Photo: Thomas Koehler/photothek/picture alliance
Based on data collected overnight in a sleep lab, artificial intelligence (AI) is able to estimate the risk of developing more than 130 diseases, from Parkinson’s disease to heart attack to breast cancer. But the program doesn’t identify causes, only correlations. The new AI model analyzes brain waves, heart rate, breathing, muscle and other activity of the human body during a night’s sleep.
And AI can do this years before the first symptoms of the disease appear, emphasizes James Zou, associate professor at Stanford University and one of the authors of the study published in the journal Nature Medicine. The new AI model, called SleepFM, was developed by a team led by Rahul Thapa, a biomedical data scientist at Stanford University, and trained on hundreds of thousands of hours of data from sleep labs.
From sleep signaling to disease prediction
The study of patients’ sleep in a special laboratory is called polysomnography. During this diagnostic study, which, as a rule, is carried out only one night, the features of the work of various body systems during sleep are monitored: the brain, heart, breathing, muscle tension, as well as eye and leg movements, and others. To “train” SleepFM, the team used about 585,000 hours of such recordings, obtained from about 65,000 people from different groups who were examined, mainly at the Center for Sleep Medicine at Stanford University.
During pre-training, the AI “learned” how to coordinate and statistically record data about a person’s brain, heart and breathing signals during sleep. After basic training, the SleepFM model was further refined for tasks such as determining sleep stages and diagnosing sleep apnea – as a result, its scores reached a level comparable to the results of universal models such as U-Sleep and YASA. These programs are among the best known so-called sleep stage classifiers that analyze electroencephalography (EEG) data, which provides a detailed picture of brain activity and helps diagnose various neurological and mental health conditions.
Next, the researchers matched the sleep data with electronic medical records from the past 25 years and examined what diagnoses could be inferred based on information from a single night. From more than a thousand categories, the model identified 130 diseases whose risk of developing could be predicted with moderate to high accuracy. As researcher Rahul Thapa notes, this approach shows “that routine sleep measurements open a hitherto underappreciated window for monitoring a person’s long-term health.”
It has proven particularly accurate in predicting dementia, Parkinson’s disease, myocardial infarction, heart failure, certain cancers, and overall mortality. “In principle, an artificial intelligence model can be trained for a very large range of possible predictions if there is an appropriate database for this,” says Sebastian Buschjäger, a sleep expert at the Lamarre Institute, which is one of the key research centers at the Technical University of Dortmund, who was not involved in the study.
What the AI is looking for in the body of a sleeping person
As the analysis shows, heart signals are particularly important for predicting cardiovascular disease, while brain signals are important for neurological and mental disorders. However, a combination of different signals is most informative, for example, when the EEG shows a stable sleep state but the heart appears “awake”.
Such discrepancies between the brain and heart may indicate hidden burdens or early stages of disease long before symptoms appear. “If our colleagues in the field of sleep medicine suspect a connection, we AI experts can incorporate this into a prediction system, or give indications of where connections might be found,” the Dortmund-based sleep specialist explained to DW. – However, the correlations we provide are mostly statistical. A cause-and-effect relationship has to be confirmed by experts.”
AI helps but does not replace humans
Models like SleepFM compress huge amounts of polysomnography data into compact numerical matrices that allow for faster and often more accurate analysis. “They can be used to efficiently describe stages of sleep and apnea – a very time-consuming task that is fraught with errors when performed manually. This will leave doctors with more time for patients,” Matthias Jacobs is convinced.
Dortmund-based sleep expert Sebastian Buschieger emphasizes that interdisciplinary cooperation is crucial: “Artificial intelligence can be well-trained in therapy planning, but a human – i.e. a doctor – interprets the results and chooses the therapy, often without knowing all the reasons.” Thus, AI remains a tool and an early warning system, while the responsibility for diagnosis and treatment remains with healthcare professionals.
Whether and to what extent the patterns discovered may point to underlying biological mechanisms is still an open question. But this is where researchers see great potential.
If certain signals received during sleep are consistently correlated with specific diseases, this could suggest which processes in the nervous, cardiovascular or immune system are impaired in the early stages of the disease. It will also help to draw conclusions about the health status of people outside the groups studied in sleep laboratories.
