Douwe Horsthuis 2023-01-04
This is still a work in progress. This Repo will contain the full pipeline to analyse resting state data, in Matlab using EEGlab. The pipeline will contain everything from pre-processing to creating microstates and doing a power/frequency analysis. This is work from Ana Francisco Douwe Horsthuis and Filip de Sanctis in discussion with Sophie Molholm.
Table of Contents
- About The Project
- Getting started
- Resting state project
- Pipeline
- Power Frequency Analysis
- Microstates
- Extra functions
- 2023 microstate paper
- License
- Contact
- Acknowledgement
All of these scripts are created to run in matlab. They use EEGlab
- EEGlab
- ERPlab (eeglab plugin)
- XDFimport (eeglab plugin)
- IClabel (eeglab plugin)
- MST Plugin (microstate plugin)
This pipeline is build to analyse data collected using the Restingstate paradigm in our lab. This paradigm is essentially a 5 min trial with a fixation cross with the text “Please fixate on the cross” followed by a 5 min trial that has the text “please close your eyes”. However, some of our data is collected by telling the participant instead to just look at the center of the screen, or to keep your eyes closed for 5 minutes. For this reason, the pipeline might add the triggers to keep the eyes open and eyes closed data separate.
This data is collected for a lot of different groups. The script is set-up in a way that you can analyse 1 to 3 groups in one go. However, there is room to add another group or simply to run the script again for a different group.
This is normally pretty straight forward in EEGlab. But in this case we have 3 different scripts, or variations of the same script.
The mostly used script is the A_bdf_merge_sets. This simply takes the Raw data (.bdf) and turns it into a .set file
Then there is the A_bdf_non_paradigm_merge_sets. Since some of the data is collected without a paradigm but is saved into 2 separate files without triggers, this script solves that. It loads the .bdf file that ends in _open.bdf and adds a eyes open trigger to the start. It does the same for the .bdf file ending in _closed with the exception that here it adds and eyes closed trigger. Lastly it merges the two files into one.
The third script load XDF files. These are files that have both EEG data
and Optitrack movement data.
This script loads the xdf file (1 file per participant) and deletes the
data that is not in the EEG channels or in the first channel. The first
channel has all the trigger info. The script uses this channel to add
the trigger to the EEG data.
This script is the first of the pre-processing scripts. It runs all the
people in order of their group.
One of the issues we encountered was that some participants had their
data collected using the wrong configuration file. This is taken care
of.
The data is down-sampled from 512Hz to 256 Hz.
Externals are all deleted since not everyone has externals. So we cannot
use them as a reference.
We apply a 1Hz and 50Hz filter. We add channel info to all the channel.
For this we use the following 3 files: standard-10-5-cap385, BioSemi160,
BioSemi64. The first 2 are from BESA and have the correct layout. The
3rd is needed for the MoBI data. You can find these in the Functions and
files folder (inside the src folder).
Lastly this script uses eeglab’s clean_artifacts function deletes the
bad channels and bad parts of data. Channels will get deleted by the
standard noise criteria, if they are flat over 5 seconds and the
function checks if channels are overly correlated with each other.After
that it devides the data in 0.5 second epochs. In these it looks if
there data has peaks that go over 35 standard deviantion of the channels
amplitlude. If so these get deleted. Because it’s possible that between
2 of these moments there is a very short amount of data, we make sure to
delete all the continues data if between 2 boundries there is less then
2 seconds of data. We got to the criteria of 35, because 20 is the
default setting, but because resting state data is different from task
based data, we had to find out a way that we still had enough data left,
without it being noisy.
This script plots all the data in EEGlab as continues data and allows you to delete channels manually.
This script will double check and fix any potential trigger issue we
encountered. It saves a Matrix with the information for each individual
participant. This script can be skipped It is only useful for
documenting triggers. We added the
pop_rejcont
in the next script and this deletes triggers sometimes, so we need to
double check triggers again (see G_preprocces5).
Since off the 12/6/2021 update this is not the case. The clean_artifacts
takes care of noisy continues data. But it might be good to run either
script since they are quick. If you want to save time, skip this one.
This script loads a file with all the original channels, deletes the externals and uses these file locations to interpolate the channels of the corresponding subject’s data. In the case of 160 channel data, it uses the transform_n_channels function to interpolate the remaining channels not to the original 160, but to 64 channel data so that it is the same as all the other data. For this to work Matlab needs to know the location of 2 things, the trannsform_n_channel.m file and the EEG files called 64.set and 64.fdt. We chose to interpolate before the ICA because like this we can still use the ICA weights for all the channels, and since we are setting a PCA (amount of ICs we want the ICA funtion to create) we account for the interpolated channels not being used for this.
This script will do an average reference.
This is followed by an Independent Component
Analysis.
We use the pca option to prevent rank-deficiencies, as best as we can.
We set it to be all channels before interpolation - 1. We don’t want
more because this gives “Ghost ICs” that are duplicates of existing
ones. We also exclude an extra IC because of the avg ref, hence the -1.
After his we delete eye, muscle and channel noise components by using
IClabel. IClabel will only delete the
component if it has more than 70% eye data or 80% muscle or 70% channel
noise and always less then 10% brain data. Normally we only delete eye
components (>0.8), and we arrived at this criteria after comparing (for
a different dataset) how many components we (Ana, Douwe and Filip) would
delete manually and what threshold would get the closest to that.
However, since this is different data, and it’s un epoched so its harder
to clean, we decided to also delete extra components.
In this script, we first make sure that the triggers are still in the right place. It is possible that the triggers got deleted if the corresponding continues data were too noisy. If they got deleted, the scripts calculates what the time of the onset of that deleted part of data was and uses that instead as the latency of the trigger. This should already be fixed before, but this is a double check.
After that we use the the pwelch function of Matlab and a log transformation of the results to get the power frequency results.
for now we are only using 1 pre-selected channel to save data. This data can either be saved as an excel file, which is easy to use for stats in different platforms, or as a table or individual variable in Matlab.
The main Microstate analysis follows the instructions as written On the
website of the
developer
Thomas Konig
of the microstate toolbox that we use. Using his microstate plugin, and
ragu
software
the final part was analyised. The plugin itself has a script
RunMicrostateAnalysis which we used to do the analysis for everyone.
Because this script uses both 160 and 64 channel collected with biosemi caps we needed to get the to a same format. To do this we are using the transform_n_channels function documented here. In this case it turns all the 160 channel data into 64 channels. But it will keep the original channels that are closed to the location of the corresponding 64ch cap.
All the code used for this paper is located in the 2023 paper folder. While almost everything is the same as the pipeline there are some small differences, for example, we did not have to use the XDF files nor did we use 160 channel data (only 64 channels) so no transformation has been used.
Distributed under the MIT License. See LICENSE for more information.
Ana Francisco - ana.alvesfrancisco@einsteinmed.org Douwe Horsthuis - douwehorsthuis@gmail.com Filip De Sanctis - pierfilippo.sanctis@einsteinmed.org
Project Link: resting-state-analysis-pipeline-microstates-frequency