[Eeglablist] R: ICA and dipfit: high residual variance

[Giovanni Vecchiato]

Dear Tarik and John,

Thanks a lot for your replies.
Your notes helped to obtain more plausible ICA results from the localization point of view.

Actually, I was using that processing (HP @ 0.1Hz) to match the suggestion given by Tim Mullen to perform the connectivity analysis via SIFT, which is my final goal.
He provided a sample data collected at 256Hz and just high pass filtered at 0.1Hz before performing the ICA, DIPFIT and the following connectivity analysis. 
That's why I initially opted for the same approach.

However, following your suggestion, I found benefits from a high pass filtering at 1Hz and a spatial downsampling (by eliminating the two more outer circumferences of electrodes), and a low pass at 70Hz which actually helped to obtain lower residual variance from the DIPFIT analysis. I used the PCA before computing ICA because I don't have enough trials to do it on all components; anyway, by computing ICA on the whole session (and not on trials) I did not find any improvement of residual variance.

After that, I start working on the connectivity analysis via SIFT and would like to share with the following observations.
My data are recorded with fs=500Hz and, according to the rule (M^2p)/(NtW)<0.1, I collected Nt=40 trials to estimate a Vieira-Morf model with W=0.3s, M=6 independent components and assuming a model order of 16.

By following the standard preprocessing of the tutorial, from the model order selection procedure I get the following values for the related criteria:

- hq=13
- aic=23

I set p=16 (which is the limit order I could apply according to the above rule) but the whiteness tests fail. Then, I tried to "force" the model fitting by using a p=23 but again the results of the following whiteness tests are negative. 

Then, I repeated the analysis on the downsampled dataset at fs=250Hz.
In this way, using the same parameters for the analysis, I get the following values for the model order selection criteria:

- hq=10
- aic=15

which are too high to satisfy the ratio parameters/datapoints. However, in this case when I put p=15 the whiteness tests are positive for all the windows, the model is stable for all the windows and the average consistency is above the 80%. So I assume the following connectivity patterns are valid, although the model has been estimated without respecting the parameters/datapoints ratio.

Therefore, my issues are the following:

1) Should I avoid to use fs=500Hz when working with SIFT? From the tutorial I understood that there could be problems with this analysis with fs>500 but at this point I assume that also having fs=500Hz is risky.

2) When working with fs=250Hz, can I rely on the goodness of the model validation results (whiteness, consistency, stability) even if the model has been estimated without respecting the ratio parameters/datapoints?

Sorry for this long email, but hope to have been clear enough.
Thanks again for your help.

Best,
Giovanni


-----Messaggio originale-----
Da: Tarik S Bel-Bahar [mailto:tarikbelbahar at gmail.com] 
Inviato: mercoledì 21 settembre 2016 20:05
A: Giovanni Vecchiato
Cc: eeglablist
Oggetto: Re: [Eeglablist] ICA and dipfit: high residual variance

Hello Giovanni, some notes below for you that should help a little.
Let the list know of your future success for dipfit-residual-variance-reduction!





******************************************************************************************************
Your main question should probably be "are my ICs and the dipfit results" computed correctly and accurate ? After you're sure about that, then  think about "ways" to "decrease" residual variance of dipfits. If you haven't yet, please be sure to walk through the online eeglab tutorial doing the same steps with the eeglab tutorial data.
This is useful for understanding how things should look and what to expect.

Review published articles, and poll current researchers, about what is the usual, acceptable, and OK levels of residual variance.

You should be careful about using both PCA and ICA. My understanding is the recommendation from eeglab is to just use ICA. This might be impacting your ICA-results quality.

Make sure you are not trying to get "lower" residual variance of dipole fits with "bad ICs".

Review the normal residual variance of dipfit for the ICs in the online eeglab tutorial datasets. Compare to your expectations. Compare to published dipfit variance results.

To get lower residual variance per IC, your IC scalp maps need to be "better" in terms of being more bipolar, etc..

You have a low amount of time in your recording, so you may want to downsample in channels.

I would say review your continuous pre-ICA and post-ICA data to make sure there is not noise or artifacts that can be removed that are being included, possibly influencing your IC quality (and thus dipfit).

I believe that Makoto's pipeline recommends a 1hz high-pass for ICA.
Also, it's often recommended to filter at 1-50 or 1-40 hz before doing ICA.

Probably best to test multiple settings on a few files, rather than a few settings on many kinds of files.
*******************************************************************************************************