Detection of pediatric obstructive sleep apnea using a multilayer perceptron model based on single-channel oxygen saturation or clinical features

Abstract

Purpose: This study was performed to develop and evaluate a method of detecting pediatric obstructive sleep apnea (OSA) using a multilayer perceptron (MLP) model based on single-channel nocturnal oxygen saturation (SpO₂) with or without clinical data.

Methods: Polysomnography data for 888 children with OSA and 417 unaffected children were included. An MLP model was proposed based on the features obtained from SpO₂ and combined features of SpO₂ and clinical data to screen symptomatic children for OSA. The performance of the overall classification was evaluated with the receiver operating characteristics curve and the metrics of sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (LR+), negative likelihood ratio (LR-), and accuracy.

Results: The sensitivity, specificity, PPV, NPV, LR+, LR-, and accuracy of the MLP model for SpO₂ of an obstructive apnea-hypopnea index (OAHI) cutoff value of 1, 5, and 10 were 0.62-0.96, 0.11-0.97, 0.70-0.81, 0.55-0.93, 1.08-21.0, 0.39-0.39, and 0.69-0.91, respectively. The area under the receiver operating characteristics curve of an OAHI cutoff value of 1, 5, and 10 was 0.720, 0.842, and 0.922, respectively. After adding the clinical data of age, sex, body mass index, weight category, adenoid grade, or tonsil scale, the performance of the MLP model was basically at the same level as only single-channel SpO₂.

Conclusions: Application of this MLP model using single-channel SpO₂ in children with snoring has high accuracy in the diagnosis of moderate to severe OSA but a poor effect in the diagnosis of mild OSA. The combination of clinical data did not significantly improve the diagnostic performance of the MLP model.

Keywords: Child; Diagnosis; Neural network; Obstructive; Sleep apnea.