Abstracts

Rationale: Human brain is considered a complex dynamic system that intermittently diverges from normal activity in patients with epilepsy. Characterization of the dominant epileptic brain networks over time may provide a better understanding of seizures’ underlying dynamics. Analyzing the dynamic evolution of these networks can help gain insights into defining the areas of seizure onset and spread. Through mapping the brain network underlying ictal onset, we can enhance the delineation of the epileptogenic zone (EZ), the brain area whose removal is necessary for seizure freedom. Here, we aim to characterize seizure dynamics of intracranial EEG (iEEG) data in patients with drug resistant epilepsy (DRE) by extracting the dynamic networks during seizure onset and assessing their ability to delineate the EZ using an artificial intelligence approach.

Methods: We processed ictal iEEG data obtained from 31 patients with DRE undergoing neurosurgical planning. The patients were grouped into having good (22 patients; Engel 1) or poor (9 patients; Engel 2-4) outcome, with post-surgical follow-ups of at least one year after surgery. For each patient, the multichannel iEEG data was dissected into short windows, each of which was modeled using the dynamic mode decomposition method to extract dynamic spatial-spectral coherence features (Fig. 1A). The non-negative matrix factorization technique was then used to identify the dominant patterns that characterize the network dynamics during seizure evolution (Fig. 1B). We then employed the clinically defined seizure onset and termination times to label the seizure onset networks (SON) (Fig. 2A) and construct a temporal evolution map for each patient (Fig. 2B-C). By estimating the area under the curve (AUC) from the receiver operating characteristic curve, we assessed the ability of the identified network to delineate the clinically defined SOZ and resection. We finally assessed whether resecting the SON predicted outcome.

Results: We found that the proposed AI-based methodology can characterize the evolution of ictal dynamics with a finite number of networks. The SON had higher power inside the SOZ, and resection (p< 0.001) compared to outside (Fig. 2D). The SON in good outcome patients had higher overlap with resection (p< 0.05) and was closer to resection (p< 0.05) compared to poor outcome patients (Fig. 2E). We found that the identified SON can predict the SOZ in good outcome patients with an average AUC of 0.84 (Fig. 2F). Finally, resection of the SON predicted outcome (p< 0.05) (Fig. 2G).