MASleepNet: A Sleep Staging Model Integrating Multi-Scale Convolution and Attention Mechanisms.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Sleep disorders are increasingly prevalent due to modern lifestyle pressures, impacting cardiovascular and psychiatric health.
• Accurate sleep staging is crucial for early detection and treatment, but traditional manual methods are subjective and time-consuming.
• Deep learning offers promising automated solutions for sleep staging, addressing limitations of manual analysis.

Purpose of the Study:

• To develop and evaluate MASleepNet, a novel deep learning model for automated sleep staging.
• To integrate multimodal deep features from Polysomnography (PSG) signals for enhanced sleep staging accuracy.
• To leverage attention mechanisms for adaptive feature fusion and temporal feature extraction.

Main Methods:

• MASleepNet utilizes multi-channel PSG signals (EEG, EOG, EMG) as input.
• A multi-scale convolutional module extracts features at various time scales.
• Channel-wise and temporal attention mechanisms are employed for adaptive feature fusion and identification of key temporal segments.
• A Bidirectional Long Short-Term Memory (BiLSTM) network encodes temporal dependencies.

Main Results:

• The MASleepNet model achieved classification accuracies of 82.56% and 84.53% on the Sleep-EDF-78 and Sleep-EDF-20 datasets, respectively.
• The integration of multimodal signals and attention mechanisms demonstrated superior performance.
• The model effectively extracts and fuses features from different signal modalities and time scales.

Conclusions:

• Deep learning models integrating multimodal signals and attention mechanisms can significantly enhance automatic sleep staging efficiency.
• MASleepNet presents a viable and effective approach for automated sleep staging, outperforming existing methods.
• Further research in this area holds promise for improved diagnosis and management of sleep disorders.