Proceedings of the Cognitive Neuroscience Society meeting
April 18-24 - 2004 - San Francisco (CA, USA)
 
Physiological plausibility and stability of independent component analysis
 
Scott Makeig, Julie Onton, Arnaud Delorme
 

Institute for Neural Computation, University of California San Diego

scott@salk.edu, julie@sccn.ucsd.edu, arno@salk.edu

 
    Here, we tested the stability of independent component analysis (ICA) decomposition of the human eleectroencephalogram (EEG) using a new metric: the 'cumulative dipolarity histogram' (CDH) of the set of scalp maps associated with the recovered components. The projection to the scalp electrodes of synchronous activity within a cortical patch should highly resemble the projection of an equivalent current dipole. Thus the variance of the scalp map associated with the independent components from their best-fitting equivalent dipole maps is a statistical metric of the plausibility the identified decomposition isolates actual cortical and artifactual EEG sources. Twenty-two blind decomposition algorithms (all freely available on the Internet as Matlab code) were used to decompose 14 69-channel, 300k-point EEG data sets. The mean CDH profiles varied from far less dipolar to considerably more dipolar than random raw EEG maps. The latter category included some (but not all) well-known ICA algorithms. Near-dipolar components returned by these algorithms were strongly correlated. Repeat decompositions on random dataset subsets using runica() (sccn.ucsd.edu/eeglab) demonstrated the relative stability of the identified near-dipolar components. Independent Component Analysis (ICA) can separate the volume conducted activities of electroencephalographic (EEG) source signals (e.g., synchronous activities within oriented cortical patches, plus non-brain artifact sources) from their mixtures recorded at individual scalp electrodes (Makeig et al., NIPS, 1996) by multiplying the data by an unmixing matrix that it adapts so as to make the output signals maximally independent. Each output represents the time course of activity of a separated source. Results support the use of ICA for investigating event-related EEG brain dynamics.