Recurring functional interactions predict network architecture of interictal and ictal states in neocortical epilepsy

Abstract

Human epilepsy patients suffer from spontaneous seizures, which originate in brain regions that also subserve normal function. Prior studies demonstrate focal, neocortical epilepsy is associated with dysfunction, several hours before seizures. How does the epileptic network perpetuate dysfunction during baseline periods? To address this question, we developed an unsupervised machine learning technique to disentangle patterns of functional interactions between brain regions, or subgraphs, from dynamic functional networks constructed from approximately 100 hours of intracranial recordings in each of 22 neocortical epilepsy patients. Using this approach, we found: (i) subgraphs from ictal (seizure) and interictal (baseline) epochs are topologically similar, (ii) interictal subgraph topology and dynamics can predict brain regions that generate seizures, and (iii) subgraphs undergo slower and more coordinated fluctuations during ictal epochs compared to interictal epochs. Our observations suggest that seizures mark a critical shift away from interictal states that is driven by changes in the dynamical expression of strongly interacting components of the epileptic network.

Significance Statement Localization-related epilepsy is a debilitating condition where seizures begin in dysfunctional brain regions, and is often resistant to medication. The challenge for treating patients is mapping dysfunction in brain networks that also subserve normal function several hours before seizures. Localizing brain regions that generate seizures is critical for improving seizure freedom rates following invasive surgery. We develop new methods to identify clusters of functionally interacting brain regions from approximately 100-hour intracranial, neocortical recordings per epilepsy patient. Our results indicate seizure-generating brain regions: (i) can be predicted before seizures and (ii) may kindle dysfunction through interactions with non-seizure generating brain regions. These findings may have clinical implications for targeting specific brain regions to control seizures several hours before they occur.