Sleep staging based on single-channel EEG and EOG with Tiny U-Net.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Current sleep staging algorithms lack generalization, limiting practical application.
• Heterogeneous datasets are crucial for developing robust sleep analysis tools.
• Addressing class imbalance is vital for accurate sleep stage classification, especially for minority stages like N1.

Purpose of the Study:

• To develop a lightweight, generalizable automatic sleep staging architecture (TinyUStaging).
• To improve the recognition of minority and difficult-to-classify sleep stages (N1, N3), particularly in OSA patients.
• To validate the model's performance on large-scale, imbalanced, and heterogeneous sleep data.

Main Methods:

• Utilized seven heterogeneous sleep datasets (9970 records, 20k+ hours, 7226 subjects).
• Developed TinyUStaging, a U-Net architecture with Channel and Spatial Joint Attention (CSJA) and Squeeze and Excitation (SE) blocks.
• Implemented probability-compensated sampling strategies and a class-aware Sparse Weighted Dice and Focal (SWDF) loss function.

Main Results:

• Achieved average overall accuracy of 84.62%, macro F1-score (MF1) of 79.6%, and kappa of 0.764 on heterogeneous datasets.
• Demonstrated superior performance, especially in N1 classification, outperforming existing methods.
• Showcased model stability with an overall MF1 standard deviation within 0.175 across cross-validation folds.

Conclusions:

• TinyUStaging offers a robust and generalizable solution for automatic sleep staging using single-lead EEG and EOG.
• The model provides a foundation for effective out-of-hospital sleep monitoring, even with imbalanced and diverse data.
• The proposed methods effectively address class imbalance and improve classification accuracy for critical sleep stages.