Continuous Seizure Detection Based on Transformer and Long-Term iEEG.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Refractory epilepsy necessitates reliable automatic seizure detection algorithms.
• Existing methods often use discontinuous intracranial electroencephalograms (iEEG) and neglect channel-specific information.

Purpose of the Study:

• To evaluate a novel algorithm using continuous, long-term iEEG for clinical applicability.
• To integrate transformer networks for channel attention in seizure detection.

Main Methods:

• Developed an end-to-end convolutional and transformer-based model for multi-channel iEEG.
• The model processes raw time-series data without requiring feature engineering.
• Evaluated performance on the SWEC-ETHZ and TJU-HH iEEG datasets.

Main Results:

• Achieved high event-based sensitivity (97.5% on SWEC-ETHZ, 98.1% on TJU-HH) with low false detection rates (0.06/h and 0.22/h, respectively).
• Demonstrated improved performance and reduced latency (13.7s and 9.9s) compared to existing methods.
• Transformer attention highlighted channels near seizure onset zones, enhancing model explainability.

Conclusions:

• The proposed transformer-based model significantly enhances automatic seizure detection performance and explainability.
• This approach offers a promising tool for clinical routine monitoring of patients with epilepsy.
• The model's ability to focus on relevant channels improves understanding of seizure detection mechanisms.