Domain adaptation for EEG-based, cross-subject epileptic seizure prediction

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epileptic seizure prediction is vital for patient safety and preventing complications.
• Patient-specific models struggle with data variability, limiting generalization.
• Existing approaches often fail when applied to new patients due to inter-patient data differences.

Purpose of the Study:

• To develop deep learning models capable of handling patient data variability for improved seizure prediction.
• To introduce novel cross-subject and multi-subject prediction models for enhanced generalization.
• To investigate domain adaptation techniques to boost the performance of cross-subject models.

Main Methods:

• Developed multi-subject deep learning models incorporating data from multiple patients.
• Adapted neural network architectures for cross-subject seizure prediction.
• Applied and evaluated three domain adaptation methods to improve cross-subject model performance.

Main Results:

• The multi-subject model demonstrated superior performance over existing methods on CHB-MIT and SIENA datasets.
• Cross-subject prediction models faced challenges in generalizing to unseen patients.
• Domain adaptation significantly improved model accuracy, by 10.30% on CHB-MIT and 7.4% on SIENA.

Conclusions:

• Deep learning models, particularly with domain adaptation, show promise for generalized epileptic seizure prediction.
• Multi-subject and cross-subject approaches offer improved generalization compared to purely patient-specific methods.
• Further research into domain adaptation is warranted to overcome cross-subject prediction challenges.