Abstract
Correctly identifying sleep stages is essential for assessing sleep quality and treating sleep disorders. However, current sleep staging methods face several challenges: manual feature extraction requires professional knowledge and is time-consuming; similarity between stage features complicates classification; and multi-channel data acquisition requires complex equipment. This paper proposes a novel feature relearning method for automatic sleep staging based on single-channel electroencephalography (EEG). We design a bottom–up and top–down network using an attention mechanism to fully learn EEG information. A cascading step with an imbalanced strategy further improves classification performance. Experimental results on the Sleep-EDF public dataset demonstrate that our method outperforms state-of-the-art approaches.
Introduction
Sleep staging—the categorization of sleep into phases such as Wake, REM, and various NREM stages—is the gold standard for diagnosing sleep disorders. Traditionally, this process relies on Polysomnography (PSG), which requires a complex setup of EEG, EOG, and EMG sensors. While single-channel EEG is more accessible for wearable applications, automated classification remains difficult due to three primary hurdles: the expertise gap in manual feature extraction, high feature similarity between stages like N1 and REM, and significant class imbalance in typical sleep data.
To address these issues, we developed a two-stage "feature relearning" neural network. By first grouping similar sleep stages and subsequently employing a specialized cascaded network to differentiate them, the model overcomes the pitfalls of signal similarity that often plague traditional automated systems.
Methodology
The proposed architecture utilizes a hierarchical "divide and conquer" strategy, divided into two distinct parts.
Part 1: The Generalist (4-Class Classification)
This initial network combines Convolutional Neural Networks (CNNs) for local feature extraction with Long Short-Term Memory (LSTM) units to capture the temporal flow of sleep stages. It incorporates a Convolutional Block Attention Module (CBAM) to focus on critical signal components in both time and frequency domains. Crucially, this stage merges N1 and REM into a single category, as they are often too similar to distinguish reliably at a general level.
Part 2: The Specialist (Feature Relearning)
When Part 1 identifies a segment as "N1-REM," the data is passed to a second, specialized CNN. This network is trained on a balanced dataset of N1 and REM samples to "relearn" the subtle nuances required to distinguish the two. This bottom-up and top-down design mimics a feature pyramid, ensuring high-level semantic features are fused with low-level structural features.
Dataset and Pre-processing
The study utilized the Sleep-EDF public dataset, using raw 30-second single-channel EEG segments (Fpz-Cz and Pz-Oz channels) sampled at 100Hz. Unlike traditional methods, this model requires minimal pre-processing, working largely on raw data without complex filtering.
Results and Discussion
The "Merge then Split" strategy proved highly effective. By merging N1 and REM initially, the model achieves high accuracy on "easier" stages (Wake, N2, N3) before focusing computational resources on the difficult N1 vs. REM distinction. The inclusion of CBAM allows the model to adaptively weight frequency bands, similar to how a human expert identifies sleep spindles or K-complexes.
The proposed method achieved state-of-the-art performance on the Sleep-EDF dataset:
- Accuracy: 82.7% on the Fpz-Cz channel.
- Macro F1-Score (MF1): 76.8%, indicating robust performance across all stages, including minority classes.
- Ablation Studies: Removing the attention mechanism or top-down connections led to measurable performance drops, validating the architectural design.
Conclusion
This research presents a significant advancement in single-channel EEG sleep staging. While the hierarchical structure introduces some computational overhead, it successfully addresses the long-standing challenge of N1 and REM confusion. Future work will focus on optimizing model complexity for real-time applications and further refining the morphological analysis of N1 stage similarities with other phases to enhance generalization across diverse, noisy data from consumer wearables.
Summary updated