Temporal dependency in automatic sleep scoring via deep learning based architectures: An empirical study.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automatic sleep scoring from polysomnography (PSG) is crucial for diagnosing sleep disorders.
• Deep learning models show promise in analyzing complex PSG data, but their ability to capture temporal dependencies needs further evaluation.
• Understanding temporal encoding is key to improving the accuracy and clinical relevance of automated sleep scoring systems.

Purpose of the Study:

• To evaluate the effectiveness of various deep learning architectures in encoding temporal dependencies from raw polysomnography (PSG) signals for sleep scoring.
• To compare the performance of different neural networks, including state-of-the-art models, using single-channel EEG data.
• To introduce and utilize a novel metric, δ_norm, for assessing the temporal encoding capabilities of these models.

Main Methods:

• Utilized the open-source Sleep-EDF expanded database for training and testing deep learning models.
• Evaluated a comprehensive range of neural network architectures, from simple feed-forward networks to complex state-of-the-art models.
• Assessed model performance using standard metrics (accuracy, Cohen's kappa, F1-score) and a newly proposed metric (δ_norm) for temporal dependency encoding.

Main Results:

• The best performing model achieved an overall accuracy of 85.2% and a Cohen's kappa of 0.8.
• The F1-score for sleep stage N1 was 50.2%, highlighting challenges in scoring lighter sleep stages.
• A simple feed-forward architecture demonstrated comparable performance to complex models and superior encoding of continuous temporal sleep characteristics, as measured by δ_norm.

Conclusions:

• Deep learning models can effectively encode temporal dependencies in sleep data, crucial for accurate automatic sleep scoring.
• Simpler architectures, like feed-forward networks, can be highly effective and may offer advantages in capturing the continuous nature of sleep.
• Further understanding of temporal pattern encoding by neural networks can significantly improve the clinical utility of automated sleep scoring systems.