A foundation model for sleep-based risk stratification and clinical outcomes

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Cardiovascular and Neurological Disease Research

Background:

• Clinical sleep studies generate extensive physiologic data, but interpretation often relies on limited metrics like the apnea-hypopnea index.
• Current metrics possess restricted prognostic value for predicting long-term health outcomes.
• There is a need for advanced analytical methods to extract comprehensive risk information from sleep recordings.

Purpose of the Study:

• To develop and validate a foundation model for analyzing complex sleep physiology data.
• To identify latent risk structures within sleep recordings that are not captured by conventional metrics.
• To stratify patients into distinct risk groups based on sleep physiology for improved health outcome prediction.

Main Methods:

• Utilized a foundation model trained on over 10,000 clinical sleep recordings linked to electronic medical records.
• The model learned rich representations of sleep physiology to uncover hidden risk patterns.
• Validated the model's generalizability on an independent dataset (Sleep Heart Health Study) with lower-resolution data.

Main Results:

• Identified five distinct patient risk groups with significantly different trajectories for mortality, cardiovascular, and neurological diseases.
• The highest-risk group exhibited more than double the mortality risk compared to the lowest-risk group.
• Apnea-hypopnea index severity categories demonstrated limited predictive value compared to the foundation model's risk stratification.

Conclusions:

• Foundation models can extract clinically meaningful risk information from routine sleep recordings that conventional metrics miss.
• This approach offers a scalable pathway toward precision sleep medicine.
• Sleep physiology contains complex risk structures that can be leveraged for improved patient prognostication and tailored interventions.