Cross-Subject Seizure Detection via Unsupervised Domain-Adaptation

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Automatic seizure detection from Electroencephalography (EEG) is crucial for epilepsy diagnosis and treatment.
• Current patient-specific methods lack scalability for broader clinical application.

Purpose of the Study:

• To develop a cross-subject seizure detection method for Electroencephalography (EEG) data.
• To overcome the limitations of patient-specific models using unsupervised domain adaptation.

Main Methods:

• Utilized Convolutional Neural Network (CNN) for shallow feature extraction.
• Applied Multi-Kernel Maximum Mean Discrepancies (MK-MMD) to minimize shallow feature distribution gaps.
• Employed adversarial learning for deep feature alignment and generalizability.

Main Results:

• Demonstrated the feasibility of cross-subject seizure detection.
• Validated the method's effectiveness in diminishing domain disparities between patients.
• Achieved positive results in epoch-based and event-based experiments on CHB-MIT and Siena datasets.

Conclusions:

• The proposed unsupervised domain adaptation method enhances the scalability of EEG-based seizure detection.
• Feature alignment techniques effectively bridge domain gaps across different patients.
• This approach holds promise for more generalized and accessible epilepsy management tools.