| title | tags | authors | affiliations | date | bibliography | ||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
meeg-tools: A framework for EEG/MEG data processing and analysis |
|
|
|
11 January 2022 |
paper.bib |
The meeg-tools package provides a semiautomatic and reproducible framework for
electroencephalography (EEG) and magnetoencephalography (MEG) data processing and analysis.
Here, we present a semiautomatic and reproducible framework for preprocessing and
analyzing EEG/MEG signals. It aids researchers from the cognitive neuroscience community
to analyze data without having a lot of expertise in working with EEG/MEG signals.
The preprocessing tutorial reduces the manual steps required to clean the data based on
visual inspection, and the subjectivity in rejecting segments of data or interpolation
of electrodes. The analysis pipeline is an executable Python script that provides a
complete solution to perform high-level time-frequency or connectivity analysis on clean data.
This package utilizes modules from MNE-Python [@10.3389/fnins.2013.00267], a popular open-source
Python package for working with neurophysiological data. For automated rejection of
bad data spans and interpolation of bad electrodes it uses the Autoreject [@Jas2017]
and the Random Sample Consensus (RANSAC) [@10.3389/fninf.2015.00016] packages.
The general use-case of the package is to use it from a Jupyter notebook.
The tutorials folder contains notebooks that demonstrate data operations and
that are described in the Background section.
In order to remove high-frequency artefacts and low-frequency drifts, a
zero-phase band-pass filter (0.5 - 45 Hz) is applied to the continuous data
using MNE-Python [@10.3389/fnins.2013.00267]. This temporal filter adapts the filter
length and transition band size based on the cutoff frequencies.
The lower and upper cutoff frequencies can be changed in the configuration
file (config.py) located at the utils folder.
Subsequently, the filtered data is segmented into non-overlapping segments ( epochs) to facilitate analyses. The default duration of epochs is one seconds, however it can be changed in the configuration file.
The removal of bad data segments is performed in three steps. First, epochs are
rejected based on a global threshold on the z-score (> 3) of the epoch variance
and amplitude range. To further facilitate the signal-to-noise ratio,
independent components analysis (ICA) is applied to the epochs. ICA is a
source-separation technique that decomposes the data into a set of components
that are unique sources of variance in the data. The number of components and
the algorithm to use can be specified in the configuration file. The default
method is the infomax algorithm that finds independent signals by maximizing
entropy as described by [@Bell1995; @doi:10.1142/9789812818041_0008].
Components containing blink artefacts are automatically identified using
MNE-Python. The interactive visualization of ICA sources lets the user decide
which components should be rejected based on their topographies, time-courses
or frequency spectra. The number of components that were removed from the data
are documented in the “description” field of the epochs instance “info”
structure. The final step of epoch rejection is to apply Autoreject [@Jas2017]
on the ICA cleaned data. Autoreject uses unsupervised learning to
estimate the rejection threshold for the epochs. In order to reduce computation
time that increases with the number of segments and channels, autoreject can be
fitted on a representative subset of epochs (25% of total epochs). Once the
parameters are learned, the solution can be applied to any data that contains
channels that were used during fit.
The final step of preprocessing is to find and interpolate outlier channels.
The Random Sample Consensus (RANSAC) algorithm [@10.1145/358669.358692]
selects a random subsample of good channels to make predictions of each channel
in small non-overlapping 4 seconds long time windows. It uses a method of
spherical splines [@Perrin1989] to interpolate the bad sensors.
Additionally, the EEG/MEG reference can be changed to a “virtual reference” that
is the average of all channels using MNE-Python.
This research was supported by the National Brain Research Program (project 2017-1.2.1-NKP-2017-00002); Hungarian Scientific Research Fund (NKFIH-OTKA K 128016, to D.N.); IDEXLYON Fellowship of the University of Lyon as part of the Programme Investissements d'Avenir (ANR-16-IDEX-0005) (to D.N.). We thank Zsofia Zavecz for her useful comments and suggestions.