[Mne_analysis] summary performance for ICA

[Chen, Yu-han]

Hi Dennis,

Thanks a lot for youre reply. Please see my answers below:

From: Denis-Alexander Engemann <denis.engemann at gmail.com<mailto:denis.engemann at gmail.com>>
Date: Monday, November 3, 2014 2:43 PM
To: The Children's Hospital of Philadelphia The Children's Hospital of Philadelphia <cheny4 at email.chop.edu<mailto:cheny4 at email.chop.edu>>, "<Mne_analysis at nmr.mgh.harvard.edu<mailto:Mne_analysis at nmr.mgh.harvard.edu>>" <Mne_analysis at nmr.mgh.harvard.edu<mailto:Mne_analysis at nmr.mgh.harvard.edu>>
Subject: Re: [Mne_analysis] summary performance for ICA

Hi Yuhan,

I include the mailing list in this part of our conversation because this might be interesting for other people as well.


2014-11-03 20:30 GMT+01:00 Chen, Yu-han <CHENY4 at email.chop.edu<mailto:CHENY4 at email.chop.edu>>:
>
> Hi Dennis,
>
> I (...) figured out that the SSP was turned on during epoching. Our original problem is that when we tried to epoch raw data, ICA cleaned data, ECG signal, the averaged ECG epoch is flat (zeros across the entire epoch). When we set proj = False for mne.Epoch, the problem is resolved. Is it true that by default, Epoch.py will set the SSP on?
>

Yes, you need to take care of this. Anyways I'm surprised you're getting flat signals. What kind of SSPs did you include?

*** I didn't intentionally include any SSPs when I read in the data so I don't know what type of SSP was included.


>
> The data I feed in is the Elekta data with maxfiltered tsss and movement correction applied.
>

Normally you have SSPS included in the raws if you don't use maxfilter. But think there's no point in using those if you have maxfiltered your data.
If you aim to identify biological artifacts using ICA there is no point in having ECG/EOG projections.
When using maxfilter and ICA, you should not necessarily use SSP in addition.
FYI I'd recommend to look up the number of dimensions estimated by maxfilter from the log or recover it by estimating the rank of the data and using this as input to ICA(n_components).

*** I wasn't planning to use SSP. In my case the SSPS was off when I read in maxfiltered raw data. For some reason when I epoch the raw data, by default the SSP was set on. So I think the Epoch.py function just set SSP on. Should I always assume the SSP is on when I epoch any type of raw Elekta data no matter the maxfilter was applied or not?

*** the n_components I used for ICA is 95% assuming 5% of our data is noise. I realized I don't have the number of dimensions extimated from the maxfilter log. Is this something I should specifically set when I run maxfilter? Also could you please tell e how to get the estimated ran of the data?




HTH,
Denis

PS: what about the performance metric, still interested in contributing it? ;-)

*** I'm still using the exact same code from Juergen and I thought that was already shared and uploaded to gitHub?


> Thanks!
> Yuhan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20141103/e2a7aed5/attachment.html