Unsupervised domain adaptation for cross-patient seizure classification.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epileptic seizure classification relies on electroencephalogram (EEG) data.
• Challenges in cross-patient EEG analysis include low signal-to-noise ratio, non-stationarity, and individual variability.
• Developing generalizable seizure classification models across patients remains difficult.

Purpose of the Study:

• To address the challenge of cross-patient EEG-based seizure classification using multi-source unsupervised domain adaptation.
• To develop a novel approach that leverages labeled EEG data from multiple source patients to classify seizures in a new target patient.
• To enhance the efficiency and accuracy of automated seizure diagnostics.

Main Methods:

• Proposed a Source Domain Selection (SDS)-Global Domain Adaptation (GDA)-Target Agent Subdomain Adaptation (TASA) framework.
• SDS filters out dissimilar source patient data.
• GDA aligns overall data distributions between selected sources and the target domain.
• TASA identifies the most relevant source domain for label transfer.

Main Results:

• The SDS-GDA-TASA approach demonstrated superior performance in unsupervised cross-patient seizure classification.
• Outperformed 13 existing methods on two public seizure datasets.
• Validated the effectiveness of the proposed domain adaptation strategy.

Conclusions:

• The developed SDS-GDA-TASA method significantly improves unsupervised cross-patient seizure classification accuracy.
• This approach has the potential to substantially reduce the time and effort required for manual EEG data labeling by clinicians.
• Increases the overall efficiency of seizure diagnostics for epilepsy patients.