Authors :
Presenting Author: Georgios Ntolkeras, MD – Boston Children's Hospital, Department of Neurology, HMS
Navaneethakrishna Makaram, PhD – Newborn Medicine – Boston Children's Hospital, FNNDSC, Harvard Medical School; Joseph Madsen, MD – Boston Children's Hospital, Department of Neurosurgery, HMS; Scellig Stone, MD – Boston Children's Hospital, Department of Neurosurgery, HMS; Phillip Pearl, MD – Boston Children's Hospital, Department of Neurology, HMS; Christos Papadelis, PhD – Cook Children's Hospital, Texas Christian University School of Medicine; Ellen Grant, MD – Boston Children's Hospital, FNNDSC, Harvard Medical School; Eleonora Tamilia, PhD – Boston Children's Hospital, FNNDSC, Harvard Medical School
Rationale:
For children with drug-resistant epilepsy (DRE), neurosurgery can provide seizure control if successful in targeting the epileptogenic zone (EZ). This is estimated via a multimodal presurgical process, during which conventional scalp EEG is key, being the most widely available tool for seizure characterization in all epilepsy centers. Yet, traditional interictal EEG biomarkers (i.e., spikes) may lack specificity to the EZ and add limited value to the presurgical workup.
As epilepsy is increasingly theorized as a brain network disorder, applying source functional connectivity (FC) analysis to routine scalp EEG allows to extract network data linked to the underlying epileptogenicity. Several FC features have been proposed as potential epilepsy biomarkers; yet literature lacks data-driven machine-learning (ML) attempts to optimize their use for patient-specific EZ localization before surgery.
Here, we aim to develop an EEG method to deconstruct the epileptogenic networks of children with DRE, reveal FC biomarkers of the EZ and develop ML models to estimate the EZ using brief EEG data, with or without epileptiform activity. We hypothesize that interictal EEG data can quantify the specific increase in FC that characterizes the EZ compared to other areas.
Methods: (Figure 1). We analyzed scalp EEG (5 min) from 32 children with DRE who became seizure-free after surgery. We placed ~1,200 virtual sensors (VSs) across their cortex, reconstructed their electrical activity (electrical source imaging, ESI), and computed FC between them during spike or silent EEG epochs separately (i.e., with or without spikes). We defined each child’s EZ (i.e., resection) along with three non-epileptogenic zones (NEZs) on their cortex, and compared their FC (in 6 frequency-bands and 3 spatial ranges of connections) to identify EEG signatures of the EZ. We then tested the ability of our FC metrics to distinguish the EZ among all the VSs via cross-validated ML models, which were also compared with interictal spike localization.
Results:
(Figure 2). Region-specific FC differed between EZ and NEZs (Wilcoxon sign-rank; p < 0.05) during silent and spike epochs, showing higher FC in the EZ than its homotopic contralateral NEZ for various spatial ranges and frequencies. During spike epochs, FC of the NEZ in the epileptogenic hemisphere was often higher than its contralateral NEZ. ML classifiers reached 75% accuracy (91% sensitivity; 74% specificity) in identifying the EZ using spike epochs in the gamma and beta frequency bands, and 62% accuracy using silent epochs in the broad band (1-70 Hz); both outperformed spike localization (accuracy=47%; p< 0.05).