STDP-based adaptive graph convolutional networks for automatic sleep staging.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Automatic sleep staging is crucial for diagnosing and treating sleep disorders.
• Machine learning models offer a promising approach, but require explainability for clinical trust.
• Existing methods may struggle with individual physiological signal variations.

Purpose of the Study:

• To propose an explainable model for automatic sleep staging.
• To leverage neuroscience principles for improved feature extraction from Polysomnography (PSG) signals.
• To enhance model robustness against individual differences in physiological data.

Main Methods:

• Developed STDP-GCN, an adaptive Graph Convolutional Network (GCN) inspired by Spike-Timing-Dependent Plasticity (STDP).
• Modeled PSG channels as neurons and used STDP for synapse strength, forming a graph structure.
• Employed GCN for spatial feature extraction, temporal convolution for stage transitions, and domain adversarial training for robustness.

Main Results:

• The STDP-GCN model demonstrated performance comparable to current state-of-the-art methods.
• The model effectively extracts relevant features from PSG signals.
• Domain adversarial training minimized the impact of individual physiological signal discrepancies.

Conclusions:

• The proposed STDP-GCN model offers an explainable and robust solution for automatic sleep staging.
• Integrating neuroscience principles like STDP enhances machine learning model interpretability and performance.
• This approach holds potential for improving sleep disorder diagnosis and treatment.