Standardization of EEG recordings collected across days in a within-subject analysis

[Nvar]

Hi everyone,

Great to be here and meet everyone. This is my first post and as the title suggests, I am seeking advice on standardizing EEG recordings collected from the same participant but across multiple days/sessions. A few words about the experimental design:

I am dealing with a patient cohort that is pretty rare, which makes data availability pretty tough. Patients are tested on a set of cognitive tasks, longitudinally across days. And this detail makes things complicated…

As a first step, I am computing time-frequency spectra and I am trying to assess if condition A is significantly different from condition B. Any significant differences will then be source-localized using beamforming methods. As this is a within-subject design, for argument’s sake let’s assume there are 4 recordings in total and each recording contains both conditions. I am trying to find an appropriate way to standardize the data, so that I can perform statistical analyses.

I am really unsure how to standardize my data prior to pooling and then testing condition A vs condition B. In a group-level design, I would compute session-specific dB-normalized maps, then average those, get a subject-specific time-frequency map, and then be on my way to SPSS. However, for a within-subject design, things get… complicated.

How do you approach such issues in within-subject designs? Is there an “ideal” way to standardize data prior to pooling and hypothesis testing?

Thank you a bunch for your input!

[tmedani]

Hi @Nvar & welcome to the BST forum

Brainstorm can cover most of what you’re trying to do.

For multi-session / within-subject designs, the recommended approach is to work with your data as a “mini-group analysis”, except that your “group” consists of the different sessions.

You can follow these brainstorming workflows:

• Do baseline normalization separately for each session’s time-frequency maps: Tutorial → Time-Frequency → Normalized time-frequency maps
  You can also check the group level analysis tutorial: https://neuroimage.usc.edu/brainstorm/Tutorials/GroupAnalysis

• Within-subject A vs B contrasts: For each session, average condition A and B, then compute A − B or any other metrics. You can also try doing statistics across the trials of each condition across sessions.
  You can refer to this tutorial:
  https://neuroimage.usc.edu/brainstorm/Tutorials/Workflows

  You can also refer to the following workshop: PracticalMeeg2025: Brainstorm hands-on tutorial

Best,
Takfarinas

[Nvar]

Hi Takfarinas,

I just got around to your message. Thank you for the response, that clears up things somewhat. I guess I was looking for guidance in the case where N (subjects), K (number of runs), or L (number of trials within run) is limited. For example, N=3 subjects with K=3 independent runs across days, each consisting of L=30 trials per condition. If I analyze individually the data per subject and compare means of conditions A-B, then I can only get a minimum p-value of p=0.125. I can compute a weighted average across conditions (i.e. by L) and get a grand-average per participant, but that route holds the same limitations (N=3 participants).

I’m fundamentally limited by the rare cohort of subjects analyzed in my domain, which is why I leaned towards standardizing data across runs prior to computing stats non-parametrically, on the trial and electrode level. Any suggestions that you might have are greatly appreciated!

Thank you!