This is the authors' abstract. We don't add key points for this paper.
Automated sleep staging from polysomnography (PSG) is a core component of objective sleep analysis and plays a crucial role in enabling reliable clinical assessment and downstream physiological interpretation, yet the non-stationary nature of physiological signals poses significant challenges for compact models with fixed temporal receptive fields that cannot adapt to time-varying sleep rhythms. To address this challenge, we propose MG-STNet, a compact multi-granularity spatiotemporal network for adaptive sleep staging. The model first utilizes a Multi-Granularity Feature Learning Module (MFLM) to extract multi-resolution representations via parallel downsampling branches, thereby reducing sequence length and computational redundancy. Building on this compact representation, the Global–Local Feature Recalibration (GLFR) module recalibrates multi-channel PSG features by jointly leveraging global statistics and local temporal interactions, while the Adaptive Spatiotemporal Convolution (ASTC) module dynamically routes among complementary temporal branches to capture evolving sleep rhythms. Furthermore, a Temporal Attention Encoding Module (TAEM) is integrated to model long-range dependencies within each epoch. We evaluate MG-STNet on Sleep-EDF-20 and Sleep-EDF-78 using subject-independent k-fold cross-validation. The four-channel model achieves accuracies of 87.3% and 84.5% with 46.6K parameters and 17.2M FLOPs, while its single-channel counterpart achieves 86.6% and 83.2% with 48.6K parameters and 9.8M FLOPs. Experimental results indicate that MG-STNet achieves competitive accuracy with a substantially smaller parameter footprint than mainstream deep sleep-staging models, offering a favorable accuracy–efficiency trade-off for automatic sleep staging under computational constraints.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: