Hello
I created two spectral density plots in EEGLAB from the same participant's data. One corresponds to a resting-state control condition, and the other to a meditation condition. Both recordings were 5 minutes long with eyes closed, and they underwent identical pre-processing steps, which included:
1. Manual rejection of bad channels and major artifacts (e.g., movements).
2. Filtering (0.5–75 Hz).
3. Re-referencing to the averaged signal.
4. Independent Component Analysis (ICA) decomposition.
5. Rejection of artifact components using EEGLAB’s default flagging function, followed by visual inspection.
The control spectrum was processed on Computer A, which has the Signal Processing and Statistics and Machine Learning toolboxes installed, while the meditation spectrum was processed on Computer B, which does not have these toolboxes.
As expected, the two spectra are similar in shape, but their power density scales differ drastically. Specifically:
- The control condition processed on Computer A shows a shift down in power density, with some values even being negative (which is physically implausible).
- This discrepancy is not due to the different conditions, as it is not observed in other participants.
The issue seems to arise from differences between the two computers or the toolboxes available. Has anyone encountered a similar issue or can provide insight into the potential cause of this problem?
Hi! Here you are inspecting two conditions (1, 2) x two toolbox types (A, B) x two signal processing paths (not the same artefacts might have been removed for conditions 1 and 2). I would suggest taking the exact same condition (after identifcal manual rejection), and only perform steps 2 and 3 afterwards, once on computer A and once on computer B. Then you would get evidence regarding what the effects of the computer/toolboxes are. Otherwise, there are too many confounds to isolate the cause of the differences.
Despite that, it is interesting to see that the relative levels (~-5 to 10 dB in Fig. 1, vs. 25 to 40 dB in Fig. 2) are similar.
Hi Suddha Sourav thank you for your answer, you had a very good point. I'm now sharing the spectrograms of the same recording (same person, same condition) done on computer A and B after completing the same pre-processing path. As you can see the problem is still present with this aberrant scale in Fig1. If you have an idea what the problem could be, I'd be happy to hear.
Dear Fabio Carbone , I have now tested that even on the same computer and with the same file, using signal processing toolbox vs. not using it creates an almost constant level difference of about 25 dB (As in your case, the topographies are similar). Two exceptions are the start and the end points.
Furthermore, it seems that with and without the signal processing toolbox (I set "usepwelch" to 0 in spectopo.m), the frequency axis is also not exactly the same. For the difference plot I had to interpolate them to the same frequency axis.
I think the scaling for the two different methods might not be proper, and that you might have uncovered a bug here. Would you like to report it at https://github.com/sccn/eeglab/issues?
Meanwhile, it would be crucial I think to use the version with the signal processing toolbox!
Dear Suddha Sourav thank you so much for your help!
I think I found the origin of the problem. As indicated here https://github.com/sccn/eeglab/blob/develop/functions/sigprocfunc/spectopo.m the function spectopo.m is used by default if the signal processing toolbox is installed and returns the power spectral density. Otherwise the function spec.m is used, the output of which is the power spectrum and not the PSD. The units are then different but the topography should be similar. See also https://github.com/sccn/eeglab/issues/172
Thanks again!
- Badges
- Science topic
- Similar topics
- Books