UC San Diego

Electrocorticography (ECoG), also known as intracranial electroencephalography (iEEG), is the practice of recording electrical potentials on the cerebral cortex via electrodes placed on the exposed brain surface. ECoG has been a critical component of epilepsy medical treatment protocols involving neurosurgery for more than half a century. More recently, ECoG has emerged as a promising recording modality for brain-machine interfaces and neuroscience research. The BRAIN Initiative is representative of a renewed and concerted effort to push the boundaries of possibility in medical care and technology, and to expand our understanding of brain function. Concomitant with this new drive is a need for techniques that address the challenges posed by high-channel count ECoG signal analysis as well as by neural data collection limited due to the invasiveness of ECoG. In this dissertation, we introduce the use of a discrete-state based probabilistic method for modeling ECoG-derived signals, and contrast this method with previously existing analogous probabilistic models without a discrete component. We then explore a class of discrete-state based probabilistic models, and identify spatial and temporal model constraints that were advantageous in the analysis of a high-channel count ECoG dataset. Finally, we introduce another probabilistic model that we use for unsupervised learning of ECoG trial spatiotemporal structure, and clustering of the ECoG trials in a data-limited context.