Question about referencing and connectivity

[Robin Thomas]

Hello Mike,

Our seminar is getting ready to work through the Laplacian chapter (Ch. 22) in your book and revisit computing coherence and inter site phase clustering to compare the results with and without Laplacian transformation.  However, I just encountered this paper by Thatcher (Thatcher, 2012) which seems to argue that rereferencing using either the average or (especially) the Laplacian is invalid and eliminates true coupling.  I am worried now what is truth and I do not yet have the skills to ferret this out myself.  Have you seen this paper and should we (me and my students) be concerned?  Thanks.  - Robin

[Mike X Cohen]

Hi Robin. 

Robert Thatcher seems to feel pretty strongly against using spatial filtering (e.g., Laplacian or ICA), although he does not provide (in my mind) a compelling argument, particularly for the neurophysiological validity. His argument seems to rest on the observation that phase values are changed after filtering. This is true. It's also of filtering in the time domain. From there it seems he draws the conclusion that this deprives the data of neurophysiological relevance, although this line of reasoning is not explained and no relevant physiological data are presented.

One could turn the argument around and state that raw EEG measures a spatially mixed signal, and that this signal contains distorted phases due to volume conduction, smearing from the skull, etc. Thus, spatial filtering is a sensible approximation to obtaining the un-mixed signal, and therefore is more physiologically interpretable. One could also, I am sure, present a mathematical model that would show this. This fact -- that voltage EEG data are a mixture of spatially broad signals and the measured phase values are a distorted mix of local activities -- is quite apparent when looking at intracranial EEG.

To me, the validity of a method lies not in results from an oversimplified set of assumptions, but rather in its reproducibility, relevance to other studies using other methodologies, support for predictions from a theory, and ability to inspire new, testable, and biologically plausible hypotheses. In these senses, I am satisfied with Laplacian and other spatial filters. It is by no means a perfect solution, and has its own set of limitations, as any method does. We and others have obtained results using Laplacian across different paradigms, recording setups, and countries, that provide convergent, replicable, and theoretically sensible results that are consistent with findings from fMRI and invasive recordings. To be sure, further work should be done to validate or invalidate other spatial filtering approaches, but the argument provided in this paper is neither compelling nor convincing.

You might also consider looking through a recent special issue on the Laplacian in EEG research: http://www.sciencedirect.com/science/journal/01678760/97/3. Perhaps the most relevant paper for your question is on page 285. There have been several other simulation studies showing that the Laplacian is a great technique for recovering true phase synchronization, e.g., https://www.ncbi.nlm.nih.gov/pubmed/25234308

More generally, I should say that (1) neuroscience is a big and complicated field, (2) all methods entail uncertainty, (3) the "ground truth" is difficult or impossible to determine, (4) simulations are OK but are always imperfect and sometimes simply wrong, and, perhaps most importantly (5) neuroscientists tend to be a bunch of strongly opinionated loud-mouths (including yours truly) who are sometimes too convinced of their own personal opinions based on their own personal experiences to appreciate they they are wrong or that their view is correct only in some narrow situations.

I hope that helps, or was at least entertaining.

Mike

On Thu, Apr 21, 2016 at 3:05 PM, Robin Thomas <thom...@miamioh.edu> wrote:

Hello Mike,

Our seminar is getting ready to work through the Laplacian chapter (Ch. 22) in your book and revisit computing coherence and inter site phase clustering to compare the results with and without Laplacian transformation.  However, I just encountered this paper by Thatcher (Thatcher, 2012) which seems to argue that rereferencing using either the average or (especially) the Laplacian is invalid and eliminates true coupling.  I am worried now what is truth and I do not yet have the skills to ferret this out myself.  Have you seen this paper and should we (me and my students) be concerned?  Thanks.  - Robin

--
You received this message because you are subscribed to the Google Groups "AnalyzingNeuralTimeSeriesData" group.
To unsubscribe from this group and stop receiving emails from it, send an email to analyzingneuraltimes...@googlegroups.com.
Visit this group at https://groups.google.com/group/analyzingneuraltimeseriesdata.
For more options, visit https://groups.google.com/d/optout.

--

Mike X Cohen, PhD
mikexcohen.com

[Robin Thomas]

Thank you so much for a very helpful response.  I will definitely look through the special issue articles you suggest.  My own perspective on methods validity trend towards yours regarding reproducibility and theory supported so I am grateful for your information regarding the Laplacian on that front.  This might serve as a good teaching opportunity for my students as well so I am glad to see such detail in your response.

If I have not said so already, I, and my students, value very much your texts and your online resources.  It has been on my bucket list for some time to really get into signal processing and I had always been thwarted with other sources and their impenetrability or incompleteness.  We are humming along nicely using your books, videos, and code to learn these concepts.  This type of work is so important for training the next generation and I am glad folks like you are willing to take that on.

Robin

[Mike X Cohen]

Regarding the teaching opportunity, it might be an interesting idea to have students debate the use of the Laplacian (or perhaps other spatial filters more broadly, including dipole fitting, beamforming, etc.). The Laplacian is simply a spatial filter, and it is not guaranteed to make the data better in all situations. A discussion of its pros and cons could be informative (although it's a bit of a lopsided debate -- there aren't too many papers arguing against it).

I'm glad you are finding the book and online material useful. I think I can make public that I have a new book coming out. It will be called "Matlab for brain and cognitive scientists" and is a textbook for Matlab programming. Not surprisingly, there is a big focus on data analysis, but more on the implementation side than on theory side. There are already several great intro-Matlab books, but I think they are all too beginner-level. This one is written for PhD/postdoc level. It should be in print in early 2017, also from MIT Press.

Mike

--

You received this message because you are subscribed to the Google Groups "AnalyzingNeuralTimeSeriesData" group.
To unsubscribe from this group and stop receiving emails from it, send an email to analyzingneuraltimes...@googlegroups.com.
Visit this group at https://groups.google.com/group/analyzingneuraltimeseriesdata.
For more options, visit https://groups.google.com/d/optout.