Explainable multiscale temporal convolutional neural network model for sleep stage detection based on...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Sleep is vital for metabolism and overall health; sleep deprivation has serious health consequences.
• Accurate sleep stage detection is crucial for research and clinical applications.
• Existing automatic sleep staging models often lack explainability and require further performance improvements.

Purpose of the Study:

• To develop an explainable automatic sleep stage detection model.
• To improve the performance and efficiency of sleep staging algorithms.
• To address the limitations of current deep learning models in sleep analysis.

Main Methods:

• Implemented a Multiscale Temporal Convolutional Neural Network (MTCNN) using neurophysiology-mimicking kernels.
• Captured electroencephalogram (EEG) activities across various frequencies and temporal scales.
• Evaluated MTCNN performance on the open-source Sleep-EDF Database Expanded (153 days of polysomnogram data).

Main Results:

• MTCNN effectively identified EEG features specific to each sleep stage (e.g., K-complexes, sawtooth waves).
• Achieved high accuracy (91.12% OAcc, 0.86 kappa) in cross-subject analysis.
• Demonstrated strong performance (88.24% OAcc, 0.80 kappa) in leave-few-days-out analysis.
• Outperformed existing deep learning models, achieving 85.62% OAcc and 0.75 kappa with only 16% of EEG data.

Conclusions:

• The proposed MTCNN model offers enhanced explainability in sleep stage detection.
• MTCNN demonstrates high accuracy and efficiency, outperforming current deep learning methods.
• The model's ability to train with limited data makes it suitable for real-world applications where large datasets are scarce.