A Foundation Model for Sleep-Based Risk Stratification and Clinical Outcomes

Area of Science:

• Biomedical Informatics
• Sleep Medicine
• Artificial Intelligence

Background:

• Sleep disorders contribute significantly to morbidity and mortality.
• Polysomnography (PSG) data is underutilized in clinical practice.
• Enhanced characterization of sleep dysfunction can improve patient outcomes.

Purpose of the Study:

• To develop a novel sleep foundation model integrating PSG time-series signals and electronic medical record data.
• To leverage data-driven representations for identifying patient subpopulations with differential health trajectories.
• To create a clinically applicable framework for risk stratification and health outcome prediction.

Main Methods:

• Utilized a transformer-based foundation model.
• Integrated PSG time-series signals with electronic medical record data.
• Analyzed a diverse dataset of 10,000 patients with a mean observation period of 14.5 years.
• Clustered model-generated representations to identify subpopulations.
• Externally validated findings in a National Sleep Research Resource cohort.

Main Results:

• Identified distinct subpopulations with differential health trajectories.
• The highest-risk group showed strong correlations with all-cause mortality (HR 4.83), cardiovascular, and neurological outcomes.
• These predictions remained significant after accounting for traditional measures.
• External validation confirmed the model's predictive capabilities.

Conclusions:

• A novel framework effectively leverages information-dense PSG data for enhanced risk stratification.
• The foundation model predicts health outcomes beyond traditional methods.
• This approach offers a clinically applicable tool for improving patient care in sleep medicine.