Bayesian nonparametric modeling of epileptic events

Epilepsy is a common neurological disorder that today plagues over 50 million people worldwide. The 20-40% of patients whose seizures are unable to be controlled with pharmacological treatments commonly receive scalp and intracranial electroencephalogram (EEG) monitoring to determine whether surgical treatment is appropriate. Epileptic events like large, clinical seizures and small, sub-clinical bursts recorded on the EEG are of primary diagnostic interest, but these events—which usually range from a few seconds to a few minutes across tens or hundreds of individual EEG channels—are very complex and high-dimensional. Human epileptologists are well-trained in analyzing individual epileptic events but their ability to generalize across and compare many such events is limited due to the complex, high-dimensional nature of these EEG event recordings. In this work, we develop and apply statistical models for analyzing and understanding large numbers of these events. Our Bayesian nonparametric models naturally incorporate available prior knowledge and uncertainty about these events. While motivated by these epileptic event data, our models generalize to large class of application domains. We first develop and validate a model for describing seizures that intelligently shares information across the seizures of the same patient and those of other similar patients. We then develop and validate a model for producing a fine-grained parsing of both shorter burst and longer seizure events, allowing for straightforward comparisons between the two. Finally, we apply this later model to large datasets of hundreds of epileptic bursts and seizures, finding that the bursts often display large similarities with the onsets of seizures. These results show the benefit of well-motivated, straightforward Bayesian modeling and the large impact it can have in the quantitative analysis of epileptic events.