Independent Component Analysis (ICA)
Independent component analysis (ICA) is a signal processing method to
separate Independent sources linearly mixed in several sensors. The first ICA
paper was published in 1986 and has been applied to EEG since 1996. At
Integrated Neuroscience Services we apply ICA to your EEG data using the
MATLAB based EEGLAB toolbox, developed by Arnaud Delorme and Scott
Makeig, to produce maximal statistically independent components that
present specific patterns of Coherence. Which can then be used to generate
a specialized protocol for your clients. Utilizing ICA with EEG data
distinguishes the issue of source localization from that of source
identification
ICA works by unmixing brain sources and artifacts detected by sensors
to produce the true underlying source components or waveforms. From these
source components, brain components are retained and artifact components
are removed, producing a clean EEG. This process is necessary because the
waveforms observed at each electrode site is a weighted sum of many
source waveforms. The wave form observed at a given electrode site is a
sum of each source waveform multiplied by the weighting factor between
each component and that electrode site. This weighting factor is derived
from the position of the source relative to the electrode, and the conductivity
of the tissue. The distribution of source waveforms on the scalp can then be
understood as a matrix of weights, Source X Electrodes. Noise or artifact
which are independent from neurogenic source waveforms are independent
sources included in the mixing matrix.
After the weight matrix is applied to the EEG, ICA generates the time-course
of ICs, scalp topography maps and dipole locations for each IC. The time
course indicates the strength of each component at any given time in
microvolts, the scalp topography describes the strength of the IC at each
electrode site, and the dipole shows the modeled location and positon of ICs
in the brain. ICs will often display a dipolar distribution on scalp topography,
which is indicative of a good component. This dipole is different from but
consistent with the current dipole, which represents the activity of functional
brain regions. These images in combination with other information allows us
to distinguish IC components reflecting underlying true components of
functional brain regions from artifacts. The corrected EEG is produced by
multiplying the time-course of the ICs by the inverse of the weight matrix.
The artifacts are removed by zeroing out their time-course in the
multiplication process.
ICA unmixes these sources by forming a weight matrix, Components X
Electrodes. This matrix produces a set of weighted values which is then
multiplied by the EEG time-course to produce statistically independent
components (ICs) that may or may not correspond to true neural sources.
These IC waveforms are analogous to the original source waveforms pre-
mixing at sensors. However, because ICs are set equal to the number of
channels in the EEG, multiple source components can be collapsed into one
IC and one source component can be split into multiple ICs. Hence, the
quality of data at the recording of the EEG is important to reduce the
collapsing and splitting of artifacts into ICs and to maximize production of
true components.
The corrected EEG is then ran through a connectivity analysis of
coherence (multivariate grange-causality) and graph theory analysis of
connectivity; using MVGC MATLAB toolbox, which is based on advanced
vector auto-regression theory. Granger- Causality is a statistical notion of
causality applied to time-series data, whereby, cause precedes, and helps
predict effect. Defined in both time and frequency domains and allows for
conditioning out of common influences. Otherwise, a test for determining
whether one time-series is useful in predicting another. Modern neuroscience
takes a network-centric approach to describing brain function and cognition.
In which information, flows between multi-leveled, evolving network
structures. Using the dipoles produced by ICA as vectors in a network model,
multivariate granger-causality provides information about the level of
functional communication between regions; the flow of information, its effect
size, and the strength of communication between vectors.
An additional means of measuring brain connectivity is graph theory
analysis. When applied to EEG we can better understand the directional
connectivity of brain networks. We define the nodes of the graph as the
dipoles and the edges as the functional connectivity between these dipoles.
The graph describes the direction of communication between nodes, and
weights are applied to the edges to represent the strength of the links
between the nodes. Furthermore, network measures such as cluster
coefficient, and path length are extracted from the graph theory for further
interpretation. All of these data and images better allow us to produce a
specialized protocol for your clients. Interpretations of these data and images
are presented and reviewed during the online video consolation.
View a sample report below!