Nonnegative partial least squares for metaanalytic parcellation: A functional atlas for the human brain 
 Abstract  Algorithms can automatically analyze human brain mapping studies represented in a neuroinformatics database. We describe one such metaanalytic method that relies on a combination of text mining, functional volumes modeling, partial least squares and nonnegative matrix factorization (NMF): NMF decomposes the product between a bagofwords matrix, constructed from abstract words, and a voxelization matrix constructed by kernel density modeling of the stereotaxic coordinates in the scientific papers contained in a database. The components found allows us to construct a functional atlas where voxels and words get loaded on components interpretable as brain functions. When applied on the Brede Database with 186 papers we find components such as memory, emotion, pain and audition. Furthermore, we present a cluster validation procedure based on permutation and cluster matching that quantifies the variability of the functional atlas.  Keywords  Neuroinformatics, text mining, functional volumes modeling, nonnegative matrix factorization, partial least squares, human brain, positron emission tomography, functional magnetic resonance imaging (fMRI), metaanalysis  Type  Journal paper [Submitted]  Year  2006 Month September  Electronic version(s)  [pdf]  BibTeX data  [bibtex]  IMM Group(s)  Intelligent Signal Processing 
