Stage Prototypes for Transition-Aware Temporal Modeling in Single-Channel Sleep Staging

Automatic sleep staging from polysomnography (PSG) remains challenging under subject-independent evaluation, particularly around stage transitions where EEG patterns are inherently ambiguous. We present \textbf{SleepTSP}, a temporal model that incorporates boundary-aware attention within a sliding-window framework for single-channel EEG. A learned boundary signal modulates temporal interactions, while stage-level representations provide global contextual anchoring. On the EDF-20 benchmark, we observe that errors are concentrated near stage transitions, and show that explicitly modeling transition structure improves robustness in these regions without increasing inference complexity. These findings highlight the importance of transition-aware temporal modeling for stable sleep staging.