Unsupervised seizure identification on EEG.

Area of Science:

• Neurology
• Machine Learning
• Signal Processing

Background:

• Epilepsy is a common neurological disorder often detected by seizures.
• Electroencephalogram (EEG) is crucial for seizure identification but manual analysis is time-consuming.
• Existing automated methods predominantly rely on supervised learning, requiring difficult-to-obtain expert labels.

Purpose of the Study:

• To develop a fully unsupervised deep learning method for seizure identification on raw EEG.
• To overcome the limitations of supervised methods that require expert-annotated seizure data.
• To provide an automated tool for efficient and accurate seizure detection.

Main Methods:

• A variational autoencoder (VAE) was trained on non-seizure EEG data.
• Seizures were identified based on reconstruction errors during inference.
• The VAE training loss was modified to mitigate EEG artifacts.
• The method requires no manual feature extraction or ground-truth labels.

Main Results:

• The unsupervised VAE method achieved up to 0.83 Area Under the Curve (AUC) on intracranial EEG datasets.
• The approach demonstrated the ability to distinguish seizure from non-seizure EEG activity.
• The algorithm shows potential for real-time inference, processing at least 10 seconds of EEG per second.

Conclusions:

• This work presents the first successful deep learning-based unsupervised seizure identification on raw EEG.
• The method can reduce the workload for clinical experts in analyzing EEG.
• Early seizure identification using this approach may improve epilepsy detection and treatment initiation.