A Semi-Supervised Few-Shot Learning Model for Epileptic Seizure Detection.

Area of Science:

• Neurology
• Machine Learning
• Biomedical Signal Processing

Background:

• Machine learning significantly enhances epileptic seizure detection from Electroencephalography (EEG) data.
• Limited availability of labeled EEG data due to high annotation costs hinders model training.
• Existing methods often require extensive neurologist-labeled datasets.

Purpose of the Study:

• To develop a cost-effective, one-step semi-supervised system for epilepsy detection.
• To leverage unlabeled EEG data to improve model robustness and accuracy.
• To reduce the reliance on expensive, neurologist-annotated datasets.

Main Methods:

• Proposed a novel neural network training strategy for semi-supervised learning.
• Implemented a one-step system to fully utilize unlabeled EEG data.
• Enforced prediction consistency on unlabeled data to refine decision boundaries.

Main Results:

• Achieved a 10.3% higher Area Under the Receiver Operating Characteristic (AUROC) curve on the CHB-MIT dataset compared to supervised methods.
• Demonstrated a 4.9% higher AUROC on the Kaggle dataset versus supervised approaches.
• Validated the effectiveness of the semi-supervised strategy in improving epilepsy detection performance.

Conclusions:

• The proposed semi-supervised system effectively utilizes unlabeled data for improved epilepsy detection.
• The method offers a significant reduction in labeling costs while enhancing model accuracy.
• This approach presents a promising direction for developing more accessible and efficient epilepsy diagnostic tools.