Groupwise independent component decomposition of EEG data and partial least square analysis

This paper focuses on two methodological developments for analysis of neuroimaging data. The first is the derivation of robust spatiotemporal activity patterns across a group of subjects using a combination of principal component analysis (PCA) and independent component analysis (ICA). In applications to ERP data, the space dimension is typically represented in terms of scalp electrodes. The signal recorded by high density electrode caps is known to be highly correlated due in part to volume conduction. Consequently, this redundancy is also reflected in spatiotemporal patterns characterizing signal differences across experimental conditions. We present an alternative spatial representation and signal compression based on PCA for dimensionality reduction and ICA conducted across all subjects and conditions simultaneously. The second advancement is the use of partial least squares (PLS) analysis to assess task-dependent changes in the expression of the independent components. In an application to empirical ERP data, we derive an efficient number of independent component maps. Comparative PLS analysis on the independent components versus original electrode data shows that task effects are not only preserved under compression, but also enhanced statistically.