Content uploaded by Andrew X Stewart
Author content
All content in this area was uploaded by Andrew X Stewart on Apr 24, 2018
Content may be subject to copyright.
Content uploaded by Andrew X Stewart
Author content
All content in this area was uploaded by Andrew X Stewart on Apr 13, 2018
Content may be subject to copyright.
1
The LIMO EEG toolbox:
extending the statistical analysis of EEG data to the spectral domain
Andrew Stewart1, Guillaume Rousselet2 & Cyril Pernet1
1 Brain Research Imaging Centre, University of Edinburgh, Edinburgh, UK
2 Centre for Cognitive Neuroimaging, Institute of Neuroscience and Psychology, University
of Glasgow, Glasgow, UK
Introduction
LIMO EEG is an open source Matlab toolbox, EEGLAB compatible, which provides a suite
of tools based around statistical LInear MOdelling and robust estimators of EEG data [1].
The main idea behind the toolbox is to analyse the entire information-space, in time or
frequency domain.
In the 1st version of the toolbox, ERP could be analysed at all electrodes and time points.
Here, we report the extension of these tools into the spectral domain: (i) for power spectra,
showing what frequencies are represented across trials, (ii) for Event-Related Spectral
Perturbations (ERSP), showing the power spectra in time and frequency spaces.
Stable version: https://gforge.dcn.ed.ac.uk/gf/project/limo_eeg/
Beta version: https://github.com/LIMO-EEG-Toolbox
Features of the toolbox
Four key features distinguish the LIMO EEG toolbox:
(1) Hierarchical modelling: within and between subjects’ analyses (Figure 1).
(2) Robust 2nd level statistical tests: between subjects’ analyses rely on trimmed means,
iterative reweighted least-squares and multivariate approach for repeated measures.
(3) Multiple Comparison Correction methods based on bootstrap estimates of the maximum
statistics (corrected T/F values, cluster-mass, threshold free cluster-enhancement [2]).
(4) Visualization tools for the whole data space and detailed course plots in 2D and 3D
(Figures 3, 4, 5).
2
Figure 1. Illustration of the hierarchical linear modelling approach used in LIMO EEG.
Illustration and data analysis
The LIMO EEG data set was analysed in the time and frequency domains. Briefly, 14
subjects ranging from 21 to 68 years of age performed a 2AFC task, choosing between 2
faces presented at different noise levels ([3], [4], Figure 2). Subjects performed over 1000
trials and 128 channels EEG data recorded. The statistical analysis consisted in modelling,
per subject, the EEG response with 1 categorical variable (face A or face B) and 1 continuous
variable (noise level). At the group level, a one-sample t-test and a regression analysis were
computed to test the effect of noise level across subjects.
3
Figure 2. Experiment used for the LIMO EEG data set. Subjects discriminated between 2 faces with various
levels of noise. The grand mean ERP show a strong modulation of the N170/P200. Analyses were carried out by
modelling, for each subject, the EEG response as a function of stimulus type and noise level across stimuli.
ERP Analyses
Each subject showed significant modulations of the N170/P200 components by the level of
stimulus noise, and at the group level, significant effects were observed very early (within
100 ms) and in the N170 time-window (Figure 3). A regression analysis showed that the
effect of stimulus noise was stronger with age later on (P200), i.e. it takes longer for older
subjects to process stimulus information [4].
4
Figure 3. Results showing the effect of noise on face ERP. The top panels show the effects in one subject and
the bottom panels show the effect of age on the modulatory effect of noise (regression). Maps show the F values
at p<0.05 corrected for multiple comparisons using TFCE.
Power spectrum Analyses
Significant modulations of the power spectrum were observed in most subjects, but no
significant effect emerged at the group level (Figure 4). This can be explained by the change
in ERP shape / stimulus noise relationship with age, which shows up as a significant
modulation of power at ~17 Hz by stimulus noise.
5
Figure 4. Results showing the effect of noise on face evoked Power. The top panels show the effects in one
subject and the bottom panels show the effect of age on the modulatory effect of noise (regression). Maps show
the F values at p<0.05 corrected for multiple comparisons using TFCE.
ERSP Analysis
The time-frequency decomposition showed again significant effects at the subject level
(Figure 5) but failed to show significant effect across subjects. Non-corrected results suggest
modulation effect in the alpha range, as expected from ERP and Power analyses.
6
Figure 5. New LIMO EEG interface for time-frequency analysis. Results show the effect of noise on ERSP to
faces in one subject. Maps show the F values at p<0.05, corrected for multiple comparisons using TFCE. The
topographic plot shows the F value distribution on the scalp at 14Hz. Under the map, the blue trace shows the
frequency distribution at electrode B8.
Discussion
LIMO EEG offers a unique combination of tools to analyze ERP, Power and ERSP through a
GUI, or batch mode, or command line. The data import is done automatically using the
EEGLAB format but the code is flexible enough to process data from any software. The only
‘major’ problem for now is that Matlab is memory hungry. While ERP and power analysis
can be performed on almost any new computer, processing the full time-frequency space
requires large memory (8Gb minimum), preferably in combination with solid state hard
drives (store and retrieve matrices from ~1Gb to ~4Gb). Similarly, bootstrapping can be slow
when the search space is large. Future work will incorporate the analyses of component form
ICA so that again, the full data space can be analysed.
References
[1] Pernet et al. (2011) LIMO EEG: a toolbox for hierarchical LInear Modeling of
EletroEncephaloGraphic data. Computational Intelligence and Neuroscience, Volume 2011,
Article ID 831409.
[2] Pernet, et al. (2014) Cluster-based computational methods for mass univariate analyses of
event-related brain potentials/fields: a simulation study. Journal of Neuroscience Method,
submitted
[3] Rousselet et al. (2008). Parametric study of EEG sensitivity to phase noise during face
processing. BMC Neurosience, 9, 98-120.
[4] Rousselet et al. (2009). Age-related delay in information accrual for faces: Evidence from
a parametric, single-trial EEG approach BMC Neuroscience, 10: 114
