The field of sleep medicine has long been constrained by task-specific models that focus narrowly on localized micro-structure features, often neglecting the rich, multi-modal context of Polysomnography (PSG) and failing to capture the global macro-structure of a full night's sleep. According to the paper "SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures" published on arXiv, researchers have developed a new foundation model called SleepMaMi to bridge this gap. SleepMaMi is engineered to master both hour-long sleep architectures and fine-grained signal morphologies from PSG data.
The Problem with Current Sleep Models
Traditional approaches in sleep analysis are typically optimized for specific tasks — such as sleep stage classification or event detection — and they rarely leverage the full spectrum of information available in overnight PSG recordings. The paper notes that while foundation models have revolutionized other deep learning domains, sleep medicine has remained largely restricted to these task-specific models. This results in poor generalizability and inefficient use of data, particularly when labeled data is scarce.
How SleepMaMi Works
SleepMaMi introduces a hierarchical dual-encoder design to address these limitations:
- Macro-Encoder: Models full-night temporal dependencies. It is trained via Demographic-Guided Contrastive Learning, which aligns overnight sleep patterns with objective subject metadata such as age, sex, and BMI to refine global representations.
- Micro-Encoder: Captures short-term characteristics from biosignals. It is optimized using a hybrid Masked Autoencoder (MAE) and multi-modal contrastive objective.
The dual-encoder architecture allows SleepMaMi to simultaneously understand the big picture of a night's sleep and the subtle details within individual signal segments.
Training and Performance
SleepMaMi was pre-trained on a massive corpus of >20,000 PSG recordings, totaling 158,000 hours of sleep data. According to the paper, this extensive pre-training enables the model to outperform or match state-of-the-art existing foundation models across a diverse suite of downstream tasks, demonstrating superior generalizability and label-efficient adaptation for clinical sleep analysis.
| Component | Encoder Type | Training Objective | Input/Output |
|---|---|---|---|
| Macro-Encoder | Full-night temporal | Demographic-Guided Contrastive Learning | Subject metadata (age, sex, BMI) |
| Micro-Encoder | Short-term biosignal | Hybrid MAE + multi-modal contrastive | PSG signal segments |
Implications for Healthcare AI
SleepMaMi represents a shift toward unified foundation models in sleep medicine. By integrating macro- and micro-structures, it reduces the need for large labeled datasets — a critical advantage in clinical settings where annotated sleep data is expensive and time-consuming to produce. The model's ability to leverage demographic information further personalizes sleep analysis, potentially improving diagnostics for conditions such as sleep apnea, insomnia, and circadian rhythm disorders.
The paper concludes that SleepMaMi sets a new benchmark for label-efficient, generalizable sleep analysis. For enterprise technology leaders in healthcare, this foundation model approach could accelerate the development of AI-powered diagnostic tools, reduce costs, and improve patient outcomes through more accurate sleep assessments.