SleepSIM: Conditional GAN-based non-REM sleep EEG Signal Generator.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Generative networks, particularly Generative Adversarial Networks (GANs), excel at synthetic image generation.
• Time series data generation, especially for biological signals like electroencephalography (EEG), remains an area requiring further research and development.
• Accurate simulation of sleep EEG signals is crucial for training and skill enhancement in sleep medicine.

Purpose of the Study:

• To propose and evaluate a Conditional GAN (CGAN)-based system for generating unique, realistic samples of non-REM sleep EEG signals.
• To leverage deep learning to model complex data distributions inherent in EEG signals.
• To provide a tool for enhancing the training of sleep medicine fellows and technicians.

Main Methods:

• Development of a CGAN model incorporating a 1-D Convolutional Neural Network architecture.
• Training the CGAN model using real EEG data acquired from healthy control subjects.
• Utilizing the trained model to generate artificial 30-second epochs of non-REM sleep EEG.

Main Results:

• The developed CGAN model successfully generated artificial non-REM sleep EEG epochs.
• The generated synthetic EEG signals exhibited power spectrum characteristics identical to real sleep EEG signals.
• The model demonstrated the capability to learn and replicate complex EEG signal distributions.

Conclusions:

• The proposed CGAN-based system effectively generates realistic synthetic non-REM sleep EEG signals.
• This synthetic data generation method offers a valuable tool for sleep medicine education and research.
• Deep learning, specifically CGANs, provides a powerful approach to modeling and simulating complex biological time-series data like EEG.