AI Model Uses Sleep Data to Predict Risk of Over 130 Diseases
Researchers have developed a powerful artificial intelligence model that can assess an individual’s risk of developing more than 130 health conditions by analysing sleep data alone. The system, named SleepFM, marks a significant advance in predictive medicine and early disease detection.
The study, published in the medical journal Nature Medicine, involved scientists from multiple institutions, including Stanford University.
Trained on Massive Sleep Dataset
SleepFM was trained using nearly 600,000 hours of sleep data collected from around 65,000 participants. The data was gathered through clinical sleep studies, offering detailed insights into physiological activity during sleep.
Initially, the AI model was evaluated on conventional sleep-related tasks such as identifying sleep stages and assessing the severity of sleep apnoea.
After demonstrating strong performance in these areas, researchers expanded its use to predict long-term health outcomes.
Predicting Disease Before Symptoms Appear
The research team paired sleep recordings with anonymised medical records from a sleep clinic to examine whether patterns during sleep could signal future illness.
More than 1,000 disease categories were analysed, and the model successfully predicted the onset of 130 conditions with notable accuracy.
According to the researchers, sleep provides a uniquely rich window into overall health.
“Sleep captures an extraordinary range of physiological signals over several uninterrupted hours, making it ideal for predictive analysis,” said senior author Emmanuel Mignot, a professor of sleep medicine at Stanford.
How the AI Reads the Body During Sleep
Sleep data was collected using polysomnography, considered the gold standard in sleep research.
This technique records brain activity, heart rhythms, muscle movement, breathing patterns, blood oxygen levels, and eye movements.
ALSO READ: IIT Bombay Launches 5-Month Certificate Course in Generative AI with Great Learning
SleepFM was designed to integrate multiple streams of biological data, including electroencephalography (EEG), electrocardiography (ECG), electromyography (EMG), pulse readings, and airflow measurements.
The model learns how these signals interact rather than analysing them in isolation.
To train the AI more effectively, the researchers introduced a novel method known as “leave-one-out” contrastive learning.
In this approach, one data stream is hidden, and the model is challenged to reconstruct it using the remaining signals, strengthening its ability to understand complex physiological relationships.
Strong Results Across Critical Health Conditions
The AI demonstrated particularly strong predictive performance for serious illnesses. Its accuracy was measured using the concordance index, or C-index, which evaluates how well a model can predict which individual will experience a health event sooner.
SleepFM achieved C-index scores above 0.8 for several major conditions, including cancer, cardiovascular disease, pregnancy-related complications, and mental health disorders.
It also showed strong predictive ability for dementia, heart attacks, strokes, chronic kidney disease, atrial fibrillation, and overall mortality risk.
ALSO READ: New Monoclonal Antibody Shows Promise for Rare Liver Disease
Researchers also noted promising results in identifying the risk of Parkinson’s disease, where sleep disturbances often appear years before clinical diagnosis, as well as developmental delays and neuropsychiatric disorders.
A Step Toward Preventive Healthcare
The findings suggest that sleep-based AI tools like SleepFM could one day play a key role in preventive medicine, enabling earlier intervention long before symptoms emerge.
While further validation is required, researchers believe the model highlights the untapped potential of sleep data in understanding long-term health trajectories.
ALSO READ: Elon Musk Reacts as Google AI Overview Gets Basic Calendar Question Wrong
