Bayesian Calibration of Seizure Risk Forecasting using Electrocorticographic Network Features

Seizures in patients with epilepsy are associated with pathological neural synchronization. Counterintuitively, the preictal period immediately prior to seizure onset often shows reduced spatial synchronization compared with normal cognition. This aligns with the brain behaving as a complex dynamical system, in which reduced dimensionality and loss of resilience can precede a phase transition. The Critical Brain Hypothesis also associates the loss of healthy scale-free behavior with various disorders, including epilepsy. In this study, we investigate preictal changes using network features. These features include the mean node degree and mean clustering coefficient, constructed from thresholded correlation matrices derived from patient intracranial electrocorticographic electrode data. In the minutes leading up to seizure onset, we observe a loss of intermittent high-synchronization 'excursions' within the phase space. The constriction of hypervolume explored during the preictal state suggests a breakdown in the brain's ability to maintain normal coherence. We utilize these preictal changes to predict the probability of seizure onset using a Support Vector Machine algorithm. These discrete predictions can then be used to calibrate a real-time, continuous seizure risk forecast via Bayesian updating. This novel approach may offer significant advancement over static predictions, providing a more adaptable, quantitative, and interpretable tool for seizure management.