ABSTRACT
Functional magnetic resonance imaging (FMRI) is a designed experiment which often consists of a patient being given a sequence of stimuli AB or ABC. Imaging takes place while the patient is responding either passively or actively to these stimuli. A model is used which views the observed time courses as being made up of a linear trend, responses due to the presentation of the stimuli, and other cognitive activites that are typically termed random and grouped into the error term. The utility of the Bayesian Source Separation model for FMRI can be motivated by returning to the classic “cocktail party” proble. The Bayesian Source Separation model decomposes the observed time course in a voxel into a linear trend and a linear combination of unobserved component sequences. The choice of the reference functions in computing the activation of FMRI has been somewhat arbitrary and subjective.
