An End-to-End Multi-Channel Convolutional Bi-LSTM Network for Automatic Sleep Stage Detection

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Automatic sleep stage detection from polysomnography (PSG) is crucial for sleep quality monitoring.
• Single-channel deep learning models face data inefficiency and skewness challenges.
• Multi-channel approaches offer improved performance but demand significant computational resources.

Purpose of the Study:

• To introduce a computationally efficient, multi-channel convolutional bidirectional long short-term memory (Bi-LSTM) network for automatic sleep stage detection.
• To leverage transfer learning by fusing pre-trained dual-channel modules for enhanced performance.
• To address the tradeoff between model performance and computational cost in sleep stage classification.

Main Methods:

• A four-channel convolutional Bi-LSTM network was designed, utilizing EEG Fpz-Cz, EEG Pz-Oz, EOG, and EMG data.
• Dual-channel convolutional Bi-LSTM modules were pre-trained on pairs of PSG channels.
• Transfer learning was applied by fusing two pre-trained dual-channel modules.
• A two-layer convolutional neural network extracted spatial features, coupled for Bi-LSTM input to capture temporal features.

Main Results:

• The proposed model achieved high accuracy on the Sleep EDF-20 dataset (ACC: 91.44%, Kp: 0.89, F1: 88.69%) using EEG Fpz-Cz + EOG and EEG Fpz-Cz + EMG modules.
• On the Sleep EDF-78 dataset, the best performance (ACC: 90.21%, Kp: 0.86, F1: 87.02%) was obtained with EEG Fpz-Cz + EMG and EEG Pz-Oz + EOG modules.
• A comparative analysis demonstrated the proposed model's efficacy against existing literature.

Conclusions:

• The developed multi-channel convolutional Bi-LSTM network effectively extracts spatiotemporal features for accurate automatic sleep stage detection.
• The transfer learning approach using fused dual-channel modules offers a balance between performance and computational efficiency.
• The model shows significant potential for improving sleep monitoring systems.