This is the authors' abstract. We don't add key points for this paper.
Automatic sleep staging is a key technology for the precise diagnosis and treatment of sleep disorders and for long-term home sleep monitoring. Due to its convenience of use, single-channel electroencephalogram (EEG) has become a research focus in home sleep monitoring. However, existing algorithms still face several practical bottlenecks. For example, models with a large number of parameters are prone to overfitting on small labeled datasets; the hardware heterogeneity of customized portable sensors makes datasets not directly shareable; large-scale labeled samples are hard to obtain in real scenarios; there is low recognition accuracy for difficult-to-distinguish sleep stages such as N1 and REM; and there is a lack of suitable research methods for evaluating the required scale of training data for lightweight deployment scenarios. In this paper we propose GamSleepNet, a low-parameter-count automatic sleep staging framework for single-channel EEG. The framework combines a feature extraction module (FEB), which integrates improved Gabor kernels with a learnable filter branch, with a standard Mamba selective state-space model (SSM) used for temporal modeling. A teacher–student feature alignment loss (FAL) and a two-stage training strategy are introduced to enhance the discriminative ability for difficult sleep stages. In addition, following the established paradigm of varying the training dataset’s size and observing the performance curve, we make four adaptations tailored to sleep staging, thereby determining the appropriate training data volume for GamSleepNet and providing a methodological reference for determining the training data scale required for deploying other models. Under the limited-labeled-subject setting of the Sleep-EDF dataset, the model has 30.86 k parameters and achieves an overall accuracy of 87.86% and a macro F1 score of 82.62%—the highest values among the compared models—and improves the recognition of difficult REM and N1 samples. The model still maintains stable staging performance under low signal-to-noise ratio and label-scarce scenarios.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Wei et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: