Answers are grounded in this paper and cite their sources.
Sleep stage classification is pivotal in evaluating sleep quality and diagnosing sleep-related disorders. Recent advancements in automated single-channel electroencephalogram (EEG)--based classification have gained traction due to their cost-effectiveness and portability. However, the inherent non-stationarity of EEG signals and inter-class imbalance pose significant challenges for model design. This paper proposes MultiScaleSleepNet, an enhanced deep learning architecture that addresses these limitations through dynamic multi-scale feature fusion and residual structural optimizations. Our contributions are threefold: (1) A selective kernel convolution module (SKConv) that dynamically integrates multi-branch convolutional features (kernel sizes: 3, 5, 7) via attention mechanisms to adaptively capture frequency-specific patterns in EEG signals; (2) A residual multi-branch downsampling module that mitigates information loss while preserving high-frequency details for minority-stage classification; (3) Comprehensive experiments on the Sleep-EDF-20 dataset demonstrate superior performance, achieving a macro F1-score (MF1) of 79.6%—a 1.5% improvement over baseline models—with notable gains in classifying the N1 stage (F1-score: 47.0%, +4.4% relative improvement). Quantitative ablation studies validate the efficacy of SKConv and residual connections in enhancing feature discriminability. This study delivers a robust single-channel EEG-based sleep analysis framework, demonstrating significant clinical applicability in resource-constrained settings.
No takes yet. Share an insight, caveat, or question.
Jiajun Zhong (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: