Hypergraphs in Phase-Space: A New Method for Predicting Epileptic Seizures

Epilepsy is a chronic disorder characterized by recurrent seizures. Prolonged seizure can evolve into status epilepticus, which can lead to injury or death. We propose a seizure prediction algorithm using a hyper-graph approach to phase-space analysis. Objective indications of seizure onset are derived via time delay embedding of minimally invasive time-serial scalp EEG. The approach considers the brain as a complex nonlinear dynamical system whose states can be characterized to determine change in brain dynamics related to epileptic seizure activity. Our method extracts phase-space graphs via nonlinear time series analysis and time delay embedding and partitions the phase-space graphs to form hypergraphs. The features of the hypergraphs are evaluated using spectral analysis to form biomarkers for seizure prediction. The algorithm correlates historical degrees of change in hypergraph spectra from repeated measurements and makes accurate pre-dictions of seizure onset. Our method yields statistically significant results on scalp EEG data, with a training accuracy of 93% and testing accuracy of 80%.