[FieldTrip] Different results with source reconstruction from July 2018 and onwards
Schoffelen, J.M. (Jan Mathijs)
janmathijs.schoffelen at donders.ru.nl
Thu Dec 9 08:21:45 CET 2021
What you describe and show in the TFR figures looks like what can be seen on: https://www.fieldtriptoolbox.org/faq/why_does_my_tfr_look_strange/
Thus I think you are facing spectral leakage from (very) large amplitude low frequency signal components, that somehow are more prominent depending on the software version. This suggests that DC-offset removal is insufficient to avoid it in your data. If the time courses that you showed in an earlier e-mail are examplary for what the signals look like, I can imagine that the strong transients are causing this. (just as a side note: are you sure that those transients are a property of the data, and not an artifact of the analysis pipeline, e.g. trying to ‘glue’ together disjoint - yet temporally contiguous - segments which have passed through their own processing (filtering etc)?).
Anyway, this does not bring us any closer to an answer to the question why the two reported FT versions behave (slightly) differently. It’s annoying that you have been stuck on this for some time, but I would like to ask first why you want to stick with >3 year old versions? Is there a strong reason against upgrading to a much more recent version of the code base, and evaluate whether that performs well enough? If that is satisfactory, I think it would outweigh the intellectual satisfaction of solving the ‘different results puzzle’.
On 7 Dec 2021, at 16:43, Frans Nord via fieldtrip <fieldtrip at science.ru.nl<mailto:fieldtrip at science.ru.nl>> wrote:
Thank you so much for your quick response Jan-Mathijs!
I agree, the change is not huge when looking at individual level but it is definitely there and it makes a big difference later on in our analysis pipeline. Maybe this is a sign that our analysis method is weak but let's not get there yet.
I get this result when running the analysis on either an OS:X machine or a machine with Windows 10. The difference in output between the fieldtrip versions is there in both OS's.
I'll attach a calculated power spectrum on group level for pre and post commit so it becomes clearer. These are calculated with the same config file in ft_freqanalysis and then contrasted in the same way with the same classes that we look at. Sure, it's the same overall shape but there should not be any difference at all imo. To triple check what I'm doing I downloaded the two versions again today and ran them through our analysis and I still get the same difference.
This has been bugging me for close to a month now and I can't seem to figure it out on my own so your help is much appreciated!
On Mon, Dec 6, 2021 at 6:49 PM Schoffelen, J.M. (Jan Mathijs) via fieldtrip <fieldtrip at science.ru.nl<mailto:fieldtrip at science.ru.nl>> wrote:
Why do you say that there’s a ‘huge’ change? I hardly see the difference, that is the most prominent features in the time series are pretty comparable I’d say.
Just to be clear on this: are you comparing the results across operating systems? At this point I don’t see any changes in the code from 11 to 12 July 2018 that at first sight would affect your analysis. The ‘guilty’ commit is a merge of a PR by Sarang that pertains to (bookkeeping of data needed for) forward computations using openmeeg as a method. I don’t see how that would affect your analysis pipeline, if it is sufficiently close to the one on OSF that you refer to.
On 6 Dec 2021, at 16:59, Frans Nord via fieldtrip <fieldtrip at science.ru.nl<mailto:fieldtrip at science.ru.nl>> wrote:
I get different results with the same source reconstruction code when using Fieldtrip before 11 July 2018 compared to 12 July 2018. The results I get from before the twelfth seems to be more in line with previous found intracranial results and measurements on sensor level. I'm using eLORETA and the code is basically the same as this: https://osf.io/bgfs5/?view_only=ebb85532e1064d128db79bacb6be27f8<https://urldefense.com/v3/__https://osf.io/bgfs5/?view_only=ebb85532e1064d128db79bacb6be27f8__;!!HJOPV4FYYWzcc1jazlU!_AyLxWa-VXiG3g1vjZDuQdq7Ze9eST4_qjhy9H4t0vaBg2DIzlz-a283UfUGtsZL3SLJzRwNkeK-Jwyb22bpLfLvJvxdqgnfvxtpvQ$>
The 'pre' attached file is how the first trial for the first participant look with the version before july 11 and the 'post' is after july 11.
- Operating System: Windows 10 and Mac os:x Mojave
- MATLAB version: R2019a
- FieldTrip version: fieldtrip-171478a3e1c71f499ad0cf01f661dbb257ff5eb5 and fieldtrip-d1aa34272c61454dbbd9684f836c68b2dd26f337
What was it that justified such a huge change in the model? Can I read about it somewhere? Note that this is not the problem described in https://github.com/fieldtrip/fieldtrip/issues/1020<https://urldefense.com/v3/__https://github.com/fieldtrip/fieldtrip/issues/1020__;!!HJOPV4FYYWzcc1jazlU!_AyLxWa-VXiG3g1vjZDuQdq7Ze9eST4_qjhy9H4t0vaBg2DIzlz-a283UfUGtsZL3SLJzRwNkeK-Jwyb22bpLfLvJvxdqgk-MKNj-g$> but appeared in an earlier version.
fieldtrip mailing list
fieldtrip mailing list
fieldtrip mailing list
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the fieldtrip