Exploratory Data Analysis Methods for fMRI - Biomedical Signal Analysis: Contemporary Methods and Applications - page 261

8.3 Other Analysis Models

Before continuing to other biomedical applications, we briefly want to
review other recent work of the authors in this field.
TheconceptofwindowICAcanbeusedfortheanalysisoffMRI
data[133]. The basic idea is to apply spatial ICA in sliding time windows;
this approach avoids the problems related to the high number of signals
and the resulting issues with dimension reduction methods. Moreover,
it gives some insight into small changes during the experiment which
are otherwise not encoded in changes in the component maps. We
demonstrated the usefulness of the proposed approach in an experiment
where a subject listened to auditory stimuli consisting of sinusoidal
sounds (beeps) and words in varying proportions. Here, the window ICA
algorithm was able to find different auditory activation patterns related
to the beeps (respectively, the words).
An interesting model for activity maps in the brain is given by sparse
coding; after all, the component maps are always implicitly assumed
to show only strongly focused regions of activation. Hence we asked
whether specific sparse modeling approaches could be applied to fMRI
data. We showed a successful application to the above visual-stimulus
experiment in [90]. Again, we were able to show that with only five
components, the stimulus-related activity in the visual cortex could be
nicely reconstructed.
A similar question of model generalization was posed in [263]. There
we proposed to study the post-nonlinear mixing model in the context of
fMRI data. We derived an algorithm for blindly estimating the sensor
characteristics of such a multisensor network. From the observed sensor
outputs, the nonlinearities are recovered using a well-known Gaussian-
ization procedure. The underlying sources are then reconstructed using
spatial decorrelation as proposed by Ziehe et al. [296]. Application of
this robust algorithm to data sets acquired through fMRI leads to the
detection of a distinctive bump of the BOLD effect at larger activations,
which may be interpreted as an inherent BOLD-related nonlinearity.
The concept of dependent component analysis (see chapter 5) in
the context of fMRI data analysis is discussed in [174], [175]. It can be
shown that dependencies can be detected by finding clusters of depen-
dent components; algorithmically, it is interesting to compare this with
tree-dependent [12] and topographic ICA [122]. For the fMRI data, a