Standardized image-based polysomnography database and deep learning algorithm for sleep-stage classification

Area of Science:

• Sleep medicine
• Artificial intelligence
• Biomedical data science

Background:

• Polysomnography (PSG) scoring is challenging due to labor intensity, subjectivity, and ambiguity.
• Existing deep learning (DL) models for automated sleep scoring are limited by fixed input channel and resolution requirements.
• Data heterogeneity from various PSG devices and lab environments complicates automated analysis.

Purpose of the Study:

• To develop a standardized image-based dataset for polysomnography (PSG) data.
• To create and validate an image-based deep learning (DL) model for automated sleep staging.
• To overcome the limitations of raw signal heterogeneity in PSG analysis.

Main Methods:

• Converted European data format files containing raw PSG signals into standardized images.
• Developed an image-based DL model for automatic sleep staging.
• Compared the image-based DL model against a signal-based model and validated it on an external dataset.

Main Results:

• Constructed a dataset of 10,253 image-based PSG records.
• The image-based DL model achieved over 80% accuracy, comparable to signal-based models.
• Demonstrated explainable AI (DL) in sleep medicine using Eigen-class activation maps and achieved good external validation performance.

Conclusions:

• A standardized image-based PSG dataset has been successfully created.
• The DL model shows robustness to changes in data sampling rate or sensor count, with minor performance variations.
• This approach offers a flexible and potentially more adaptable solution for automated sleep scoring.