Comparative evaluation of Independent Components Analysis algorithms for isolating target-relevant information in brain-signal classification
Jeremy Hill, Michael Schroeder, Thomas Navin Lal and Bernhard Schölkopf
In: Brain-Computer Interface Technology: Third International Meeting, 14-19 June 2005, Rensselaerville, NY.
We present results from two binary BCI experiments. The first is an motor imagery experiment with 5 subjects (classification based on frequency features of the signal---essentially mu activity) and the second is a paradigm based on covert shifts of attention to auditory stimuli with 15 subjects (classification based on averaged segments of the time series---essentially ERPs). We have previously reported good classification performance using Support Vector Machines in both cases (Lal et al 2004, Hill et al 2005). In addition, Recursive Channel Elimination (Lal et al 2004) has been reported to be an effective feature-selection tool for reducing the number of necessary features on both data sets, and in the motor-imagery experiment it reliably identifies the channels near the motor cortex as most important.
Here we demonstrate that, prior to classification, blind source-separation using Independent Components Analysis (ICA) consistently improves classification accuracy by roughly 10%. In addition, after recursive elimination of the Independent Components by a method analogous to that of Lal et al, the number of features required for classification is considerably smaller. For the motor imagery experiment, we compare a number of different variants of ICA. The differences between variants were slight in terms of error rate and feature reduction. Across different random subsets of the signal, we also examined the stability of the algorithms, in terms of variation in the spatial weighting of electrodes in the components ranked as most important by recursive elimination. InfoMax ICA (see Makeig et al, 1996) proved to be markedly the most stable of the variants tested, particularly for the best-performing subjects, reliably weighting the channels near the motor cortex most highly and in a consistent pattern. Its classification and feature reduction performance was comparable to or better than the other variants.
|EPrint Type:||Conference or Workshop Item (Poster)|
|Project Keyword:||Project Keyword UNSPECIFIED|
|Subjects:||Brain Computer Interfaces|
|Deposited By:||Jeremy Hill|
|Deposited On:||24 June 2006|