Apply or not baseline correction for Time-Frequency analysis

Hi everyone,

I’m working on a time-frequency EEG analysis with MNE-Python with quite long epochs (9.6 seconds). I was wondering if baseline correction is recommended in this context, and if yes, at which point in the analysis is it best to do so: before or after the time-frequency transformation ?

This is what I know so far about baseline use:

  • In evoked potentials (ERPs), baseline correction helps to eliminate variations in the initial baseline potential, ensure a stable reference and centers ERP variations around zero. It allow to enable deflections to be interpreted more easily.

  • In time-frequency (TF) analysis, it seems that baselines are not always necessary. This is because oscillations are continuous, not strictly dependent on a resting potential, and spectral power measures relative amplitudes rather than absolute variations in potential. That said, a baseline can sometimes be used in TF to normalize spectral power or to control power variations between trials.

I understand that this can depend on the nature of the data, but given the length of my epochs, I’d like to know what the best practices are for reliable time-frequency analysis with long epochs.

Thanks for your feedback and advice!
Johan

Hey,
this is a very late reply, but maybe it will help someone.

Opinions on baseline correction in both ERPs and TF analysis are quite diverse (even in the realm of ERPs, opinions are diverse, see around the 8th minute here: https://www.youtube.com/watch?v=2wS7-XILNso).

You are right that baseline correction is not always necessary. Actually, baseline correction in TF analysis is usually used for visualization only, statistical analysis is still performed on non-baseline corrected data.

Baseline correction in TF analysis is used due to the 1/f nature of EEG spectra. Without baseline correction, it is difficult to see differences over time because the power of low frequencies is usually much greater than the power of high frequencies. However, you can separate 1/f (aperiodic) activity from oscillatory (periodic) activity by using specparam. See our recent paper for the idea (Modulation of aperiodic EEG activity provides sensitive index of cognitive state changes during working memory task, Figure 2 & 3).
Time-resolved FOOOF decomposition is now also available in specparam:
07: Fitting Models over Time — specparam 2.0.0rc3 documentation

Alternatively, if you are specifically interested in oscillations, you can use a measure of rhytmicity instead of power: Rhythmicity of neuronal oscillations delineates their cortical and spectral architecture | Communications Biology