SeqSleepNet: End-to-End Hierarchical Recurrent Neural Network for Sequence-to-Sequence Automatic Sleep Staging.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Computational Neuroscience

Background:

• Automatic sleep staging typically classifies individual polysomnography epochs independently.
• This approach overlooks the inherent sequential nature of sleep architecture.

Purpose of the Study:

• To develop a novel sequence-to-sequence classification model for automatic sleep staging.
• To improve the accuracy and efficiency of sleep stage classification using deep learning.

Main Methods:

• Proposed SeqSleepNet, a hierarchical recurrent neural network.
• Incorporated a filterbank layer for frequency-domain preprocessing and an attention-based recurrent layer for short-term modeling.
• Utilized a top recurrent layer for long-term sequential modeling and end-to-end training.

Main Results:

• SeqSleepNet achieved state-of-the-art performance on a public dataset.
• Obtained an overall accuracy of 87.1%, macro F1-score of 83.3%, and Cohen's kappa of 0.815.
• Demonstrated superior performance compared to existing automatic sleep staging methods.

Conclusions:

• The sequence-to-sequence approach with SeqSleepNet effectively captures the temporal dependencies in sleep data.
• The hierarchical architecture and end-to-end training strategy enable robust and accurate automatic sleep staging.