Single-trial baseline normalization for similarity analysis with EEG

Hello Mike,

First of all, thank you so much for all of the resources including books, youtube lectures and, this discussion list. It is always really helpful when I’m stuck.

Currently, I’m working on a similarity analysis with EEG data, which is not that common I think.
I’m interested in Encoding-Retrieval similarity, so I’m calculating the similarity between an encoding trial and its matched trial in a retrieval block. The similarity value is estimated by correlation and Euclidian distance. I converted the raw EEG to time-frequency domain and averaged the power values in four frequency bands (theta, alpha, beta and, gamma). And the similarity is estimated between power vectors in encoding trial A and retrieval trial A, respectively for each frequency band (i.e., alpha power vector in encoding trial A and alpha power vector in retrieval trial A is used to calculate similarity values.)

So far, I did not apply any baseline correction on my dataset, but I’m wondering if it is necessary. And if it’s necessary, should it be applied on single trials which seems tricky. Could you share your insight about this issue?

So, my questions are

  1. Is baseline correction is necessary for this similarity analysis between single trials?
  2. If so, what kind of baseline correction would be the best for single trials? (dB, z-score, linear subtraction…).

It would be really appreciated if you can share your idea. Please let me know if you need any further information about this analysis. Thank you.

Best,
Kyoungeun

Hi Kyoungeun. There have been several posts here about single-trial baselining, and I’ve also written about it in my ANTS book. In general, I’m not a fan. There is a paper about it, but in my hands it’s usually not a very stable approach.

That said, if you’re using correlation or any normalized metric, you don’t need to worry about it.