Hierarchical Bayesian Modeling (HBM) in EEG and MEG source analysis Carsten Wolters Institut für...

Hierarchical Bayesian Modeling (HBM) in EEG and MEG source analysis

Carsten WoltersCarsten WoltersCarsten WoltersCarsten Wolters

Institut für Biomagnetismus und Biosignalanalyse, Westfälische Wilhelms-Universität MünsterInstitut für Biomagnetismus und Biosignalanalyse, Westfälische Wilhelms-Universität Münster

Vorlesung, 6.Mai 2014Vorlesung, 6.Mai 2014

Carsten.wolters@uni-münster.deCarsten.wolters@uni-münster.de

[Lucka, Burger, Pursiainen & Wolters, NeuroImage, 2012] [Lucka, Burger, Pursiainen & Wolters, Biomag2012, 2012]

[Lucka, Diploma thesis in Mathematics, March 2011]

[Lucka, Burger, Pursiainen & Wolters, NeuroImage, 2012] [Lucka, Burger, Pursiainen & Wolters, Biomag2012, 2012]

[Lucka, Burger, Pursiainen & Wolters, NeuroImage, 2012]

Hierarchical Bayesian Modeling (HBM):Mathematics: The likelihood model

• Central to Bayesian approach: Accounting for each uncertainty concerning the Central to Bayesian approach: Accounting for each uncertainty concerning the

value of a variable explicitly: The variable is modeled as a random variablevalue of a variable explicitly: The variable is modeled as a random variable• In this study, we model the additive measurement noise by a Gaussian random In this study, we model the additive measurement noise by a Gaussian random

variablevariable

• For EEG/MEG, this leads to the following likelihood model:For EEG/MEG, this leads to the following likelihood model:

Hierarchical Bayesian Modeling (HBM):Mathematics: The likelihood model

[Lucka, Burger, Pursiainen & Wolters, NeuroImage, in revision] [Lucka, Burger, Pursiainen & Wolters, Biomed.Eng., 2011]

• The conditional probability density of B given S is called likelihood density, in our The conditional probability density of B given S is called likelihood density, in our

(Gaussian) case, it is thus:(Gaussian) case, it is thus:

Hierarchical Bayesian Modeling (HBM):Mathematics: Prior and Bayes rule

• Due to the ill-posedness, inference about S given B is not feasible like that, we Due to the ill-posedness, inference about S given B is not feasible like that, we

need to encode a-priori information about S in its density pneed to encode a-priori information about S in its density pprpr(s), which is called (s), which is called priorprior

• We call the conditional density of S given B the We call the conditional density of S given B the posteriorposterior: p: ppost post (s|b)(s|b)

•Then, the model can be inverted via Bayes rule:Then, the model can be inverted via Bayes rule:

• The term p(b) is called The term p(b) is called model evidencemodel evidence, (see , (see Sato et al., 2004; Trujillo-Barreto et

al., 2004; Henson et al., 2009, 2010). Here, it is just a normalizing constant and not ). Here, it is just a normalizing constant and not

important for the inference presented nowimportant for the inference presented now

Hierarchical Bayesian Modeling (HBM):Mathematics: MAP and CM

• The common way to exploit the information contained in the posterior is to infer a The common way to exploit the information contained in the posterior is to infer a

point-estimate for the value of S out of itpoint-estimate for the value of S out of it• There are two popular choices, the There are two popular choices, the Maximum A-PosterioriMaximum A-Posteriori (MAP, the highest (MAP, the highest

mode of the posterior) and the mode of the posterior) and the Conditional MeanConditional Mean (CM, the expected value of the (CM, the expected value of the

posterior):posterior):

• Practically, the MAP is a high-dimensional optimization problem and the CM is a Practically, the MAP is a high-dimensional optimization problem and the CM is a

high-dimensional integration problemhigh-dimensional integration problem

Hierarchical Bayesian Modeling (HBM):Mathematics: Specific priors used in EEG/MEG

• To revisit some commonly known inverse methods, we consider Gibbs To revisit some commonly known inverse methods, we consider Gibbs

distribution as prior:distribution as prior:

• Here, P(s) is an energy functional penalizing unwanted features of sHere, P(s) is an energy functional penalizing unwanted features of s• The MAP-estimate is then given by:The MAP-estimate is then given by:

Hierarchical Bayesian Modeling (HBM):Mathematics: Some choices for P(s) used in EEG/MEG

• Minimum Norm Estimation Minimum Norm Estimation (MNE), see (MNE), see

Hämäläinen and Ilmoniemi, 1984

Hierarchical Bayesian Modeling (HBM):Mathematics: Some choices for P(s) used in EEG/MEG

• Weighted Minimum Norm Estimation Weighted Minimum Norm Estimation

(WMNE), see (WMNE), see Dale and Sereno, 1993

• Specific choices for WMNE: Specific choices for WMNE: Fuchs et al., 1999

Hierarchical Bayesian Modeling (HBM):Mathematics: sLORETA

• standardized LOw REsolution electromagnetic TomogrAphy standardized LOw REsolution electromagnetic TomogrAphy (sLORETA), see (sLORETA), see

Pascual-Marqui, 2002• The MAP estimate (which is the MNE) is standardized by the posterior

covariance, yielding a pseudo-statistic of F-type for the source amplitude at a

source space node

Hierarchical Bayesian Modeling (HBM):Mathematics:

• Brain activity is a complex process comprising many different spatial patterns• No fixed prior can model all of these phenomena without becoming

uninformative, that is, not able to deliver the needed additional a-priori

information• This problem can be solved by introducing an adaptive, data-driven element into

the estimation process

• The idea of Hierarchical Bayesian Modeling (HBM) is to let the same data

determine the appropriate model used for the inversion of these data by extending

the model by a new level of inference: The prior on S is not fixed but random,

determined by values of additional parameters called hyperparameters• The hyperparameters follow an a-priori assumed distribution (the so-called

hyperprior phpr()) and are subject to estimation schemes, too.

• As this construction follows a top-down scheme, it is called hierarchical modeling:

Hierarchical Bayesian Modeling (HBM):Mathematics for EEG/MEG application

• The hierarchical model used in most methods for EEG/MEG relies on a The hierarchical model used in most methods for EEG/MEG relies on a

special construction of the prior called special construction of the prior called Gaussian scale mixtureGaussian scale mixture or or conditionally conditionally

Gaussian hypermodelGaussian hypermodel ( (Calvetti et al., 2009; Wipf and Nagarajan, 2009)

• ppr(s|) is a Gaussian density with zero mean and a covariance determined by

Hierarchical Bayesian Modeling (HBM):Mathematics: Our chosen posterior

• CM estimation: Blocked Gibbs sampling, a Markov chain Monte Carlo CM estimation: Blocked Gibbs sampling, a Markov chain Monte Carlo

(MCMC) scheme ((MCMC) scheme (Nummenmaa et al., 2007; Calvetti et al., 2009)

• MAP estimation: Iterative alternating sequential (IAS) (MAP estimation: Iterative alternating sequential (IAS) (Calvetti et al., 2009)

Hierarchical Bayesian Modeling (HBM):Goal of our study

• Step 1: Computed forward EEG for reference source (green dipole)Step 1: Computed forward EEG for reference source (green dipole)• Step 2: Computed HBM inverse solution without indicating the number of Step 2: Computed HBM inverse solution without indicating the number of

sources (yellow-orange-red current density distribution on source space)sources (yellow-orange-red current density distribution on source space)

Hierarchical Bayesian Modeling (HBM):Validation means: DLE and SP

Hierarchical Bayesian Modeling (HBM):Validation means: EMD

Hierarchical Bayesian Modeling (HBM):Validation means: Source depth

Hierarchical Bayesian Modeling (HBM):Methods: Head model

Hierarchical Bayesian Modeling (HBM):Methods: EEG sensors

Hierarchical Bayesian Modeling (HBM):Methods: Full-cap (f-cap), realistic cap (r-cap)

Hierarchical Bayesian Modeling (HBM):Methods: Source space and EEG lead field

Hierarchical Bayesian Modeling (HBM):Methods: Source space

Hierarchical Bayesian Modeling (HBM):Study 1: Single dipole reconstruction

Hierarchical Bayesian Modeling (HBM):Methods: Generation of noisy measurement data

Hierarchical Bayesian Modeling (HBM):Results: Single focal source scenario

• Step 1: Computed forward EEG for reference source (green dipole), add Step 1: Computed forward EEG for reference source (green dipole), add

noisenoise• Step 2: Computed HBM inverse solution without indicating the number of Step 2: Computed HBM inverse solution without indicating the number of

sources (yellow-orange-red current density distribution on source space)sources (yellow-orange-red current density distribution on source space)

Hierarchical Bayesian Modeling:Single focal source scenario

HBM: Conditional Mean (CM) estimateHBM: Conditional Mean (CM) estimate

HBM: CM followed by Maximum A-Posteriori estimate (MAP)HBM: CM followed by Maximum A-Posteriori estimate (MAP)

Hierarchical Bayesian Modeling:Study 1: Single focal source scenario

Hierarchical Bayesian Modeling :Study 1: Single focal source scenario

• A mark within the area underneath the A mark within the area underneath the y=x line indicates that the dipole has y=x line indicates that the dipole has been reconstructed too close to the been reconstructed too close to the surface surface • A mark above the line indicates the A mark above the line indicates the oppositeopposite• qqabab denotes the percentage of marks denotes the percentage of marks

above the y=x line minus 0.5 (optimally: above the y=x line minus 0.5 (optimally:

qqabab= 0)= 0)

Hierarchical Bayesian Modeling :Single focal source scenario

Hierarchical Bayesian Modeling (HBM):Study 2: Two sources scenario

Hierarchical Bayesian Modeling:Study 2: Two sources scenario

Hierarchical Bayesian Modeling (HBM):Study 3: Three sources scenario

Hierarchical Bayesian Modeling:Study 3: Three sources scenarioHierarchical Bayesian Modeling:Study 3: Three sources scenario

Thank you for your attention!Thank you for your attention!

Hierarchical Bayesian Modeling (HBM) in EEG and MEG source analysis Carsten Wolters Institut für...

Documents

Six degrees of syndication: Was machen Blogs in Bibliotheken? Dr. Oliver Obst Universitäts- und Landesbibliothek Westfälische Wilhelms Universität Münster

Da´wat-e Islami – a transnational movement for spiritual renewal Presentation for Werkstatt Nahost, Graduate School of Politics, Westfälische-Wilhelms-Universität

Westfälische Wilhelms-Universität Münster Westfälische Wilhelms-Universität Münster The University of Münster – Institute of Tax Law

Umgang mit zentralvenösen Kathetern Shahram Ramtin Medizinische Klinik und Poliklinik C Kardiologie, Angiologie UKM Münster Westfälische Wilhelms-Universität

Studies on a di-electron trigger for Upsilon-mesons …...Westfälische Wilhelms-Universität Münster acFhbereich Physik Institut für Kernphysik Arbeitsgruppe/Working group Johannes

Dynamical modeling of Mars deep interior Sylvaine Ferrachat Institut für Planetologie Westfälische Wilhelms-Universität Münster

Hbm solution

Impressive Design. Stunning Performance. · LG QuickWindow™ Folio Case • Bluetooth Mono Headset (HBM-260, HBM-W600, HBM-280, HBM-290) • ™Bluetooth Stereo Headset (LG TONE

RHEINISCHE FRIEDRICH-WILHELMS-UNIVERSITÄT BONN Faculty …repositorio.educacionsuperior.gob.ec/bitstream/28000/1995/1/T... · RHEINISCHE FRIEDRICH-WILHELMS-UNIVERSITÄT BONN Faculty

1 Westfälische Wilhelms-Universität Münster Prof. Dr. W. Woyke Deutschland und die NATO

cubierta kickstart 2011 okguerra/publicaciones/cubierta_kickstart_20… · Kick Start cubierta kickstart 2011 ok.qxp 14/09/2011 17:40 PÆgina 1. NC'* 'lc WESTFÄLISCHE WILHELMS-UNIVERSITÄT

Mathematical Methods for the Segmentation of Medical Images Carsten Wolters Institut für Biomagnetismus und Biosignalanalyse, Westfälische Wilhelms-Universität

HC26.11.310 HBM Bandwidth Kim Hynix Hot Chips HBM 2014 v7

Westfälische Wilhelms-Universität Münster Volkswirtschaftliche Diskussionsbeiträge · Westfälische Wilhelms-Universität Münster Volkswirtschaftliche Diskussionsbeiträge

Bachelor Thesis - uni-muenster.de...Bachelor Thesis The Effects of Different Verbal Route Instructions on Spatial Orientation Stefan Fuest Westfälische Wilhelms-Universität Münster

Course Program - uni-muenster.de · Course Program of the BSc in Physics - Dekanat des Fachbereichs Physik - April 2018. 2 2 INTRODUCTION The Westfälische Wilhelms-Universität in

The problem with Wilhelms II Germany

Courses in English, Fall 2017/18 · Courses in English, Fall 2017/18 . at the School of Business and Economics, Westfälische Wilhelms-Universität Münster . As of September 4. th,

Assessing Teaching and Assessment Competences of Biology ... · Sascha Hasse1, Cora Joachim2, Susanne Bögeholz 1, Marcus Hammann2 1 Westfälische Wilhelms-Universität Münster 2

Coordination Modes of Hydroxylamine Ligands at Elements ... · 1 Coordination Modes of Hydroxylamine Ligands at Elements Through the Periodic Table WESTFÄLISCHE WILHELMS-UNIVERSITÄT