Dual-Stage Improvement with Domain Adaptation for Cross-Subject Epileptic Seizure Prediction

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG)-based epileptic seizure prediction is crucial for managing drug-resistant epilepsy.
• Existing domain adaptation (DA) methods struggle with cross-subject EEG variability, data imbalance, and noisy pseudo-labels, hindering generalization.
• These limitations lead to biased models and poor understanding of target data distributions.

Purpose of the Study:

• To propose a novel cross-subject seizure prediction method using domain adaptation with dual-stage improvements.
• To address data imbalance and noisy pseudo-labels in EEG-based seizure prediction.
• To enhance the generalization performance of seizure prediction models.

Main Methods:

• A global context-aware generative network was developed to generate synthetic preictal samples, rectifying class imbalance.
• A common spatial pattern clustering filter with confidence-guided covariance weighting was employed to optimize spatial filters.
• A dual-filtering mechanism was implemented to iteratively eliminate noisy pseudo-labels, mitigating error propagation.

Main Results:

• The proposed method effectively mitigates interpatient domain discrepancies in EEG data.
• Experimental results on the CHB-MIT dataset demonstrate superior performance compared to leading approaches.
• The dual-stage DA approach significantly improves generalization performance in cross-subject seizure prediction.

Conclusions:

• The developed method offers a robust solution for EEG-based epileptic seizure prediction in cross-subject scenarios.
• Integrating data balancing and pseudo-label optimization within DA framework is effective.
• This approach holds promise for improving adjunctive treatment strategies for epilepsy.