Single-Channel EEG Sleep Staging

• Single-channel EEG sleep staging is a technique that uses data from a single EEG derivation to automatically classify sleep stages, offering simplicity and cost-effectiveness.
• It employs classical signal processing features alongside advanced deep learning architectures like CNNs, LSTMs, and attention mechanisms to capture detailed sleep patterns.
• Current systems achieve 80–92% accuracy with strong performance for most stages while addressing challenges in N1 detection and device adaptability.

Single-channel EEG sleep staging refers to the automated classification of standard sleep stages—Wake (W), N1, N2, N3, and REM—using data from a single EEG derivation. Unlike conventional polysomnography (PSG), which leverages multiple EEG, EOG, and EMG channels, single-channel approaches prioritize simplicity, cost-effectiveness, comfort, and real-time feasibility, supporting both clinical and ambulatory scenarios. The field has progressed rapidly, with deep neural architectures now rivaling multi-channel and expert-level performance in large-scale validation, while enabling efficient on-device or wearable deployment and extending to advanced interpretability and domain adaptation (Li et al., 2024, Koushik et al., 2018, Vakili et al., 28 Dec 2025, Liao et al., 2024).

1. Signal Acquisition, Datasets, and Preprocessing

Single-channel EEG for sleep staging typically uses bipolar derivations such as Fpz–Cz (Sleep-EDFx/EDF, 100 Hz), C3–A2 (ISRUC, 200–256 Hz), Cz–A1 (DREAMS-SUB, 200 Hz), or C4–A1 (SHHS, 125 Hz), adhering to the 10–20 system and referenced to mastoid/earlobe. Consumer and wearable devices (e.g., Muse, ear-EEG) increasingly support high-fidelity single-channel acquisition (Koushik et al., 2018, Nakamura et al., 2017).

Public datasets employed include Sleep-EDFx/EDF, MASS, ISRUC, SHHS, and Physio2018, providing 30 s epochs scored per AASM or Rechtschaffen–Kales standards. Preprocessing pipelines standardize sampling rates (typically 100 Hz), segment continuous EEG into non-overlapping 30 s epochs, and apply band-pass (0.5–40 Hz) and notch (50/60 Hz) filters to suppress slow drifts and mains interference. Artifact rejection—via amplitude thresholds, label exclusion, or advanced blind source separation for single-channel data—is often implemented (Koushik et al., 2018, Li et al., 2024, Nakamura et al., 2017).

Normalization is crucial for cross-subject and cross-device robustness: Z-score normalization per epoch or via a baseline wake epoch standard deviation $\sigma_{\text{wake}}$ has shown effectiveness, especially for on-device adaptation and transfer learning (Koushik et al., 2018, Liao et al., 2024, Vakili et al., 28 Dec 2025).

2. Feature Extraction and Network Architectures

Approaches for extracting discriminative sleep-related patterns span traditional signal processing and advanced deep learning paradigms:

Classical Features:

• Time-domain: mean, variance, zero-crossing rate, Hjorth parameters.
• Frequency-domain: band powers (δ, θ, α, β), relative power ratios, spectral edge frequency (SEF), spectral entropy.
• Nonlinear/time-frequency: wavelet coefficients, sample entropy, Hilbert–Huang transforms (Nakamura et al., 2017, Li et al., 2024).

Deep Learning Feature Extractors:

• CNN-based models: 1D convolutional nets learn time-domain filters aligned with physiological microstructures; multi-scale conv blocks (kernels of 3, 5, 7, or larger) capture both slow-wave and spindle/sawtooth phenomena (Wang et al., 2021, Vakili et al., 28 Dec 2025, Humayun et al., 2019).
• Residual and Dense Architectures: Pre-activation ResNets (34-layer) and dense blocks permit very deep stacks, stabilizing gradient flow and enhancing feature selectivity (Humayun et al., 2019, Wang et al., 2021, Sharma et al., 2023).
• Attention Mechanisms: Squeeze-and-Excitation (SE), channel/spatial or dual-attention modules provide channel- or time-point-wise recalibration, increasing model sensitivity to subtle staging cues (Wang et al., 2021, Sharma et al., 2023, Vakili et al., 28 Dec 2025).
• Temporal Modeling: Bidirectional LSTMs, Bi-GRUs, and Transformer encoders capture long-range inter-epoch dependencies. Hierarchical and multi-scale temporal encoders (e.g., context-aware multi-scale + hierarchical BiLSTM, feature pyramid + Transformer) have demonstrated improvement in challenging classes (especially N1/REM) and calibration (Vakili et al., 28 Dec 2025, Lee et al., 2022, Sharma et al., 2023, Sadik et al., 2023).

Table 1. Representative architectures and key components (abbreviated selection):

Architecture	Key Components	Context Modeling
DeepSleepNet	Dual-branch CNN (small/large kernels) + Bi-LSTM	Sequence Bi-LSTM
MSDAN	Multi-scale CNN + Dual (channel/spatial) Attention	-
ContextAware 2025	Multi-scale + SE + Dilated CNN + BiLSTM + Attention	Hierarchical
DenseRTSleep-II	CNN (Dense) + Transformer + BiLSTM	Transformer-BiLSTM

3. Training Strategies, Class Imbalance, and Evaluation

Supervised training employs categorical cross-entropy, sometimes augmented with class-weighted losses, focal loss, or specialized terms (e.g., MFE/MSFE) to counter severe imbalance (notably low prevalence of N1) (Wang et al., 2021, Mousavi et al., 2019, Vakili et al., 28 Dec 2025). Data augmentation—amplitude scaling, Gaussian noise, time-shifting, temporal masking—is used to enrich minority classes (Vakili et al., 28 Dec 2025, Zhang et al., 2024).

Optimization is typically performed using Adam or RMSProp, with dropout, L2 weight decay, and batch or adaptive normalization to prevent overfitting. Learning rate schedules (ReduceLROnPlateau, cyclic LR) and early stopping are standard practice (Koushik et al., 2018, Vakili et al., 28 Dec 2025, Liao et al., 2024). Stratified group k-fold cross-validation by subject ensures robust out-of-sample performance estimation.

Metrics: Overall accuracy, macro-F1, per-class F1, and Cohen’s κ (agreement with human scorers, κ≈0.7–0.83 considered competitive). Per-class recall highlights persistent challenges: Wake and N2/N3 reproducibly achieve >85% F1; N1 remains systematically lower (typically 40–62%), driven by its transitional/overlap status and single-channel limitations (Vakili et al., 28 Dec 2025, Sharma et al., 2023, Koushik et al., 2018, Seo et al., 2019).

4. Knowledge Distillation, Domain Adaptation, and Semi-/Self-Supervision

Sophisticated transfer and distillation methods leverage abundant multi-channel data and/or teacher-student schemes to improve single-channel generalization:

• MCMD knowledge distillation: Multi-channel, multi-domain pre-training on PSG (MASS dataset), teacher outputs (including EMG/EOG knowledge) transferred via intermediate and output alignment to a single-channel student model; observed 2% accuracy gain and negligible student-teacher gap (Zhang et al., 2024).
• Self-supervised and contrastive pretraining: Hybrid contrastive + masked auto-encoding on unlabeled EEG, followed by supervised fine-tuning, enables performance parity with fully supervised models (NeuroNet + Mamba TCM) and strong cross-dataset generalization (Lee et al., 2024).
• Adaptive normalization and gradient re-weighting: Subject-wise AdaBN plus gradient density-based loss reweighting allow rapid, unsupervised personalization to novel subjects with no retraining, supporting on-device deployment (Liao et al., 2024).
• Self-distillation (MoME Transformers): Mixture-of-expert transformer architectures with multimodal teacher, EEG-only student, and cross-modal interaction yield state-of-the-art single-channel accuracy (88% on mouse dataset) (Chen et al., 27 Jan 2025).

5. Interpretability and Clinical Relevance

Explainability modules are integrated into modern single-channel sleep staging systems:

• Attention-based visualization: GradCAM adapted for 1D convolutional architectures, channel-wise SE, temporal attention, and segment importance heatmaps reliably align high-activation regions with established sleep microstructures (alpha, spindles, K-complexes, delta, low-amplitude REM) (Sharma et al., 2023, Vakili et al., 28 Dec 2025, Guo, 2024).
• Calibration analysis: Reliability curves quantify probability accuracy for rare classes (N1), with state-of-the-art systems achieving ECE ≈ 0.010 (Vakili et al., 28 Dec 2025).
• Resource usage and real-time feasibility: Lightweight architectures (DetectsleepNet-tiny, LightSleepNet) achieve ≲0.05M parameters and ≲50 MFLOPs, attaining >80% accuracy in edge/on-device settings (Guo, 2024, Liao et al., 2024).
• Clinical deployment: Smartphone integration (TensorFlow Lite interpreter, Bluetooth EEG streaming, sub-10 ms latency) enables fully mobile, real-time monitoring, with performance comparable to clinical PSG (ACC ≈83% vs. gold standard ≈85–90%) (Koushik et al., 2018).

6. Current Performance, Challenges, and Future Directions

Single-channel systems on Sleep-EDF, MASS, SHHS, and Physio2018 datasets now routinely achieve accuracy 80–92%, macro-F1 >75%, and κ up to 0.87–0.89, rivaling expert agreement (Vakili et al., 28 Dec 2025, Humayun et al., 2019, Wang et al., 2021, Lee et al., 2022, Guo, 2024, Sharma et al., 2023). Performance on minority stages (N1) has advanced, with F1 ≈54–62% reported by models incorporating multi-scale, temporal, and attention modules (Vakili et al., 28 Dec 2025, Wang et al., 2021, Sharma et al., 2023). Systems such as Context-Aware Temporal Modeling, MSDAN, and MCMD-KD narrow the gap to multi-channel approaches.

Outstanding challenges include cross-dataset generalization, robust artifact attenuation in real-world conditions, and overcoming single-channel limitations for microstate detection. Interpretability, subject/device adaptation, and sequence modeling remain active areas. Emerging directions focus on transformer/graph-based temporal modeling (Vakili et al., 28 Dec 2025, Chen et al., 27 Jan 2025), self-supervised and semi-supervised learning for scalable annotation, and integration with multi-modal generative augmentation (YOAS) for virtual channel enrichment (Li et al., 2024, Vakili et al., 28 Dec 2025, Zhang et al., 2024).

7. Comparative Table of Core Results

Model/Approach	Dataset	ACC (%)	MF1 (%)	κ	N1 F1 (%)	Reference
Context-Aware + HSL + SE	SleepEDF-20	89.7	85.5	85.9	61.7	(Vakili et al., 28 Dec 2025)
MSDAN (Multi-scale Attention)	Sleep-EDF	91.7	82.3	87.2	54.4	(Wang et al., 2021)
SE-ResNet+Bi-LSTM+1DGradCAM	SleepEDF-20	87.5	82.5	0.82	56.9	(Sharma et al., 2023)
MCMD KD (KD from multi-chan)	Sleep-EDF	86.5	80.9	0.82	–	(Zhang et al., 2024)
DetectsleepNet-Tiny	SHHS	87.4	80.7	0.831	48.5	(Guo, 2024)
IITNet (BiLSTM, sub-epoch RF)	SleepEDF	83.9	77.6	0.78	–	(Seo et al., 2019)
DeepSleepNet	Sleep-EDF	82.0	76.9	0.76	46.6	(Supratak et al., 2017)
LightSleepNet (on-device)	Sleep-EDF	83.8	75.3	0.78	31.0	(Liao et al., 2024)

Performance is consistently robust for majority classes (Wake, N2/N3, REM), while N1 remains the primary limitation. All recent systems demonstrate substantial gains in efficiency, interpretability, or transferability, propelling single-channel EEG to the frontier of automated sleep medicine.