Mamba-CAM-Sleep: A Mamba-based Channel Attention Model for Sleep Staging Classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Accurate sleep stage classification is critical for diagnosing sleep disorders.
• Polysomnography (PSG) records are the gold standard for sleep analysis.
• Existing methods face challenges in capturing complex sleep stage transitions and feature extraction.

Purpose of the Study:

• To introduce Mamba-CAM-Sleep, a novel framework for automated sleep staging.
• To leverage a state-space backbone (Mamba) and channel attention graph neural network for improved classification.
• To enhance the analysis of long-range temporal dependencies in PSG data.

Main Methods:

• Utilized polysomnography (PSG) records for sleep staging.
• Developed a novel framework integrating a Mamba backbone for temporal dependencies and a channel attention graph neural network for feature extraction.
• Validated the model on the DOD-H benchmark dataset.

Main Results:

• Achieved an F1 score of 80.4% and Cohen's kappa of 76.3%.
• Outperformed existing methods like LSTM, SimpleSleepNet, RobustSleepNet, and DeepSleepNet by 3.0-6.0% (F1) and 2.2-10.8% (kappa).
• Demonstrated superior robustness with lower standard deviations (±2.1 F1, ±2.5 kappa) and excelled in classifying challenging stages N1 (53.9%) and N3 (74.7%).

Conclusions:

• Mamba-CAM-Sleep offers a scalable and reliable solution for clinical sleep staging.
• The framework balances high accuracy with practical deployability.
• The proposed architectural innovations significantly improve sleep staging performance.