Research Article: Protocol-aware dual-view single-channel EEG sleep staging with improved N1 discrimination on public sleep-EDF and SHHS benchmarks
Abstract:
To develop and evaluate DualViewSeqSleepNet, a protocol-aware single-channel EEG model for automatic sleep staging under reproducible public benchmark settings.
The model integrates raw-waveform and spectral encoders, gated epoch-level fusion, lightweight bidirectional cross-epoch aggregation, and an auxiliary loss targeting Wake–N1 and N1–N2 transitions. Primary testing used the exact Sleep-EDF SC-subset protocol of SeriesSleepNet. Secondary no-retuning evaluation used SHHS-1 healthy C4-A1 data.
On Sleep-EDF, DualViewSeqSleepNet achieved macro- F 1 0.821, accuracy 0.869, and Cohen's kappa 0.808, exceeding the SeriesSleepNet anchor and improving N1 F 1 from 0.506 to 0.535. On SHHS-1, it reached macro- F 1 0.791 and accuracy 0.854 without architecture-family retuning. Ablation and robustness analyses supported the contributions of dual-view fusion, cross-epoch aggregation, and N1-aware supervision.
DualViewSeqSleepNet provides a modest, protocol-aligned improvement for public single-channel sleep staging, especially near Wake–N1–N2 boundaries, but does not establish external clinical validation or autonomous deployment readiness.
Introduction:
Sleep stage scoring remains foundational in sleep medicine because sleep architecture shapes the interpretation of normal sleep physiology, sleep onset, fragmentation, and many disorder-relevant phenotypes ( 1 , 2 ). Manual scoring under American Academy of Sleep Medicine criteria, however, is labor intensive and only moderately reproducible, with longstanding inter-scorer variability that is most pronounced around transitional stages such as N1 ( 3 – 5 ). Automated staging is therefore clinically relevant not…
Read more