I am working with data that is very messy such that, for a given file, the decomposition will contain a noise artifact IC that accounts for >70% of the variance. Out of curiosity, I reconstituted the data without that high variance noise IC (the reconstituted data still looked very messy), then reran the reconstituted data through ICA and achieved a much more reasonable decomposition and fairly clean reconstituted data. Any thoughts as to whether this is an allowable method for cleaning data?
Hi Victor,
Have you posted this question on the EEG lab forum? Pretty much all questions are answered, and this would probably be relevant for other researchers too.
Best
Giorgia
Victor - without understanding ICA in full depth, I think that it is problematic inasmuch as by excluding one or several components, you reduce the dimensionality of the data. When you backproject this dimension-reduced data from component space to electrode space, however, your second ICA will be oblivious to this reduced dimensionality. In practice, when you have 64 electrodes or even more, this should not be such a big deal, but mathematically speaking, it is problematic, and it will probably come back to haunt you once you try to publish your work and get a critical reviewer.
An alternative approach that I have chosen in the past is to run a first ICA, identify the data segments that are particularly noisy (much easier to see in component space than in electrode space!), exclude those segments, and re-run the ICA.
HTH,
Marius
Giorgia Michelini Good idea! I'll be sure to do that.
Marius Keute Interesting! Can you speak more to why it would be mathematically problematic (or point me in the direction of some relevant literature)? I understand what you are saying about reducing the dimensionality of the data, but I don't quite understand why running an ICA oblivious to the dimension reduction would result in issues. Thanks for your reply!
As I said, I don't have a deep mathematical understanding of ICA, but I think the critical point is this: when you start with 64 channels, then your data has a maximum of 64 dimensions, i. e. 64 components will definitely be enough to convey the full information contained in your data. When you remove one (say, noise) component, only 63 dimensions are left. But when you back-project this to electrode space, you pretend that your data is 64-dimensional again, bc you want to restore your 64 channels. I think the mathematical expression would be that your data matrix is rank-deficient now (not sure if it's the correct term though). In the second ICA, you extract 64 components again, while there are really only 63 dimensions left in your data.
Anyway, I think that you can get a more mathematically substantiated answer if you repost this to the EEGLAB forum or FieldTrip mailing list, as Giorgia suggests.
Cheers,
Marius
- Badges
- Science topic
- Similar topics
- Philosophy
- Philosophy Of Science
- Reasoning