[Mne_analysis] Discussion: robust, linear and polynomial detrending

[Marijn van Vliet]

        External Email - Use Caution        

Dear Dip,

It is not that hard to create Raw and Epochs data structures with custom 
data. You do this using the RawArray and EpochsArray classes, see here:

https://mne.tools/stable/auto_tutorials/simulation/plot_creating_data_structures.html

Note that you can just re-use the .info from the original raw and can 
skip the parts on how to create a custom Info object.

Once you get comfortable creating these datastructures, you can explore 
different ways to de-trend the data as you please, with the usual 
numpy/scipy functions.

best,
Marijn.


On 24/02/2020 10:54, Diptyajit Das wrote:
>
>         External Email - Use Caution
>
> Dear MNE community,
>
> I would like to discuss the following issues:
>
> Data background: Our data (MEG-EEG) is often composed with slow drifts 
> (mostly due to street artifacts). We are interested in finding out 
> cortical activity in lower frequency states, therefor a high pass 
> filter option (>.5-1hz) is not a optimal solution for us since it 
> removes most of the signals that we are interest in.
>
> Some of the highlighted points:
> 1. Current high filter setting has set to (.1 hz as lower cutoff), we 
> can see the spatial map of the drift components in the ICA 
> decomposition but ICA decomposition for other artifacts (eye 
> blinks/cardiac) is not stable which we already expect at this point. 
> Removing the drift components directly from ICA and then rerun the ICA 
> for the second time to remove other biological artifacts sound 
> suspicious to me. I am not an expert of ICA decomposition but as far I 
> know, removing PCA components will alter the linear projection of the 
> sources. How valid is the idea of ruining ICA for two times?
>
> 2. MNE offers a dc and a liner detrending option to remove show drifts 
> at the epochs/evoked level. But often I have seen linear detrending 
> removes activity that is coming from cortical sources  (i.e., not a 
> optimal fit for our data), in fact  a lower order polynomial fit is 
> actually more reasonable.  A recent work by de Cheveigné A 
> <https://www.ncbi.nlm.nih.gov/pubmed/?term=de%20Cheveign%C3%A9%20A%5BAuthor%5D&cauthor=true&cauthor_uid=29448077> 
> (https://www.ncbi.nlm.nih.gov/pubmed/29448077) [method: weighted 
> robust detrending] is a nice idea to tackle this kind of issue. The 
> algorithm has been adapted in 'MEEGKIT' package 
> [https://nbara.github.io/python-meegkit/auto_examples/example_detrend.html#sphx-glr-auto-examples-example-detrend-py] 
> but lately I have figured it out that it still needs more testing. 
> Now, coming back to my question: Is there any way to implement such 
> algorithm directly at the raw data? So far I am not able to modified 
> the raw mne structure with synthetic values (after removing the trend 
> with a polynomial fit) while keeping the same info directory. Any 
> advice or suggestion would be helpful at this stage.
>
> 3. Is there any way to apply a linear detrending only to a sub set of 
> the data (either meg/eeg even though they are combined at epochs 
> level) ?, so far its implemented in evoked level as I have seen. It 
> would be ideal that if I can somehow apply the linear detrending only 
> to a subset of the data at the epochs level (i.e., meg in my case) 
> instead of doing it at evoked level. I am willing to contribute in 
> this case.
>
>
> Best,
> Dip
>
> _______________________________________________
> Mne_analysis mailing list
> Mne_analysis at nmr.mgh.harvard.edu
> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/mne_analysis
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20200224/d62a30ab/attachment.html