SleepGCN: A transition rule learning model based on Graph Convolutional Network for sleep staging.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Sleep Medicine

Background:

• Automatic sleep staging is crucial for diagnosing sleep disorders affecting millions globally.
• Existing models often overlook critical sleep transition rules used by experts.
• There is a need for advanced models that incorporate these transition dynamics.

Purpose of the Study:

• To develop an automatic sleep staging model, SleepGCN, that effectively captures sleep transition rules.
• To improve the accuracy and reliability of automatic sleep staging for clinical applications.

Main Methods:

• SleepGCN utilizes a Sleep Representation Learning (SRL) module with ResNet and LSTM for deep feature extraction from two-channel EEG-EOG signals.
• A Sleep Transition Rule Learning (STRL) module with Graph Convolutional Network (GCN) captures inter-stage transition dynamics.
• The model integrates SRL features with STRL-derived transition rules for sleep stage identification.

Main Results:

• SleepGCN achieved high accuracy across five public datasets, including 89.70% on SleepEDF-20 and 86.16% on SHHS.
• Macro-average F1-scores ranged from 72.44% (DOD-H) to 85.20% (SleepEDF-20), demonstrating robust performance.
• Ablation studies confirmed the significant contributions of both SRL and STRL modules.

Conclusions:

• SleepGCN significantly outperforms existing models in automatic sleep staging.
• The model's effectiveness is validated by its superior performance on multiple datasets.
• Utilizing two-channel EEG-EOG data enhances sleep staging accuracy compared to single-channel approaches.