Pediatric Sleep Staging from Clinical Polysomnography: EEG Channel Relevance, Multimodal Fusion, and Adult-to-Pediatric Transfer

Automatic sleep stage classification has been extensively studied in adults but remains less explored in children because clinical pediatric polysomnography (PSG) datasets are limited and sleep physiology changes throughout development. In this work, we present a clinical pediatric PSG dataset collected from the archives of Victor Gomoiu Children’s Hospital (VGCH), consisting of 20 overnight recordings from children aged 9–17 years. Using AttnSleep, we assess EEG channel relevance, the contribution of EOG and EMG signals, and adult-to-pediatric transfer learning from SHHS and Sleep-EDFx datasets. Among the individual EEG derivations, P4 achieves the highest cross-validation (CV) means, with 79.5% accuracy and 70.8% balanced accuracy, while several temporal and posterior derivations perform similarly. Adding both EOG channels increases mean accuracy to 82.9%, macro-F1 to 74.9%, and balanced accuracy to 75.0%, whereas EMG produces smaller changes. The highest performance is obtained after Sleep-EDFx EEG+EOG pretraining and full fine-tuning on VGCH, reaching 84.0% accuracy, 76.4% macro-F1, and 76.4% balanced accuracy. Within this cohort of 20 recordings, the observed patterns highlight the potential value of combining EEG and EOG and support further evaluation of adult-to-pediatric transfer learning for pediatric sleep-stage classification.