Using data collected during just one night in a sleep laboratory, artificial intelligence (AI) can assess the risk of developing more than 130 diseases — ranging from Parkinson’s disease to heart attacks and breast cancer. However, the program identifies correlations rather than causes. The new AI model analyzes brain waves, heart rate, breathing, muscle activity, and other bodily signals during nighttime sleep.

Moreover, the AI can make such predictions years before the first symptoms appear, emphasizes James Zou, associate professor at Stanford University and one of the authors of the study published in Nature Medicine. The new AI model, called SleepFM, was developed by a team led by Rahul Tapa, a biomedical data scientist at Stanford, and trained on hundreds of thousands of hours of sleep lab data.

The sleep study conducted in a specialized laboratory is known as polysomnography. During this diagnostic procedure — usually lasting just one night — various bodily systems are monitored during sleep, including the brain, heart, breathing, muscle tension, as well as eye and leg movements, among others. To “train” SleepFM, the team used about 585,000 hours of such recordings from roughly 65,000 people across different groups, mostly examined at Stanford’s Sleep Medicine Center.

During pre-training, the AI “learned” to coordinate and statistically register signals from the brain, heart, and breathing during sleep. After this base training, the SleepFM model was fine-tuned for tasks such as identifying sleep stages and diagnosing sleep apnea. Its performance reached levels comparable to established universal models such as U-Sleep and YASA. These programs are among the most well-known sleep-stage classifiers, analyzing electroencephalography (EEG) data, which provide detailed insight into brain activity and help diagnose various neurological conditions and mental health disorders.

Next, the researchers linked sleep data with electronic medical records from the past 25 years and examined which diagnoses could be predicted based on information from just one night. Out of more than a thousand disease categories, the model identified 130 conditions whose risk could be predicted with moderate to high accuracy. As Rahul Tapa notes, this approach shows that “routine sleep measurements open a previously underestimated ‘window’ into monitoring a person’s long-term health.”

Predictions were particularly accurate for dementia, Parkinson’s disease, myocardial infarction, heart failure, certain types of cancer, and overall mortality.

“In principle, the AI model can be trained for a very wide range of possible predictions, provided that appropriate datasets are available,” says Sebastian Buschjäger, a sleep expert at the Lamarr Institute — one of the key research centers at Dortmund University of Technology — who was not involved in the study.

The analysis shows that heart signals are especially important for predicting cardiovascular diseases, while brain signals are more relevant for neurological and psychiatric disorders. However, the most informative results come from combining different signals — for example, when EEG indicates stable sleep while the heart appears to be “awake.”

Such mismatches between brain and heart activity may point to hidden stress or early stages of disease long before symptoms appear.

“If our colleagues in sleep medicine suspect a connection, we AI specialists can incorporate it into prediction systems or indicate where relationships may emerge,” explained the Dortmund sleep expert to DW. “However, the correlations we provide are primarily statistical. Causal relationships must be confirmed by experts.”

The model is based mainly on data from sleep laboratories — meaning from people who were typically referred due to sleep problems and who lived in relatively affluent regions with access to high-tech healthcare. The researchers integrated several American and European groups, known as cohorts.

They are additionally testing the model in independent studies, but people without sleep problems and those from regions with less developed healthcare systems remain underrepresented.

The researchers stress that SleepFM does not identify causes of disease, only correlations: the AI recognizes statistical patterns in sleep that may be associated with future diagnoses.

“Most artificial intelligence methods cannot identify cause-and-effect relationships,” explains computer scientist Matthias Jacobs from Dortmund University of Technology, who studies AI and machine learning methods for analyzing sleep data and was not involved in the research.

Machine learning methods are computational techniques that allow computers to learn patterns from example data and make predictions without being explicitly programmed for every rule. At the same time, Jacobs sees “potential for diagnosis and therapy even if only statistical correlations are used,” he told DW.

Follow NEWS.am Medicine on Facebook and Twitter