Deep Learning for Non-Invasive Obstructive Sleep Apnea Screening in Children: A Systematic Comparison of Seven Architectures

Abstract

Overnight polysomnography remains the diagnostic gold standard for pediatric obstructive sleep apnea, yet is widely inaccessible due to cost and infrastructure constraints. This study benchmarks seven deep learning architectures on the Cleveland Children's Sleep and Health Study cohort of 517 children, using 59 non-invasive features spanning craniofacial proxies, anthropometric measurements, and symptom-based questionnaire responses. While attention-based models led cross-validation performance, ResNet achieved the highest test AUC of 0.7375, underscoring the generalization advantage of architecturally parsimonious models in low-sample clinical settings. A weighted ensemble further yielded a three-tier risk stratification with OSA prevalence rising from 6.0% to 50.2% across strata, offering a practical triage instrument for primary care. These findings establish a reproducible non-invasive screening benchmark and advocate for generalization-aware model selection in pediatric clinical machine learning.