Iām currently working on single-trial GLM analysis of SEEG data in the high-gamma band. My initial steps involve extracting epochs and calculating the time-frequency representation (TFR) for each trial. However, I have some uncertainties regarding the baseline correction process.
Currently, I epoch the data from -1.5s to 2s around stimulus onset and apply baseline correction from -1.5s to 0s for each trial (this is inherently done in mne.Epoch, which subtracts the mean amplitude of the pre-stimulus EEG signal I guess). After this, I perform Morlet wavelet TFR for each trial and crop the data to the -1s to 1.5s window. I then apply apply_baseline(baseline=(None, 0), mode=??) to the trial-level TFR.
The TFR data would be used for subsequent GLM analysis on frequency band power. More specifically, I extract the mean time series from the high-gamma band frequencies and compute the mean time series for each trial and contact. I then conduct GLM on each time point, using high-gamma activity as the dependent variable and behavioral indices as independent variables.
For my purpose, I have a few questions:
Is my procedure for TFR analysis appropriate in this context?
I know that MNE offers multiple baseline correction modes, including mean, ratio, logratio, percent, zscore, zlogratio. Which one would be most suitable for my analysis, and why? Most iEEG papers seem to use mean or zscore, but their descriptions are often ambiguous, leaving me uncertain about the rationale.
Should I crop the data before or after applying baseline correction?
Should I consider the baseline mode when epoching the data?
I havenāt worked with SEEG, but I have with MEG. Hereās my advice:
I would not recommend using apply_baseline(baseline=(None, 0), mode=??) to the trial-level TFR. (see: General question: Applying baseline). I believe itās best not to over-clean the data, as we risk losing valuable signals. Iād only clean them if there were clear artifacts. However, keep in mind that opinions vary and some believe baseline correction can increase statistical sensitivity.
I suggest using mode=āratioā or āzratioā instead of āmean.ā This makes data unitless, and easy to compare between modalities like MEG and EEG, Also, it deals better with the brainās 1/f effect (see: [FieldTrip] TFR Baseline Correction Type).
During epoching of your data with mne.Epoch, I think you canāt specify a mode. I think it automatically subtracts the mean amplitude of the pre-stimulus EEG signal, as you already pointed out.
Thank you for your detailed reply; it was very helpful!
Since most SEEG studies use baseline correction, so Iāve applied it here as well. However, Iām still unclear about the different baseline correction methodsāspecifically, how they generally impact the results and when to choose one over the other.
Here are my two figures from my TFR-based GLM result with ratio (top) or zlogratio (bottom) correction for TFR. The ratio based result looks quite unusual. Do you have any insights
on this?
I would not say that the āratioā based result looks unusual. Just different from the ālogratioā one. With a robust analysis both methods should give the same results. So something should be off in your results
Note that the two methods have different values on the y axes. But this is expected. If you plan to contrast the two conditions (ācurrentā vs āpreviousā) the value on the y axis is irrelevant. You only care about the difference between the y-values.
I would suggest to plot the TFR (y:freq., x:time, colorcode:power) and see if the two methods match. Then we can further discuss how this might affect your statistics/regression coefficients.
A thought I had is that you use GLM which is a linear model. This linear model will change depending on the way you calculated your power (log vs ratio - note that log and ratio are both non-linear methods). Maybe for your analysis the best mode to use is the āmeanā which is a linear method.
Just as a note, in MEG/EEG we do not recommend applying baseline correction per trial since one trial has a lot of variance and by doing baseline correction you introduce only variance to your data.
It is better to average across trials and then perform baseline correction.
