IEEE Trans Neural Syst Rehabil Eng. 2026 Oct 29;PP. doi: 10.1109/TNSRE.2026.3747551. Online ahead of print.
ABSTRACT
Slow wave sleep (SWS) is vital for neural recovery and cardiovascular stability in children following congenital heart disease (CHD) surgery. However, pediatric CHD sleep patterns are often disrupted by clinical stressors and exhibit substantial domain discrepancies relative to adult data distributions. To address these challenges, this study establishes the Pediatric Cardiac Post-surgical Sleep (PCPS) dataset, a self-collected and expert-annotated pediatric polysomnography (PSG) dataset specifically targeting the early post-CHD surgical period. We further propose PCPS-SleepNet, an end-to-end deep learning framework for cross-domain sleep staging from single-lead EEG, enabling effective knowledge transfer from public sleep data to pediatric post-surgical CHD data. The architecture employs a multi-band input representation across six sub-bands and leverages a Four-Stage Progressive Encoder integrating Multi-Scale Depthwise Convolution (MSDC) for local features, self-attention for global density modeling, and a Conformer block for hybrid refinement. To mitigate limited pediatric data availability and domain shift, this network is pre-trained on public datasets and fine-tuned on the PCPS dataset. Subject-independent validation showed competitive performance, with an overall accuracy of 75.65 ± 0.58%, an AUROC of 89.06 ± 0.18%, an AUPRC of 83.49 ± 0.13%, and a macro F1-score of 76.68 ± 0.70%. For SWS detection, the model achieved an AUPRC of 87.89 ± 0.20% and an F1-score of 81.55 ± 0.74%. Architectural ablations and frequency-band input contribution analysis supported the model design and suggested the complementary discriminative value of slow-wave frequency-band inputs. These results suggest a practical and patient-friendly sleep detection solution for individualized sleep assessment and rehabilitation in vulnerable pediatric populations.
PMID:42955757 | DOI:10.1109/TNSRE.2026.3747551