AI Model Predicts Long-Term Health Risks from Routine Sleep Study Data

A new AI foundation model developed by Cleveland Clinic and IBM analyzes routine sleep study data to predict long-term health risks. Trained on over 10,000 polysomnograms from the STARLIT registry, the model identifies five distinct patient risk groups. Patients in the highest-risk group face twice the five-year mortality risk of those in the lowest-risk group. This approach successfully predicts cardiovascular and cognitive decline risks across both men and women, outperforming the traditional apnea-hypopnea index.

Sleep health foundation model

• ▪Researchers developed a transformer-based artificial intelligence foundation model designed to learn general representations of sleep physiology from routine polysomnography data
• ▪The artificial intelligence model was trained using data from the Cleveland Clinic Sleep Signals, Testing, and Reports Linked to Patient Traits registry, which contains over 10,000 clinical sleep recordings
• ▪The artificial intelligence model was developed by a collaborative team of researchers through the Discovery Accelerator, a 10-year research partnership between Cleveland Clinic and IBM

PSG embedding risk stratification

• ▪The artificial intelligence model grouped patients into five distinct risk categories based on latent physiological features extracted from overnight sleep studies
• ▪Patients classified in the highest-risk group faced twice the risk of dying within five years compared to those in the lowest-risk category
• ▪The model's risk stratification performance was independently confirmed in a nationwide patient cohort, the Sleep Heart Health Study

Cardiovascular outcome prediction

• ▪The model's high-dimensional embeddings of sleep physiology predicted cardiovascular events with equal accuracy across both male and female patients
• ▪The artificial intelligence model's sleep embeddings yielded risk groups that showed strong, monotonic associations with incident cardiovascular outcomes

Neurologic outcome prediction

• ▪The model's risk stratification for neurological outcomes outperformed conventional diagnostic measures by identifying high-risk patients across different sleep apnea severity levels
• ▪The artificial intelligence model identified hidden sleep patterns and latent physiological features linked to cognitive decline and neurological disease

AHI diagnostic limitations

• ▪The apnea-hypopnea index historically performs better in assessing risk for men than for women, whereas the new artificial intelligence model achieved balanced predictive accuracy across both sexes
• ▪The standard clinical measure for sleep apnea, the apnea-hypopnea index, failed to capture the long-term mortality and disease risks identified by the artificial intelligence model

Precision sleep medicine implementation

• ▪Researchers emphasized that further validation in more diverse patient populations is required before the artificial intelligence model can be routinely deployed in clinical care
• ▪An estimated 1 million to 4 million polysomnograms are conducted annually in the United States, representing a vast source of underutilized clinical data