Enhancing Cross-Patient Seizure Detection with Test-Time Adaptation

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Neurology

Background:

• Automated seizure detection is crucial for epilepsy management.
• Patient-specific models struggle with generalization due to data distribution shifts.
• Existing methods primarily focus on training-phase generalization.

Purpose of the Study:

• To develop a test-time adaptation strategy for improving cross-patient seizure detection.
• To dynamically adjust model parameters during testing using incoming test samples.
• To enhance the generalization capabilities of automated seizure detection models.

Main Methods:

• A ResNet-18 based architecture was employed.
• A learnable consistency loss was incorporated as an auxiliary training objective.
• Adaptive blocks were introduced for dynamic parameter updating during the testing phase.

Main Results:

• On the CHB-MIT dataset, the method achieved 95.24% accuracy, 94.69% sensitivity, and 95.85% specificity.
• On the Siena dataset, the method achieved 91.88% accuracy, 92.32% sensitivity, and 91.59% specificity.
• The test-time adaptation strategy demonstrated strong performance across key metrics.

Conclusions:

• Test-time adaptation shows significant promise for cross-patient seizure detection.
• The proposed method offers a viable solution to the generalization problem in automated seizure detection.
• This research provides valuable insights for future advancements in epilepsy monitoring technology.