Abstract
While sleep is fundamental to human health, sleep disturbances reduce quality of life and constitute risk factors for neurodegenerative diseases including Parkinson's and Alzheimer's. Automated sleep staging networks achieve human-level performance on multimodal physiological signals, but they operate as black boxes, limiting clinical trust and preventing the discovery and validation of sleep biomarkers linked to human health status.We propose ProtoSleepNet (PSN), the fist prototype-based sequence-to-sequence sleep staging architecture that achieves human-level sleep staging accuracy while providing interpretability through an intrinsic codebook of learned prototypes. Each prototype captures distinctive sleep microstructure patterns, visualized as physiologically meaningful features across EEG, EOG, and EMG channels. We validate PSN against state-of-the-art approaches on over 10,000 subject recordings across 10 benchmark datasets, demonstrating in-line or superior sleep staging performance, robustness to channel occlusion attacks, and interpretability through a novel explainability framework that translates abstract prototypes into clinically aligned natural-language matching rules.Finally, we show that prototype sequences (prototype-grams) from individual patients encode clinically relevant information: without any disease-specific training, prototype-grams effectively discriminate Parkinson's and Alzheimer's disease patients from healthy controls, revealing disease-specific sleep microstructure alterations aligned with known pathophysiology.