14.02.2013 Views

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

Mathematics in Independent Component Analysis

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

1.6. Applications to biomedical data analysis 45<br />

1 2 3<br />

4 5 6<br />

(a) recovered component maps<br />

1 cc: −0.14 2 cc: −0.13 3 cc: −0.22<br />

4 cc: 0.89 5 cc: 0.12 6 cc: 0.09<br />

(b) time courses<br />

Figure 1.23: Extracted ICA-components of fMRI record<strong>in</strong>gs. (a) shows the spatial and (b)<br />

the correspond<strong>in</strong>g temporal activation patterns, where <strong>in</strong> (b) the grey bars <strong>in</strong>dicate stimulus<br />

activity. <strong>Component</strong> 4 conta<strong>in</strong>s the (<strong>in</strong>dependent) visual task, active <strong>in</strong> the visual cortex (white<br />

po<strong>in</strong>ts <strong>in</strong> (a)). It correlates well with the stimulus activity (b), 4.<br />

with the known stimulus task, see figure 1.24. The absolute correspond<strong>in</strong>g autocorrelations<br />

are 0.84 (stNSS), 0.91 (stSOBI with one-dimensional autocorrelations), 0.58 (stICA applied to<br />

separation provided by stSOBI), 0.53 (stICA) and 0.51 (fastICA). The observation that neither<br />

Stone’s spatiotemporal ICA algorithm (Stone et al., 2002) nor the popular fastICA algorithm<br />

(Hyvär<strong>in</strong>en and Oja, 1997) could recover the sources showed that spatiotemporal models can<br />

use the additional data structure efficiently <strong>in</strong> contrast to spatial-only models, and that the<br />

parameter-free jo<strong>in</strong>t-diagonalization-based algorithms are robust aga<strong>in</strong>st convergence issues.<br />

Other analysis models<br />

Before cont<strong>in</strong>u<strong>in</strong>g to other biomedical applications, we shortly want to review other recent work<br />

of the author <strong>in</strong> this field.<br />

In Karvanen and Theis (2004), we proposed the concept of w<strong>in</strong>dow ICA for the analysis of<br />

fMRI data. The basic idea was to apply, spatial ICA <strong>in</strong> slid<strong>in</strong>g time w<strong>in</strong>dows; this approach<br />

avoided the problems related to the high number of signals and the result<strong>in</strong>g issues with dimension<br />

reduction methods, and moreover gave some <strong>in</strong>sight <strong>in</strong>to small changes happen<strong>in</strong>g dur<strong>in</strong>g<br />

the experiment, which are otherwise not encoded <strong>in</strong> changes <strong>in</strong> the component maps. We demonstrated<br />

the usefulness of the proposed approach <strong>in</strong> an experiment where a subject listened to<br />

auditory stimuli consist<strong>in</strong>g of s<strong>in</strong>usoidal sounds (beeps) and words <strong>in</strong> vary<strong>in</strong>g proportions. Here,<br />

the w<strong>in</strong>dow ICA algorithm was able to f<strong>in</strong>d different auditory activations patterns related to the<br />

beeps respectively the words.<br />

An <strong>in</strong>terest<strong>in</strong>g model for activity maps <strong>in</strong> the bra<strong>in</strong> is given by sparse cod<strong>in</strong>g; after all, the<br />

component maps are always implicitly assumed to show only strongly focused regions of activation.<br />

Hence we asked the question whether the sparse models proposed <strong>in</strong> section 1.4.1 could be

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!