EEG power spectra parameterization and adaptive channel selection towards semi-supervised seizure prediction

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy seizure prediction algorithms often require high-density electroencephalogram (EEG) and extensive labeled data, posing challenges for daily patient monitoring and clinical application.
• Current methods are burdensome due to the need for numerous EEG channels and significant training data, demanding considerable time from epileptologists.

Purpose of the Study:

• To develop an adaptive channel selection strategy to reduce the number of EEG channels required for seizure prediction.
• To implement a semi-supervised deep learning model to minimize the amount of labeled data needed for accurate seizure prediction.
• To create a computationally efficient and practical system for daily epilepsy monitoring.

Main Methods:

• An adaptive channel selection module based on EEG power spectra parameterization to identify seizure-associated channels.
• A semi-supervised deep learning model integrating generative adversarial networks (GANs) and bidirectional long short-term memory (BiLSTM) networks.
• Evaluation on the CHB-MIT and Siena epilepsy datasets.

Main Results:

• The proposed method achieved high performance using only 4 EEG channels, with an Area Under the Curve (AUC) of 93.15% on the CHB-MIT dataset and 88.98% on the Siena dataset.
• The adaptive channel selection approach significantly reduced model parameters and training time.
• Demonstrated outstanding seizure prediction accuracy with reduced data requirements.

Conclusions:

• Adaptive channel selection combined with semi-supervised learning provides a foundation for lightweight and efficient seizure prediction systems.
• This approach enhances the practicality of daily epilepsy monitoring, potentially improving patients' quality of life.
• The developed method addresses key limitations of current seizure prediction technologies.