Automatic Seizure Classification Based on Domain-Invariant Deep Representation of EEG

Area of Science:

• * Artificial Intelligence in Neuroscience
• * Medical Signal Processing
• * Machine Learning for Healthcare

Background:

• * Accurate seizure identification is crucial for effective epilepsy treatment and drug prescription.
• * Current artificial intelligence (AI) methods for electroencephalography (EEG) analysis face challenges due to the personalized nature of seizure representations.
• * Existing research often yields unsatisfactory results for clinical applications.

Purpose of the Study:

• * To propose a domain-invariant deep feature representation method driven by adversarial learning.
• * To enhance the clinical adaptability and reliability of hybrid deep networks (HDN) for seizure type identification.
• * To improve the accuracy of automated EEG analysis for epilepsy diagnosis.

Main Methods:

• * Utilized squeeze-and-excitation networks (SENet) for short-term EEG feature extraction and long short-term memory networks (LSTM) for long-term feature extraction.
• * Employed adversarial learning between LSTM and a clustering subnet to align feature distributions between patient-specific EEG and database EEG.
• * Developed a domain adaptive deep feature representation method for improved classification.

Main Results:

• * The proposed method achieved a 5% improvement in classification accuracy on the target dataset.
• * Demonstrated enhanced reliability in identifying seizure types using hybrid deep models.
• * Validated the approach on the TUH EEG Seizure Corpus and CHB-MIT seizure database.

Conclusions:

• * The adversarial learning-driven domain-invariant deep feature representation significantly improves seizure classification accuracy.
• * The method enhances the clinical applicability of automated EEG analysis equipment.
• * This approach offers a promising solution for personalized and accurate epilepsy diagnosis.