Source-linked AI summary
Automatic Sleep Stage Scoring with Single-Channel EEG Using Convolutional Neural Networks
Orestis Tsinalis, Paul M. Matthews, Yike Guo, Stefanos Zafeiriou
TL;DR
Single-channel EEG sleep staging needs an accessible automatic scoring method, while CNNs have been less widely applied to one-dimensional biosignals. This paper develops an end-to-end CNN that learns task-specific filters without AASM prior knowledge and achieves performance comparable to a state-of-the-art hand-engineered approach, with interpretable filters consistent with AASM guidelines.
Problem
Accessible, reliable automatic sleep-stage scoring from portable single-channel EEG remains needed for unsupervised, at-home sleep monitoring.
Method
The paper uses an end-to-end CNN for single-channel EEG sleep-stage scoring, with convolutional filtering, pooling, stacking, fully connected layers, and class-balanced training for imbalanced stages.
Results
The CNN achieves performance comparable to a state-of-the-art hand-engineered feature approach using single-channel EEG, while learning filters that closely capture AASM frequency guidelines.
Takeaways & Limitations
CNN filters can be interpretable in terms of human sleep-scoring rules and remain consistent across cross-validation folds without using AASM prior knowledge.
Takeaways & Limitations
The comparison between automatic filter learning and AASM-based hand-engineered features may require a larger dataset to clarify performance differences.
Abstract
from arXiv · showhide
We used convolutional neural networks (CNNs) for automatic sleep stage scoring based on single-channel electroencephalography (EEG) to learn task-specific filters for classification without using prior domain knowledge. We used an openly available dataset from 20 healthy young adults for evaluation and applied 20-fold cross-validation. We used class-balanced random sampling within the stochastic gradient descent (SGD) optimization of the CNN to avoid skewed performance in favor of the most represented sleep stages. We achieved high mean F1-score (81%, range 79-83%), mean accuracy across individual sleep stages (82%, range 80-84%) and overall accuracy (74%, range 71-76%) over all subjects. By analyzing and visualizing the filters that our CNN learns, we found that rules learned by the filters correspond to sleep scoring criteria in the American Academy of Sleep Medicine (AASM) manual that human experts follow. Our method's performance is balanced across classes and our results are comparable to state-of-the-art methods with hand-engineered features. We show that, without using prior domain knowledge, a CNN can automatically learn to distinguish among different normal sleep stages.
1 Introduction
The paper motivates accessible automatic sleep-stage scoring from single-channel EEG and presents CNNs as a way to learn task-specific filters without prior domain knowledge.
- CNN motivation: CNNs integrate feature extraction and classification while learning task-specific filters directly from raw input.This end-to-end approach avoids specifying domain-informed features in advance.
- Paper objective: The authors present a CNN architecture for automatic sleep-stage scoring using a single EEG channel.The method targets wearable, affordable, portable, and unobtrusive monitoring applications.
- Motivation for sleep scoring: Sleep-stage scoring is important because disrupted sleep is associated with health consequences and neurodegenerative disease risk.The paper highlights potential value in detecting disruption early and treating sleep pathologies.
- Comparison with prior work: The study compares automatic filter learning with a previous approach using hand-engineered Morlet-wavelet features and stacked sparse autoencoders.The prior method used the Fpz-Cz electrode and features tuned to sleep-stage-specific signal characteristics.
2 Materials and Methods
The study evaluates a raw single-channel EEG CNN using a publicly available sleep dataset, context windows, a stacked convolutional architecture, class balancing, and subject-wise cross-validation.
- Dataset: The evaluation uses a publicly available PhysioNet sleep PSG dataset with recordings from 20 healthy young adults.The study removes Movement and Not Scored epochs and merges N3 and N4 into N3.
- CNN architecture: The architecture combines two convolutional-pooling pairs, a stacking layer, two fully connected layers, and a softmax output.The stacking layer rearranges 20 filtered and subsampled signals into a 2D input for the second convolutional layer.
- Training: Class-balanced random sampling is incorporated into SGD training to prevent performance from favoring overrepresented sleep stages.The dataset contains many more epochs for some stages, especially N2, than for W and N1.
- Input representation: The CNN receives raw EEG from the current epoch plus two preceding and two succeeding epochs as a 15,000-timepoint input.The surrounding context addresses transition-dependent distinctions among N1, N2, and R.
- Evaluation: The study uses 20-fold subject-wise cross-validation and evaluates precision, sensitivity, F1-score, per-stage accuracy, and overall accuracy.Metrics are derived from a normalized confusion matrix and reported across five one-vs-all sleep-stage classifications.
- Additional analyses: Sleep efficiency and transitional-epoch percentage are measured to assess whether scoring performance varies with sleep quality or epoch ambiguity.The analysis computes R2 and p-values for associations between these recording-level measures and scoring performance.
3 Results
Across cross-validation, the CNN produced balanced sleep-stage classification, with performance largely independent of sleep quality and transitional-epoch prevalence. Learned filters showed consistent stage-associated frequency patterns across folds.
- Sleep stage scoring performance: Around 90% of N3 epochs were correctly classified, compared with around 75% for R and N2, around 70% for W, and 60% for N1.
- Sleep stage scoring performance: N1-W and N1-R were the most frequent confusion pairs at about 15%, while N1-N3, N3-R, and N3-W errors were close to zero.
- Sleep stage scoring performance: False negatives were approximately balanced within most stage pairs, indicating that class-imbalance-related misclassification errors were mitigated.
- Sleep stage scoring performance: Mean F1-score was 79% (range 81–83%), mean accuracy across individual sleep stages was 80% (range 82–84%), and overall accuracy was 74% (range 71–76%).
- Sleep stage scoring performance: Regression tests found no significant relationship between scoring performance and sleep efficiency or transitional-epoch percentage, with all coefficient p-values above 0.25 and R2 below 0.05.
- CNN filter analysis and visualization: Across folds, learned filters showed recurring frequency associations with stages N3, N2, W, R, and N1, including distinct low-, mid-, and high-frequency patterns.
4 Discussion
The CNN learned filters with characteristics aligned to AASM sleep-scoring criteria, while end-to-end CNN performance was intermediate between the SAE and fixed-filter M-CNN models. Remaining errors likely reflect information unavailable from single-channel EEG, and several model differences were not statistically significant.
- Model comparison: SAE outperformed CNN, which outperformed M-CNN, although confidence intervals overlapped across nearly all metrics.The reported differences across subjects were not statistically significant overall, except for mean and worst-stage accuracy between SAE and M-CNN.
- Model comparison: Hand-engineered AASM-based features may outperform automatic filter learning, but the dataset was insufficient to establish a statistically significant difference.The authors suggest a larger dataset could clarify performance differences and benefit CNN training.
- Model comparison: Fixed hand-engineered filters produced worse performance than end-to-end CNN learning, although this difference was not statistically significant.The M-CNN architecture was not optimized for its fixed filters.
- Limitations: Balanced performance across sleep stages remained limited by patterns related to EOG and EMG that single-channel EEG cannot readily capture.Adding more than 20 filters did not improve results and sometimes worsened them.
- Learned filters: CNN filters showed AASM-consistent patterns, including low-frequency N3 activity and spindle-related 12.5-14 Hz activity with little N2 activation.These filters suggest the CNN captured N3 epochs in which sleep spindles persist.
- Learned filters: R-stage filters consistently captured 2-5 Hz activity, approximately 14 Hz activity, and absent high-frequency activity, while N1 filters were less consistent across folds.N1 was the most misclassified class, and one recurring N1 filter combined activity near 7 and 9 Hz with 20-50 Hz activity.
5 Conclusion
The CNN achieved sleep-stage scoring comparable to a state-of-the-art hand-engineered approach using single-channel EEG without AASM prior knowledge. Its learned filters were interpretable, closely matching AASM frequency guidelines and remaining consistent across cross-validation folds.
- The CNN matched a state-of-the-art hand-engineered feature approach for single-channel EEG sleep-stage scoring without using AASM prior knowledge.
- Learned filters closely captured AASM guidelines through their frequency characteristics for each sleep stage.
- Filter interpretations were consistent across cross-validation folds.
- The filter-analysis approach may support understanding in other EEG- and ECG-based classification applications with limited domain knowledge.