Automatic Sleep Stage Scoring with Single-Channel EEG Using Convolutional Neural Networks (1610.01683v1)

Abstract: 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.

• The paper introduces a CNN that processes raw EEG data using preceding, current, and succeeding epochs to classify sleep stages.
• The study achieved a mean F1-score of 81% and 82% accuracy, demonstrating balanced performance despite class imbalance.
• The model eliminates manual feature extraction, paving the way for more accessible sleep monitoring and automated clinical analyses.

Automatic Sleep Stage Scoring with Single-Channel EEG Using Convolutional Neural Networks

The paper presents a convolutional neural network (CNN) architecture for the automated scoring of sleep stages using single-channel electroencephalography (EEG) signals. It evaluates the efficacy of CNNs in learning task-specific filters for classifying different sleep stages without leveraging prior domain knowledge. The research utilizes a publicly accessible dataset comprising recordings from 20 healthy young adults to measure performance, deploying 20-fold cross-validation to assess generalizability.

Methodology and Findings

The paper implements a CNN that processes raw EEG input data, eschewing the need for feature extraction. The architecture consists of convolutional, pooling, and fully connected layers, culminating in a softmax layer for classification. The most notable methodological innovation involves using a combination of the EEG signal from preceding, current, and succeeding epochs as input, aligning with the sequential nature of sleep stages. The authors address class imbalance—a common issue in sleep stage datasets—by implementing class-balanced random sampling during stochastic gradient descent (SGD) optimization.

Performance outcomes indicate high mean $F_1$-score (81\%), with mean accuracy across individual sleep stages at 82%. The CNN model demonstrates commendable balance across different classes, reducing typical bias towards more frequently occurring stages such as N2. Misclassifications were notably prevalent between stages N1-W and N1-R, attributed to EEG-based scoring lacking comprehensive multimodal input present in full polysomnography analyses.

Comparison with Other Methods

The proposed CNN approach is weighed against previous methods involving hand-engineered features and stacked sparse autoencoders, alongside a hybrid system combining hand-engineered wavelets with CNNs. Results suggest that while hand-engineered features marginally outperform the CNN in certain metrics, differences do not achieve statistical significance. This finding underscores the CNN's potential utility, particularly given its elimination of extensive prior knowledge and manual feature design requirements. The end-to-end learning capacity of the CNN model allows it to automatically derive deep feature representations, which are normally unattainable or labor-intensive through manual engineering.

Implications and Future Directions

This paper contributes to the burgeoning field of automated sleep scoring by validating CNNs as a viable tool for biomedical signal processing, particularly when using single-channel EEG. The implications extend to developing more accessible sleep monitoring technologies, potentially enhancing early-stage detection of sleep-related disorders linked to neurodegenerative diseases. Future investigations might focus on expanding dataset sizes and diversity to elucidate CNN capabilities further, especially under varying conditions of sleep quality and transition dynamics.

A broader implication lies in the CNN model's interpretability. The filter analysis and visualization show that learned filters are consistent with established sleep scoring guidelines and exhibit robust cross-validation performance. This presents an opportunity to utilize CNN architectures across other domains within biosignal processing where domain-specific knowledge is sparse, allowing researchers to develop performance-driven models with minimal manual intervention.

In conclusion, the paper exemplifies the application of CNNs in a traditionally manual and heuristic-driven domain, showcasing the strength of deep learning methodologies to autonomously learn and apply complex classification tasks. As the field continues to evolve, further integration of CNN approaches in biomedical engineering is anticipated, offering a confluence of practical utility and theoretical advancement.