[Eeglablist] Group level brain connectivity analyses in sourcelevel

Thu Jul 14 12:21:41 PDT 2016

One question: why not simply interpolate between sources on the brain and then morph all of the individual source spaces into a common source space? This is what MNE does. I personally then build my connectivity matrices etc. in the common source space.

From: Makoto Miyakoshi <mmiyakoshi at ucsd.edu<mailto:mmiyakoshi at ucsd.edu>>
Reply-To: <mmiyakoshi at ucsd.edu<mailto:mmiyakoshi at ucsd.edu>>
Date: Thursday, July 14, 2016 2:25 PM
To: Aleksandra Vuckovic <Aleksandra.Vuckovic at glasgow.ac.uk<mailto:Aleksandra.Vuckovic at glasgow.ac.uk>>
Cc: EEGLAB List <eeglablist at sccn.ucsd.edu<mailto:eeglablist at sccn.ucsd.edu>>, James Jones-Rounds <jj324 at cornell.edu<mailto:jj324 at cornell.edu>>
Subject: Re: [Eeglablist] Group level brain connectivity analyses in sourcelevel

Dear Alex,

In theory yes, but instead of dipole density you need to have dipole pair density. All of this minor difference from Measure Projection is summarized as Network Projection which is unpublished but presented in Nima's PhD thesis.

I chose, however, not to create domains. Creating 'domains' based on measure similarity smells like double dipping to me. Instead, I separate the dipole pair density cloud into anatomical regions (note that this was also prepared by Nima, so I'm still using his option). Using anatomical ROIs is as good as creating across-subject domains to prepare common data structure across subjects.

It uses AAL (Tzourio-Mazoyer et al., 2002) that has 88 anatomical regions, but I merged upper and lower basal regions such as caudate, thalamus, amygdala, hippocampus, etc. Thus the current scheme uses 76 regions, and the connectivity matrix size is 76x76. Each of the graph edge has time x freq x subject data.

Makoto

On Thu, Jul 14, 2016 at 11:10 AM, Aleksandra Vuckovic <Aleksandra.Vuckovic at glasgow.ac.uk<mailto:Aleksandra.Vuckovic at glasgow.ac.uk>> wrote:
Hi
Does it mean that if you do MPT analysis and create domains you extend to group connectivity analysis between domains?
Regards
Alex

Sent from my iPhone

On 14 Jul 2016, at 18:49, Makoto Miyakoshi <mmiyakoshi at ucsd.edu<mailto:mmiyakoshi at ucsd.edu>> wrote:

Dear James,

I'm surprised you know a lot about development of the group-level SIFT. Let me tell you a little bit about the current development. When I started it, there were two possibilities for me.
1. Wes & Tim's 'Bayesian Hierarchical Modeling' with interpolating 'missing' ICs acorss subjects.
2. Nima's 'Network Projection' which blurs existing ICs and average across subjects.
I chose the second plan since Bayesian Hierarchical Modeling was too fancy. Nima's concept of 'Network Projection', which is a variation of Measure Projection, was technically more straightforward and familiar to me. So I bothered Nima a lot to learn how to make it work.

SCCN has been struggling with IC inconsistency across subjects, and the difficulty in the solution is actually only a variation of the same issue... it's always there as long as you use ICA on individual data, which is Scott's philosophy and I don't want to give it up.

Thus, the solution I've been developing is not Tim's. What I do is Nima's 'blur, weight, and average' approach using his code. It would be interesting if we could compare the results from the two methods.

Makoto

On Thu, Jul 14, 2016 at 6:31 AM, James Jones-Rounds <jj324 at cornell.edu<mailto:jj324 at cornell.edu>> wrote:
I have been struggling with this issue for a couple years, also, and have two responses to your questions, Iman.

First, some of the (fomer) researchers at UCSD, including Tim Mullen (who developed the SIFT connectivity toolbox), were working hard on a group-level, source-space connectivity toolbox, that would use markov chain monte carlo estimations of posterior distribution densities to deal with the missing data problem that is inherent to multi-subject connectivity analyses. Here is a link to a poster they presented a couple years ago on it.

http://sccn.ucsd.edu/~scott/pdf/Thompson_and_Mullen_Poster_ICONXI.pdf

However, due to some extenuating circumstances, they were not able to complete it. I am so thankful Makoto has taken up the charge to address the issue with his own toolbox!

Second, a work-around that I have been using has been to conduct k-means clustering (using the standard STUDY function in EEGLAB), with a heavy weight on "dipole location" in the pre-clustering step. Then, I use these across-subject clusters (choosing the clusters that have the highest recruitment, i.e. have IC's from as many subjects as possible) to determine which IC's I will measure connectivity between. In other words, if Cluster 5 and Cluster 9 both recruited IC's from >90% of my subjects, then for each subject that has IC's in both clusters, I will measure the connectivity between the components, for that subject, that are in both clusters.

Then i average the connectivity values accordingly, and simply ignore any subjects that are not co-occurring in both clusters of interest.

This is not ideal but does generate plausible results.

I'm so glad you brought this issue back onto the front burner, Iman!

James

From: Makoto Miyakoshi<mailto:mmiyakoshi at ucsd.edu>
Sent: Wednesday, July 13, 2016 12:00 AM
To: Iman Mohammad-Rezazadeh<mailto:irezazadeh at ucdavis.edu>
Cc: EEGLAB List<mailto:eeglablist at sccn.ucsd.edu>; Scott Makeig<mailto:smakeig at ucsd.edu>; s.sanei at surrey.ac.uk<mailto:s.sanei at surrey.ac.uk>
Subject: Re: [Eeglablist] Group level brain connectivity analyses in sourcelevel

Dear Iman,

I have an alpha version. I'll present it in SfN this year (see below for poster info). I'm currently working on ULCA data sets using it. I plan to release a beta version of the tool this summer (of course, it's a pluging for EEGLAB). I'll meet my colleagues in Qusp soon to report finalization of it and discuss publication.

Makoto

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Your abstract, "Group-level statistics on EEG effective source connectivity," has been accepted into the program as a traditional poster presentation

Session Type: Poster
Session Number: 851
Session Title: Computational Tools for Human Data II
Date and Time: Wednesday Nov 16, 2016 1:00 PM - 5:00 PM
Location: San Diego Convention Center: Halls B-H
Abstract Control Number: 9417

Multivariate connectivity measures in EEG have been gathering attention to investigate causal information flow in dynamics across signals. There are known issues in applying this method on scalp-recorded channel EEG data, such as volume conductance and scalp mixing, which makes the original scalp channel data highly correlated. To address these issues, applying independent component analysis (ICA) as preprocess is effective. It finds a linear transform to obtain effective source signals that are temporally maximally independent to each other. Thanks to this, the issues raising from volume conductance and scalp mixing are both addressed cleverly without estimating any parameters in electrophysiological forward model. However, because ICA reveals individual differences, it creates problems in the group-level statistics. For example, in the conventional EEG analysis, group-level statistics was straightforward; selecting a channel e.g. Cz from all the subjects was regarded to be sufficient. But after ICA preprocessing, there is no exact common independent comonent across all the subjects. Hence we developed a following statistical framework: 1. Preprocess individual data with ICA, estimate equivalent current dipoles, and apply multivariate connectivity measures across all pairs of ICs; 2. Compute dipole density that distributes within MNI brain space by applying 3-D Gaussian kernel; 3. Segment dipole density into anatomical regions; 4. Compute region-to-region pairwise dipole density that is weighted by connectivity measures; 5, Repeat above process for all subjects; 6, Perform statistics between conditions using variance across subjects. We developed a free, open-source source toolbox that plugs into EEGLAB. With this solution, effective EEG source connectivity can be evaluated in time-frequency domain at the group-level statistics with multiple comparison corrections. It is also generate a movie to visualize information flow. We expect it will also impact MEG, ECoG, and other electrophysiological data analysis.

On Tue, Jul 12, 2016 at 4:11 PM, Iman Mohammad-Rezazadeh <irezazadeh at ucdavis.edu<mailto:irezazadeh at ucdavis.edu>> wrote:
Hi EEGLABers,

My concern is because in any experiment subjects don’t necessarily share common EEG sources  ( nodes in the network) in terms of their locations, number of sources , etc.

1-      Does anyone has experience on doing group level connectivity analyses on source  ( not channel level).

2-      In a single subject design ( pre-/ post-treatment , for example) how can we compare the brain connectivity in source level by considering the above concern?

I would greatly appreciate any input and also please share any resources that you might think it is helpful.

Thanks
Iman
-------------------------------------------------------------
Iman Rezazadeh, Ph.D
Project Scientist  | Semel Intitute, UCLA , Los Angeles, CA
Adjunct Researcher | Center for Mind and Brain, Davis, CA

--
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego

--
James Jones-Rounds
Laboratory Manager
Human Development EEG and Psychophysiology (HEP) Laboratory,
Department of Human Development,
--------------------------------------------
Cornell University | Ithaca, NY
607-255-9883<tel:607-255-9883>
eeg at cornell.edu<mailto:eeg at cornell.edu>

_______________________________________________
Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu<mailto:eeglablist-unsubscribe at sccn.ucsd.edu>
For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu<mailto:eeglablist-request at sccn.ucsd.edu>

--
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
_______________________________________________
Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu<mailto:eeglablist-unsubscribe at sccn.ucsd.edu>
For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu<mailto:eeglablist-request at sccn.ucsd.edu>

--
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego