SomnoNet: A Lightweight and Interpretable Framework for Sleep Staging Using Single-Channel EEG.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Medicine

Background:

• Scalable sleep assessment is crucial for clinical practice and research.
• Single-channel electroencephalography (EEG) offers a practical approach for widespread sleep monitoring.
• Existing automated sleep staging methods face challenges in balancing accuracy, efficiency, and interpretability.

Purpose of the Study:

• To develop an automated sleep staging framework using hierarchical raw-EEG analysis.
• To create a compact and efficient model variant for deployment in resource-constrained environments.
• To enhance the clinical interpretability of automated sleep staging predictions.

Main Methods:

• Proposed SomnoNet, a hierarchical framework modeling multi-scale EEG rhythms and temporal dependencies.
• Developed SomnoNet-Nano, a lightweight variant with a frozen encoder and simplified temporal stack.
• Utilized two large public datasets (Physio2018, SHHS) for model validation.
• Implemented rhythm-aware decision analysis for visualizing model evidence.

Main Results:

• SomnoNet achieved high accuracy (80.9% Physio2018, 88.0% SHHS) and macro-F1 scores.
• SomnoNet-Nano demonstrated remarkable efficiency (29.49 ms/epoch) with minimal accuracy loss (retaining 99.5% and 99.3%).
• Rhythm-aware analysis provided clinically meaningful insights into model predictions.

Conclusions:

• SomnoNet offers a robust and accurate solution for single-channel EEG sleep staging.
• SomnoNet-Nano provides an efficient and compact alternative suitable for practical deployment.
• The framework enhances transparency and clinical relevance in automated sleep analysis.