PCA-Preprocessing of fMRI Data Adversely Affects the Results of ICA

Independent Component Analysis (ICA) is a new technique for analyzing fMRI data. Unfortunately, the size of fMRI datasets sometimes renders this technique computationally intractable, and certain compromises must be made to perform the analysis. One such compromise is to project the dataset onto a lower-dimensional subspace using a Principal Components Analysis (PCA). This subspace, which in some sense captures the essence of the data, is then used as the input to ICA. It is demonstrated herein, however, that an ICA analysis of a PCA-preprocessed dataset will tend to favor Gaussian and nearGaussian distributions and can miss task-related activation components.