Biomedical Engineering Reference
In-Depth Information
Reliability of Connectivity Measures
One important question when analyzing connectivity is which connectivity measures
can capture “true” underlying interactions between or among signals. Although all
measures have been initially tested in simulations and applied to neurophysiological
data, the answer to this question remains unclear because of the lack of a thorough
systematic comparison. A recent development tested different multivariate effective
connectivity measures in combination with different numerical signifi cance compu-
tational approaches on simulated data with various noise levels, data lengths, model
orders, and coupling strengths (Florin et al.
2011
). The results suggest that the
squared magnitude of PDC combined with LOOM is the most reliable and robust
choice. However, it is important to keep in mind that LOOM is problematic for data
with a small number of trials. Nonetheless, choices of features and parameters are
crucial, and great caution should be taken in practice.
3.3
Online Analysis
The goal of online analysis is to map activity features (observable variables) to the
underlying neural processes of interest (predicted variables) in real time, which is
designed primarily for BMI applications. Similar to offl ine analysis, one crucial
step of online analysis is
feature selection
, i.e., to identify the relevant subset of
activity features. Another important step is the design of a
decoder
(regressor or
classifi er) that can effi ciently translate the selected features to continuous variables
(e.g., movement trajectories) or discrete variables (e.g., cognitive states). Based on
different strategies of feature selection and decoder design, the following approaches
are usually implemented for high-dimensional data: the fi lter approach (Sect.
3.3.1
),
the wrapper approach (Sect.
3.2
), and the integration approach (Sect.
3.3.3
).
3.3.1
Filter Approach
In the fi lter approach, activity features are selected independent of decoding, and the
selected features are then used in the design of a decoder. The features can be
selected according to their relevance, such as correlation between each feature and
the predicted variables (as in an example of ECoG-based hand-trajectory decoding
Schalk et al.
2007
), or by their statistical signifi cance, as in the offl ine analysis
described in Sect.
3.2
. Typically, the fi lter approach is applied to univariate mea-
sures, such as TFRs; therefore, it does not consider the relationships between fea-
tures while selecting them. As ECoG data are inherently multivariate, with strong
spatial correlation between activities from different brain areas, it is also useful to
select features based on connectivity measures. The fact that this method does not
consider the performance of the decoder during feature selection is a drawback of
the fi lter approach.
