Time-frequency problems with aperiodic control

Discussions
, , ,
1

Hi everyone,
I'm working on resting-source-space EEG data, and extracting the powers in different individual frequency bands (defined from the iAPF). To do this, I use a custom Matlab script that runs on files generated by Brainstorm's Specparam plugin, specifically :

  • Peaks for the periodic components,
  • Offset and Exponent for the aperiodic.

Here are the main features of my pipeline:

  • Source reconstruction with sLORETA, MNE and unconstrained sources,
  • Atlas: Destrieux,
  • Extraction of periodic powers per band from "Peaks" files only,
  • Control of aperiodic activity via Offset/Exponent files.

My concern: for all the bands, I found powers very close to 1 (between 0.90 and 1.10), with little inter-subject variability. As if the data had been normalised somewhere, or the informative spectral component flattened.

Does Brainstorm apply automatic normalisation to files from Specparam?

Is this expected when extracting power from "Peaks" files?

If I want to recover raw powers, or at least powers that are comparable between subjects, do I have to start from another type of file or modify something in the Specparam pipeline?

Thank you very much in advance for your feedback!
I can share an extract from the script if required.

2

No, there is not normalization.

Yes it is. The amplitude of the peaks field in the FOOOF parameter are given in Bells i.e., log10(power). But this data is Bells that will be added to the periodic component.

Check this post and linked code were we go on detail on how to assemble the FOOOF parameters to get the parametrized PSD:

3

Hi Raymundo,
Following up on this, about the fact that there is no normalization of the PSD from Specparam, is there an alternative to achieve this using Brainstorm (IRASA-type)?
Thank you

4

Hi @thhinault, By normalization, I meant that the PSD is not scaled by the total power before running specparam (aka FOOOF).

Well, this is specparam, by modeling the PSD modelling with it, it is possible to distinguish between the pariodic and aperiodic compoments of the PSD:
https://neuroimage.usc.edu/brainstorm/Tutorials/Fooof

If you want to normalize the PSD, you can use the process Stardardize > Spectrum normalization to normalize to relative power (scale by the total power), or apply the 1/f compensation.

I hope this makes it clearer.

5

So, just to confirm, to analyse periodic activity in the signal, is it the correct approach to:
-Run specparam on the PSD
-Extract the peak field
-Convert the Bells data
If yes, what is the most appropriate convertion of Bells data?
Thank you
Thomas

6

This is correct. Please check this post and the script that is linked, it uses the parameters from the computed FOOOF model to reconstruct the aperiodic and periodic components
Questions about fooof - #8 by Raymundo.Cassani

power [Bell] = log10(power [u^2/Hz])

power [u^2/Hz] = 10^power [Bell]