SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures

Researchers introduce SleepMaMi, a foundation model designed to analyze sleep patterns by capturing both hour-long sleep architecture and fine-grained biosignal features. Trained on over 20,000 polysomnography recordings, the model outperforms existing approaches and demonstrates superior generalizability for clinical sleep analysis applications.

SleepMaMi represents a significant shift in sleep medicine research from fragmented, task-specific models to a unified foundation model approach. The innovation addresses a critical gap in medical AI: while foundation models have transformed computer vision and natural language processing, sleep medicine has remained segmented, with models typically focusing either on macro-level sleep cycles or micro-level signal morphologies—rarely both simultaneously. The dual-encoder architecture elegantly solves this by separating concerns: the Macro-Encoder captures full-night temporal dependencies aligned with demographic factors like age and BMI, while the Micro-Encoder extracts fine-grained biosignal characteristics through masked autoencoders and contrastive learning.

The scale of training data—158,000 hours of polysomnography recordings—is substantial for the medical AI domain, enabling the model to learn robust, generalizable representations. Demographic-guided contrastive learning is particularly noteworthy; anchoring overnight patterns to objective subject metadata creates clinically meaningful embeddings that transcend individual recording variations.

For the healthcare and medical AI sectors, this development signals accelerating adoption of foundation models in specialized clinical domains. The demonstrated label-efficiency and generalizability suggest potential cost reductions in sleep disorder diagnosis and screening, while improving diagnostic accuracy. Medical device manufacturers and clinical decision support vendors may integrate similar architectures. The research validates that foundation models, when thoughtfully designed for domain-specific requirements, outperform siloed approaches even in narrow medical specialties. Future work likely involves multi-center validation, regulatory pathway exploration, and real-world clinical deployment studies.

• →SleepMaMi combines macro-level sleep architecture and micro-level biosignal analysis through a hierarchical dual-encoder design, unifying previously fragmented approaches.
• →Training on 158,000 hours of polysomnography data enables superior generalizability and label-efficient adaptation compared to existing foundation models.
• →Demographic-guided contrastive learning grounds the model in clinically relevant subject metadata, improving representation quality for medical applications.
• →The foundation model approach demonstrates significant potential for reducing diagnostic costs and improving accuracy in sleep medicine.
• →This work exemplifies how domain-specific foundation models can advance specialized medical fields beyond generic deep learning approaches.