Abstracts

Rationale:
Localizing brain networks where seizures begin and spread is essential to controlling drug-resistant epilepsy. Intracranial EEG (iEEG) attempts to localize these tissues in neurosurgery or neuromodulation therapy candidates. iEEG recordings are analyzed qualitatively for seizures, epileptic spikes, and abnormal oscillations in sleep and wake states. In this study, we seek to quantify abnormality across states of consciousness and localize epileptiform brain tissue.

Methods:
A total of 93 patients with drug-resistant epilepsy (female = 49, mean age = 36.2) underwent iEEG to localize epileptic tissue. We identified baseline iEEG recordings during wake, N2, N3, and REM sleep states using the SleepSEEG tool. All recordings were bipolar re-referenced and localized to anatomical brain regions using an automatic registration pipeline. We integrated recordings from 106 patients at a second epilepsy center. For each state, we extracted spectral power in canonical frequency bands. These features were z-scored by channels deemed normal that met the following criteria: 1) outside the clinically defined seizure onset zone, 2) less than one spike per hour, 3) outside of the resection zone in patients who underwent surgery, 4) localized to the same brain region. At the channel level, the z-scored features were transformed into abnormality by using a 10-fold cross-validated Random Forests model that classified channels as normal or those that were resected in patients who were seizure free after surgery. At the patient level, we evaluated the degree of separability between surgically spared and resected channels’ abnormality using the area under the precision-recall curve (AUPRC)—deemed separability—to account for class imbalance. We then built logistic regression models to predict surgical success (Engel 1 vs < 1) from separability in each sleep stage.