Abstracts

Rationale: Patients with medically refractory focal epilepsy often undergo extensive diagnostic workups to localize areas of ictal onset. This process is long and morbid with patients often spending weeks in the hospital undergoing intracranial recording to capture seizures. A few groups have used single pulse electrical stimulation (SPES) to create cortico-cortical evoked potentials (CCEP) in distant gray matter regions in an effort to localize ictal onset (Matsumoto et al. Seizure 2017; 44; 27-36). However, the important localizing features of CCEPs remain unclear, and ictogenic localization ability has been limited (Prime et al. Journal of Neuroscience Methods 2020; 334; 108559). A potential reason for limited efficacy could be that important CCEP features could be complex and non-linear. Neural networks are particularly suited to identifying complex highly non-linear features beyond human comprehension. Thus, we aimed to leverage this sophisticated feature extraction ability of neural networks to identify the ictal onset zone.

Methods: We included four patients with medically refractory focal epilepsy who were admitted to the hospital for intracranial monitoring. We consented each patient to perform SPES on all gray matter adjacent bipolar pairs of contacts. We conducted the stimulation session during anti-epileptic drug medication weaning, and at least four hours after any electroclinical seizures. We used a 1 Hz, 10-second pulse train of 300 microsecond biphasic pulses at 1, 3, and 5 mA (1, 2, and 3 mA for mesial regions). For each stimulation pair, we divided the 10-second stimulation recording in to one-second post-stimulation epochs and applied bandpass filters at 1-59 Hz, 61-119 Hz and 121-159 Hz. Recordings from electrode contacts that were within 20 mm from the stimulation were discarded. We divided the remaining data into random sets of 40 one-second post-stimulation epochs per patient. The total dataset had approximately 1.2 million unique 40-channel inputs. We then used a one-dimensional multi-channel convolutional neural network to categorize whether each set of 40 signals were generated from stimulation of an ictogenic or non-ictogenic region (as determined by clinical evaluation according to the attending epileptologist). To assess generalization, we implemented a leave-one-out testing across all four patients.

Results: The confusion matrices for the leave-one-out validation results are outlined in Fig. 1. The trained models performed well for three of the four patients withheld (Patients 1, 2, and 4)—average sensitivity of 72% and specificity of 77%. The model performed poorly for Patient 3 withheld. This patient was later discovered to have had a high burden of subclinical seizures near the time of data collection.

Conclusions: This work serves as evidence that a neural network can be trained with multi-channel SPES data to categorize ictogenic versus non-ictogenic regions. This work is significant because it outlines the possibility that a neural network trained on a larger dataset could be used during intracranial monitoring in real time at the bedside or in the operating room to categorize brain regions as ictogenic.

Funding: Please list any funding that was received in support of this abstract.: R00NS09761805, R01NS11225202, and T32GM734740.