Abstracts

Rationale: Accurate delineation of the epileptogenic zone (EZ) in children with drug-resistant epilepsy (DRE) is crucial for controlling and eliminating seizures following resective neurosurgery. Currently, human experts visually analyze ictal intracranial electroencephalography (iEEG) data to identify the onset of clinical seizures; yet this process is time-consuming and prone to errors and biases. Since interictal epileptiform activity is also linked to the EZ, we propose here the development of an artificial intelligence (AI) framework that transforms interictal iEEG data into coherent networks and temporal maps. These networks automatically identify the EZ and predict surgical outcome in children with DRE while their corresponding temporal maps are concordant with the occurrence of interictal epileptiform discharges (IEDs).


Methods: We retrospectively analyzed iEEG from 45 children with DRE who had neurosurgery classified as good (n=28, Engel I) and poor (n=17, Engel ≥II) outcome (follow-up two years). For each patient, multimodal brain images were preprocessed to define the clinical seizure onset zone (SOZ) and resection volume (Fig. 1A). Next, interictal iEEG was dissected into short windows, each modeled using the dynamic mode decomposition to find the spectral powers of coherent channels in different physiologically relevant frequency bands. We used the non-negative matrix factorization to find the coherent networks and their corresponding temporal maps (Fig. 1B). The networks were then categorized as epileptogenic (less frequently active) and background. We studied the concordance of the active segments of the networks with IEDs (1-70 Hz) and ripples (80-250 Hz). We defined for each network: focality as the reciprocal of the average distance between electrodes, overlap as the percentage overlap within 10 mm of resection, and distance as the Euclidean distance from resection. Using the SOZ and resection as gold standards, we evaluated the ability of the epileptogenic network to delineate the EZ and predict surgical outcome.


Results: In good outcome patients, the epileptogenic network had higher focality and overlap, lower distance from resection, and higher area under the receiver operating curve (AUC) with resection (and SOZ) compared to the background. Additionally, the temporal activity of the epileptogenic network was concordant with IEDs (Fig. 2A). We found that the epileptogenic network had higher power inside (compared to outside) resection (p< 0.05) only in good outcome patients (Fig. 2B). In good outcome patients, the epileptogenic network delineated resection and the SOZ in the θ band with an AUC > 0.74 (Fig. 2C). The epileptogenic network in good outcome patients had higher focality, higher overlap, and lower distance from resection (P< 0.01) compared to poor (Fig. 2D). Finally, the epileptogenic network in the θ band predicted outcome (P< 0.05) (Fig. 2E).


Conclusions: Our AI-based framework automatically identifies the EZ from interictal iEEG data and predicts surgical outcome. Such an automated process would enhance the presurgical planning and improve the outcome of patients with DRE undergoing resective neurosurgery.


Funding: RO1NS104116-01A1 and R01NS134944-01A1 by NINDS