After a stroke, many patients slowly recover functions corresponding to the permanently damaged region, by recruiting different areas of the brain to compensate. I will present analyses of fMRI images of stroke patients with impaired hand movement, obtained on four occasions during the first six months post-stroke. A simple examination of mean differences between earlier and later images shows consistent changes in brain activation over the recovery period. We are interested in latent variable models of this data, to give “filtered”, higher-order descriptions of the changes that occur, and ideally find predictors of good stroke recovery. More generally, such models of fMRI data may inform understanding of functional networks in the brain, and assist in predicting variables of clinical interest. In this talk, I will present generative and discriminative models based on restricted Boltzmann machines and independent component analysis.
