Semantic Segmentation of Sleep Events for High-Resolution Sleep Decoding

Sleep is structured by brief, recurring EEG waveforms-such as slow waves, K-complexes, and spindles-that underpin sleep architecture and link to cognition, aging, and disease. Yet event-level analysis in sleep science remains constrained by reliance on labor-intensive manual annotation and the absence of automated, multi-event detection methods. Here, we present a unified, high-resolution frame-work for detecting multiple EEG-based sleep events continuously across the night. Integrating self-supervised and active learning to guide expert annotation, we constructed a cross-dataset, large-scale resource comprising 276,404 sleep events spanning seven physiologically and clinically relevant types. Leveraging this resource, we developed a sleep semantic segmentation model that decodes single-channel EEG into millisecond-level probability distributions for each event type. We demonstrated the versatility of the model across diverse applications in sleep science: real-time forecasting of imminent events to enable sleep interventions, automated sleep staging with state-of-the-art performance, and interpretable disease classification from whole-night EEG. By shifting sleep analysis from coarse staging to continuous, event-centric decoding, this study establishes a foundation for scalable, mechanistic, and clinically translatable sleep research.