MARA Tutorial
April 23, 2014
Contents
1 Introduction 2
1.1 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Citing MARA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 MARA in brief 3
2.1 Process Chain for ICA Artifact Rejection . . . . . . . . . . . . . . . . 3
2.2 MARA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Installation 5
4 Tutorial 6
4.1 Loading a sample EEG set . . . . . . . . . . . . . . . . . . . . . . . . 6
4.2 Starting MARA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.3 MARA’s options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.3.1 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.3.2 After classification . . . . . . . . . . . . . . . . . . . . . . . . 7
4.4 Running MARA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.5 Visualising components with EEGLAB . . . . . . . . . . . . . . . . . 8
4.6 Visualizing components with MARA . . . . . . . . . . . . . . . . . . 10
4.7 Removing rejected components . . . . . . . . . . . . . . . . . . . . . 13
4.8 Automatic Mode of MARA . . . . . . . . . . . . . . . . . . . . . . . 14
4.9 Command Line Usage . . . . . . . . . . . . . . . . . . . . . . . . . . 14
5 Contact and Support 14
1
1 Introduction
This is the tutorial for MARA, an open-source EEGLAB plugin which automatizes
the process of hand-labeling independent components for artifact rejection [8]. The
core of MARA is a supervised machine learning algorithm that learns from expert
ratings of 1290 components by extracting six features from the spatial, the spectral
and the temporal domain. Features were optimized to solve the binary classifica-
tion problem ”reject vs. accept”. Thus, MARA is not limited to a specific type
of artifact, and should be able to handle eye artifacts, muscular artifacts and loose
electrodes equally well.
MARA is written in MATLAB and makes extensive use of EEGLAB’s function-
ality including its excellent visualization tools [2].
1.1 Requirements
In addition to the requirements of EEGLAB (http://sccn.ucsd.edu/eeglab/),
MARA needs the Matlab Statistics Toolbox, the Optimization Toolbox and the Sig-
nal Processing Toolbox.
MARA requires channel locations with labels in the form FPz, F3, Fz, F4, Cz,
Oz, . . ., and components need to be at least 15s long for the computation of the
temporal feature. Note that MARA might lead to suboptimal results on narrow
band-passed filtered data, as the spectrum features are calculated on the power
spectrum between 2 Hz and 39 Hz. (You can deactivate the frequency features if
you are familiar with MATLAB scripting.)
1.2 Citing MARA
Like EEGLAB, MARA is free software distributed under the GNU General Public
License [1]. If you find this toolbox useful, please refer EEGLAB and the following
publication:
I. Winkler, S. Haufe, and M. Tangermann. Automatic classification of
artifactual ICA-components for artifact removal in EEG signals. Behav-
ioral and Brain Functions, 7:30, 2011.
2
2 MARA in brief
Artifact rejection with Independent Component Analysis (ICA) is based on the idea
that artifactual components and neuronal activiy are generated independently. The
EEG signal is decomposed into independent source components (ICs) in the hope
that artifactual components are separable from brain signals. MARA’s goal is to
identify the components that are predominantly driven by artifacts. The EEG can
then be reconstructed without the artifactual components.
2.1 Process Chain for ICA Artifact Rejection
The typical process chain for ICA artifact rejection consists of the following steps:
1. A rough pre-cleaning of the data by e.g. channel rejection and trial rejec-
tion may be performed. This step is usually helpful for obtaining a good
ICA decomposition. EEGLAB uses statistical thresholding to suggest epochs
to reject from analysis and provides a rich class of functions to do so (such
as Tools>Automatic Channel Rejection or Tools>Automatic Epoch Rejec-
tion). For more information, see http://sccn.ucsd.edu/wiki/Chapter_01:
_Rejecting_Artifacts. EEGLAB authors suggests to run ICA twice, first
to reject epochs based on the IC time courses, second to obtain a good ICA
decomposition for the cleaner data.
2. As ICA decomposition is known to be sensitive to slow drifts, high pass filtering
the data (at 0.5 Hz or even 2 Hz) can sometimes improve the quality of the
decomposition. In EEGLAB, this can be achieved using Tools>Filter the
data>Basic FIR filter. For more information, see http://sccn.ucsd.edu/
wiki/Chapter_04:_Preprocessing_Tools.
Note that MARA might lead to suboptimal results on narrow band-passed fil-
tered data, because its spectrum features are calculated on the power spectrum
between 2 Hz and 39 Hz.
3. Tools>Run ICA calculates ICA decompostion (it might take a while), see
http://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA.
The option ’pca’ can be set to perform a dimensionaliy reduction prior to IC
computation. Such a step may be helpful, in order to reduces the noise level
and avoid an unnatural splitting of sources. (It also makes IC computation
faster and reduces and the number of components that have to be labeled.)
4. To study and identify artifactual componets by visual inspection, EEGLAB
provides the Tools>Reject data using ICA>Reject components by map func-
tion. EEGLAB also comes with a powerful class of plotting functions under
the Plot menu.
This is the step where you might want to use MARA, as it automatizes the
identification of artifactual components.
3
5. After artifactual ICs have been identified, the EEG can be reconstructed with-
out them. In EEGLAB, this can be called via Tools>Remove components.
2.2 MARA
MARA realizes automatic IC classification by a linear pre-trained classifier. It is
based on the following six features which were determined in a feature selection
procedure described in [8]. One feature aims to detect outliers in the time series
of an IC, three features are extracted from the spectrum, and two feature extract
information form the scalp map of an IC.
1. Current Density Norm: ICA itself does not provide information about the
locations of the independent source components. However, ICA scalp maps
can be interpreted as EEG potentials for which the location of the sources
can be estimated. MARA considers 2142 locations arranged in a 1 cm spaced
3D-grid and seeks the source distribution with minimal l2 -norm [5, 6] (i.e. the
simplest solution of the forward problem formulated according to [3, 4, 7]).
Since this source distribution can model cerebral sources only, it is natural
that artifactual signals originating outside the brain can only be modeled by
rather complicated sources. Those are characterized by a large l2 -norm, which
MARA uses as a feature.
2. Range Within Pattern: The logarithm of the di↵erence between the mini-
mal and the maximal activation in a scalp map. Spatially localized scalp maps
stemming from e.g muscle artifacts or loose electrodes are typically character-
ized by a high Range Within Pattern.
3. Mean Local Skewness: The mean absolute local skewness of time intervals
of 15 s duration. This feature aims to detect outliers in the time series.
4. and Fit Error: These two features describe the deviation of a compo-
nent’s spectrum from a prototypical 1/f curve and its shape. The parameters
k1 , , k2 > 0 of the curve
k1
f 7! k2 (1)
f
are determined by six points of the log spectrum: (1) the log power at 2 Hz, (2)
the log power at 3 Hz (3) the point of the local minimum in the band 5–13 Hz,
(4) the point one Hertz below the third point of support (5) the point of the
local minimum in the band 33–39 Hz, and (6) the point one Hertz below the
fifth point of support. MARA uses the logarithm of and the mean squared
error of the approximation of f to the real spectrum in the 8–15 Hz range as
features.
The spectrum of muscle artifacts, characterized by unusual high values in the
20–50 Hz range, is approximated by a comparatively steep curve with high .
5. 8-13 Hz: The average log band power of the ↵ band (8–13 Hz). This feature
aims to detect the typical ↵ peak in components of neural origin.
4
3 Installation
To install MARA, download the compressed plug-in file, uncompress it, and paste
it into your EEGLAB plugins folder. EEGLAB will now automatically recognise
the plugin. The next time you start EEGLAB, the following line should appear in
your matlab command window:
EEGLAB: adding plugin function "eegplugin_MARA"
After loading a data set, MARA can now be found in the tools menu, Tools>IC
Artifact Classification (MARA) :
5
4 Tutorial
This tutorial takes you step by step through the MARA plugin. For this tutorial,
we use the sample data EEGLAB data epochs ica.set that you should find in the
sample data folder provided with your EEGLAB distribution. Decomposition into
IC components has already been performed on this data set.
We assume that you are familiar with EEGLAB and ICA decomposition, which
are very well documented in http://sccn.ucsd.edu/eeglab/. More information
on ICA decomposition in EEGLAB in particular can be found in http://sccn.
ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA.
4.1 Loading a sample EEG set
Select File>Load exisiting dataset and browse to your sample data folder. Select
EEGLAB data epochs ica.set, press ’Open’ and the sample data is loaded.
4.2 Starting MARA
To start MARA, go to Tools>IC Artifact Classification (MARA)>MARA classifi-
cation
6
The following main menu will open:
4.3 MARA’s options
4.3.1 Pre-processing
If you have not pre-processed your data or run ICA yet, you can select to do so by
ticking the corresponding checkboxes in the MARA main menu, ’Filter the data’ and
’Run ICA’. Note that obtaining a good ICA decomposition is a difficult problem,
and you might consider more elaborate pre-processing as outlined in Section 2.1.
1. ’Filter the data’ calls the same function as Tools>Filter the data>Basic FIR
filter. A pop up appears asking you to provide filtering options. After filtering,
you will be asked a name for the new data set.
Note that MARA might yield suboptimal classification results on narrow-
bandpass filtered data as some of its features are based on the power spectrum
between 2 Hz and 39 Hz.
2. ’Run ICA’ calls the same function as Tools>Run ICA. A pop up appears
asking you to provide ICA options. After ICA computation (which might take
a while), you will be asked for the name of the new data set.
ICA is already performed on the sample data set we analyse in this tutorial, so
we will not go through these steps.
4.3.2 After classification
MARA o↵ers the following three options that concern the course of action after
MARA classified the components:
7
• The default option is to use MARA only to mark components for rejection.
The components are not removed and not visualized. You can use standard
EEGLAB functions for further investigation and re-labeling of the components.
• If you check ’Plot and select components for removal’ MARA will provide its
own visualisation of the IC components which allows you to quickly change
the classifier’s proposed rating. The components will not be removed, you can
later inspect components with EEGLAB’s plotting functions.
If you check ’Visualize Classification Features’, an additional window will dis-
play the six features that MARA classification is based on for each component.
• If you check ’Automatically remove components’ you accecpt MARA’s decision
blindly. The components will be automatically removed.
4.4 Running MARA
Pressing ’Ok’ on the MARA main menu will start MARA classification and produce
the following output on the Matlab command window:
MARA is computing features. Please wait
.................................
Features ready
MARA marked the following components for rejection:
3 11 21 23 24 26 28 31
That’s it. If you click Tools>Reject data using ICA>Reject components by map
you can see the components marked for rejection in red.
4.5 Visualising components with EEGLAB
After MARA component classification, you can study the selected (and not selected)
components using EEGLAB functionaliy that is well described in the EEGLAB Tu-
torial http://sccn.ucsd.edu/wiki/Chapter_09:_Decomposing_Data_Using_ICA.
Here, we just give a very brief overview.
Go to Tools>Reject data using ICA>Reject components by map.
8
The following window will pop up, where components selected for rejection are
marked with red buttons:
You can plot the properties of each component and change its label (Accept /
Reject) by clicking on the rectangular button above each component’s scalp map:
In the examples above, IC3 is a typical eye blink artifact (strong frontal activi-
tion, steep power spectrum), IC7 is a neural component (alpha peak around 10 Hz,
scalp map indicates an occipital brain source), and IC31 is a typical muscle artifact
(spatially localized acitivity, high power above 20 Hz).
9
To identify artifacts by visual inspection, it is also helpful to inspect the compo-
nents time courses, which you can access via Plot>Component activations (scroll).
4.6 Visualizing components with MARA
MARA provides a visualization function that allows for a quick change of the clas-
sifiers proposed ratings and gives an overview of each component’s scalp map and
power spectrum. After MARA computation, you can call it from the menu by
Tools>IC Artifact Classification (MARA)> Visualize Components.
Alternatively, you can perform both MARA and its visualization by checking ’Plot
and select components for removal’ in the MARA main menu.
10
The visualization MARA provides shows each component scalp map, its spectrum,
the current label of the component (artifact or not), as well as the probability of each
component of being an artifact as computed by MARA. You can change the label
of each component by checking or un-checking the checkbox above each component.
Mark all the ICs you wish to reject from the data and press ’OK’. You can now
continue inspecting the ICs with EEGLAB functions as described in Section 4.5. To
substract their activiy from your data, go to Tools>Remove components as outlined
in the follwoing section.
11
MARA features
For information purposes, windows that display for each component the six features
that MARA classification is based on will pop up:
Features that contribute towards a component to be marked as an artifact are
plotted as red bars, features that indicate the component contains neuronal activity
are marked in blue.
These features may be helpful in understanding MARA’s decision. Recall the
following from Section 2.2: A high FitError and a high 8-13Hz feature indicate
that the component contains an ↵ peak and thus neuronal activity. A high value
indicates a spectrum with high values in the 20-50 Hz range, which is a sign that the
component contains muscle activity. A high Local Skewness value indicates outliers
in the component’s time series. A high Range in Pattern and Current Density Norm
indicates a scalp map that is unusual for a neuronal component.
Lets consider the example components we visualized in Section 4.5.
1. The eye blink artefact component IC3 is characterized by strong outliers in
the time series (rare blinks) and contains no ↵ peak. MARA labels IC3 as an
artifact due to its low FitError, low 8-13Hz, and high Local Skewness.
2. The neuronal component IC7 contains an ↵ peak, no strong outliers in the
time series and a smooth, dipole-like pattern. All features of MARA indicate
it is a neuronal component.
3. The muscle artifact IC31 is recognized by MARA due to its lack of ↵ peak
(low FitError, low 8-13Hz ) and its very localized scalp map (high Range in
Pattern and high Current Density Norm )
12
4.7 Removing rejected components
To remove the rejected components, go to Tools>Remove components
Press ’Ok’ to remove the ICs from data.
EEGLAB will ask for confirmation and a name to the new data set.
13
4.8 Automatic Mode of MARA
You can automatically remove components without inspection by clicking ’Automat-
ically remove components’ in the MARA main menu. In this case, there will be no
visualization of the components, and you will be directly forwarded to EEGLAB’s
remove components menu.
EEGLAB will ask for a confirmation and a name of the new data set.
4.9 Command Line Usage
You can start MARA from the command line as follows:
[ALLEEG,EEG,CURRENTSET] = processMARA(ALLEEG,EEG,CURRENTSET)
For more information, please see its help function.
5 Contact and Support
If you have questions, suggestions or bug reports, please contact irene.winkler@tu-
berlin.de
[email protected] 14
References
[1] http://www.gnu.org/licenses/gpl.txt.
[2] A. Delorme and S. Makeig. EEGLAB: an open source toolbox for analysis of
single-trial EEG dynamics including independent component analysis. Journal
of Neuroscience Methods, 134(1):9 – 21, 2004.
[3] V. Fonov, A. Evans, R. McKinstry, C. Almli, and D. Collins. Unbiased nonlinear
average age-appropriate brain templates from birth to adulthood. NeuroImage,
47, Supplement 1(0):S102 –, 2009.
[4] V. Fonov, A. C. Evans, K. Botteron, C. R. Almli, R. C. McKinstry, and D. L.
Collins. Unbiased average age-appropriate atlases for pediatric studies. Neu-
roImage, 54(1):313 – 327, 2011.
[5] M. S. Hämäläinen and R. Ilmoniemi. Interpreting magnetic fields of the brain:
minimum-norm estimates. Med. Biol. Eng. Comput., 32:35–42, 1994.
[6] S. Haufe, V. V. Nikulin, A. Ziehe, K.-R. Müller, and G. Nolte. Combining spar-
sity and rotational invariance in EEG/MEG source reconstruction. NeuroImage,
42:726–738, 2008.
[7] G. Nolte and G. Dassios. Analytic expansion of the EEG lead field for realistic
volume conductors. Physics in Medicine and Biology, 50(16):3807, 2005.
[8] I. Winkler, S. Haufe, and M. Tangermann. Automatic classification of artifactual
ICA-components for artifact removal in EEG signals. Behavioral and Brain
Functions, 7:30, 2011.
15
