New AI model can predict disease risk while you sleep

Polysomnography is considered the “gold standard” for studying sleep and is performed in a laboratory setting under the overnight supervision of specialists. As scientists have found, such records represent a huge, previously largely untapped source of data on human health. To make effective use of the vast amount of sleep data, scientists have created a basic model – a type of AI that learns from large amounts of information and can then be applied to a wide variety of tasks. An example of such systems are large language models like ChatGPT, which are trained on colossal sets of texts.

SleepFM was trained on 585,000 hours of polysomnographic recordings collected from patients in various sleep laboratories. This data was broken down into five-second chunks, similar to how language models break text into words. The model learned to combine several types of signals – electroencephalogram, electrocardiogram, electromyogram, pulse and respiratory flow indicators – and understand the connections between them.

To achieve this, researchers have proposed a new learning approach called leave-one-out contrast learning. The essence of the method is that one category of data is deliberately hidden, and the model must “invent” the missing information based on other signals. After preliminary training, SleepFM was further configured to solve specific problems. First, it was tested on familiar tasks—recognizing sleep stages and determining the severity of apnea. The results were on par or even better than those of modern specialized models.

The team then moved on to the more challenging goal of predicting diseases years in advance based on overnight sleep recordings. To do this, it was necessary to compare the polysomnography data with long-term medical information about the same people. Fortunately, the researchers had access to the sleep clinic’s medical archives, which were maintained for more than 50 years.

The SleepFM model analyzed more than a thousand types of diseases listed in medical records and identified 130 conditions whose risk could be reliably predicted from one night of sleep data. The model was particularly accurate in predicting cancer, pregnancy complications, heart disease and mental disorders – its performance reached a consistency index (C-index) above 0.8.

The C-index is a common metric that shows how accurately a model can predict which of two people in the same group will experience an event first. SleepFM showed impressive results in predicting Parkinson’s disease (C-index 0.89), dementia (0.85), coronary heart disease (0.84), myocardial infarction (0.81), prostate cancer (0.89), breast cancer (0.87), and risk of death (0.84).