Society for Neuroscience Meeting
3-7 November - 2007 - San Diego (California, USA)
 
Comparing results of algorithms implementing blind source separation of eeg data
 
Arnaud Delorme (1&2), Jason Palmer (2), Robert Oostenveld (3), Julie Onton (2), Scott Makeig (2)
 

(1) Centre de Recherche Cerveau & Cognition, Toulouse, France

(2) Swartz Center for Computational Neuroscience, UCSD, La Jolla, USA

(3) FC Donders, Nijmegen, The Netherlands

arno@salk.edu

 
    Independent component analysis is now used widely for separating artifacts from EEG data. Not only can ICA separate brain EEG activity from non-brain artifacts, increasingly it is used to study brain activities themselves. ICA separates signals from multi-channel data whose time courses are maximally independent from each other. It remains unclear however which ICA algorithm returns the most dipolar components when applied to EEG data. To date, the three ICA algorithms applied most often to process EEG data are Infomax ICA, SOBI, and FastICA. However, there are a large variety of other available ICA algorithms that may possibly be useful for EEG decomposition. Here, we tested this using a total of 23 linear decomposition algorithms, 20 ICA algorithms plus principal component analysis (PCA).
    Fourteen subjects (7 male, 7 female) participated in a steinberg memory study where subject had to remember letter of a specific color. EEG data were collected from 71 channels. Continuous data was first high-pass filtered above 0.5 Hz using a FIR filter and epochs were extracted. A few noisy data epochs were also removed. The number of data samples in each dataset was amount to more than 250,000. After computing all 23 decompositions for each of the 14 EEG datasets, we localized a best-fitting single equivalent dipole corresponding to each component using a single equivalent dipole in the default spherical 4-shell head model in the DIPFIT plugin of the EEGLAB toolbox.
    Extended Infomax returned the largest number of such components among all ICA algorithms, although Pearson, simple (super-Gaussian) Infomax, and ERICA gave near-equivalent results. All these algorithms use natural gradient descent. SOBI, another widely used algorithm for decomposing EEG data, and the other time-dependent algorithms did not return as many dipolar components as the natural gradient-based algorithms. As expected, Principal Component Analysis returned only a small number of dipolar components (those accounting mostly for large eye artifacts).
    We compared ICA algorithms using the Amari measure of distance by computing the pairwise distance between all pairs of decompositions for all datasets. ICA algorithms were clustered into two groups, the first group consisting of the instantaneous ICA algorithms, the second group consists of algorithms that depend on time-domain relationships. To test whether the most dipolar components returned by these two classes differed, we repeated the analysis using only the scalp maps of the 10 most dipolar components for each algorithm. Clearly, (apart from PCA) ICA and other blind source decompositions seem to return similar dipolar components.