EEG-Based Sleep Stage Classification via Neural Architecture Search.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Improving quality of life increases focus on sleep quality and sleep disorders.
• Electroencephalogram (EEG)-based sleep stage classification is crucial for assessing sleep health.
• Current automatic sleep staging relies on time-consuming, expert-designed neural networks.

Purpose of the Study:

• To propose a novel neural architecture search (NAS) framework for automated EEG-based sleep stage classification.
• To develop an efficient and accurate method for designing neural networks for sleep staging.
• To reduce the manual effort required in designing sleep classification models.

Main Methods:

• A novel NAS framework utilizing bilevel optimization approximation for EEG sleep stage classification.
• Search space approximation and regularization with shared parameters among cells for model optimization.
• Evaluation on Sleep-EDF-20, Sleep-EDF-78, and SHHS datasets.

Main Results:

• The NAS-searched model achieved an average accuracy of 82.7% on Sleep-EDF-20.
• The model obtained an average accuracy of 80.0% on Sleep-EDF-78.
• An average accuracy of 81.9% was achieved on the SHHS dataset.

Conclusions:

• The proposed NAS framework offers an effective approach for automatic neural network design in sleep classification.
• The method provides a valuable reference for future automated network design in sleep analysis.
• Demonstrates the potential of NAS in advancing sleep disorder diagnosis and management.