Epileptic Seizure Detection Based on Stockwell Transform and Bidirectional Long Short-Term Memory

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Epilepsy diagnosis and monitoring rely heavily on accurate seizure detection.
• Intracranial electroencephalogram (EEG) recordings provide detailed neural activity for analysis.
• Existing automatic seizure detection methods face challenges in accuracy and efficiency.

Purpose of the Study:

• To develop an efficient automatic seizure detection method for intracranial EEG.
• To leverage Stockwell transform and bidirectional long short-term memory (BiLSTM) networks for improved detection.
• To validate the proposed method's performance using extensive clinical data.

Main Methods:

• Applying Stockwell transform (S-transform) to raw EEG segments to generate time-frequency representations.
• Utilizing bidirectional long short-term memory (BiLSTM) neural networks for feature extraction and classification of EEG segments.
• Implementing postprocessing techniques including moving average filter, threshold judgment, multichannel fusion, and collar technique to refine detection.

Main Results:

• The system achieved a segment-based sensitivity of 98.09% and specificity of 98.69% on 689 hours of intracranial EEG data from 20 patients.
• Event-based evaluation yielded a sensitivity of 96.3% and a low false detection rate of 0.24/h.
• The proposed method demonstrated robust performance in automatic seizure detection.

Conclusions:

• The developed automatic seizure detection method, integrating S-transform and BiLSTM, is highly effective.
• The system shows significant potential for practical application in clinical epilepsy monitoring and diagnosis.
• Further validation and implementation in clinical settings are warranted.