Unsupervised Learning and Mapping of Brain fMRI Signals Based on Hidden Semi-Markov Event Sequence Models

Most methods used in functional MRI (fMRI) brain mapping require restrictive prior knowledge about the shape of the active blood-oxygenation-level-dependent (BOLD) response, thus leading to suboptimal or invalid inference. To solve this problem, we propose to assess local neural activity in terms of time alignment between the sequence of BOLD dynamics changes of interest and an Hidden Semi-Markov Event Sequence Model (HSMESM) of brain activation. The topology of the HSMESM is built from the deterministic transitions of the input stimulation paradigm and its parameters are automatically and iteratively learned from all intracranial fMRI signals. The brain mapping results achieved by HSMESMs in language processing demonstrate the relevance of such models in BOLD fMRI, especially to cope with strong variabilities of the active BOLD signal across time, brain, experiments and subjects.