One of the biggest opportunities for artificial intelligence (AI) machine learning is in the field of health and disease diagnosis. A groundbreaking new study demonstrates how AI can predict a person’s risk of developing more than a hundred serious diseases from data collected non-invasively during a single night of sleep.
“This study underscores the potential of sleep-based baseline models for risk stratification and longitudinal health monitoring,” wrote Stanford University co-authors James Zou and Emmanuel Mignot, in collaboration with co-authors Rahul Thapa, Magnus Ruud Kjaer, Bryan He, Ian Covert, Hyatt Moore IV, Umaer Hanif, Gauri Ganjoo, M. Brandon Westover, Poul Jennum, and Andreas Brink-Kjaer.
Why sleep?
Sleep is essential for maintaining not only physical health, but also psychological well-being, as it affects emotional regulation, cognition, resilience, concentration and memory. An estimated 50% of insomnia cases are related to psychological stress, anxiety or depression, and obsessive-compulsive disorder (OCD) is often linked to lack of sleep, according to the National Alliance on Mental Illness (NAMI).
According to the Cleveland Clinic, there are more than 80 types of sleep disorders. The most common include chronic insomnia, obstructive sleep apnea, restless legs syndrome, REM sleep behavior disorder, narcolepsy, delayed sleep phase syndrome, and shift work sleep disorder.
Who doesn’t sleep well?
Many people do not sleep well around the world. By 2034, the sleep disorders market is expected to reach $72 billion, with a compound annual growth rate of 10% between 2025 and 2034, according to Global Market Insights. The American Brain Foundation estimates that between 50 and 70 million people in the U.S. alone suffer from sleep-wake disorders. About 1 in 3 American adults reported not getting enough rest or sleep on a daily basis, according to the U.S. Centers for Disease Control and Prevention. Globally, there are nearly 1 billion adults ages 30 to 69 with sleep apnea, according to a 2019 study by Benjafield et al. and published in The Lancet Respiratory Medicine. Sleep apnea is just one of the sleep disorders.
Nighttime sleep: A goldmine in AI data
The researchers created a multimodal AI model called SleepFM, trained on polysomnography (PSG) data, data captured non-invasively during an overnight sleep study. The Greek prefix “poly” means “many,” and polysomnography records numerous physiological signals.
During PSG, brain waves are recorded non-invasively and painlessly using an electroencephalogram (EEG), as well as blood oxygen levels using pulse oximetry, eye movements using an electrooculogram, heart rate using an electrocardiogram, and breathing and leg movements using an electromyogram. Polysomnography is the gold standard for diagnosing sleep behaviors such as sleepwalking, sleep apnea, other sleep-related breathing disorders, chronic insomnia, periodic limb movement disorder, narcolepsy, and REM sleep behavior disorder.
In terms of AI model performance, having massive datasets with high-quality training data can improve overall accuracy. In this study, the AI model was trained with polysomnography data from approximately 65,000 participants from multiple cohorts, with more than 585,000 hours of selected recordings. Cohorts include polysomnography data from the Stanford Sleep Clinic (SSC), Outcomes of Sleep Disorders in Older Men (MrOS), the Multi-Ethnic Study of Atherosclerosis (MESA), and BioSerenity. Data from the Sleep Heart Health Study (SHHS) were used to refine the algorithm.
“Our model uses 5 to 25 times more data than pretrained supervised sleep or biosignal models,” the researchers wrote.
To train SleepFM, the team used a self-supervised learning algorithm that did not require labeled data. The researchers then tested the AI with more than a thousand disease phenotypes. The AI model performed especially well in predicting Alzheimer’s disease and Parkinson’s disease, both neurodegenerative diseases. According to the scientists, their AI model accurately predicted 130 conditions, including dementia, stroke, heart failure, chronic kidney disease, myocardial infarction, atrial fibrillation and all-cause mortality.
Artificial Intelligence Essential Readings
“This work demonstrates that base models can learn sleep language from multimodal sleep recordings, enabling scalable and efficient analysis in labeling as well as disease prediction,” the researchers concluded.
Copyright © 2026 Cami Rosso. All rights reserved.
