Source-linked AI summary
Automatic Sleep Staging of EEG Signals: Recent Development, Challenges, and Future Directions
Huy Phan, Kaare Mikkelsen
TL;DR
Manual sleep scoring is costly, difficult to scale, and constrained by scorer subjectivity and limited PSG recordings, motivating automatic sleep staging. This review synthesizes sequence-to-sequence deep-learning methods and their clinical challenges; these systems achieve expert-like performance mainly on healthy subjects, but broader clinical usefulness remains unresolved.
Problem
Manual sleep scoring is labor-intensive and difficult to scale, while variation between scorers and limited PSG recordings constrain clinical monitoring.
Method
The review surveys recent automatic EEG sleep-staging methods, emphasizing sequence-to-sequence architectures that jointly encode epochs and their temporal context.
Results
Cohen’s kappa of ≥0.81 is reported on databases dominated by healthy subjects, whereas performance remains substandard on pathological databases.
Takeaways & Limitations
Automatic staging has comparable performance to sleep experts on healthy individuals, but clinical usefulness still requires addressing sleep disorders, data heterogeneity, explainability, and subjectivity.
Takeaways & Limitations
Mobile sleep monitoring has lower inter-scorer reliability than PSG, with state-of-the-art Cohen’s kappa around 0.75 versus above 0.8 for PSG.
Abstract
from arXiv · showhide
Modern deep learning holds a great potential to transform clinical practice on human sleep. Teaching a machine to carry out routine tasks would be a tremendous reduction in workload for clinicians. Sleep staging, a fundamental step in sleep practice, is a suitable task for this and will be the focus in this article. Recently, automatic sleep staging systems have been trained to mimic manual scoring, leading to similar performance to human sleep experts, at least on scoring of healthy subjects. Despite tremendous progress, we have not seen automatic sleep scoring adopted widely in clinical environments. This review aims to give a shared view of the authors on the most recent state-of-the-art development in automatic sleep staging, the challenges that still need to be addressed, and the future directions for automatic sleep scoring to achieve clinical value.
1. Introduction
Sleep staging assigns a sleep stage to every 30-second epoch of an overnight polysomnogram, supporting clinical assessment and diagnosis. Because manual scoring is labor-intensive and difficult to scale, recent deep-learning work has pursued automated scoring using large public datasets.
- Clinical role: Sleep staging assigns each 30-second epoch of an overnight polysomnogram to a sleep stage.The resulting sequence supports measurement of sleep cycles, stage durations, latency, and wake after sleep onset.
- Clinical role: Sleep-stage sequences can provide diagnostic markers for disorders ranging from obstructive sleep apnea to narcolepsy.
- Motivation for automation: Manual scoring is labor-intensive, time-consuming, and poorly suited to large-scale data or the needs of millions with sleep disorders.
- Motivation for automation: Automated staging could support longitudinal home monitoring while reducing clinician workload and expanding access to sleep assessment and diagnosis.The paper describes machine scoring as especially suitable because sleep staging follows predefined rules.
- Review scope: This review surveys recent automatic EEG sleep-staging methods, limits its scope to fine-grained five-stage PSG or brain-activity modalities, and discusses remaining challenges and future directions.It excludes ECG/PPG, actigraphy, audio, video, and radar-based studies.
2. Clinical context
Manual sleep scoring is standardized and versatile but remains a costly local optimum because it is time-consuming, subject to scorer variation, and tied to invasive PSG recordings. Automation could reduce existing costs and enable longitudinal monitoring and new clinical uses that current workflows make infeasible.
- Manual scoring: Manual scoring is reliable, versatile, interpretable, and globally standardized, but the paper characterizes it as a suboptimal local optimum.
- Manual scoring: Scoring an entire night manually is time-consuming and expensive, especially when sleep events must also be annotated.
- Manual scoring: Scorer variation persists despite a standardized sleep-scoring manual.
- Manual scoring: PSG-based scoring is unwieldy and invasive, so clinicians usually collect only one or at most two nights that may not represent usual sleep.
- Clinical adoption: Existing manual-scoring practices can make the benefits of automatic scoring with other sensor setups difficult to recognize in clinical settings.
- Clinical opportunities: Automatic scoring could reduce PSG costs and enable population screening, routine post-hospital follow-up, and longer recordings for discarding low-quality nights.
3. The state-of-the-art sleep scoring
Deep-learning sleep staging evolved from short-context standalone networks toward sequence-to-sequence architectures that jointly learn epoch features and long-term context. Public-database results reach almost-perfect agreement on mostly healthy cohorts, but remain substandard for pathological cohorts and are not directly comparable across differing setups.
- Early methods: Early deep-learning systems used short contexts around a target epoch with standalone DNN, CNN, or RNN architectures and achieved limited success.
- Long-term modelling: Long-term modelling improved automatic staging because sleep experts also use broad context when assigning labels.
- Sequence-to-sequence framework: Sequence-to-sequence staging transforms each epoch into a feature vector and then models interactions across the sequence using an epoch encoder and sequence encoder.
- Sequence-to-sequence framework: The framework jointly optimizes both encoders end-to-end and classifies every epoch in an input sequence rather than only the last target epoch.
- Architectures: Most reviewed systems use CNNs for epoch encoding, with alternatives including ResNet, U-Net, U2-Net, LSTM, and GRU architectures.
- Evaluation: Cohen’s kappa of ≥0.81 was achieved on databases dominated by healthy subjects, while performance remained substandard on pathological databases such as ISRUC and CAP.
- Evaluation: Table results should not be used for direct system comparisons because datasets, channel counts, adaptation strategies, and learning settings differ.
4. Challenges and future directions
Healthy-subject PSG staging has nearly reached an agreement ceiling, but clinical deployment still requires robustness to pathological sleep phenotypes and practical wearable monitoring. Future progress therefore depends on addressing disease-related signal changes and the constraints of long-term home recording.
- Current status: Automatic PSG staging on healthy adults and children has basically been solved, with little room for meaningful accuracy improvement within the sequence-to-sequence framework.
- Clinical PSG: Clinical tools must handle non-textbook sleep phenotypes whose disease-related artifacts or altered EEG may require specialized routines or make correct scoring impossible.
- Wearable EEG: Wearable EEG could enable weeks of comfortable monitoring, producing data more representative of patients’ actual sleep and supporting screening, diagnosis, and follow-up.
APPLICATIONS CHALLENGES FUTURE DIRECTIONS
Automatic scoring makes long-term sleep monitoring more feasible and could support personalized models, adaptive tracking, change detection, and sleep-based biomarkers.
- Applications: Automatic scoring combined with self-applied, unobtrusive hardware makes longitudinal monitoring over weeks more feasible.Conventional PSG scored by hand is too expensive and inconvenient for repeated nightly recordings.
- Future directions: Large numbers of nights could support personal models that adapt to changing sleep patterns while resisting catastrophic forgetting.The proposed models would retain knowledge of infrequent sleep types or events.
- Future directions: Weeks of recordings before and after major events could enable detection of serious changes in an individual’s sleep.Examples include surgery, medication changes, and disease onset.
- Future directions: Sufficient recordings may transform individual sleep patterns into reliable biomarkers for detecting changes or disease-related sleep characteristics.The paper links this possibility to trait-like characteristics that remain specific to individuals despite night-to-night variation.
4.2. Low signal quality - device limitations for mobile sleep monitoring
Mobile sleep monitoring is promising but remains constrained by lower signal quality, subject and device variability, and limited training data; personalization and transfer learning are common responses.
- Device limitations: Mobile devices achieve lower sleep-scoring reliability than PSG, with Cohen’s kappa about 0.75 versus above 0.8 for PSG.The lower signal-to-noise ratio of mobile recordings is identified as a likely cause.
- Device limitations: Mobile data can vary across recording days because device mounting may differ, adding day-to-day variability to subject-dependent signal loss.The paper models mobile recordings as lower-dimensional projections of PSG data with information loss.
- Personalization: Subject-specific information improves sleep-scoring performance, although sufficiently broad cohorts might eventually make personalized and general algorithms similar.The benefit appears to come from including maximally relevant data rather than merely excluding unrelated data.
- Personalization: Personalized models can be built from a single or few nights using random forests, while neural networks commonly use fine-tuning of population models.These approaches address the limited data available for individuals and new devices.
- Data limitations: New mobile devices usually have small training sets because they lack established user bases, may change across generations, and require paired device and PSG recordings.Deep-learning models generally reach expert-level performance only with hundreds or thousands of subjects.
- Transfer learning: Transfer learning usually improves performance over scratch training, but its gains can be modest when adapting pretrained models to small new datasets.The approach fine-tunes a model trained on a large sleep dataset, often based on PSG recordings.
4.4. Privacy preservation: a note for sleep monitoring
EEG contains sensitive personal information, making privacy protection a central concern for centrally deployed sleep-staging systems; federated learning is presented as one possible response.
- Privacy risks: EEG can reveal identity, age, gender, emotion, preferences, and personality, creating privacy and legal-compliance challenges.The paper specifically cites GDPR in the European Union and the Consumer Privacy Bill of Rights in the United States.
- Privacy risks: Centralized training, adaptation, and deployment may require transmitting EEG signals, raising data privacy and security concerns.The review notes that addressing these concerns is beyond its scope.
- Privacy-preserving methods: Federated learning is identified as a promising algorithmic solution to privacy and locality problems in sleep monitoring.The passage presents it as an emerging approach rather than a completed solution.
4.5. Disorders affecting sleep structure
Severe neurological disorders can alter sleep structure, timing, stage physiology, and transition patterns, making both manual scoring and model training harder. The review highlights disorder-resistant and more flexible approaches, while emphasizing unresolved interpretability and trust requirements for clinical adoption.
- Severe neurological disorders can alter sleep structure, timing, outward characteristics, and physiological features of sleep stages.
- These changes can make manual scoring harder, which in turn complicates training and validation of sleep-staging models.
- When sleep stages become less defined, continuous sleep-depth estimation may fit better than categorical staging.
- ASEEGA achieves an average Cohen’s kappa of 0.8 for healthy adults with full PSG and 0.75–0.80 across sleep disorders.
- Clinical adoption also requires interpretability because sleep stages are ambiguous, human scorers disagree, and deep models remain difficult to decipher.
- Objective evaluation and comparison of explainability remain open questions, especially across scalp and wearable EEG settings.
4.7. Data mismatch due to distributional shifts between datasets/cohorts
Sleep datasets differ across institutions, demographics, diseases, modalities, devices, and acquisition conditions, violating the independent-and-identically-distributed assumption. The review discusses adaptation, mapping, and invariant representations as responses to these distributional shifts.
- Sleep data mismatches across institutions, demographics, diseases, modalities, devices, and acquisition conditions violate the i.i.d. assumption.
- Collecting training data across every foreseeable condition is expensive, time-consuming, infeasible, and insufficient for novel setups.
- Domain adaptation migrates pretrained models to target cohorts with unseen conditions by modifying model parameters.
- Data mapping instead learns a function that transforms target-domain data toward the source domain before evaluation.
- Domain-invariant representations aim to preserve sleep-stage information while remaining agnostic to source- or target-specific factors.
4.8. Heterogeneity: a challenge beyond data mismatch
Heterogeneity in channel layouts and recording modalities complicates sleep staging beyond ordinary distributional mismatch. Integrating available channels through an intermediate interface could make heterogeneous inputs compatible with existing architectures.
- PSG recordings vary in channel number, modality, layout, and recording setup across sources and subjects.
- Different sleep markers are best captured by different channels, including alpha rhythm occipitally, REM sawtooth waves centrally, and N2 K-complexes centrally.
- Consolidating all available PSG channels could improve staging performance and support model building from larger multi-source datasets.
- An intermediate layer could amalgamate heterogeneous channels and modalities into a fixed-channel input for existing network architectures.
4.9. Subjectivity in model building
Manual sleep-stage labels are subjective, especially for ambiguous epochs, and single-scorer training can make models reproduce one scorer’s style. The review therefore points toward multi-scorer supervision and interactive personalization with end users.
- Human scorers reach a consensus of about Cohen’s kappa 0.76, with poorer agreement for mixed-stage, boundary, N1, and fragmented-sleep epochs.
- Finer temporal resolution does not remove subjectivity when supervision still derives from 30-second epoch annotations.
- Training with multiple scorers’ supervision could let models adapt to a cohort’s scoring style rather than averaging annotations into one label.
- Single-scorer training may overfit that scorer’s style, causing disagreement with clinicians who use the system.
- Scorer personalization requires closed-loop model–user interaction to identify disagreements, support corrections, and learn from end-user feedback.
5. Conclusions
Deep learning systems now achieve sleep-scoring performance comparable to experts in healthy individuals, marking an important methodological milestone. Clinical usefulness remains unresolved because challenges involving sleep disorders, data heterogeneity, explainability, subjectivity, robustness, privacy, and longitudinal monitoring still require attention.
- Comparable performance to sleep experts has been achieved for automatic sleep scoring, at least in healthy individuals.
- This progress reflects decades of research supported by large de-identified sleep datasets and open-source tools and techniques.
- Clinical usefulness still requires addressing sleep disorders, data heterogeneity, model explainability, and subjectivity.
- Deployment in daily living environments additionally requires prospective studies addressing low-quality mobile EEG, limited training data, privacy, and longitudinal monitoring.