Sleep decoding is central to understanding the neurophysiology of sleep and supporting clinical applications. Yet existing approaches remain dominated by coarse epoch-based staging that largely overlooks sleep microstructures. Here we introduce a microstructure-centric framework that represents sleep as a continuous composition of recurring electrophysiological microstructures. Using a hybrid self-supervised and active learning pipeline to guide expert annotation, we assembled a large-scale, cross-dataset repository of sleep microstructures, comprising 276,404 manually validated microstructures across seven classes from four publicly available electroencephalography (EEG) datasets. Leveraging this resource, we developed the Sleep Semantic Segmentation Model (SSSM), which converts single-channel EEG into millisecond-resolution probability trajectories over sleep microstructure classes, capturing the fine-grained temporal evolution of sleep microstructures. Across evaluations, the framework supports near–real-time forecasting of imminent microstructures, achieves state-of-the-art performance in cross-dataset automatic sleep staging, and enables interpretable classification of sleep-related disorders. Together, this work advances sleep decoding beyond coarse-grained staging toward an microstructure-centric representation, providing a scalable foundation for mechanistic sleep research and precision sleep medicine.