Automatic sleep scoring: A deep learning architecture for multi-modality time series.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Automatic sleep scoring is critical for sleep research due to the time-consuming nature of manual scoring.
• There is a growing need for efficient and accurate automated sleep scoring solutions.
• Polysomnography (PSG) recordings are the standard for sleep analysis.

Purpose of the Study:

• To develop a versatile deep learning architecture for automating sleep scoring.
• To utilize raw polysomnography recordings for sleep stage prediction.
• To create a model applicable to diverse sleep research scenarios.

Main Methods:

• A deep learning architecture combining 2D convolutional neural networks (CNNs) and long short-term memory (LSTM) modules.
• Feature recalibration using a 'squeeze and excitation' block.
• A linear function to handle variable input numbers, enhancing model applicability.

Main Results:

• Achieved high performance across three public datasets (SHHS, ISRUC, Sleep-EDF) with accuracy ranging from 0.86 to 0.87.
• Demonstrated outstanding performance on both healthy subjects and patients with sleep disorders.
• Highest accuracy was obtained by fusing multiple PSG signals.

Conclusions:

• The proposed model demonstrates comparable or superior performance to state-of-the-art methods with lower computational cost.
• The model exhibits excellent transferability across datasets without architectural changes, facilitating transfer learning.
• The versatile and effective model can be integrated into various PSG systems for clinical and routine sleep monitoring.