SSSM - Sleep Semantic Segmentation Model

Sleep is organized by transient, heterogeneous events (e.g., spindles, slow waves, K-complexes) that shape cognition and disease. Yet most automated analyses are constrained by narrow event definitions, coarse temporal resolution, and the lack of large-scale, multi-event datasets. This project addresses these gaps by reframing sleep analysis as semantic segmentation of EEG, delivering high-resolution event probabilities and actionable downstream outputs.

What SSSM provides

• Large-scale multi-event dataset: 276,404 events across 7 types (K-complex, slow wave, sleep spindle, micro-arousal, sawtooth wave, vertex sharp wave, background), built efficiently via self-supervised and active learning and validated with independent datasets/algorithms.
• Sleep Semantic Segmentation Model (SSSM): The first model to decode sleep EEG into fine-grained events at the sampling-point level (100 Hz), showing efficiency and strong generalizability across datasets, sampling rates, and electrode montages.
• Versatile applications: (a) Sleep event forecasting with an iTransformer, enabling up to 2 s look-ahead for real-time interventions; (b) Scalable automatic sleep staging by integrating a Transformer encoder into SSSM, achieving state-of-the-art performance across 14 clinical datasets; (c) Disease diagnosis by embedding a Mamba model into SSSM, detecting disorder-related patterns (e.g., sleep apnea, depression, REM behavior disorder).

Citation

If you use this tool or its results in your research, please cite:

@article {
           Bao2025.09.25.25336636,
           author = {Xiaoyu, Bao and Li, Man and Chen, Di and Wang, Zijian and Huang, Qiyun and Wen, Zhenfu and Li, Yuanqing},
           title = {Semantic Segmentation of Sleep Events for High-Resolution Sleep Decoding},
           elocation-id = {2025.09.25.25336636},
           year = {2025},
           doi = {10.1101/2025.09.25.25336636},
           publisher = {Cold Spring Harbor Laboratory Press},
           URL = {https://www.medrxiv.org/content/early/2025/09/29/2025.09.25.25336636},
           eprint = {https://www.medrxiv.org/content/early/2025/09/29/2025.09.25.25336636.full.pdf},
           journal = {medRxiv}
       }