Re-reference and ICA

I have EEG recordings with 19 channels of 70 children in resting-state eyes opened/closed conditions. When I re-reference, I get rank of 18, and ICA purposes me rank 18 for its analysis. I thought about skipping re-referencing, so I can have 19 components for ICA. I used two reference electrodes on earlobes during recordings (sum 19 + 2 reference =21 electrodes). Is re-referencing necessary?

Hi Ninka. It’s best to run ICA in the same reference that you want to use for analyzing the data. I don’t think knocking off 1 rank will hurt the decomposition. Usually, matrix decompositions suffer in severely reduced-rank data, for example if you had 19 channels and rank=4.

An alternative approach is to use PCA to compress the data down to 18 dimensions. Then the ICA will be on the full-rank data, and you can pass the IC weights back through the PCA to get back into channel space. I believe most ICA functions in eeglab have an option for that. Definitely the jade algorithm supports that.

1 Like

Great, if you think that 18 rank would not hurt decomposition, than I will just stick to the ICA proposition. Thank you very much for your answer!

Sorry, it looks like I still have some confusions… My preprocessing pipeline is:

  1. Set high-pass filter to 1hz, and a low-pass filter to 30hz (I’m going to analyze only alpha band, even if I use the same data later, I would not have the need for frequencies below 21hz and under 4hz).
  2. import channel locations
    3 re-reference
    4 reject big artifacts by visual inspection
    5 run ICA

I now added one more step, by applying CleanLine on 30hz. After this step my rank is back to 19. If I re-reference again it will go to 18 again, but can I just run ICA without re-referencing again? So my pipeline would look like this:

1 set filters
2 set channel locations
3 re-reference
4 remove big artifacts
5 CleanLine

That pipeline looks good to me! I would add a step 2.5 to check for bad electrodes. If an entire electrode is recording pure noise, then re-referencing would put that electrode’s noise into all the other electrodes.

Thank you very much for your answer and advice about bad electrodes!