Cognitive Neuroscience Society
San Francisco, CA
April 18-24, 2004
David Groppe1, Laura Kemmer1, Marta Kutas1, Scott Makeig2
1Cognitive Science Department, 2Institute for Neural Computation, University of California San Diego
Within-subject reliability of independent component analysis of the electroencephalogram in two text processing tasks
Independent component analysis (ICA) is a relatively new electroencephalogram (EEG) analysis technique that linearly decomposes scalp electrical potentials into temporally independent "components" that may account for the activity of individual EEG generators or artifact sources (e.g., blinks). Variability between decompositions from different subjects can complicate the interpretation of results. Makeig et al. (2002) identified eight classes of components from 15 subjects via automated clustering. On average, each cluster contained components from only 9.6 (SD 1.7) subjects. Such variability may reflect anatomic or physiological variations or may indicate the inability of ICA to reliably decompose the data. To test the latter possibility, we applied infomax ICA (sccn.ucsd.edu/eeglab) to 30-channel EEG data from eight subjects who participated in two experiments that presumably shared some EEG sources (i.e., common components) as both involved text processing and expectation violation. In one experiment, subjects read (un)grammatical sentences, and in the other they counted letter or word targets in three different visual oddball tasks. ICA was applied separately to the two data sets and components automatically paired across the decompositions according to their scalp distributions. On average, 10.4 (SD 3.5) components (out of approximately 30 possible) per subject appeared to replicate across data sets (i.e., their scalp distributions were correlated at r >= .9). These component pairs typically appeared to account for neural activity, had highly correlated power spectra and moderately correlated event-related potentials. Comparisons between-subjects/within-tasks find fewer highly correlated components. These results show that some non-artifactual ICA components are reliable despite between subject variability.