Blind Separation Reveals Hemodynamic and Electrophysiological Dynamics
Cognitive Science Student Seminar Series
University of California San Diego
May 21, 2001
Scott Makeig
The Salk Institute and
Institute for Neural Computation
University of California San Diego
http://www.sccn.ucsd.edu/~scott
Both the current primary human neuroimaging techniques, fMRI and high-density EEG/MEG, record complex changes in brain activity through time. Unfortunately, in both modalities it is difficult to separate the data into functionally distinct sources, which inevitably include several types of artifacts. Analysis techniques for functional magnetic resonance imaging (fMRI) data typically require a priori knowledge of the time course of the hemodynamic response (HR) and assume the homogeneity of HR and noise across trials, brain regions and subjects. In practice, none of these assumptions may be valid. Standard EEG averaging methods often ignore the fact the data sum the projections, through volume conduction, of synchronous activity within several cortical areas, and unquestioningly assume independence between ongoing and event-related EEG activity.
Since 1995, I and colleagues at Salk have been exploring the applications of blind source separation methods based on Independent Component Analysis (ICA) for detecting unforeseen HRs in event-related fMRI experiments and/or unforeseen spatiotemporal phenomena in event-related EEG experiments. Since its practical development about 1995, ICA has rapidly become a widely explored technique in modern signal processing, with new or potential applications in a wide range of fields. This is occurring because the idea behind ICA is in fact quite simple. Rather than finding phenomena in multi-dimensional data that most closely match some a priori model, ICA blindly separates the given data into the sum of maximally independent sources of information (in a particular sense). This proves to be possible under a range of conditions, and coincides (in a more general sense) with what one wants to be able to do with data, i.e. to separately examine functionally distinct phenomena of interest that the data contain.
I will show some current results of analysis of EEG and fMRI data that combine ICA with other visualization and signal processing methods. fMRI results suggest that HRs from fRMI experiments can be site, subject and task dependent and may vary widely from trial to trial. The causes of this variability probably include variations in subject cognitive strategy or state. Use of data-driven HR analysis methods such as ICA appears essential to identify and explore time-varying features. Likewise, the EEG results suggest new ways of understanding macroscopic event-related brain dynamics, including event-related potentials (ERPs).