[Eeglablist] 答复：Re: Tensor Decomposition of EEG Signals

Wed Jun 3 17:55:52 PDT 2015

Dear Fengyu,

Thank you for the detailed response.

> I have found that the variability of single-trial MMN-EEG data among
multiple trials of one subject is stronger than the variability of averaged
MMN-EEG data of multiple subjects.

That's our observation too, which is why we have a dogma that we should
look into single trials... but you are right, our current ICA does not
fully account the non-stationarity of the data, therefore as you say it is
still a kind of group analysis (group-trial analysis). My colleagues are
working of multiple model ICAs and some kind of sliding window method (like
online ICA).

Speaking of ICA, I would particularly highly evaluate the work done in
Onton and Makeig (2006) in which Scott proposed a physiological model of
ICA application on scalp-recorded EEG data. I believe it made the
application of ICA qualitatively different from other signal processing
methods... which never stops to fascinate me.

Thank you.

Makoto

On Wed, Jun 3, 2015 at 5:28 PM, CONG Fengyu 丛丰裕 <fengyu.cong at aliyun.com>
wrote:

> Dear Makoto
>
>
> Glad to hear from you! To my surprise, you know the meaning of my first
> name in Chinese characters!
>
>
> I answer your second question as the following:
>
> It is new to me that applying three-way data analysis on individual
> subject's independent components. I have not done that before. Tensor
> decomposition is a kind of methods for blind source separation (BSS). ICA
> is also a method for BSS. So, I need more time to digest your suggestion.
>
>
> Your first question is hard to answer.
>
>
> In an ERP data set, there are modes including time, space (channel),
> trial, stimulus, and subject (even group). Therefore, we have the following
> matrices to denote the ERP data:
>
> A) the data in the level of single-trial EEG
>
> M1) the matrix of ERP data of one single trial of one stimulus of one
> subject
>
> M2) the matrix of ERP data of concatenated single trials of multiple
> stimuli of one subject
>
> M3) the matrix of ERP data of concatenated single trials of multiple
> stimuli of multiple subjects
>
> B) the data in the level of averaged EEG over single trials.
>
> M4) the matrix of averaged ERP data of one stimulus of one subject
>
> M5) the matrix of averaged ERP data of multiple stimuli of one subject
>
> M6) the matrix of averaged ERP data of multiple stimuli of multiple
> subjects
>
>
> I have found that ICA has been applied on each matrix listed above.
> Usually, ICA is applied on M2, M3, and M6. Using EEGLAB ICA tends to be
> performed on M2, which is the most frequently applied.
>
>
> For M2, I think it is also a kind of group ICA since the variability of
> EEG data of multiple trials does exist. Without investigation, it is hard
> to know whether the variability of single-trial EEG data of multiple trials
> of one subject is smaller or bigger than the variability of averaged EEG
> data of multiple subjects.
>
>
> I mainly work on MMN data. MMN is usually produced by the difference wave
> in terms of the averaged EEG over single trials. Moreover, in light of the
> several MMN datasets I have worked on, I have found that the variability of
> single-trial MMN-EEG data among multiple trials of one subject is stronger
> than the variability of averaged MMN-EEG data of multiple subjects.
> Therefore, I use ICA in terms of M4 and M6.
>
>
> For M6, it is easy to do group-level analysis since the ERP data are
> grouped together. However, there is strong assumption on different
> subjects.  For M4, I usually do the following way:
>
> 1) Filtering the averaged EEG data using a wavelet filter to reduce the
> number of sources (please check the frequency response of the filter in
> Cong et al., NSM2014, and the effect in reducing the number of sources in
> Cong et al., NSM2013&2011). Since the number of sources in the data is
> reduced, the requirement of number of samples of ICA to converge is also
> reduced.
>
> 2) Extracting R components from the wavelet-filtered EEG data using
> ICASSO.
>
> 3) Choosing the components of interest in terms of the temporal and
> spatial properties of ERPs.
>
> 4) Projecting the selected components of interest back to the electrode
> field to correct the variance and polarity indeterminacies. It is possible
> that the correction fails due to the poor ICA decomposition (please read
> the reason in Cong et al., BSPC2011). My current suggestion is to re-do the
> steps 3&4 again. If it also fails, I would not like to suggest using ICA
> for decomposing the ERP data.
>
> 5) I do steps 1-4 for M4 for each stimulus and each subject. Then, I
> analyze the new ERP data by ICA like the conventional ERP data.
>
>
> The references mentioned above are as the following:
>
> *Fengyu Cong, *Qiu-Hua Lin, Piia Astikainen, and Tapani Ristaniemi，*How
> to Validate Similarity in Linear Transform Models of Event-related
> Potentials between Experimental Conditions? *Journal of Neuroscience
> Methods, 2014, 236:76-85.
>
> *Fengyu Cong, *Zhaoshui He, Jarmo Hämäläinen, Paavo HT Leppänen, Heikki
> Lyytinen, Andrzej Cichocki, Tapani Ristaniemi, *Validating Rationale of
> Group-level Component Analysis based on Estimating Number of Sources in EEG
> through Model Order Selection*, Journal of Neuroscience Methods, 2013,
> 212(1): 165–172.
>
> *Fengyu Cong*, Paavo H.T. Leppänen, Piia Astikainen, Jarmo Hämäläinen,
> Jari K. Hietanen, Tapani Ristaniemi, *Dimension Reduction: Additional
> Benefit of an Optimal Filter for Independent Component Analysis to Extract
> Event-related Potentials*, Journal of Neuroscience Methods, 2011, 201(1):
> 269-280.
>
> *Fengyu Cong*, Igor Kalyakin, Tapani Ristaniemi, *Can Back-Projection
> Fully Resolve Polarity Indeterminacy of ICA in Study of ERP?* Biomedical
> Signal Processing and Control, 2011, 6(4): 422-426.
>
>
> Again, I welcome your comments and suggestions.
>
> Best regards,
>
> Fengyu
>
>
>
> ------------------------------------------------------------------
>
>
>
>
> *From:* Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
>
> *Date:* 2015-06-04 02:28
>
> *To:* CONG Fengyu 丛丰裕 <fengyu.cong at aliyun.com>
>
> *CC:* EEGLAB <eeglablist at sccn.ucsd.edu>
>
> *Subject:* Re: [Eeglablist] Tensor Decomposition of EEG Signals
>
>
> Congrats for publication! Without much reading, some thoughts and comments.
>
>
> > Electroencephalography (EEG) is one fundamental tool for functional
> brain imaging.
>
>
> I'm sure Scott will like this phrase.
>
>
> > However, the mostly applied computing tools for brain research are
> oriented for one-way or two-way data.
>
>
> We do not apply ICA on concatenated multiple-subject data (which is called
> group ICA) because we appreciate individual differences across subjects.
> This sounds good, but it's also a curse because the group-level analysis is
> so complicated. How do you integrate thousands of ICs from dozens of
> subjects in the final group-level analysis? We have been struggling against
> the issue over ten years.
>
>
> By the way please tell me if it is possible to apply your three-way data
> analysis on individual subject's independent components. I don't have
> math/physic background.
>
>
> Makoto
>
>
> Dear Fengyu (rich and rich),
>
>
> On Fri, May 29, 2015 at 3:55 PM, CONG Fengyu 丛丰裕 <fengyu.cong at aliyun.com>
> wrote:
>
>> Dear EEGLablist members,
>>
>> I would like to introduce my new review paper to you. It is about Tensor
>> Decomposition of EEG Signals and published in Journal of Neuroscience
>> Methods. The paper is open access and can be downloaded via
>>
>> http://www.sciencedirect.com/science/article/pii/S0165027015001016. The
>> demo of MATLAB codes and ERP data can be downloaded via
>> http://www.escience.cn/people/cong/AdvancedSP_ERP.html
>>
>> EEG signals tend to be represented by a vector or a matrix to facilitate
>> data processing and analysis with generally understood methodologies like
>> time-series analysis, spectral analysis and matrix decomposition.
>>
>> Indeed, EEG signals are often naturally born with more than two modes of
>> time and space, and they can be denoted by a multi-way array called as
>> tensor. This review summarizes the current progress of tensor decomposition
>> of EEG signals with three aspects. The first is about the existing modes
>> and tensors of EEG signals. Second, two fundamental tensor decomposition
>> models, canonical polyadic decomposition (CPD, it is also called parallel
>> factor analysis-PARAFAC) and Tucker decomposition, are introduced and
>> compared. Moreover, the applications of the two models for EEG signals are
>> addressed. Particularly, the determination of the number of components for
>> each mode is discussed. Finally, the N-way partial least square and
>> higher-order partial least square are described for a potential trend to
>> process and analyze brain signals of two modalities simultaneously.
>>
>> Looking forward to your comments and suggestions on tensor decomposition
>> of EEG signals.
>>
>> Sincerely,
>>
>> Fengyu
>>
>> P.S.: The bibliography of the paper is as the following:
>>
>> Fengyu Cong, Qiu-Hua Lin, Li-Dan Kuang, Xiao-Feng Gong, Piia Astikainen,
>> Tapani Ristaniemi, Tensor Decomposition of EEG Signals: A Brief Review,
>> Journal of Neuroscience Methods 248: 59–69, 2015.
>>
>>
>> --
>>
>> Fengyu Cong, Ph.D., IEEE Senior Member
>>
>>
>> Professor, Department of Biomedical Engineering, Dalian University of
>> Technology, China
>>
>>
>> Docent (Adjunct Associate Professor), University of Jyvaskyla, Finland
>>
>>
>> Email: cong at dlut.edu.cn, fengyu.cong at aliyun.com
>>
>>
>> Homepage: http://www.escience.cn/people/cong/index.html
>>
>>
>>
>>
>> _______________________________________________
>>
>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>>
>> To unsubscribe, send an empty email to
>> 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
>>
>
>
>
> --
>
> Makoto Miyakoshi
>
> Swartz Center for Computational Neuroscience
>
> Institute for Neural Computation, University of California San Diego
>
>

-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20150603/638fc9fe/attachment.html>