Abstracts

Automatic Electrophysiological Seizure Patterns Identification in Long-term Ambulatory Intracranial Recordings via a Deep Learning Model

Abstract number : 3.333
Submission category : 9. Surgery / 9C. All Ages
Year : 2022
Submission ID : 2204669
Source : www.aesnet.org
Presentation date : 12/5/2022 12:00:00 PM
Published date : Nov 22, 2022, 05:26 AM

Authors :
Vasileios Kokkinos, PhD, PhD – Massachusetts General Hospital; Victoria Peterson, Phd – Department of Neurosurgery – Massachusetts General Hospital; Enzo Ferrante, Phd – Research Institute for Signals, Systems and Computational Intelligence – UNL-CONICET; Ashley Walton, Phd – Department of Philosophy – Dartmouth College; Amir Hadanny, MD – Department of Neurosurgery – Massachusetts General Hospital; Varun Saravanan, PhD – Department of Brain and Cognitive Sciences – Massachusetts Institute of Technology; Nathaniel Sisterson, MD – Department of Neurosurgery – Massachusetts General Hospital; Naoir Zaher, MD – Department of Neurology – Epilepsy Center at Orlando Health; Alexandra Urban, MD – Department of Neurology – University of Pittsburgh Comprehensive Epilepsy Center; R. Mark Richardson, MD, PhD – Department of Neurosurgery – Massachusetts General Hospital

Rationale: Managing the progress of drug-resistant epilepsy patients implanted with the responsive neurostimulation (RNS) system requires the manual evaluation of hundreds of hours of brief epochs of continuous intracranial recordings.1 The generation of these large amounts of valuable data and the scarcity of experts’ time for evaluation, necessitates the development of automatic tools to detect intracranial electroencephalographic seizure patterns (iESPs) with expert-level accuracy. We developed an deep learning system for identifying the presence and onset time of iESPs in intracranial EEG (iEEG) recordings from the RNS device.

Methods: Data consisted of 36,293 recordings segments obtained from an IRB-approved registry of 24 patients with refractory focal epilepsy who underwent RNS System implantation (NeuroPace, Mountain View, California, U.S.). Lead implantation sites varied between patients and were grouped as Neocortex, Hippocampus, Neocortex-Hippocampus, or Malformations of Cortical Development. We developed a deep learning architecture (iESPnet) for identifying the probability of seizure onset at each sample point of intracranial RNS recordings. Seizure classification performance was measured by the balanced accuracy value, while the Mean Absolute Error (MAE) between the expert-evaluated and the iESPnet estimated onset time was used to assess time error performance. The model was evaluated in a hold-one-out strategy, simulating real clinical scenarios. The detection made by the net was compared with those made by the device-intrinsic algorithm.

Results: iESPnet detected the presence of an iESP with mean balanced accuracy value of 90%, with an onset time error prediction of approximately 3.4 seconds. There was no relationship between electrode location and prediction outcome (Figure 1). The intra-rater reliability between iESPnet and the RNS detection was of 0.11, demonstrating that the iESPnet detections are independent to those made by the RNS device. iESPnet showed the lowest prediction time found in the literature with competitive detection values (Table 1).

Conclusions: iESPnet successfully assessed the probability of a seizure onset at each point sampled in chronic iEEG RNS recordings, with accuracy that is acceptable for use in clinical practice, establishing this method as a valuable tool that can be used to alleviate the time-consuming manual inspection of iESPs and facilitate evaluation of therapeutic response in RNS System patients.

Funding: There are no external funding sources to report.
Surgery