Reliable automatic sleep stage classification based on hybrid intelligence.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Manual sleep staging is time-consuming and prone to errors.
• Current automatic sleep staging models lack sufficient accuracy and interpretability for clinical use.
• Accurate sleep staging is crucial for diagnosing sleep disorders and assessing sleep quality.

Purpose of the Study:

• To develop a hybrid intelligent model for automatic sleep staging that balances accuracy, interpretability, and generalizability.
• To improve the clinical utility of automatic sleep staging by addressing limitations of existing models.
• To enhance the efficiency of sleep assessment for sleep physicians.

Main Methods:

• A hybrid intelligent model integrating data and knowledge intelligence was developed.
• The model utilizes electroencephalography (EEG) and electrooculography (EOG) channels.
• It incorporates a temporal fully convolutional network (U-Net architecture) and a multi-task feature mapping structure.
• Knowledge intelligence was applied to refine sleep stage transitions and correct coarse sleep graphs.

Main Results:

• The model achieved a Macro-F1 score of 0.804 on the ISRUC dataset and 0.780 on the Sleep-EDFx dataset.
• Compared to existing interpretable models, the proposed hybrid model demonstrated superior performance.
• Knowledge intelligence effectively addressed issues of excessive jumps and unreasonable transitions in sleep graphs.

Conclusions:

• The hybrid intelligent model offers a promising solution for accurate and interpretable automatic sleep staging.
• This approach has significant potential to improve the efficiency of clinical sleep staging.
• The model provides valuable support for sleep physicians in sleep assessment.