Skip to content
logo The magazine for fitness, health and nutrition
Diseases Sleep All topics
Revolutionary AI Model

Researchers Can Predict Risk for 1,000 Diseases Based on Sleep

Sleep is linked to health and, conversely, to diseases.
Sleep is more than just recovery for the next day—it's an early warning system for our health. Photo: Getty Images
Share article

January 9, 2026, 4:52 pm | Read time: 7 minutes

Could our nightly sleep provide hidden clues about serious illnesses years before symptoms appear? And could artificial intelligence help decipher these for affected individuals? A recent study shows: Yes. Researchers have developed an AI model that remarkably accurately detects the risk of more than 1,000 diseases from the diverse signals of a sleep lab. This is made possible by a massive database with 585,000 hours of sleep recordings. Read on FITBOOK to find out exactly what the Stanford University model can do and what this could mean for disease prevention.

Sleep is not just rest—it also reflects our health status. Sleep disorders are increasingly considered early warning signs for diseases such as dementia, Parkinson’s, or heart problems.1,2,3 Studies show this, although they mostly focus on individual diseases—with limited datasets and a lot of manual work.

The study conducted at Stanford University takes a more comprehensive approach. The research team developed “SleepFM,” a so-called foundation model—a very large AI system—that analyzes multimodal data from professional sleep studies (polysomnographies, or PSG for short). These include brain waves (EEG), eye movements (EOG), heart activity (ECG), muscle tension (EMG), and breathing.4

Can Sleep Lab Data Be Used for Disease Prevention?

Sleep lab studies are mostly used for patients seeking help due to existing sleep disorders. Experts in sleep medicine use the data obtained during sleep recordings to track both the type of sleep disorder and the possible physical causes. Just as sleep disorders can be early warning signs of future diseases, pre-existing conditions can also cause disturbed sleep.

But can the recordings of nightly body functions reveal even more about the sleeping person? And be used to detect the risk of a variety of diseases early—ideally before the first symptoms appear? The search for answers to these questions prompted Stanford researchers to develop their AI model. The researchers did not focus on a single disease but wanted to systematically test how strongly sleep data is associated with over 1,000 different diagnoses.

Also interesting: Does the risk of death increase with too little REM sleep?

AI Analyzed Brain Waves, Eye Movements, Heart Activity, Muscle Tension, and Breathing

The study and the analysis of the AI model are based on one of the largest sleep databases worldwide. A total of 35,052 datasets, each with about eight hours of material, were available. This amounted to more than 585,000 hours of PSG data from over 65,000 individuals from five independent cohorts—including the Stanford Sleep Clinic, BioSerenity, MESA, MrOS, and the SHHS study. The data covers a wide age range from children to the very elderly. The SleepFM model was trained using a self-developed algorithm called LOO-CL (Leave-One-Out Contrastive Learning). This method allows different biosignals like EEG, ECG, or breathing data to be linked.

Training the AI Model

Initially, the SleepFM model underwent pre-training with practice data, followed by fine-tuning for specific tasks such as age estimation, gender inclusion, sleep stage recognition, and especially disease prediction.

In the next step, the scientists fed the AI model with electronic health records from the aforementioned cohorts, where diagnoses with timestamps were available. They considered disease indications that appeared at least seven days and at most six years after a person’s sleep lab examination. This allowed testing whether sleep data could predict diseases that did not exist before the sleep lab visit and might only be diagnosed months or years later. The AI model could not access classic lab values or imaging diagnostics. This ensured that any predictions would be based solely on sleep signals. Established metrics such as AUROC (Area Under Receiver Operating Characteristic Curve) and Harrell’s C-Index were used for model evaluation.

Association Between Sleep Data and More Than 1,000 Diseases

The capabilities of SleepFM captured in the study are remarkable. For 130 diseases, the model achieved a prediction accuracy (AUROC and C-Index) of at least 0.75. This is considered a high value in medical diagnostics. The model was particularly precise for neurodegenerative diseases such as Parkinson’s (AUROC 0.93), dementia (0.87), and cognitive disorders (0.84). It also accurately identified cardiovascular diseases like hypertensive heart disease (0.88), heart failure (0.83), and stroke (0.81).

Surprisingly, there was also a high predictive power for certain types of cancer: breast and prostate cancer, as well as skin melanomas, achieved AUROC values of 0.90 and 0.83, respectively. Even the risk of death could be reliably estimated from a single night’s sleep data (AUROC 0.84).

Also interesting: Is the clock the enemy of our sleep?

Sleep as an Early Warning System

Just one night of professional sleep measurement was enough for SleepFM to extract extensive clues about future diseases. The fact that it recognized the risk for slowly developing and often late-diagnosed diseases like dementia, Parkinson’s, or heart failure offers hope for progress in their prevention. Instead of waiting for symptoms, at-risk patients could be identified early and treated preventively.

A sleep study could become not only a diagnostic aid for snoring or sleep apnea but also a screening for many other diseases. For research, SleepFM is a milestone, as it opens new ways to use large, complex biosignals with modern AI.

More on the topic

Assessment of the Study and Possible Limitations

SleepFM is the most comprehensive application of a foundation model in sleep medicine to date. The combination of a massive database, multimodal signal processing, and AI learning sets it apart from previous studies.

However, despite the impressive results, there are also limitations. The datasets mostly come from individuals referred to sleep labs due to sleep problems or other illnesses. Healthy people without complaints are underrepresented—this limits the generalizability to the general population.

Another drawback is that the model’s predictions are difficult to interpret. The AI provides risk values but no easily understandable explanations for them. Causal relationships between specific sleep signals and specific diseases were not established. Although analyses were conducted by sleep phases and signal types (e.g., REM sleep particularly relevant for neurodegenerative diseases), no actionable recommendations for clinical practice can be derived from this.

Further studies are needed before SleepFM or similar models can be integrated into everyday medical care—such as additional diagnostics in sleep labs or potentially with wearable devices.

Step Toward Personalized, Predictive Medicine

Even though questions remain and the path to practical applicability in medicine is still long, the Stanford researchers present a revolutionary AI tool for health prognosis—based solely on sleep data.

From over 585,000 hours of PSG recordings, the model learned to predict more than 1,000 diseases with high accuracy—including those that are incurable, affect millions of people, and heavily burden health systems worldwide.

Despite open questions about everyday usability, SleepFM is a promising step toward personalized, predictive medicine.

This article is a machine translation of the original German version of FITBOOK and has been reviewed for accuracy and quality by a native speaker. For feedback, please contact us at info@fitbook.de.

Sources

You have successfully withdrawn your consent to the processing of personal data through tracking and advertising when using this website. You can now consent to data processing again or object to legitimate interests.