A Generative Model to Synthesize EEG Data for Epileptic Seizure Prediction

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate epilepsy seizure prediction is hindered by the scarcity of high-quality electroencephalography (EEG) data.
• Deep learning models require substantial, diverse datasets for optimal performance in complex tasks like seizure prediction.

Purpose of the Study:

• To develop a method for generating synthetic EEG data using a deep convolutional generative adversarial network (DCGAN).
• To evaluate the effectiveness of transfer learning (TL) with deep learning (DL) models for epileptic seizure prediction using augmented EEG data.

Main Methods:

• A DCGAN was trained on real EEG data for patient-specific synthetic data generation.
• Synthetic data quality was validated using one-class SVM and a novel Convolutional Epileptic Seizure Predictor (CESP).
• Four DL models (VGG16, VGG19, ResNet50, Inceptionv3) were evaluated using TL on augmented data.

Main Results:

• The CESP model demonstrated strong performance, achieving high sensitivity and low false prediction rates on synthesized and real datasets.
• Inceptionv3, utilizing TL and augmented data, achieved the highest accuracy with 90.03% sensitivity and 0.03 false prediction rate per hour.
• The proposed data augmentation method improved prediction results by 4-5% compared to existing techniques.

Conclusions:

• The DCGAN effectively generates synthetic EEG data, addressing the scarcity of quality data and improving seizure prediction performance.
• The CESP model's performance indicates that synthetic data successfully captures relevant feature-label associations.
• Augmented data significantly enhances predictive capabilities, outperforming chance levels for seizure prediction.