A Spatial-Temporal Transformer Architecture Using Multi-Channel Signals for Sleep Stage Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Sleep stage classification is crucial for diagnosing sleep disorders.
• Current methods often neglect spatial-temporal signal dynamics and suffer from limited data.
• Addressing these challenges is vital for improving sleep medicine.

Purpose of the Study:

• To develop an end-to-end framework for multi-channel polysomnogram (PSG) signal processing.
• To leverage Vision Transformer architectures for enhanced sleep stage classification.
• To overcome data insufficiency in training sleep analysis models.

Main Methods:

• A Vision Transformer-based architecture with spatial and temporal encoders was employed.
• Multi-channel PSG signals were processed to capture spatial-temporal features.
• A tailored image generation technique facilitated transfer learning and feature extraction.
• The framework was validated on three independent sleep datasets.

Main Results:

• The proposed method achieved superior performance compared to state-of-the-art algorithms.
• It effectively explored the spatial-temporal relationships within PSG signals.
• The approach demonstrated robustness in addressing data scarcity issues.

Conclusions:

• The Vision Transformer-based architecture offers a powerful solution for sleep stage classification.
• This method enhances the understanding of brain region interactions during sleep.
• The framework's adaptability suggests potential applications in other 1D signal analysis tasks.