Explainable attention-augmented CNN-BiGRU hybrid model for automated sleep stage classification in pediatric subjects with obstructive sleep apnea

Abstract Purpose Reliable classification of pediatric sleep stages remains a challenge due to the complexity of polysomnographic (PSG) data and variability in manual scoring. This study investigates whether an attention-augmented CNN-BiGRU hybrid architecture can achieve accurate and interpretable four-class sleep stage classification using multi-channel EEG, while also identifying redundant channels to support sensor simplification. Methods We developed a hybrid architecture integrating channel attention, convolutional layers, and bidirectional gated recurrent units (BiGRUs) for four-class sleep staging (Wake, Stage 1, Stage 2, REM) using seven EEG channels from the NCH Sleep Databank. Data preprocessing included Butterworth filtering, 30-second epoch segmentation, z-score normalization, temporal augmentation, and class relabeling tailored to pediatric sleep. An ablation study examined the contribution of individual EEG channels, and explainability methods (SHAP, Grad-CAM) were employed to enhance model interpretability. Results The proposed model achieved 84.41% accuracy, a macro F1-score of 0.84, and an AUC of 0.9615. Ablation analysis indicated redundancy in the C3-M2 and O2-M1 channels, highlighting the feasibility of reduced-channel acquisition. Explainability techniques identified the most influential EEG channels and temporal regions relevant to classification. Conclusions The CNN-BiGRU hybrid framework provides accurate and interpretable pediatric sleep staging from multi-channel EEG. The approach supports scalable automation of PSG analysis, reduces dependency on redundant sensors, and offers clinical utility for diagnosing pediatric sleep disorders such as OSA.