Evaluation of automated pediatric sleep stage classification using U-Sleep: a convolutional neural network

Area of Science:

• Sleep Medicine
• Artificial Intelligence in Healthcare
• Pediatric Sleep Analysis

Background:

• U-Sleep is a publicly available automated sleep staging tool.
• Independent validation in pediatric populations is lacking.
• This study addresses the need for validation in children.

Purpose of the Study:

• To validate U-Sleep's performance against trained human scorers using pediatric polysomnograms.
• To identify demographic and clinical factors influencing U-Sleep accuracy in children.
• To assess the equivalence of U-Sleep to expert human sleep staging.

Main Methods:

• Utilized a concordance dataset (50 pediatric polysomnogram excerpts) and a clinical dataset (3,114 polysomnograms).
• Compared U-Sleep's 5-stage sleep staging to "gold" standard scoring using Cohen's kappa.
• Employed Wilcoxon 2 1-sided test for equivalence testing and multivariable regression for factor analysis.

Main Results:

• U-Sleep achieved statistically equivalent performance to trained humans (kappa = 0.79 vs 0.78) in the concordance dataset.
• Median kappa agreement in the clinical dataset was 0.69.
• Lower accuracy was observed in children under 2 years, those with comorbidities, decreased sleep efficiency, or sleep-disordered breathing.

Conclusions:

• U-Sleep demonstrates statistically equivalent performance to trained scorers for pediatric sleep staging.
• Accuracy is reduced in specific pediatric subgroups, including infants and those with sleep-disordered breathing or comorbidities.
• U-Sleep is suitable for clinical use in pediatrics when supplemented by expert clinician review.