Sparse PCA, a new method for unsupervised analyses of fMRI data

Exploratory analysis of functional MRI data aims at revealing known as well as unknown properties in a data-driven manner devoid of hypotheses on the time course of the hemodynamic response. This uncommitted approach usually precedes confirmatory modeling and may point to unexpected results that otherwise would be lost. Common approaches include clustering methods, principal component analysis (PCA) and in particular independent component analysis (ICA). ICA assumes that the measured activity patterns consist of linear combinations of a set of statistically independent source signals. Under favorable circumstances, one of more of these signals describe activation patterns, while others model noise and other nuisance factors. This work introduces a competing method for fMRI analysis known as sparse principal component analysis (SPCA). We argue that SPCA is less committed than ICA and show that similar results, with better suppression of noise, are obtained.