Dynamic Electroencephalography Representation Learning for Improved Epileptic Seizure Detection

Zihang Liu, Haishuai Wang

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Electroencephalography, Epileptic, Seizure, Efficient, Neuroscience

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Efficient dynamic neural network for low-latency seizure detection from streaming EEG

Abstract: Epileptic seizure is an abnormal brain activity that affects millions of people worldwide. Effectively detecting seizures from electroencephalography (EEG) signals with automated algorithms is essential for seizure diagnosis and treatment. Although much research has been proposed to learn EEG representations, most of them neglect the detection latency when it comes to real-world clinical scenarios where the inputs are streaming. Moreover, they fail to either capture the underlying dynamics of brain activities or localize seizure regions. To this end, we propose an improved seizure detection task named onset detection, which identifies both the presence and the specific timestamps of seizure events, and two new metrics to quantify the timeliness of detection methods. We further introduce the Dynamic Seizure Network, a framework for EEG representation learning, which models the evolutionary brain states and dynamic brain connectivity efficiently. Theoretical analysis and experimental results on three real-world seizure datasets demonstrate that our method outperforms baselines with low time and space complexity. Our method can also provide visualizations to assist clinicians in localizing abnormal brain activities for further diagnosis.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1074