A Lightweight Neural Network Based on Memory and Transition Probability for Accurate Real-Time Sleep Stage...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Wearable systems require efficient sleep staging algorithms.
• Single-channel electroencephalography (EEG) presents a practical configuration for wearable sleep monitoring.
• Existing methods may lack interpretability or computational efficiency for real-time applications.

Purpose of the Study:

• To develop a lightweight hybrid framework for sleep stage classification using a single frontopolar EEG channel.
• To enhance sleep staging accuracy by integrating memory and sleep stage transition probabilities.
• To create a computationally efficient model suitable for battery-powered wearable devices.

Main Methods:

• A feedforward neural network framework utilizing a single frontopolar EEG channel.
• Incorporation of prior epoch information and a transition-aware feature derived from a learned stage transition matrix.
• A confidence-driven fallback strategy combining predictions from memory-based and no-memory networks.

Main Results:

• Achieved up to 85.4% accuracy and 0.79 Cohen's kappa using only single 30s epochs.
• Outperformed convolutional neural networks, recurrent neural networks, and decision tree approaches on a single frontopolar channel.
• Confidence-based rejection improved reliability, particularly for epochs with sleep stage transitions.

Conclusions:

• The proposed method offers a balance of performance, interpretability, and computational efficiency.
• Demonstrates suitability for real-time clinical and wearable sleep staging applications.
• Highlights the potential of lightweight hybrid models for resource-constrained environments.