Biomedical Engineering Reference
In-Depth Information
3.3.2
Wrapper Approach
The wrapper approach, in contrast to the fi lter approach, uses methods in which
activity features are selected to maximize the performance of a decoder. Typically,
features are included or excluded recursively, until the maximal performance of a
predetermined decoder is reached, as in an example of ECoG-based motor imagery
classifi cation (Ince et al.
2009
). The observations that the thresholds used to select
features are usually arbitrary and the fact that different data sets may require differ-
ent settings of thresholds are weaknesses of this approach (De Martino et al.
2008
).
Furthermore, the decoder needs repeated training in selecting features, which could
be very time consuming.
3.3.3
Integration Approach
The integration approach simultaneously addresses the problem of feature selection
and decoder design. In this strategy, feature selection is included as part of decoder
design, ensuring effi cient use of data and faster computation time, as the decoder
does not need to be trained repeatedly during feature selection. This is particularly
advantageous for a large number of features and limited training data.
Here, we will outline three multivariate approaches that do not require user-
specifi ed thresholds for feature selection and have demonstrated potential for
decoding high-dimensional ECoG data. The fi rst one is partial least squares
regression (PLS-R), which aims to extract the latent structures in both observed
and predicted variables, so that their relation is maximized in terms of the cova-
riance structure. PLS-R can be implemented using the standard MATLAB
Statistics toolbox. By using PLS-R on scalograms (see Sect. 3.2.2) of ECoG
signals, we have successfully decoded high-dimensional continuous arm move-
ments with high accuracy, which was comparable to SUA-based decoding (Chao
et al.
2010
) ([data are freely available at neurotycho.org Nagasaka et al.
2011
).
The second approach is sparse logistic regression (SL-R), which imposes spar-
sity constraints to obtain a small subset of nonzero model coeffi cients. This
method has been used to classify whole-brain fMRI data (Ryali et al.
2010
;
Yamashita et al.
2008
) and to reconstruct auditory stimuli from our 128-channel
ECoG system (data not shown). SL-R can be implemented using the Sparse
Logistic Regression (SLR) toolbox (Yamashita et al.
2008
). The last, but not the
least, potential approach is variational Bayesian least squares regression
(VBLS-R), which is computationally effi cient and suitable for large amounts of
very high-dimensional data. VBLS-R was used to predict EMG activity and end-
effector velocity from motor cortical activity (Ting et al.
2008
). VBLS-R can be
implemented using MATLAB codes downloadable from
http://www-clmc.usc.
