Abstracts

Rationale: The diagnosis of epilepsy is strongly connected to the interpretation of the EEG. In the recent years, deep-learning has been minimally applied in intracranial EEG data to facilitate seizure detection in adult and neonatal populations, as well as to identify interictal EEG features. However, such studies have been limited by the availability of sufficiently large training data sets. The recent establishment of closed-loop stimulation (RNS) presents an opportunity to overcome this limitation, as the system provides large datasets of intracranially recorded seizures for individual patients. The goal of this study was to develop machine-learning-mediated intracranial seizure detection using RNS data. Methods: We collected a large human intracranial seizure dataset to date, comprised of 5,385 ictal events recorded from 22 RNS patients, using custom built software. Each recording was manually processed and ictal events were marked by an expert. We designed and developed a 23-layer deep convolutional neural network architecture that provides personalized seizure annotations for a patient given limited training data for that patient. The network takes as input a time-voltage series and outputs two predictions conditioned to a particular patient dataset: the probability that the recording contains a seizure and the estimated seizure onset time. Results: First, we tested the network’s ability to detect seizures of newly implanted patients (new to the database, without significant volume of prior data), by training it with a dataset containing the seizures of all other patients and only a few from the new patient. In a second scenario, we tested the network’s ability to detect seizures in patients with long-term recording history (new to the database, with significant repository of seizures over the years), by training it with all other patients’ seizures plus a random sample of the new patient’s seizure repository. Our method achieved an average accuracy of 83 ± 4.7% area under precision-recall curve for identifying seizures in RNS data, given 10 positive and 10 negative training intracranial seizure examples for a novel patient. The network reached consensus with the marked dataset on an average of 88%, which is higher than the current inter-rater reliability score for RNS data (79%). Conclusions: Convolutional neural networks can provide expert-level accuracy for seizure detection, even if they are provided with a limited training data from novel patients. The key to our network’s performance is the large pool of training material that allows it to develop adequate expertise. This is a unique dataset that RNS can provide due to its ability to sample and record ECoG over long periods of time. We envision that this deep learning methodology will ultimately improve the management of RNS patients and enable the development of tools that take advantage of highly accurate intracranial seizure detection. Funding: No funding