I have an Motor Imagery EEG dataset with size of (576, 751, 22) where it is (num. of trial, time points, num. of channel). I am doing research on reducing the number of channels. I would like to find a similarity measure between original data and channel-reduced data before performing any classification or deep learning method. For example, I would like to select 8 channels out of 22 channels, and then I want to know how similar or how mathematically close the original data (576, 751, 22) to channel-reduced data (576, 751, 8). How can I do this? Thank you.
Hi Ali. You can compute the similarity between two datasets with different numbers of channels through a canonical correlation analysis (CCA).
But I think a better approach would be to use PCA to compress the 22 channel data down to 8 components. That will preserve the most amount of variance in the data.
Thanks for your response.
In the second paragraph, do you mean that: 1) first apply PCA to reduce channel from 22 to 8, and then use CCA method to calculate the similarity between original data (selected 8 channels) and PCA applied data.
Or do you mean: 2) just use PCA to reduce channels and construct new virtual channels, do not reduce channels manually.
I am trying to reduce the number of channels because I want the BCI system run faster by reducing the number of channels manually. So I am trying to select which channels carry most information from original data.
Ah, I see. I initially meant your second interpretation. But from your clarification, manually picking channels according to what will be available in the production EEG setup is the best idea.