Automatic Sleep Staging Algorithm Based on Time Attention Mechanism.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Artificial Intelligence in Healthcare

Background:

• Accurate sleep staging is crucial for diagnosing sleep-related diseases and assessing sleep quality.
• Existing automatic sleep staging methods face challenges, including low recognition rates for certain sleep stages like N1 and data imbalance.

Purpose of the Study:

• To develop and validate an automatic sleep staging algorithm utilizing a time attention mechanism.
• To enhance the recognition rate of the N1 sleep stage and address data imbalance issues in sleep staging.

Main Methods:

• Extraction and normalization of time-frequency and non-linear features from six-channel physiological signals.
• Implementation of a time attention mechanism with a two-way bi-directional gated recurrent unit (GRU) for efficiency.
• Integration of a conditional random field (CRF) for improved tag information and a novel balancing strategy for N1 stage recognition.

Main Results:

• Achieved high accuracy (0.9218), WF1 (0.9177), and Kappa (0.8751) on the Sleep-EDF dataset.
• Outperformed the latest algorithms on a custom dataset with accuracy (0.9006), WF1 (0.8991), and Kappa (0.8664).
• Significantly improved N1 stage recognition (SEN-N1: 0.7) and overall accuracy (0.86) using the proposed balancing strategy.

Conclusions:

• The proposed time attention-based algorithm offers superior performance for automatic sleep staging compared to existing methods.
• The novel balancing strategy effectively addresses the under-recognition of the N1 sleep stage, a common challenge in sleep analysis.
• The algorithm demonstrates robustness, achieving considerable accuracy even without utilizing the electroencephalogram (EEG) channel.