A poor night's sleep often feels like a temporary inconvenience, but it can also be an early sign of health problems that only manifest themselves years later. Researchers at Stanford Medicine, together with international partners, have developed a new artificial intelligence model that can predict the risk of developing more than a hundred conditions based on physiological data from a single night's sleep. The model, SleepFM, shows how sleep data can become a powerful building block for predictive and personalised medicine.
SleepFM has been trained on nearly 600,000 hours of polysomnography data from approximately 65,000 participants. Polysomnography is the gold standard in sleep research and records a wide range of physiological signals throughout the night, including brain activity, heart activity, breathing patterns, muscle activity and eye movements.
‘We record an enormous amount of signals during sleep. It's eight hours of general physiology in someone who is being fully monitored. That makes it extremely data-intensive,’ said Emmanuel Mignot, professor of sleep medicine at Stanford and co-senior author of the study published in Nature Medicine.
Complex datasets
To date, only a limited portion of this information has been utilised in clinical practice and research. With the emergence of advanced AI techniques, it is now possible to extract much more coherence and meaning from these complex datasets. SleepFM is designed as a so-called foundation model, a type of AI that trains itself on large amounts of data and can then apply that knowledge to a variety of tasks.
Where large language models are trained on text, SleepFM is trained on sleep data. The night-time measurements are divided into five-second intervals, similar to words in a language model. ‘SleepFM is essentially learning the language of sleep,’ says co-senior author James Zou, professor of biomedical data science.
Multiple data streams
An important technical innovation within the project is that the model can combine and interpret multiple data streams simultaneously. To make this possible, the researchers developed a new training method, in which one type of signal is temporarily “hidden” and the model is challenged to reconstruct this missing component based on the other signals. In this way, the system learns how brain activity, heart rate, breathing and muscle activity relate to each other and where abnormalities occur.
After the training phase, SleepFM was first validated on existing clinical applications, such as automatically recognising sleep stages and determining the severity of sleep apnoea. The model performed at least as well as the current state-of-the-art algorithms on these tasks. This confirmed that the model can adequately analyse sleep according to existing clinical standards.
Predicting disease during sleep
The researchers then turned their attention to a more ambitious goal: predicting future disease based on sleep data. To do this, they linked polysomnography data to long-term health data from the same patients. The Stanford Sleep Medicine Centre, which has been conducting systematic sleep research since 1970, played a unique role in this. For a large patient cohort, consisting of approximately 35,000 people aged 2 to 96, both sleep measurements and electronic patient records were available, with up to 25 years of follow-up in some cases.
SleepFM analysed more than a thousand disease categories and identified 130 conditions for which the risk could be predicted with reasonable to high accuracy based on sleep data. The predictions were particularly strong for neurological disorders such as Parkinson's disease and dementia, cardiovascular disease, various forms of cancer and mortality. Performance was expressed in the so-called concordance index, a commonly used measure for predictive models. Values above 0.8, as achieved for several conditions, are considered very high in clinical research.
Combination of signals decisive
It is striking that no single physiological signal proved decisive, but rather the combination of different signals. Although brain activity played a greater role in predicting neurological disorders and heart signals in cardiovascular diseases, combining all data modalities yielded the most accurate results.
According to Mignot, it is mainly subtle disruptions in the interrelationships that indicate an increased risk, for example when the brain shows signals associated with sleep, while the heart shows signs of alertness.
The researchers are now working on follow-up steps to further improve the model and better explain how predictions are made. Although SleepFM cannot explain its conclusions in understandable language, interpretation techniques are being developed to gain insight into which patterns the model recognises for specific disease risks. They are also looking at expanding the model with data from wearables, so that sleep measurements outside the sleep laboratory can be utilised.
SleepFM underlines the potential of sleep as a window into future health. Combining sleep data with AI creates a new perspective on early risk detection, even before clinical symptoms become apparent. This research is part of a broader movement within digital healthcare, in which continuous monitoring, data integration and predictive models are becoming increasingly central in the transition from reactive to proactive and personalised care.
SleepFM distinguishes itself from other solutions, whether AI-powered or not, for measuring and analysing sleep data. Until now, we have mainly seen apps and other tools that determine the quality of our sleep based on data collected by, for example, smartwatches.
One development that comes close to SleepFM is the research described last month into tracking sleep patterns with unobtrusive wearables with the aim of detecting rare sleep disorders at an earlier stage.
