[Eeglablist] Any tips for unsupervised (sic) ICA ECG / EOG component rejection?

[Panagiotis Tsiatsis]

Dear all,

I am using ICA with fieldtrip and eeglab in order to remove ECG 
components from my CTF MEG data. I have a separate EKG channel just like 
in the wonderful fieldtrip tutorial as well as vertical and horizontal 
EOG channels.

The question is a rather practical one: Some times I have to reanalyze 
my data (different filtering, trial duration settings etc.) and because 
I have something like 30 subject with 10 sessions of ~7mins each, it is 
kind of time consuming to redo the ICA component rejection manually each 
time. *Of course eventually this has to be done by inspecting the 
components*, but I was just wondering whether I could use some kind of 
automatized criterion until I reach this level where I can say that I do 
not need to apply ICA again and therefore I will do the rejection 
manually. So please excuse my audacity to talk about automation in a 
procedure which is traditionally considered "supervised".


A. Concerning ICA ECG artifact removal, I have the following questions:

So, I will just describe you a couple of ideas for heuristics and a few 
concerns so that maybe you could tell me how (in)sane they sound to you, 
as I have little experience with ICA.

1. One criterion could be to use the average absolute value of coherence 
of each component when compared to the ECG channels (as calculated in 
the fieldtrip ECG removal tutorial) within a certain and predefined 
range of frequncies. In my case I used the range of [2-40]Hz. Then I 
convert all the "scores" to z-values and set a threshold of 3-4 to 
identify components that are highly coherent with the ECG . Of course, 
this z -value threshold (and all the z-scores) depends on all components 
so I am not sure whether it is a criterion that can be fully justified.

2. Similarly to (1), another criterion could be the absolute value of 
the correlation in time of the ECG channel to all the other components. 
Maybe here the frequency range should be restricted as well. Again, the 
absolute correlation values would be transformed to z-scores and the 
artifactual components would be selecting based on a z-value threshold. 
I am just not sure whether correlation has to offer something 
complementary to coherence in this case.


3. A third criterion would be the combination of 1 and 2 - calculate 
z-values for coherence and correlation and then reject the intersection 
of the components marked as artifactuals from both 1 and 2.  This might 
be a bit conservative, especially because the correlation criterion does 
not seem as robust as the coherence criterion.

[Of course the rejected components would have to be evaluated 
a-posteriori in any case.]


On another note, my biggest concern at the time, stemming from the 
incomplete understanding of ICA and the use of an automated procedure, 
is the following: Let's say that in the decomposition, the ECGconsists 
of 2 (or 3 sometimes) components. The question is whether removing one 
component of the ECG only (the most prominent one) can actually do more 
harm that good since (and please correct me if I am wrong because I am 
not very confident here) the components that when appropriately combined 
according to the the mixing matrix would result in the ECG artifact 
might contain parts that "cancel out" when combined together. Therefore, 
if not all the "real" components that would constitute the ECG artifact 
are removed, there might  be "artificial" deflections that are removed 
along with the most prominent component, which would alter the data and 
induce again artifacts. I know that the way that I am phrasing this is 
not really clear and I might have misunderstood the mechanics of ICA so 
please feel free to correct me.


B. Concerning ICA EOG artifact removal, I wanted to ask you whether 
there is any way to apply similar techniques as the ones that you apply 
for ICA ECG removal to EOG components (i.e. the coherence criterion)? I 
tried to apply the coherence criterion by using the ft_artifact_eog 
function to detect the eyeblinks, followed by alignment of the detected 
EOG artifacts in the EOG channel (after extending them so that they have 
the same length) and their averaging in order to calculate the coherence 
of this "average" EOG blink to each of  the ICA components. It seemed to 
kind of work for the few datasets that I tried but it was not as robust 
as the ECG method - one of the reasons for this might be the variability 
of the blink types (ie. vertical blinks of different speed/duration and 
"depth", the existence of horizontal eye movements and the 
non-orthogonality of horizontal and vertical EOG traces etc.) Any 
intuitions again on this would be really useful.


My final question would be whether you think that it could be possible 
to improve things in terms of unsupervised artifact rejection via ICA?  
Maybe having some spatial templates that resemble common artifacts with 
robust topology and comparing the topologies of the ICA components to 
these templates in terms of spatial correlation would be feasible?


Apologies for raising so many issues in just one (rather long) e-mail, I 
am looking forward to hearing from you about your experience on the 
above issues.


Thank you in advance!


Kind Regards,
Panagiotis


-- 
Panagiotis S. Tsiatsis
Max Planck Institute for Biogical Cybernetics
Cognitive NeuroImaging Group
Tuebingen, Germany