Semi-supervised Seizure Prediction with Generative Adversarial Networks.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Seizure prediction from electroencephalogram (EEG) signals is challenging due to low signal intensity and artifacts.
• Existing methods often rely heavily on labeled EEG data, which is less accessible.

Purpose of the Study:

• To develop a semi-supervised seizure prediction method utilizing both labeled and unlabeled EEG data.
• To enhance prediction accuracy through data fusion with cardiogram signals, body temperature, and time.

Main Methods:

• Employed short-time Fourier transform for EEG pre-processing.
• Utilized a generative adversarial network (GAN) trained in an unsupervised manner.
• Used the GAN's Discriminator as a feature extractor, followed by classification.

Main Results:

• Achieved an area under the operating characteristic curve (AUC) of 77.68% on the CHBMIT scalp EEG dataset.
• Achieved an AUC of 75.47% on the Freiburg Hospital intracranial EEG dataset.
• Demonstrated the effectiveness of unsupervised training and data fusion for seizure prediction.

Conclusions:

• The proposed semi-supervised method effectively utilizes unlabeled EEG data, reducing the need for extensive labeling.
• Data fusion with physiological signals improves seizure prediction accuracy.
• The unsupervised training aspect allows for potential real-time application and reduces patient-specific feature engineering efforts.