Epileptic Seizure Prediction Using Multi-Strategy Data Augmentation and Hierarchical Contrastive Learning.

Area of Science:

• Neurology
• Machine Learning
• Biomedical Signal Processing

Background:

• Accurate early epilepsy seizure prediction is vital for patient well-being.
• Current methods struggle with generalization and real-time performance due to large data needs.

Purpose of the Study:

• To develop an efficient seizure prediction framework requiring less labeled data.
• To enhance the distinction between interictal and preictal states for improved seizure detection.

Main Methods:

• Implemented a data augmentation strategy including wavelet-based frequency mixing and masking.
• Introduced a hierarchical contrastive loss function for improved preictal pattern capture.
• Utilized a lightweight SE-EEGNet for efficient feature extraction and real-time prediction.

Main Results:

• Achieved 94.51% accuracy and 95.05% sensitivity on the CHB-MIT dataset with 30% labeled data.
• Reported a low false positive rate (0.024/h) and a 20.12-minute prediction time.
• Demonstrated improved performance with increased labeled data on both CHB-MIT and Siena datasets.

Conclusions:

• The proposed framework effectively predicts seizures with limited labeled data.
• Contrastive learning and data augmentation significantly enhance prediction accuracy and robustness.
• The method shows practical applicability for real-time epilepsy seizure prediction.