Multi-channel deep learning with intracranial neurostimulation can localize seizure onset zones in humans

Abstract

In drug resistant temporal lobe epilepsy, automated tools for seizure onset zone (SOZ) localization using brief interictal recordings would supplement presurgical evaluations and improve care. Thus, we sought to localize SOZs by training a multi-channel convolutional neural network on stereo-EEG (SEEG) cortico-cortical evoked potentials. We performed single pulse electrical stimulation with 10 drug resistant temporal lobe epilepsy patients implanted with SEEG. Using the 500,000 unique post-stimulation SEEG epochs, we trained a multi-channel one-dimensional convolutional neural network to determine whether an SOZ was stimulated. SOZs were classified with a mean leave-one-patient-out testing sensitivity of 78.1% and specificity of 74.6%. To achieve maximum accuracy, the model requires a 0-350 ms post stimulation time period. Post-hoc analysis revealed that the model accurately classified unilateral vs bilateral mesial temporal lobe seizure onset, and neocortical SOZs. This is the first demonstration, to our knowledge, that a deep learning framework can be used to accurately classify SOZs using cortico-cortical evoked potentials. Our findings suggest accurate classification of SOZs relies on a complex temporal evolution of evoked potentials within 350 ms of stimulation. Validation in a larger dataset could provide a practical clinical tool for the presurgical evaluation of drug resistant epilepsy.