SleepMFormer: An Efficient Attention Framework with Contrastive Learning for Single-Channel EEG Sleep Staging.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Sleep stage classification is vital for diagnosing sleep disorders and assessing sleep quality.
• Electroencephalography (EEG) is the primary method for sleep staging.
• Transformer models offer high performance but suffer from computational inefficiency.

Purpose of the Study:

• To introduce SleepMFormer, an efficient end-to-end framework for automatic sleep stage classification using single-channel EEG.
• To improve the efficiency of Transformer-based sleep staging algorithms.
• To enable practical sleep staging in resource-constrained environments.

Main Methods:

• SleepMFormer utilizes a simplified Transformer encoder for efficient attention.
• Supervised contrastive learning enhances representation robustness during training.
• The framework is designed for efficient training and inference.

Main Results:

• SleepMFormer achieves competitive performance on public datasets (Sleep-EDF, PhysioNet, SHHS).
• It reduces training and inference time by up to 33% compared to standard Transformer models.
• Performance is maintained across different feature extractors (DeepSleepNet, TinySleepNet).

Conclusions:

• SleepMFormer provides an efficient and practical solution for automatic sleep staging.
• The framework shows significant potential for clinical applications in sleep medicine.