modeling pipelines ini ially de eloped using s e eo-elec oencephalog aphy (SEEG), adap ing
hem o scalp EEG signals. This adap a ion no only b oadens he applicabili y o he me hod o
non-in asi e clinical se ings bu also in oduces new challenges ela ed o signal esolu ion and
spa ial co e age.
The me hodology is o ganized in wo main blocks. Fi s , we p esen he ools, da ase s, and
compu a ional models used in he s udy. Then, we desc ibe in de ail he pipeline de eloped o
EEG-based pe sonaliza ion o whole-b ain models.
3.1. Tools and Resou ces
A undamen al componen o he me hodology used in his s udy in ol es le e aging es ablished
ools, esou ces, and compu a ional pipelines. The e ec i eness and accu acy o ou esul s
depend signi ican ly on hese exis ing amewo ks and he obus ness o a ailable so wa e and
da a-p ocessing wo k lows.
3.1.1. Pa ien Da a and Neu oimaging Resou ces
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
22
The de elopmen o his p ojec equi ed access o comp ehensi e mul imodal
da ase s om a single pa ien diagnosed wi h d ug- esis an ocal epilepsy. These
da ase s included:
Clinical anno a ions, pa icula ly he p ecise localiza ion o he epilep ogenic zone
(EZ), which was p o ided by he clinical eam. Acco ding o hei assessmen , he EZ
was loca ed in he le occipi al co ex, speci ically in pa cels 65 and 69 o he Vi ual
Epilep ic Pa ien (VEP) a las. The ana omical pa cella ion om he VEP a las wi h he
epilep ogenic pa cels 65 and 69 speci ically ma ked is illus a ed in Appendix B.
Neu oimaging da a, including: High- esolu ion T1-weigh ed MRI o ana omical
econs uc ion and pa cella ion and Di usion MRI (dMRI) o he es ima ion o
s uc u al connec i i y ia ac og aphy.
Elec ophysiological eco dings, which we e essen ial o model i ing and alida ion:
S e eo-elec oencephalog aphy (SEEG): Dep h elec odes we e implan ed in a ge ed
b ain egions, pa icula ly a ound he le occipi al lobe. G ey ma e con ac s
con i med implan a ion nea clinically suspec ed EZ egions. A de ailed lis ing o SEEG
con ac coo dina es and ana omical labels is p o ided in Appendix A. Scalp EEG:
High-densi y EEG was eco ded using a 64-channel cap (10–10 in e na ional sys em)
a a sampling equency o 1024 Hz. The ull lis o channels and me ada a is p o ided
in Appendix C.
In addi ion o he subjec desc ibed abo e, a b ain model om ano he subjec is used
o compa ison. Da a om his new subjec a e shown in Appendix D.
3.1.2. Whole B ain Model A chi ec u e
The whole-b ain model (WBM) used in his p ojec is a pe sonalized compu a ional
amewo k de eloped wi hin he Gal ani p ojec o simula ing seizu e p opaga ion in
pa ien s wi h d ug- esis an epilepsy. This model is buil om indi idual pa ien da a
and in eg a es ana omical, s uc u al, and unc ional in o ma ion o ep oduce ealis ic
la ge-scale b ain dynamics.
Pa ien -speci ic magne ic esonance imaging (MRI) is used o gene a e a cus om
co ical su ace model, which is hen pa cella ed in o 162 ana omically and unc ionally
ele an egions using he Vi ual Epilep ic Pa ien (VEP) a las. This a las di ides he
co ex in o 73 pa cels pe hemisphe e, plus 8 subco ical egions pe hemisphe e. The
pa cella ion suppo s p ecise ana omical mapping and allows consis en elec ode- o-
pa cel assignmen s.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
29
Bo h EEG da ase s unde go iden ical p ocessing s eps: band pass il e ing, ea u e ex ac ion and
low pass il e ing. To ensu e empo al consis ency, a 30-second window is selec ed o analysis
in bo h EEG da ase s. This window spans om 10 seconds be o e seizu e onse o 20 seconds
a e onse , allowing he cap u e o p e-ic al, ic al, and pos -ic al dynamics.
3.2.2. Fea u e Ex ac ion and Func ional Connec i i y Me ics
To ex ac ele an in o ma ion om bo h empi ical and syn he ic EEG signals, a ious To
ex ac ele an in o ma ion om bo h empi ical and syn he ic EEG signals, a ious
empo al and spec al ea u es we e ini ially explo ed. Al hough mul iple op ions we e
conside ed o cap u ing neu al dynamics, he Ampli ude En elope, de i ed ia he
Hilbe T ans o m, was ul ima ely selec ed as he mos sui able ea u e o subsequen
model e alua ion and compa ison. This decision was suppo ed by bo h he empi ical
esul s ob ained and ecommenda ions om ele an li e a u e.
The Ampli ude En elope cap u es he ins an aneous magni ude o oscilla o y ac i i y and
p o ides a smoo h, ime- esol ed ep esen a ion o signal luc ua ions. This ea u e p o ed
pa icula ly e ec i e in iden i ying he spa ial dis ibu ion o epilep i o m ac i i y and
enabled obus compa isons be ween empi ical and simula ed EEG signals in he
elec ode space.
Hilbe T ans o m and Ampli ude En elope Ex ac ion:
The analy ic signal 𝑧(𝑡)is compu ed om a eal- alued EEG ime se ies
𝑥(𝑡) ( eal- alued
EEG signal) using he Hilbe ans o m 𝐻(𝑥(𝑡)) (Hilbe ans o m), as ollows:
𝑧(𝑡)=𝑥(𝑡)+𝑖𝐻[𝑥(𝑡)] (2)
The Ampli ude En elope 𝐴(𝑡) is hen ob ained as he magni ude o he analy ic signal:
𝐴(𝑡)=|𝑧(𝑡)|=√𝑥(𝑡)2+𝐻(𝑥(𝑡))2 (3)
The ea u e was ex ac ed a e il e ing he signals o he he a band (4–10 Hz), which is
o en associa ed wi h ic al and p e-ic al dynamics in ocal epilepsy [6]. The en elope
p ese es bo h ins an aneous ampli ude and phase in o ma ion and is pa icula ly sui able
o iden i ying p opaga ing seizu e pa e ns. In ou pipeline, he en elope is u he low-
pass il e ed (0.1–1 Hz) o isola e slow co ical modula ions, aligning wi h obse a ions in
p e ious epilepsy s udies and p io wo k in he GALVANI p ojec .
Func ional Connec i i y Me ics:

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
30
i. Pea son Co ela ion Coe icien (PCC): Measu es linea dependencies be ween
wo signals: 𝜑𝑥𝑦=𝑐𝑜𝑣(𝑥,𝑦)
𝜎𝑥𝜎𝑦 (4)
I is pa icula ly e ec i e in e alua ing ampli ude-based synch ony.
ii. C oss-Co ela ion (CC): E alua es he simila i y be ween wo signals as a unc ion
o ime lag: 𝑅𝑥𝑦(𝜏)=∑𝑥(𝑡)𝑦(𝑡+𝜏) (5)
This is use ul o iden i ying delays in p opaga ion.
iii. Cohe ence: Measu es o equency-dependen co ela ion be ween signals:
𝐶𝑥𝑦(𝑓)=|𝑆𝑥𝑦(𝑓)|2
𝑆𝑥𝑥(𝑓)𝑆𝑦𝑦(𝑓) (6)
I e lec s he deg ee o phase and ampli ude coupling ac oss equencies.
i . Phase Locking Value (PLV): Cap u es phase synch oniza ion ac oss ials o ime:
𝑃𝐿𝑉=|1𝑁∑𝑒𝑖(𝜃𝑥(𝑛)−𝜃𝑦(𝑛)
𝑁
𝑛=1 |(7)
Highe alues indica e consis en phase di e ences.
O he me ics such as imagina y cohe ence and phase lag index ha e also been es ed
bu a e no shown because hey we e no conside ed ele an .
To quan i y he spa ial simila i y be ween empi ical and syn he ic opog aphic maps, he
ollowing echniques we e applied:
i. Global Map Dissimila i y (GMD): S anda dize each ec o by cen e ing and scaling
o uni 𝐿2no m [22]:
𝑢′=𝑢−𝑢
‖𝑢−𝑢‖2 𝑣′=𝑣−𝑣
‖𝑣−𝑣‖2 (8)
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
31
Compu e dissimila i y:
𝐺𝑀𝐷(𝑢,𝑣)= √1𝑁∑(𝑢𝑖′−𝑣𝑖′)2
𝑁
𝑖=1 , 0≤𝐺𝑀𝐷≤2 (9)
ii. Spa ial Pea son Co ela ion: Fla en he in e pola ed alues in o ec o s o leng h
and compu e: 𝜌𝑖𝑚𝑔=𝑐𝑜𝑟𝑟({𝑢𝑖}𝑖=1
𝑀,{𝑣𝑖}𝑖=1
𝑀)(10)
iii. Roo Mean Squa e E o (RMSE): quan i ies he a e age ampli ude di e ence
be ween empi ical and syn he ic opog aphic maps:
𝑅𝑀𝑆𝐸=√1
𝑀∑(𝑢𝑗−𝑣𝑗)2
𝑀
𝑗=1 (11)
Whe e 𝑢𝑗 and 𝑣𝑗 a e he in e pola ed alues a each scalp loca ion and MMM is
he numbe o alid pixels. Lowe RMSE indica es close ampli ude ma ch, making
i complemen a y o co ela ion-based me ics.
i . Clus e -wise Me ics: Pa i ion elec odes in o p ede ined egions o in e es
(clus e s) Fo each clus e C, ex ac sub- ec o s 𝑢𝑐, 𝑣𝑐 and compu e he di e en
me ics.
Toge he , hese ea u es and me ics p o ide a obus amewo k o quan i ying he
empo al, spec al, and spa ial s uc u e o epilep i o m dynamics in bo h empi ical and
simula ed EEG da a.
F equency Bands and Spec al Analysis
A c ucial componen o EEG analysis is he iden i ica ion and in e p e a ion o speci ic
equency bands, as di e en neu al p ocesses—and pa icula ly epilep i o m dynamics—
a e p e e en ially exp essed in dis inc spec al anges. EEG signals a e commonly
decomposed in o canonical equency bands:
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
32
• Del a (0.5–4 Hz): O en associa ed wi h deep sleep s ages and slow-wa e ac i i y.
In epilepsy, inc eased del a ac i i y may be obse ed in e ic ally.
• The a (4–8 Hz): F equen ly linked o d owsiness and cogni i e p ocessing.
No ably, his band has shown ele ance in ic al and p e-ic al phases o ocal
epilepsy.
• Alpha (8–12 Hz): Typically ep esen s elaxed wake ulness and is o en
supp essed du ing seizu es.
• Be a (13–30 Hz): Associa ed wi h mo o ac i i y and ale ness. Abno mal be a
powe may eme ge in epilep ic ne wo ks.
• Gamma (>30 Hz): High- equency oscilla ions (HFOs) in his band, especially 80–
150 Hz, a e o en conside ed bioma ke s o epilep ogenic zones.
The analysis o EEG ac i i y ac oss hese bands can e eal bo h local and ne wo k-le el
al e a ions in b ain dynamics. In his wo k, special emphasis is placed on he he a band
(4–10 Hz), which was empi ically and bibliog aphically alida ed as he mos
ep esen a i e o seizu e p opaga ion in ou case s udy.
Spec al Analysis Techniques:
i. Powe Spec al Densi y (PSD): Al eady desc ibed in Sec ion 4.2, PSD e eals
how powe is dis ibu ed ac oss equencies. I is used o compa e ene gy le els
in speci ic bands be ween empi ical and syn he ic EEG signals.
ii. Spec og ams: These ep esen he e olu ion o signal powe o e bo h ime and
equency using sho - ime Fou ie ans o ms (STFT)
iii. Band Powe Es ima ion: The o al powe wi hin each canonical band is compu ed
by in eg a ing he PSD o e he co esponding equency in e al. This allows
compa ison o band-speci ic powe dis ibu ions ac oss di e en b ain s a es.
Toge he , hese echniques p o ide a comp ehensi e spec al cha ac e iza ion o EEG
dynamics. Thei in eg a ion in o he modeling pipeline allows a mul i-band alida ion o
syn he ic EEG ou pu s and suppo s he iden i ica ion o equency-dependen
abno mali ies ypical o epilep ic ne wo ks.
3.2.3. Compa ison Be ween Syn he ic and Empi ical EEG
The compa ison be ween syn he ic and empi ical EEG is conduc ed using wo
complemen a y s a egies:
i. Topog aphic compa ison: EEG ampli ude en elopes a e isualized as co ical
ac i a ion maps ( opog aphic maps), ei he dynamically o e ime o as a e ages
ac oss he selec ed window. Quali a i ely, his allows he obse a ion o spa ial
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
33
ac i a ion pa e ns and hei sp ead, aiding isual iden i ica ion o ocal onse and
p opaga ion. Quan i a i ely, he spa ial co ela ion be ween opog aphic maps is
compu ed using Pea son co ela ion. Addi ionally, he mean ampli ude ec o (one
alue pe elec ode) is ex ac ed om each EEG ype and compa ed using
Pea son co ela ion o e alua e global opog aphic simila i y.
ii. Func ional connec i i y compa ison: Func ional connec i i y ma ices a e
compu ed om bo h da ase s using mul iple me ics: Pea son co ela ion, c oss-
co ela ion, cohe ence, imagina y cohe ence, and Phase Locking Value (PLV).
These ma ices a e compa ed isually o iden i y simila i ies in ne wo k s uc u e
and clus e ing pa e ns. Quan i a i ely, ma ix- o-ma ix co ela ions a e compu ed
using Pea son co ela ion o he uppe iangula elemen s o assess he alignmen
be ween empi ical and syn he ic connec i i y.
Toge he , hese analyses p o ide a obus amewo k o assessing bo h spa ial and
ne wo k-le el co espondence be ween model-gene a ed EEG signals and eal pa ien
da a. The ou comes o hese compa isons a e discussed in he Resul s sec ion. The
esul s o hese compa isons a e p esen ed in he Resul s sec ion.
The ou come o his compa ison, quan i ied using ea u es such as opog aphic
ampli ude and unc ional connec i i y, p o ides he objec i e unc ion o he
op imiza ion algo i hms. Wi hou es ablishing his compa ison s ep, he e would be no
e e ence me ic o guide pa ame e i ing.
On he o he hand, his compa ison also ac s as a alida ion ool. Since he e e ence
model has al eady been i ed using SEEG da a, we expec i o eplica e he
spa io empo al dynamics o seizu e p opaga ion. I he syn he ic EEG gene a ed om
his model aligns well wi h he eal EEG o he same subjec —bo h in e ms o spa ial
dis ibu ion ( opog aphic maps) and ne wo k s uc u e ( unc ional connec i i y
ma ices)—i p o ides s ong e idence ha he model is ealis ic and physiologically
meaning ul.
I is also impo an o emphasize ha while he long- e m goal is o pe o m model
pe sonaliza ion solely om scalp EEG da a, he ini ial model is limi ed o he a eas
co e ed by SEEG elec odes. In ou case, SEEG eco dings we e ob ained p ima ily
om occipi al egions, based on p io clinical knowledge ega ding he seizu e ocus.
This means ha only hose b ain pa cels wi h implan ed elec odes we e o iginally
i ed. As a esul , egions wi hou SEEG co e age—such as he cen al co ex— e ain
de aul exci abili y pa ame e s and a e no accu a ely modeled.
To add ess his limi a ion, he cu en wo k p oposes using EEG da a o complemen
SEEG and imp o e model accu acy ac oss he en i e co ex. Ini ially, his is done by
e ining he SEEG-based model using EEG-de i ed ea u es. In u u e wo k, his
app oach could be ex ended o i models exclusi ely om non-in asi e EEG, which
would d ama ically expand he clinical applicabili y o whole-b ain modeling.
Th oughou he compa a i e analysis be ween empi ical and syn he ic EEG signals,
we inco po a ed a alida ion s ep using addi ional compa isons. Speci ically, empi ical
unc ional connec i i y ma ices and opog aphic dis ibu ions we e compa ed no only
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
34
wi h he subjec -speci ic syn he ic EEG da a bu also wi h syn he ic EEG da a om
ano he subjec and wi h andomly gene a ed EEG ma ices. This alida ion se ed as
a baseline o con ex ualize and in e p e he signi icance o he compu ed Pea son
co ela ion coe icien s.
Fo example, a mode a e Pea son co ela ion alue o 0.4 be ween empi ical and
subjec -speci ic syn he ic connec i i y ma ices migh ini ially appea low. Howe e , i
compa isons wi h ano he subjec 's syn he ic connec i i y yield co ela ion alues nea
ze o (e.g., 0.01), his o iginal 0.4 co ela ion gains signi icance. Con e sely, i a
seemingly high co ela ion (e.g., 0.8) is simila ly achie ed wi h syn he ic da a om an
un ela ed subjec , he in e p e a i e alue o he me ic migh be econside ed. This
s ep p o ided a con ex ually g ounded benchma k, ein o cing con idence in ou
compa a i e analyses and aiding in in e p e ing he quan i a i e me ics ob ained.
3.2.4. Pa ame e Op imiza ion S a egies
In his sec ion, we desc ibe in de ail he pa ame e op imiza ion s a egies implemen ed in
his wo k. Ini ially, he p ima y objec i e o he p ojec was o de elop me hods capable o
i ing and pe sonalizing model pa ame e s di ec ly and exclusi ely om non-in asi e EEG
da a. Howe e , as he esea ch p og essed, i became e iden ha an in e media e s ep
would be bene icial. Ra he han immedia ely eplacing SEEG da a, we i s explo ed using
EEG da a o complemen and enhance he exis ing SEEG-based model.
The main eason o his app oach is ha SEEG p o ides p ecise and localized eco dings
o co ical elec ical ac i i y, bu i s spa ial co e age is inhe en ly limi ed o a eas whe e
elec odes a e implan ed. In ou pa icula case, SEEG elec odes we e p ima ily placed
in occipi al egions, guided by p io clinical knowledge abou he seizu e ocus.
Consequen ly, only b ain pa cels di ec ly associa ed wi h hese elec odes we e ini ially
i ed, lea ing o he co ical egions—such as he cen al co ex—wi h de aul exci abili y
alues ha migh no accu a ely e lec he unde lying physiology.
To add ess his limi a ion in spa ial co e age, EEG da a—which non-in asi ely co e s he
en i e co ical su ace—was in oduced as a complemen a y sou ce o in o ma ion. This
allowed us o e ine and imp o e he ini ial SEEG-based model, achie ing a mo e accu a e
and comp ehensi e pe sonaliza ion ac oss p e iously un i ed co ical egions. Once his
hyb id i ing app oach is alida ed, u u e esea ch could aim o pe sonalize whole-b ain
models exclusi ely om EEG, signi ican ly b oadening he model’s clinical applicabili y.
The ollowing subsec ions de ail he speci ic op imiza ion echniques and algo i hms
employed o achie e his pa ame e e inemen .
❖ Mask-based Op imiza ion App oach in Sou ce Space

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
35
The i s op imiza ion s a egy implemen ed in his s udy u ilizes a mask-based app oach
wi hin he sou ce space. The gene al aim is o iden i y and e ine he exci abili y
pa ame e s (𝑊𝑒𝑥𝑐) o speci ic co ical pa cels by quan i ying di e ences be ween he
empi ical EEG da a and he syn he ic EEG signals gene a ed by he model.
S ep 1: In e se Modeling o Empi ical EEG
Ini ially, we need o ob ain sou ce-space signals om he empi ical EEG eco ded in
elec ode space. To achie e his, we employed in e se me hods, speci ically he Minimum
No m Es ima e (MNE) and S anda dized Low Resolu ion B ain Elec omagne ic
Tomog aphy (sLORETA). Bo h app oaches ely on p e iously compu ed o wa d
modeling (lead ield) solu ions and he known loca ions o elec odes and co ical pa cels
(as de ined by he Vi ual Epilep ic Pa ien , VEP, a las wi h 162 pa cels).
Minimum No m Es ima e (MNE):
MNE es ima es he sou ce ac i i y by minimizing he L2-no m o he sou ce cu en s while
accoun ing o he measu ed elec ode da a, acco ding o he equa ion:
𝑋𝑀𝑁𝐸=𝐿𝑇(𝐿𝐿𝑇+𝜆𝐶𝑛)−1𝑌 (12)
Whe e:
o 𝑋𝑀𝑁𝐸 a e he es ima ed sou ce cu en s (sou ce-space signals).
o Y is he EEG da a in elec ode space.
o L is he lead ield ma ix.
o λ is a egula iza ion pa ame e ha balances be ween da a ideli y and sou ce
complexi y.
o Cn is he noise co a iance ma ix.
S anda dized Low Resolu ion B ain Elec omagne ic Tomog aphy (sLORETA):
sLORETA is a a ia ion o MNE, which s anda dizes he cu en sou ce es ima es by
he a iance o he es ima es. This imp o es localiza ion accu acy by ensu ing
s anda dized spa ial esolu ion ac oss he co ex:
𝑋𝑠𝐿𝑂𝑅𝐸𝑇𝐴,𝑖=𝑋𝑀𝑁𝐸,𝑖
√[𝐿𝑇(𝐿𝐿𝑇+𝜆𝐶𝑛)−1𝐿]𝑖𝑖 (13)
In bo h me hods, he p e iously calcula ed lead ield ma ix L, elec ode loca ions, and
co ical pa cels om he VEP a las a e used. A e applying hese in e se me hods,
we ob ain 162 ime se ies ep esen ing neu al ac i i y o each co ical pa cel.
S ep 2: Calcula ion o he Di e ence Mask
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
36
Once empi ical EEG signals ha e been p ojec ed in o sou ce space (162 pa cels),
bo h he empi ical and syn he ic sou ce-space signals a e indi idually no malized
(scaling alues be ween 0 and 1) be o e compu ing hei di e ence. The di e ence
mask o each pa cel I is hen calcula ed using he ollowing o mula:
𝑀𝑎𝑠𝑘𝑖=||𝑋𝑒𝑚𝑝𝑖𝑟𝑖𝑐𝑎𝑙,𝑖
𝑛𝑜𝑟𝑚 − 𝑋𝑠𝑦𝑛𝑡ℎ𝑒𝑡𝑖𝑐,𝑖
𝑛𝑜𝑟𝑚 || (14)
This yields a no malized mask alue be ween 0 and 1 o each co ical pa cel:
o A mask alue close o 0 indica es s ong simila i y be ween empi ical and syn he ic
signals o ha pa cel, sugges ing ha no u he adjus men o exci abili y is
equi ed.
o Con e sely, a mask alue close o 1 highligh subs an ial di e ence, sugges ing
ha pa cel exci abili y pa ame e s (𝑊𝑒𝑥𝑐) equi e adjus men o achie e be e
model alignmen wi h empi ical da a.
S ep 3: Adjus ing Pa cel Exci abili y Pa ame e s
To adjus he exci abili y pa ame e s (𝑊𝑒𝑥𝑐) based on he mask, we apply he ollowing
o mula o each pa cel iii:
𝑊𝑖=𝑊𝑖,𝑜+𝛼·𝑀𝑎𝑠𝑘𝑖 , 𝑊𝑖,0=3 (15)
Whe e:
o 𝑊𝑖,0 ep esen s he o iginal exci abili y pa ame e alue o pa cel i, which is
ini ialized o 3 ( ep esen ing a physiologically heal hy s a e).
o 𝛼 is a global scaling pa ame e , anging be ween 0 and 7, used o modula e he
deg ee o exci abili y adjus men acco ding o he mask.
o Impo an ly, he exci abili y pa ame e s p e iously adjus ed using SEEG
eco dings a e ne e modi ied, ensu ing consis ency in a eas whe e clinical da a
al eady exis . Only pa cels no co e ed by SEEG eco dings a e subjec o his
op imiza ion p ocess.
S ep 4: I e a i e Pa ame e Op imiza ion
The op imiza ion p ocedu e sys ema ically es s mul iple α alphaα alues be ween 0
and 7. Fo each α alphaα:
i. A new exci abili y con igu a ion (𝑊𝑒𝑥𝑐) is compu ed acco ding o he mask-based
o mula abo e.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
37
ii. The new co ical exci abili y pa ame e s a e in oduced in o a new ARC
(a chi ec u e) ile, de ining he pe sonalized whole-b ain model s uc u e.
iii. The whole-b ain model is e-simula ed using he upda ed ARC con igu a ion.
i . F om he esul ing simula ion, a new syn he ic EEG is ex ac ed.
. The new syn he ic EEG is quan i a i ely compa ed wi h he empi ical EEG using
Pea son co ela ion, e alua ing how well he adjus ed model aligns wi h he eal
pa ien da a.
A e comple ing hese s eps o each α alphaα, we ob ain a quan i a i e measu e o
model pe o mance (co ela ion alues be ween syn he ic and empi ical EEGs) as a
unc ion o he scaling pa ame e α alphaα. This p oduces a co ela ion cu e
illus a ing he ela ionship be ween α alphaα alues and EEG i ing accu acy, om
which we iden i y he op imal alue o α alphaα ha maximizes simila i y be ween
simula ed and empi ical EEG signals.
The esul s o his op imiza ion p ocess, speci ically he co ela ion cu es and he
iden i ied op imal alue o α alphaα, a e p esen ed and discussed in de ail in he
Resul s sec ion o his hesis.
❖ Manual Exci abili y Adjus men Guided by Visual and Li e a u e-
Based Analysis
Ano he op imiza ion app oach explo ed in his s udy was he di ec adjus men o
co ical pa cel exci abili y (𝑊𝑒𝑥𝑐) pa ame e s guided p ima ily by isual inspec ion and
li e a u e-suppo ed analysis, independen ly om he p e iously desc ibed mask-
based app oach.
This s a egy, e med he e as "manual" adjus men , in ol ed sys ema ically inc easing
he exci abili y o speci ic co ical pa cels iden i ied as equi ing e inemen .
Impo an ly, he e m "manual" e e s speci ically o he decision-making p ocess
unde lying pa cel selec ion and he exci abili y alues assigned. While he modi ica ion
o pa ame e s was execu ed p og amma ically h ough cus om-w i en code (i.e., he
ARC iles we e no manually edi ed di ec ly), he de e mina ion o which pa cels o
modi y, and he speci ic exci abili y inc emen s applied (p ima ily es ed a ixed alues
such as 5 o 7 o p ac icali y), we e based on human judgmen a he han au oma ed
algo i hms.
The pa cel selec ion p ocess ollowed wo main c i e ia:
o Li e a u e-Based Iden i ica ion:
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
38
Pa cels associa ed wi h co ical egions iden i ied h ough a comp ehensi e
li e a u e e iew as pa icula ly ele an o seizu e p opaga ion o epilep ogenic
ac i i y.
o Visual Compa a i e Analysis:
Pa cels co esponding o co ical a eas isually iden i ied as showing subs an ial
disc epancies be ween empi ical and syn he ic EEG opog aphic maps
( opog aphic maps), sugges ing insu icien o un ealis ic simula ed neu al ac i i y
in compa ison wi h he obse ed empi ical da a.
A e iden i ying candida e pa cels h ough hese c i e ia, we p oceeded wi h he
i e a i e, code-based modi ica ion p ocedu e as ollows:
i. Pa ame e adjus men (code-based):
Exci abili y pa ame e s (𝑊𝑒𝑥𝑐) o selec ed pa cels we e sys ema ically inc eased
om he o iginal baseline o 3 o p ede e mined es ing alues ( ypically 5 o 7).
ii. Whole-b ain model e-simula ion:
A e each adjus men , a new ARC a chi ec u e ile was gene a ed
p og amma ically o e lec hese pa ame e upda es. The e ised model was
subsequen ly simula ed o gene a e a new syn he ic EEG signal.
iii. Visual and quan i a i e e alua ion:
The new syn he ic EEG signals we e compa ed o empi ical EEG da a h ough
isual inspec ion o co ical opog aphic maps and quan i a i ely assessed by
calcula ing Pea son co ela ion coe icien s. The esul s o his e alua ion we e
used o guide u he adjus men s i e a i ely.
While inhe en ly less au oma ed compa ed o ully algo i hmic op imiza ion me hods,
his app oach allowed o he u iliza ion o expe ana omical and physiological
insigh s, which can some imes be challenging o cap u e ully h ough au oma ed
echniques. I p o ided a ge ed op imiza ion ocused explici ly on isually and
clinically ele an co ical a eas, complemen ing o he au oma ed op imiza ion
s a egies employed in his wo k.
❖ Gene ic Algo i hm-Based Op imiza ion
An addi ional op imiza ion app oach explo ed in his wo k in ol es he use o gene ic
algo i hms (GAs), which a e popula ion-based op imiza ion echniques inspi ed by
e olu iona y p ocesses. The main objec i e o his GA-based op imiza ion was o
maximize he co ela ion be ween syn he ic and empi ical EEG signals wi hin he
elec ode space. This app oach is dis inc and independen om he p e iously
desc ibed mask-based and manual app oaches.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
45
A
A
B

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
46
Figu e 10. Empi ical s Syn he ic EEG compa ison. (A) EEG signals showing b ie , i egula empi ical seizu es con as ed
wi h p olonged, synch onous syn he ic seizu es. (B) En elope signals pos -p ocessing, highligh ing empi ical (le ) and
syn he ic ( igh ) di e ences in ampli ude dis ibu ion and seizu e du a ion. (C) Syn he ic signals wi hou noise.
Figu e 11. A (le ) Empi ical EEG signals (pos -p ocessing, en elope, and low-pass il e ing), highligh ing sho e , bell-
shaped seizu e pa e ns. B ( igh ) Syn he ic EEG signals p ocessed simila ly, showing p olonged pla eau-shaped seizu e
ac i i y, clea ly illus a ing he di e ence in empo al seizu e pa e ns be ween empi ical and syn he ic signals.
This disc epancy in seizu e du a ion and mo phology in oduces addi ional challenges in
subsequen co ela ion-based analyses.
EMP R CA
42
S NT ET C
A
B
C
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
47
4.2. E alua ion o Signal Fea u es and Connec i i y Me ics
4.2.1. Am li de E elo e a d Low Pass
This sec ion illus a es an example o ampli ude en elope ex ac ion using he Hilbe ans o m
and subsequen 0.1 Hz low-pass il e ing.
Figu e 12. Ampli ude En elope and Low pass in he signal il e ed be ween 4 and 10 Hz o one elec ode o Syn he ic EEG.
Figu e 13. Ampli ude En elope and Low pass in he signal il e ed be ween 4 and 10 Hz o one elec ode o Empi ical EEG
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
48
4.2.2. Connec i i y Me ics E alua ion
To e alua e connec i i y me ics, empi ical EEG da a was analyzed h ough mul iple connec i i y
measu es be o e and a e ampli ude en elope ex ac ion and low-pass il e ing.
Empi ical Da a:
Figu e 14. Empi ical EEG unc ional connec i i y ma ices a e il e ing in 4-10 Hz band and be o e en elope ex ac ion,
including c oss-co ela ion, Pea son co ela ion, cohe ence, and phase locking alue (PLV). Cohe ence isually
demons a es clea e clus e delinea ion and less sa u a ion.
Figu e 15. Empi ical EEG unc ional connec i i y ma ices a e en elope ex ac ion and low pass il e , ein o cing
cohe ence as he op imal me ic due o clea e clus e isibili y and educed noise.
Syn he ic Da a:
C oss Co ela ion Pea son Co ela ion Cohe ence P V
C oss Co ela ion Pea son Co ela ion Cohe ence P V
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
49
Figu e 16. Syn he ic EEG connec i i y ma ices (c oss-co ela ion, Pea son co ela ion, cohe ence, PLV), clea ly show
cohe ence as he mos sui able me ic due o less sa u a ion and clea e clus e delinea ion. Cohe ence pa icula ly
highligh s occipi al clus e s ma ching clinical epilep ogenic a eas.
In luence o Noise in Syn he ic EEG:
Figu e 17. A (Le ): Compa ison o syn he ic EEG connec i i y ma ices (c oss-co ela ion) be o e and a e adding noise.
Signi ican imp o emen a e noise addi ion, educing un ealis ic high synch ony ac oss channels. B (Righ ): G aphical
illus a ion o high- and low-ampli ude syn he ic EEG channels be o e(blue) and a e noise addi ion ( ed), showing noise
p ese ing epilep ogenic high-ampli ude ac i i y while e ec i ely educing synch ony in low-ampli ude signals.
C oss Co ela ion Pea son Co ela ion Cohe ence P V
Wi h noiseWi h noise
Wi h
noise
Wi h noise Wi h noise
Solu ion Add Noise
Reduce co ela ion wi h he
low-ampli ude nodes
main ains he co ela ion wi h
he high-ampli ude nodes
A
B
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
50
4.2.3. Topog aphic maps
Figu e 18. The spa ial dis ibu ion o co ical b ain ac i i y du ing seizu e p opaga ion is shown o bo h empi ical and
syn he ic cases. Di e ences in seizu e du a ion and localiza ion a e obse ed.
In he igu e abo e, he empo al e olu ion o he spa ial dis ibu ion o he signal is p esen ed
using opog aphic maps, along wi h he co esponding syn he ic o empi ical EEG signals. Fo
he empi ical signals, due o he sho e du a ion o he seizu e, i can be obse ed how he ocus
o co ical ac i i y appea s and subsequen ly disappea s. In con as , o he syn he ic signals, he
ac i i y eme ges bu does no disappea wi hin he analyzed ime window. Fu he mo e, in he
syn he ic signals, he ocus o ac i i y emains localized mainly in he le occipi al egion, whe eas
in he empi ical signals, in addi ion o his occipi al ac i i y, a seconda y high-ampli ude ac i a ion
can also be iden i ied in he cen al co ical egions.
4.2.4. F equency Band Selec ion and Spec al Analysis
A c ucial componen o EEG analysis is he iden i ica ion and in e p e a ion o speci ic equency
bands, as di e en neu al p ocesses—and pa icula ly epilep i o m dynamics—a e p e e en ially
exp essed in dis inc spec al anges. Fo his s udy, he equency band om 4 o 10 Hz (ex ended
he a band) was selec ed o signal p ocessing and analysis, a choice suppo ed by
complemen a y app oaches.
Li e a u e e iew highligh s speci ic EEG equency bands ela ed o epilep ic dynamics; del a
band (1–4 Hz) is o en associa ed wi h widesp ead seizu e p opaga ion, while he he a band (4–
8 Hz) is p edominan ly linked o ocal seizu e ac i i y. Gamma band ac i i y (>30 Hz), in con as ,
is ypically associa ed wi h ocal epilep ogenic zone localiza ion a he han p opaga ion.
Conside ing he s udy's ocus on seizu e p opaga ion a he han solely localiza ion, he li e a u e
ecommends selec ing equencies be ween del a and he a bands.
The decision o selec he he a band (4–10 Hz) was also s ongly suppo ed by di ec spec al
analysis. Compu ing he Powe Spec al Densi y (PSD) o he EEG signals consis en ly e ealed
a p ominen peak wi hin he he a equency ange, pa icula ly in channels loca ed o e
55
Seizu e Window
The a Band 4 10 z
Ampli ude En elope ilbe 0.1 z low pass l e
35
EMP R CA EEG S NT ET C EEG

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
51
epilep ogenic a eas. This empi ical inding con i med ha he maximum powe o epilep i o m
ac i i y lies p ecisely in he selec ed equency band.
Toge he , he suppo om exis ing li e a u e on equency-speci ic seizu e dynamics and he
empi ical e idence om PSD analysis s ongly jus i ied ou choice o wo king wi hin he 4–10 Hz
equency ange o all subsequen p ocessing and connec i i y analyses
Figu e 19. Spec al analysis suppo ing equency band selec ion. (A) Powe spec al densi y plo highligh ing highes
powe wi hin he 4–10 Hz band in key elec odes. (B) Ana omical ep esen a ion om he Vi ual Epilep ic Pa ien (VEP)
a las showing epilep ogenic pa cels in ed and p opaga ion pa cels in yellow, co esponding o egions wi h signi ican
spec al powe .
A
B
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
52
Figu e 20. Powe Spec al Densi y (PSD) analysis o EEG signals om an elec ode placed o e he epilep ogenic egion. A
clea peak in signal powe is obse ed wi hin he 4–10 Hz (ex ended he a) equency band, con i ming i s clinical
ele ance and sui abili y o u he analysis.
4.3 Compa a i e Analysis be ween Syn he ic and Empi ical EEG
The ollowing sec ion p esen s an analysis o he me ics de ailed abo e.
Fi s , he wo co ela ion ma ices ep esen ing unc ional connec i i y a e compa ed. Figu e 21
displays he cohe ence me ics o he p ocessed EEG da a, which has been il e ed, subjec ed
o Hilbe ans o m, and had i s ampli ude en elope ex ac ed and il e ed a 0.1 Hz.
Figu e 21. Empi ical and Syn he ic Func ional Connec i i y Ma ices using Cohe ence.
Bo h ma ices exhibi highly simila geome ies and s uc u es. They appea o show he same
clus e s, pa icula ly he one loca ed in he bo om igh , associa ed wi h occipi al and pa ie al
egions. This is consis en wi h he clinical in o ma ion and he spa ial dis ibu ion ep esen a ion
shown in he opog aphic map, indica ing ha his is he a ea whe e he epilep ic seizu e o igina es
and p opaga es. The empi ical ma ix also e eals ano he clus e associa ed wi h elec odes in
44
Syn he ic Da a
Empi ical Da a
The isual inspec ion o co ela ion ma ices con i ms pa e n simila i y
Clus e shi s in eal s. syn he ic da a.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
53
he cen al co ex, which aligns pe ec ly wi h he second a ea o high ac i i y depic ed in he
empi ical opog aphic map. Howe e , simila o he opog aphic map, his egion whe e he
epilep ic ac i i y seems o be p opaga ing is absen in he syn he ic ma ix.
Ano he no able obse a ion in he ma ices is ha in he case o syn he ic signals, mos , i no all,
a e highly synch onized. Consequen ly, signals wi h su icien ly high ampli udes o o e come
added noise will show high co ela ions in he ma ix, while hose wi h low ampli udes will be
signi ican ly a ec ed by noise and exhibi low co ela ion. This explains why occipi al elec odes
and hose hea ily in luenced by seizu e p opaga ion egions appea ed in he syn he ic ma ix.
This is no he case in he empi ical da a, whe e signals a e he e ogeneous and asynch onous,
especially hose om channels loca ed nea he epilep ic ocus. Du ing a seizu e, an abno mal
synch oniza ion o a la ge numbe o neu ons occu s, esul ing in excessi e hy hmic ac i i y.
Howe e , his ac i i y may ha e an ini ial desynch oniza ion in ce ain b ain a eas, and in addi ion
o hy hmic ac i i y, spikes (sha p wa es) and slow wa es can be obse ed. In he empi ical
ma ix, he O1 channel, which shows high co ela ion in he syn he ic ma ix, exhibi s p ac ically
ze o co ela ion. This is because slow wa es a e p edominan ly obse ed in his channel. This
signi ican di e ence in he co ela ion ma ix o elec odes di ec ly associa ed wi h he epilep ic
ocus sugges s ha i ing he model using unc ional connec i i y will be challenging.
Following he isual analysis o he ma ices, a quan i a i e analysis using Pea son's Co ela ion
Coe icien (PCC) was pe o med. The PCC alue be ween he syn he ic and empi ical ma ices
is app oxima ely 0.1. Conside ing ha 0 ep esen s no co ela ion and 1 ep esen s maximum
co ela ion, 0.1 is a e y low alue. The di e ences be ween he wo ma ices, which could
con ibu e o his low alue, ha e been p e iously men ioned. Ano he impo an di e ence is ha
clus e s in he syn he ic ma ix appea mo e smoo hed, indica ing a mo e uni o m dis ibu ion o
ne wo ks o egions wi h highe o lowe connec i i y. In con as , he empi ical ma ix is mo e
he e ogeneous, appea ing noisie , wi hou a la ge ed egion bu a he a mix u e o ed and whi e
a eas. To educe he esolu ion o he empi ical ma ix, di e en echniques such as a Gaussian
il e , node clus e ing echniques, and esolu ion educ ion wi h eigen ec o s we e employed.
These echniques imp o ed he PCC, al hough no subs an ially.
As explained in p e ious sec ions, calcula ing he PCC o a single subjec is insu icien o
de e mine he signi icance o he alue. The e o e, a b ain model o ano he pa ien wi h a di e en
a chi ec u e (ARC ile) was simula ed. The co ical simula ion and ac i i y in he sou ce space
we e gene a ed, and he syn he ic EEG was compu ed. The unc ional co ela ion ma ix was
calcula ed ollowing he same s eps as wi h he o iginal subjec , and he Pea son coe icien was
compu ed be ween his ma ix and he empi ical ma ix o he o iginal subjec . In his case, he
alue was signi ican ly lowe , a ound 0.04, indica ing ha he p e iously ob ained alue o 0.1 is
no as low as ini ially pe cei ed. The same p ocedu e was applied o a comple ely andom ma ix,
yielding a Pea son alue close o ze o, as expec ed. Figu e 22 allows o a isual compa ison o
he unc ional connec i i y co ela ion ma ices using cohe ence and he opog aphic maps o he
empi ical EEG o he o iginal subjec , he syn he ic EEG o he o iginal subjec , and he syn he ic
EEG o he new subjec . This isual analysis e eals ha he unc ional connec i i y ha migh
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
54
coincide be ween bo h subjec s is due o some ac i i y in he igh empo al pa in he empi ical
da a, which is he egion whe e he epilep ic ocus is in he o he subjec .
Figu e 22. Co ela ion ma ices using Cohe ence and opog aphic maps o empi ical da a and syn he ic da a o bo h
subjec s.
Quan i a i e co ela ion was no only calcula ed be ween he ma ices bu also ex ac ed as he
Pea son coe icien be ween he ec o s con aining he means o each channel, which is wha is
isualized in he opog aphic map. The PCC be ween ma ices measu es he simila i y be ween
unc ional connec i i y pa e ns, while he PCC be ween opog aphic maps measu es he
simila i y be ween he dis ibu ion o ampli udes in he EEG.
The PCC alues be ween he mean ec o s a e signi ican ly highe han hose be ween he
ma ices, which is easonable gi en ha hese a e signals in he elec ode space.
Table 1 shows he PCC esul s in bo h cases:
SUBJECT 1
SUBJECT 2
Mean Ampli ude
0.6438
- 0.0375
FC Ma ix
0.1062
0.041
Table 1. Compa ison Resul s be ween syn he ic mean ampli ude ec o s o bo h subjec s wi h he empi ical ec o and
esul s o PCC wi h FC ma ices.
The able indica es ha he bes me ic o compa ing bo h syn he ic and empi ical da a is he
ampli ude o he signals hemsel es, a he han unc ional connec i i y. Be o e de ini i ely
selec ing opog aphic maps as he me ic o i he model, o he echniques besides PCC we e
Syn he ic New Subjec 2 Empi ical iginal Subjec
EEG Real
WBM Dipole ac i i y Fo wa d Me hod EEG Syn he ic
E 2 E 1
- 0.03750.6438
Mean
Ampli ude
CAR
0.0410.1062FC Ma ix
0.02120.7214
Mean
Ampli ude
C
0.5360.248FC Ma ix
Syn he ic iginal Subjec 1
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
61
model's abili y o eplica e seizu e dynamics a a global le el. Addi ionally, di e ences we e
iden i ied be ween he syn he ic and empi ical signals ha posed ini ial challenges, such as
disc epancies in seizu e du a ion, signal mo phology, and synch oniza ion be ween channels.
Howe e , i was ound ha app op ia e p ocessing, including he ex ac ion o ele an ea u es
and il e ing in speci ic equency bands (such as he he a band), allowed hese di e ences o be
managed o subsequen analyses.
The nex c i ical s ep in he me hodology was he compa ison be ween he empi ical and syn he ic
EEG da a. This p ocess was app oached h ough wo main a enues. The i s in ol ed using
co ela ion ma ices o e alua e unc ional connec i i y, in his case, be ween EEG elec odes. A
signi ican challenge he e lies in analysing unc ional connec i i y in elec ode space, whe e
phenomena such as olume conduc ion and noise a ec he signal. This implies ha he
unc ional connec i i y ma ices ob ained in elec ode space a e no as op imal o di ec ly
in e p e able as hose ob ained in sou ce space. Despi e his limi a ion, he isualized ma ices
showed no able quali a i e ag eemen , eplica ing he clus e and ne wo k s uc u e obse ed in
he empi ical da a. Howe e , quan i a i ely, he co ela ion alues ob ained be ween he ma ices
we e no su icien ly high o allow o a obus model i ing based on his me ic in elec ode space.
The second a enue o compa ison was he use o he spa ial dis ibu ion o signals, ep esen ed
by opog aphic maps. This app oach p o ed o be much mo e p omising. Quali a i ely, he
syn he ic opog aphic maps gene a ed om he model we e able o con incingly eplica e he
gene a ion o he epilep ic seizu e and i s ini ial p opaga ion h ough he occipi al a ea, in
ag eemen wi h he clinical and SEEG elec ode localiza ion. Howe e , isual compa ison also
e ealed a signi ican disc epancy: while he empi ical EEG showed high-ampli ude ac i i y in a
cen al co ical egion (sugges ing p opaga ion o his a ea), his ac i i y was absen in he
syn he ic opog aphic map. Quan i a i ely, he Pea son co ela ion coe icien be ween he
empi ical and syn he ic opog aphic maps yielded e y good alues, con i ming ha his me ic
was sui able o e alua ing spa ial simila i y. Gi en he limi a ions encoun e ed wi h unc ional
connec i i y ma ices and he a ou able esul s ob ained wi h opog aphic maps, i was decided
ha he compa ison o spa ial dis ibu ions would be he p ima y me ic o guide pa ame e i ing.
The absence o simula ed ac i i y in he cen al co ical egion ha was no p e iously i ed wi h
SEEG led us o p opose a c ucial in e media e s ep be o e add essing he i ing o a comple ely
new model wi h only EEG: e i ing he exis ing model (p e iously i ed wi h SEEG) using scalp
EEG da a. The objec i e was o adjus he co ical pa cels ha had no been co e ed by SEEG
elec odes so ha he model could eplica e he p opaga ion o epilep ic ac i i y h oughou he
en i e co ex, esul ing in a ully i ed model. Once his is achie ed, he s a egy o i ing a model
om sc a ch using only EEG would be simila .
Fo pa ame e i ing, di e en echniques we e explo ed. An ini ial p oposal in ol ed using a mask
de i ed om he compa ison in sou ce space a e applying an in e se me hod o he empi ical
EEG. Al hough he in e se me hod allowed ob aining signals in sou ce space whe e epilep ic oci
and p opaga ion zones could be dis inguished, he sub le di e ences in shape and ampli ude

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
62
be ween he empi ical and syn he ic signals in his space we e no clea enough o allow o an
e ec i e model i ing using his mask, and he esul s we e no sa is ac o y. Ne e heless, his
app oach is conside ed a p omising line o wo k o u u e p ojec s, sugges ing he need o
op imiza ions in he in e se me hod o speci ic pos -p ocessing o he signals in sou ce space.
The main me hod and inally chosen o pa ame e i ing was he gene ic algo i hm, using he
co ela ion in he spa ial dis ibu ion ( opog aphic maps) as he objec i e unc ion. This app oach
yielded p omising esul s, enabling he model o gene a e simula ed ac i i y in he cen al co ical
a ea ha did no appea be o e. Howe e , a signi ican limi a ion ha a ec ed he comple e
de elopmen o his s age was he high compu a ional and empo al cos o he gene ic algo i hm.
This limi ed he numbe o simula ion gene a ions and he explo a ion o a wide ange o possible
exci abili y alues, p e en ing he comple e op imiza ion o he p ocess wi hin he p ojec imeline.
Al hough he ac i i y in he a ge egion was success ully eplica ed, some ac i i y, albei o lowe
ampli ude, was also obse ed in o he unexpec ed egions, indica ing he need o u he
e inemen o he i ing.
Summa izing he gene al conclusions, he capabili y o he p e- i ed b ain model wi h SEEG o
simula e seizu e gene a ion and ini ial p opaga ion is alida ed, as e idenced quali a i ely in he
connec i i y ma ices and mo e clea ly in he opog aphic maps. Func ional connec i i y ma ices
in elec ode space, al hough quali a i ely consis en , we e no op imal o quan i a i e i ing due
o he in luences o elec ode space. On he o he hand, opog aphic maps p o ed o be a obus
me ic bo h quali a i ely and quan i a i ely o compa ison and i ing. The s a egy o e i ing he
a eas no co e ed by SEEG using EEG and a gene ic algo i hm guided by opog aphic maps
p o ed o be a iable and p omising in e media e s ep owa ds achie ing a ully i ed model.
Despi e he compu a ional limi a ions ha p e en ed comple e op imiza ion a his s age, he
p elimina y esul s wi h he gene ic algo i hm a e encou aging, demons a ing he abili y o
gene a e he expec ed ac i i y in p e iously un i ed a eas. As his p ojec con inues beyond he
TFM submission, he nex s eps will ocus on pe ec ing he gene ic algo i hm, possibly explo ing
o he me ics and add essing he compu a ional limi a ion o achie e a mo e p ecise and speci ic
i ing. Once he model is success ully e i ed using EEG, he ul ima e goal will be o add ess he
i ing o a b ain model om sc a ch, based exclusi ely on EEG da a, which would consolida e he
p oposed non-in asi e me hodology. The possibili y o explo ing in e se modelling echniques
and unc ional connec i i y ma ix-based i ing in sou ce space in g ea e dep h in he u u e is
also conside ed, gi en hei in insic po en ial. Expanding he s udy o mul iple subjec s will be
c ucial o alida e he obus ness and gene ali y o he p oposed pipeline.
6. CONCLUSSIONS
This hesis success ully ad anced he pe sonaliza ion o whole-b ain models o epilepsy by
de eloping and alida ing a me hodology based on non-in asi e scalp EEG da a. This wo k
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
63
ep esen s a signi ican s ep owa ds o e coming he limi a ions o in asi e SEEG-based
app oaches, enhancing clinical accessibili y and he po en ial o model seizu e dynamics ac oss
he en i e co ex.
A key achie emen was he iden i ica ion o he ampli ude en elope, ex ac ed ia he Hilbe
ans o m wi hin he he a band (4-10 Hz), as a obus and clinically ele an me ic o compa ing
empi ical and syn he ic EEG. While ini ial explo a ions wi h unc ional connec i i y me ics aced
challenges in elec ode space due o olume conduc ion, opog aphic maps de i ed om
ampli ude en elopes demons a ed s ong spa ial co espondence and p o ed e ec i e o
guiding model i ing in his s udy.
The in es iga ion in o pa ame e op imiza ion echniques highligh ed he gene ic algo i hm as he
mos p omising app oach. Despi e compu a ional cons ain s limi ing i s ull explo a ion,
p elimina y esul s success ully demons a ed he algo i hm's abili y o induce simula ed epilep ic
ac i i y p opaga ion in cen al co ical egions no co e ed by he ini ial SEEG-based i ing,
aligning wi h empi ical obse a ions and add essing a c i ical limi a ion o he p e ious modeling
amewo k.
This esea ch p o ides a solid ounda ion o ully non-in asi e EEG-based pe sonaliza ion o
whole-b ain models. The me hodological amewo k de eloped has b oade clinical implica ions,
pa icula ly o pa ien s who canno unde go in asi e moni o ing.
Fu u e esea ch will build upon hese indings by ocusing on e ining he gene ic algo i hm
h ough inc eased compu a ional esou ces and explo ing al e na i e op imiza ion s a egies.
Expanding he pipeline's alida ion o mul iple pa ien da ase s is c ucial o assess i s obus ness
and gene alizabili y. The ul ima e goal emains he de elopmen o a eliable me hodology o
pe sonalizing whole-b ain models using exclusi ely non-in asi e EEG da a, pa ing he way o
imp o ed p esu gical e alua ion and pe sonalized he apeu ic in e en ions in epilepsy.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
64
7. PLANNING
The imeline s uc u e was di ided in o se e al key phases:
Oc obe – No embe 2024: Li e a u e e iew and amilia iza ion wi h pa ien da a, code
eposi o ies, and b ain model a chi ec u e.
Decembe 2024 – Janua y 2025: P ep ocessing o empi ical EEG and gene a ion o syn he ic
EEG using o wa d modelling.
Janua y – Feb ua y 2025: Compa ison o empi ical and syn he ic da a using unc ional
connec i i y me ics in elec ode space.
Feb ua y – Ma ch 2025: Compa ison based on opog aphic ampli ude dis ibu ions ( he a band),
con i ming hei sui abili y o model i ing.
Ma ch – Ap il 2025: Pa ame e op imiza ion using gene ic algo i hms, explo ing di e en
s a egies and consolida ing esul s.
May 2025: Ongoing wo k on model pe sonaliza ion and algo i hm e inemen .
June 2025: Applica ion o he pipeline o non-SEEG- i ed models, explo ing ull EEG-only
pe sonaliza ion.
The ollowing igu e illus a es his imeline isually:
Figu e 27 P ojec Timeline
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
65
This s uc u e allowed o a me hodical app oach o he a ious aspec s o he p ojec , om
backg ound e iew o implemen a ion and e alua ion o he p oposed me hodologies, es ablishing
a s ong basis o u u e esea ch.
8. ECONOMIC ASSESSMENT
This sec ion p o ides a gene al assessmen o he cos s in ol ed in de eloping his p ojec as i i
we e o be ou sou ced o eplica ed as a se ice. The analysis includes labou cos s so wa e ools
compu a ional esou ces and in as uc u e usage.
1. Pe so el os
The p ojec was ca ied ou by a biomedical enginee ing in e n o e a pe iod o app oxima ely 7
mo hs c obe 2024 – May 2025 wi h an a e age wo load o 20 ho s e week. Al hough he
in e nship compensa ion was €10/ho which is ypical o a s uden placemen a mo e ealis ic
a e o a p o essional biomedical enginee migh be highe depending on he expe ise and loca ion.
owe e o his es ima ion we will s ic o he ac ual in e n a e
• Ho s wo ked 20 hou s/wee × 28 wee s 560 hou s
• Ho ly a e €10/hou
• o al e so el cos €5,600
2. o wa e a d ools
• Py ho o ammi la a e pen-sou ce and ee → €0
• MNE-Py ho , N mPy, ciPy, a d c s om Ne oelec ics ools All open-sou ce o
in e nally de eloped → €0
• Ne oelec ics modeli la o m n e nal ools used unde he in e nship ag eemen →
€0
3. I as c e a d Elec ici y
• Pe so al com e sa e a e age powe consump ion o 60W used 4 hou s/day 5
days/wee o e 28 wee s
→ App ox. 33.6 k h wi h an es ima ed cos o €0.20/k h in Spain
o Elec ici y cos 33.6 Wh × €0.20 ≈ €6.72
• Ha dwa e de ecia io PC use o e 7 mon hs assumed minimal o sho - e m academic
usage → Ne li ible
4. P e-exis i I as c e a d Pi eli es
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
66
• The whole-b ain modeling pipeline in e se and o wa d modeling ools and pa ien -speci c
da a MR dMR SEEG EEG we e al eady a ailable wi hin he Neu oelec ics ecosys em.
These a e conside ed p e-exis ing company asse s and we e accessed a no addi ional cos
du ing he in e nship.
o al Es ima ed os :
a e o y
os (E R)
Pe sonnel n e nship
€5 600
So wa e Tools
€0
n as uc u e & Elec ici y
€7
P e-exis ing Pipelines & Da a
€0
Da a
€5 606.72
Table 4. Economic assessmen able.
9. ENVIRONMENTAL ASSESSMENT
Al hough he execu ion o his p ojec did no in ol e he use o physical ma e ials o p oduce
was e in he adi ional sense, i did equi e conside able use o compu a ional esou ces, which
inhe en ly en ails elec ici y consump ion and hus ca bon emissions.
Elec ici y Consump ion
Th oughou he de elopmen o his p ojec , a pe sonal compu e was used app oxima ely 4 hou s
pe day, 5 days a week, o e he cou se o 7 mon hs. Assuming a conse a i e a e age powe
usage o 60W, he es ima ed o al ene gy consump ion is as ollows:
• Daily usage: 4 hou s × 60 W = 0.24 kWh/day
• Weekly usage: 0.24 kWh/day × 5 days = 1.2 kWh/week
• To al usage o e 28 weeks: 1.2 kWh/week × 28 weeks ≈ 33.6 kWh
Using he Spanish go e nmen calcula o o ca bon emissions based on na ional ene gy
p oduc ion, his ansla es o an app oxima e emission o :
• 33.6 kWh → ~6.8 kg CO₂ emi ed
This emission is ela i ely low, and no addi ional a el o ma e ial esou ces we e equi ed,
minimizing he o e all en i onmen al oo p in .
T anspo a ion Emissions (Commu ing)
The p ojec in ol ed egula commu ing o he wo kplace in Ba celona om Oc obe 2024 o Ap il
2025. Commu ing p ima ily in ol ed public bus anspo a ion, wi h an es ima ed a el ime o 40
minu es pe one-way ip and an app oxima e dis ance o 6 km pe ip. Based on a ypical wo k
pa e n, bus a el occu ed app oxima ely ou days pe week o he ou bound jou ney and h ee
days pe week o he e u n jou ney (wi h one o wo e u n jou neys pe week being made by

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
67
walking). Conside ing he p ojec du a ion (app oxima ely 7 mon hs o 28 weeks, accoun ing o
holidays), he es ima ed numbe o bus ips is:
• Es ima ed wo king weeks: ∼24 weeks (7 mon hs excluding holidays/b eaks)
• Bus ips pe week: 4 (ou bound) + 3 ( e u n) = 7 ips
• To al es ima ed bus ips: 7 ips/week × 24 weeks = 168 ips
• Es ima ed dis ance pe ip: ∼6 km
• To al es ima ed dis ance by bus: 168 ips × 6 km/ ip = 1008 km
Using a ep esen a i e emission ac o o u ban public bus anspo , es ima ed a a ound 80
g ams o CO₂ pe passenge -kilome e (0.08 kg CO₂/pkm) based on da a om ele an
en i onmen al and anspo a ion au ho i ies in Spain, he o al es ima ed ca bon emissions om
bus a el a e:
• To al CO₂ emissions om bus: 1008 km × 0.08 kg CO₂/km ≈ 80.6 kg CO₂
This analysis indica es ha anspo a ion, e en by public bus, ep esen s a mo e signi ican
po ion o he p ojec 's di ec ca bon oo p in compa ed o he ene gy consump ion om he
pe sonal compu e used o compu a ional asks, al hough he o al emission om commu ing is
lowe han ini ially es ima ed wi h he p e ious dis ance assump ion.
Use o A i icial In elligence and La ge-Scale Compu ing
n addi ion o pe sonal compu ing he p ojec also in ol ed he use o a i cial in elligence
sys ems. Speci cally gene a i e language models we e selec ed o summa isa ion and
code debugging. The use was concen a ed on
• Cha GPT GPT-4/ o-se ies ≈ 100 que ies.
Pee - e iewed and indus y s udies epo ha a single Cha GPT-class que y emi s
anywhe e be ween 0.382 g C ₂e lowe -bound empi ical es ima e and 4.32 g C ₂e uppe -
bound li ecycle es ima e [22].
Mul iplying hose ac o s by he numbe o p omp s gi es a conse a i e in e al
• owe bounda y 0.382 g 100 × 0.382 g ≈ 0.0382 g C ₂e
• ppe bounda y 4.32 g 100 × 4.32 g ≈ 0.432 g C ₂e
This 0.0382 –0.432 g ange is in en ionally wide ac ual emissions a y wi h he model
e sion da a-cen e e iciency and he g id’s ca bon mix a he exac ime o each que y.
owe -bound s udies amo ise model aining ac oss billions o que ies while uppe -bound
blog es ima es assume a ossil-hea y powe supply—hence he sp ead.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
68
E en a he op end he hesis’s oo p in is oughly ha o a 2 m ip by ca . oo ing
ahead A wo loads may double da a-cen e elec ici y demand by 2030 bu hype scale s
a e also amping up enewables and ad anced cooling. Con inued mind ul use o Ms
will eep u u e esea ch oo p in s modes .
10. SOCIAL AND GENDER EQUALITY ASSESSMENT
This p ojec was de eloped using highly sensi i e clinical da a om a single pa ien , whose gende
iden i y is wi hheld o p ese e p i acy. Ne e heless, he me hodological app oach —based on
pe sonalized b ain modelling om EEG and SEEG da a— is en i ely neu al and does no
inco po a e any biases ela ed o gende , ace, cul u e, o socioeconomic s a us.
The algo i hms and pipelines designed can be equally applied o any pa ien , ega dless o hei
gende iden i y o social backg ound. Mo eo e , he use o non-in asi e EEG echnology
inc eases accessibili y o ulne able popula ions, including pa ien s who, due o medical o
economic easons, canno unde go in asi e p ocedu es such as SEEG.
The de elopmen o his me hodology di ec ly suppo s he Sus ainable De elopmen Goals
(SDGs):
SDG 3: Good Heal h and Well-Being, by p omo ing mo e p ecise, pe sonalized, and less
in asi e ea men s o people li ing wi h epilepsy.
SDG 10: Reduced Inequali ies, by acili a ing access o ad anced diagnos ic and ea men
echnologies o adi ionally unde se ed popula ions.
In summa y, his wo k has been designed om an inclusi e pe spec i e, ensu ing ha scien i ic
ad ances can bene i all indi iduals ai ly and wi hou disc imina ion.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
69
BIBLIOGRAPHY
[1]
[Online]. A ailable: h ps://www.gal ani-lab.eu/.
[2]
Wendling, F., Ru ini, G., & Ba olomei, F., “GALVANI: Con olling epilep ic b ain ne wo ks
wi h compu a ionally op imized weak elec ic ields [Resea ch p oposal, Pa B1].,” ERC
Syne gy G an , Inse m, Neu oelec ics, Aix Ma seille Uni e si é, F ancia, España y F ancia,
2018.
[3]
]. J. S. Gomes, “T ansc anial cu en s imula ion in epilepsy: A sys ema ic e iew o he
undamen al and clinical aspec s,” F on ie s in Neu oscience, ol. 16, no. 909421, 2022.
[4]
M. A. Lopes, “Compu a ional Modelling in Sou ce Space om Scalp EEG o In o m
P esu gical E alua ion o Epilepsy,” Clinical Neu ophysiology, 2020.
[5]
P. . Mie lo, “Func ional B ain Connec i i y om EEG in Epilepsy: Seizu e P edic ion and
Epilep ogenic Focus Localiza ion,” P og ess in Neu obiology, 2014.
[6]
J. Cases Gend a, “A No el Biophysical Whole-B ain Model Explains Powe Spec um
Al e a ions Unde Se o one gic Psychedelics,” 2024.
[7]
E. L.-S. Bo ja Me cadal, “Towa ds a mesoscale physical modeling amewo k o
s e eo ac ic-EEG eco dings,” 2023.
[8]
,. G. R. E. L.-S. F. Jan C. Gend a, “A No el Biophysical Whole-B ain Model Explains Powe
Spec um Al e a ions Unde Se o one gic Psychedelics Using Mul imodal Neu oimaging”.
[9]
J. Wi sich, “Concu en EEG- and MRI-de i ed unc ional connec omes exhibi linked
dynamics,” 2020.
[10]
E. Lopez-Sola, B. Me cadal, E. Lleal, R. Sal ado , F. Ba olomei, F. Wendling and G. Ru ini,
“Pe sonalized Whole-B ain Models o Seizu e P opaga ion,” bio Xi (p ep in ), 2025.
[11]
B. H. a. R. V. G. Jansen, “Elec oencephalog am and isual e oked,” Biological Cybe ne ics
, ol. 73, p. 357–366, 1995.
[12]
E. Lopez-Sola, “A pe sonalizable au onomous neu al mass model o epilep ic seizu es,”
2022.
[13]
M. Hashemi, “The Bayesian Vi ual Epilep ic Pa ien : A p obabilis ic amewo k,”
Neu oImage 217, 2020.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
70
[14]
V. K. Ji sa, “The i ual epilep ic pa ien : indi idualized whole-b ain models o epilepsy
sp ead,” Neu oimage 145 , pp. 377 - 388, 2017.
[15]
H. E. Wang, Jou nal o neu oscience me hods 348, 2021.
[16]
M. Good ellow, “Es ima ion o b ain ne wo k ic ogenici y p edic s ou come om epilepsy
su ge y,” Scien i ic epo s 6.1, 2016.
[17]
R. Sanchez-Todo, “A physical neu al mass model amewo k o he analysis o oscilla o y
gene a o s om lamina elec ophysiological eco dings,” Neu oImage 270, 2023.
[18]
H. E. Wang, “VEP a las: An ana omic and unc ional human b ain a las dedica ed o epilepsy
pa ien s,” J. Neu osci. Me hods, ol. 348, 2021.
[19]
D. M. S. S.-T. J. Casas, “EEG-in o med pe sonaliza ion o whole-b ain models in ocal
epilepsy,” 2024.
[20]
G. Ru ini, “EEG Fo wa d Modeling and Sou ce Localiza ion: Theo y and Applica ions,”
2015.
[21]
A. S.-L. G. D. S. Sánchez-Todo, “Whole-b ain modeling o seizu e p opaga ion using ealis ic
connec i i y and hyb id models,” F on ie s in Compu a ional Neu oscience, ol. 17, 2023.
[22]
D. &. S. W. Lehmann, “Re e ence- ee iden i ica ion o componen s o checke boa d-e oked
po en ial ields,” Elec oencephalog aphy and Clinical Neu ophysiology, 48(6), 609–621.,
1980.
[23]
Z. B. A. C. S. H. R. &. S. E. P. Wang, “ Image quali y assessmen : F om e o isibili y o
s uc u al simila i y.,” IEEE T ansac ions on Image P ocessing, 13(4), 600–612., 2004.
[24]
E. Lopez-Sola, “Pe sonalized Whole-B ain Models o Seizu e P opaga ion,” Ba celona,
2025.
[25]
J. H. Holland, “Adap a ion in Na u al and A i icial Sys ems,” Camb idge, MA, USA: MIT
P ess, 1992.
[26]
D. E. Goldbe g, “Gene ic Algo i hms in Sea ch, Op imiza ion, and Machine Lea ning,” 1989.
[27]
M. Mi chell, “An In oduc ion o Gene ic Algo i hms,” Camb idge, MA, USA: MIT P ess, 1998.
[28]
M. Good ellow, “Es ima ion o b ain ne wo k ic ogenici y p edic s ou come om epilepsy
su ge y,” Scien i ic Repo s, ol. 6, 2016.
[29]
S. M. Villalon, “EpiTools, A so wa e sui e o p esu gical b ain mapping in epilepsy:
In ace eb al EEG,” J. Neu osci. Me hods, 2018.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
77
The ollowing gu es show b ain ac i i y be o e and a e he seizu e.
Figu e F2. The uppe igu e shows he ac i i y be o e he seizu e and he lowe igu e shows he ac i i y du ing he seizu e.

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
78
The ollowing gu es show he signal a e p ocessing
Figu e F3. On he le is shown he il e ed signal om 4 o 10 Hz in blue and he Hilbe en elope ampli ude in ed, he
igu e on he igh shows a 30-second ime window cen e ed on he seizu e onse and il e ed a 0.1 Hz (dashed ed line).
Appendix G. Gene ic Algo i hm Pa ame e s
This Appendix p o ides a concise, ye de ailed o e iew o he gene ic algo i hm (GA) design
used o op imizing co ical exci abili y pa ame e s (W) in he pe sonalized whole-b ain model.
O e iew o Gene ic Algo i hm
Gene ic algo i hms a e op imiza ion echniques inspi ed by biological e olu ion p inciples [25] [26].
They in ol e i e a i ely e ining a se o candida e solu ions (popula ion) h ough ope a ions like
selec ion, c osso e ( ecombina ion), and mu a ion o con e ge owa d an op imal o nea -op imal
solu ion.
In he con ex o his s udy, he GA aimed o maximize he spa ial co ela ion (Pea son co ela ion
coe icien calcula ed om opog aphic images) be ween syn he ic EEG opog aphic maps and
empi ical EEG maps, by adjus ing co ical exci abili y pa ame e s in selec ed b ain pa cels [27].
GA Pa ame e s and Con igu a ion
The able below summa izes he pa ame e s, and hei ini ial alues used in his p ojec :
PARAMETER
VA E
Popula ion Size
20
Mu a ion Ra e
0.2
C osso e
0.5
Numbe o Gene a ion
10
Table 5 Gene ic Algo i hm Design
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
79
Popula ion Size e e s o he numbe o candida e solu ions gene a ed and e alua ed a each
gene a ion. A la ge popula ion size allows g ea e explo a ion o he pa ame e space bu
inc eases compu a ional demand. Mu a ion Ra e ep esen s he p obabili y o andomly
al e ing elemen s wi hin each candida e solu ion. Mu a ion in oduces a iabili y and helps
he algo i hm explo e di e se a eas o he pa ame e space p e en ing p ema u e
con e gence. C osso e Ra e de e mines he p obabili y o combining elemen s om wo
pa en solu ions o p oduce o sp ing [26]. A highe c osso e a e encou ages mixing o
p omising ai s acili a ing con e gence owa d op imal solu ions. Numbe o Gene a ions
indica es how many i e a i e cycles o selec ion c osso e and mu a ion a e pe o med.
Mo e gene a ions gene ally esul in imp o ed solu ions as he algo i hm i e a i ely e nes
i s sea ch.
o side a io s a d e Im o eme s
Due o compu a ional cons ain s he algo i hm employed ela i ely conse a i e alues o
popula ion size and numbe o gene a ions. Speci cally only 10 gene a ions we e easible
wi hin he cu en compu a ional esou ces. None heless p elimina y esul s showed
signi can imp o emen s in spa ial co ela ion.
Fo u u e i e a ions subs an ial pe o mance imp o emen s could be achie ed by
inc easing he compu a ional esou ces he eby allowing
o nc easing he numbe o gene a ions om 10 o a leas 50 enhancing he con e gence
and accu acy o he op imized solu ions.
o Expanding popula ion size o inc ease he di e si y o candida e solu ions u he
educing he li elihood o local minimum and imp o ing global sea ch capabili ies.
Such adjus men s would signi can ly enhance he algo i hm's capabili y o ne- une
co ical exci abili y pa ame e s and mo e accu a ely eplica e empi ical EEG dis ibu ions.
Appendix Gene ic Algo i hm S a egies o Co ical Pa ame e p imiza ion
This appendix p esen s he compu a ional s a egies and e alua ion me ics used in he
implemen a ion o se e al gene ic algo i hms (GAs) o op imize he co ical exci abili y pa ame e s
(W) in he pa ien -speci ic b ain model. The co e objec i e ac oss all me hods was o imp o e he
spa ial esemblance be ween he syn he ic EEG and he empi ical EEG, pa icula ly ega ding he
opog aphic dis ibu ion o ac i i y.
H.1 O e iew o Gene ic Algo i hm Implemen a ion
Each gene ic algo i hm was s uc u ed a ound a s anda d e olu iona y p ocess:
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
80
• Ini ializa ion: A popula ion o indi iduals (pa ame e se s o exci abili y W alues) was
andomly gene a ed, wi h a subse ixed o baseline exci abili y (W=3) and he emainde
assigned andom alues wi hin a ealis ic physiological ange (3–8).
• E alua ion: Each indi idual was e alua ed using a i ness unc ion, which measu ed he
simila i y be ween syn he ic and empi ical EEG da a using a ious spa ial me ics.
• Selec ion and Rep oduc ion: The bes -pe o ming indi iduals ( op_k) we e e ained, and
new indi iduals we e c ea ed h ough c osso e and mu a ion ope a ions.
• I e a ion: The p ocess was epea ed o a ixed numbe o gene a ions.
A penal y e m (L1- egula iza ion) was op ionally included in each i ness unc ion o discou age
excessi e de ia ions om he heal hy baseline exci abili y (W=3).
H.2 Pea son-Based Op imiza ion Techniques
The Pea son co ela ion was explo ed h ough h ee di e en con igu a ions:
H.2.1 Pea son Co ela ion on Mean Ampli ude Vec o s
This ea ly me hod compu ed he Pea son co ela ion be ween he 1D ec o s o a e age
ampli ude ac oss EEG channels. I p o ided a quick global me ic bu p o ed insu icien
o cap u ing opog aphic s uc u e. Resul s o en showed imp o emen in occipi al
ampli ude pa e ns bu ailed o de ec cen al ac i i y p opaga ion.
H.2.2 Pea son Co ela ion on In e pola ed Topog aphic Maps (Full Image)
This echnique in e pola ed he empi ical and syn he ic EEG alues o e a 2D head model
g id and calcula ed he Pea son co ela ion be ween he esul ing images. This app oach
cap u ed he ull spa ial dis ibu ion o b ain ac i i y.
I p oduced he bes esul s, wi h isual and quan i a i e imp o emen s in bo h occipi al
and cen al egions (e.g., = 0.34 → 0.50).
Fi ness unc ion:
𝑓(𝑤)=𝜌𝑖𝑚𝑔− 𝐿1 (19)
Whe e:
• 𝜌𝑖𝑚𝑔 is he Pea son co ela ion be ween he la ened pixel ec o s o bo h
opomaps.
• 𝐿1 is a egula iza ion e m explained below.
H.2.3 Cen al Clus e Pea son Co ela ion
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
81
He e, he Pea son co ela ion was calcula ed using only a p ede ined cen al egion o
in e es (C1, CP1, C2, CP2, C3, CP3, C4, CP4). This localized me ic enabled a be e
ocus on epilep ic p opaga ion in mo o a eas.
Fi ness unc ion:
𝑓(𝑤)=𝜌𝑐𝑒𝑛𝑡𝑟𝑎𝑙− 𝐿1 (20)
Figu e H1. Topog aphic maps o syn he ic EEG wi h he new models a e shown a e modi ying he exci abili y
pa ame e s o he ARC ile using his GA. Va ious GA esul s a e shown, demons a ing how he algo i hm e ol es.
Pea son Values ( ) : (A) = 0.7530, (B) =0.7595, (C) = 0.7582, (D) = 0.7615.
The images o he opog aphic map o he o iginal syn he ic EEG and he eal EEG a e
shown
Figu e H2. On he le , Figu e A, he opog aphic map o he o iginal syn he ic EEG is shown and, on he le , Figu e B, he
empi ical one is shown.
H.3 RMSE-Based Op imiza ion
This me hod minimized he Roo Mean Squa e E o (RMSE) be ween he ull in e pola ed
opog aphic maps. RMSE di ec ly quan i ied he a e age magni ude o pixel-wise di e ences,
p o iding a scale-sensi i e and in ui i e measu e.
Resul s showed a consis en and smoo h imp o emen o he global map alignmen (e.g., RMSE
om ~4.0 → ~3.8), hough less exp essi e in localized ac i i y changes compa ed o Pea son.
A
B
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
82
Fi ness unc ion: 𝑓(𝑤)=−𝑅𝑀𝑆𝐸(𝑢,𝑣)−𝐿1 (21)
Whe e u and a e he in e pola ed alues o he syn he ic and empi ical EEG maps, espec i ely.
Figu e H3. Topog aphic maps o syn he ic EEG wi h he new models a e shown a e modi ying he exci abili y pa ame e s
o he ARC ile using his GA. Va ious GA esul s a e shown, demons a ing how he algo i hm e ol es. Roo Mean Squa e
E o Values ( ) : (A) RMSE: 4.3536, (B) RMSE: 3.9855, (C) RMSE: 3.9825, (D) RMSE= 3.6094.
H.4 GMD-Based Mul i-Objec i e Op imiza ion
The inal s a egy u ilized Global Map Dissimila i y (GMD), a no malized spa ial dissimila i y
measu e compu ed a e sub ac ing he mean and no malizing ampli ude dis ibu ions. I was
applied independen ly o wo b ain egions:
• Cen al clus e : C1, CP1, C2, CP2, e c.
• Occipi al clus e : O1, PO3, OZ, POZ, O2, e c.
A mul i-objec i e gene ic algo i hm was employed o join ly op imize he in e se GMD (1 - GMD)
in bo h egions. Solu ions we e e alua ed using Pa e o dominance.
Fi ness unc ion: 𝑓(𝑤)=[1−𝐺𝑀𝐷𝑐𝑒𝑛𝑡𝑟𝑎𝑙,1−𝐺𝑀𝐷𝑜𝑐𝑐𝑖𝑝𝑖𝑡𝑎𝑙] (22)
The GA e u ned a Pa e o on o solu ions, balancing bo h objec i es.

De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
83
Figu e 28 Topog aphic map o he syn he ic EEG a e op imiza ion using GA using GMD as he e alua ion me hod. The
igu e shows one o he possible solu ions, he es being no be e han his one. In his case, he ollowing combina ion is
used: GMD Clus e Cen al = 0.9752, Occipi al = 0.9745
H.5. Regula iza ion Te m (L1 penal y)
All i ness unc ions inco po a ed an L1 penal y o p e en ex eme inc eases in exci abili y
alues ac oss he b ain:
Fo mula: 𝐿1=( 1
𝑁+1)∑|𝑊𝑖−𝑊ℎ𝑒𝑎𝑙𝑡ℎ𝑦|∗𝜆 (23)
Whe e:
• N is he numbe o pa cels being op imized.
• 𝑊𝑖 is he exci abili y o pa cel i.
• 𝑊ℎ𝑒𝑎𝑙𝑡ℎ𝑦 is he baseline exci abili y (W = 3).
• λ is he egula iza ion coe icien , se o 0.01 in all expe imen s.
Ano he po en ial simila i y me ic is he S uc u al Simila i y Index Measu e (SSIM), commonly
used in image analysis o e alua e pe cei ed s uc u al simila i y be ween wo images. SSIM
conside s luminance, con as , and s uc u al in o ma ion, p o iding a pe cep ually mo i a ed
sco e be ween –1 and 1. Al hough SSIM was no implemen ed in he p esen wo k, i emains a
p omising candida e o u u e s udies. I s applica ion o EEG opog aphic maps could o e a
complemen a y pe spec i e, po en ially imp o ing he sensi i i y o op imiza ion algo i hms o
sub le spa ial ea u es.
Each s a egy o e ed dis inc ad an ages and e ealed di e en aspec s o syn he ic EEG quali y.
Pea son on he ull opog aphic map eme ged as he mos e ec i e single-objec i e s a egy.
De elopmen o EEG-Based Whole-B ain Models
o Simula ing Seizu e Dynamics
84