Automatic Sleep Stage Classification of Children with Sleep-Disordered Breathing Using the Modularized Network

Area of Science:

• Pediatric Sleep Medicine
• Artificial Intelligence in Healthcare
• Biomedical Signal Processing

Background:

• Accurate sleep stage classification is crucial for diagnosing sleep disorders in children.
• Manual scoring of polysomnography (PSG) data is time-consuming and subjective.
• Developing automated methods can improve efficiency and consistency in pediatric sleep analysis.

Purpose of the Study:

• To create an automated sleep stage analysis model for pediatric patients.
• To assess the model's effectiveness in diagnosing sleep-disordered breathing (SDB).
• To compare automated sleep parameter calculations with traditional methods.

Main Methods:

• Deep neural networks were developed using electroencephalography (EEG), electrooculography (EOG), and electromyogram (EMG) data.
• The study included 344 pediatric SDB patients aged 2-18 years undergoing PSG.
• Model performance was evaluated using accuracy, precision, recall, F1-score, and Cohen's Kappa coefficient (ĸ).

Main Results:

• The best performing model ensemble achieved 83.36% accuracy (ĸ=0.7817) for 5-stage sleep classification and 96.76% accuracy (ĸ=0.8236) for 2-stage classification.
• No significant differences were found in key sleep parameters (TST, SE, SOL, etc.) between manual scoring and the automated model.
• The model demonstrated high average classification accuracies (91.94%-92.76%) on external datasets.

Conclusions:

• An automated pediatric sleep stage classification model with high reliability and generalizability was established.
• The model can be utilized for quantitative sleep parameter calculation.
• The developed model shows potential for aiding in the evaluation of SDB severity in children.