“Fix sleep schedule” features at the top of millions of New Year’s resolutions each year. In fact, it’s widely known that getting adequate rest — that is, eight hours per night — helps improve mood and cognitive performance. But how much does sleep impact health beyond energy and mood?
Turns out, a lot. Stanford researchers recently created an AI model, SleepFM, that uses sleep recordings as a predictor for disease. Led by senior co-authors James Zou and Emmanuel Mignot, the model is able to accurately predict the onset of over 130 conditions, ranging from dementia to stroke.
“We know intuitively that sleep is a very important aspect of human life,” said Zhou. “A typical individual spends one-third of our lives sleeping, but it’s still relatively under-explored from an AI perspective.”
SleepFM was trained on over 585,000 hours of sleep recordings from 65,000 participants across multiple sleep clinics. The data wasn’t contained in one type; Zou’s team specifically utilized polysomnography (PSG) recordings, which capture rich physiological signals from multiple aspects of the body.
“We’re taking very detailed sleep recordings that capture brain signals, heart signals, muscle contractions and even breathing patterns,” Zou said.
The combination of these inputs creates a multimodal dataset for the AI to learn about sleep holistically. However, a large dataset doesn’t come without its challenges. Rahul Thapa, a third-year computer science Ph.D. student and lead author on the study, described the technical hurdles in working with multimodal data. Thapa said the sheer number of signals present in the data was one of the biggest surprises.
With over eight hours of continuous recordings for each patient, the first main goal was to understand what training methods worked best at a large scale, which “took a significant amount of time and iteration,” according to Thapa.
The team found that training the AI across different body signals worked better than traditional supervised learning methods due to the variety in the dataset. They also developed a novel “leave-one-out” method, which trained the model to retain its predictive capabilities even with missing or heterogenous data.
“We’re basically trying to get AI to learn the language of sleep,” Zou said.
Thapa said the second part of the study focused on applications of the base model. By pairing their sleep data with patient electronic health records, the researchers asked whether patterns in someone’s sleep are informative about future health outcomes.
Thapa cautions that the predictions should be interpreted as estimates of relative risk and not a definitive diagnosis, since the models are not FDA-approved and have not been prospectively validated in a clinical setting.
“Our goal is to understand population-level signals and associations, rather than to provide medical decisions for individual patients,” he said.
Looking to the future, Zou and Thapa see this project extending into wearables, which are small, portable electronic devices with embedded sensors and software to collect health, fitness or performance data. With the latest models of Apple Watches even providing sleep apnea scores and ECGs, these devices are increasingly positioning themselves as the frontline in disease risk screening.
Chibuike Ukwakwe M.D. ’28 Ph.D. ’28, who researches wearable bioelectronics, praised the researchers’ creativity in designing SleepFM’s architecture. Although the model is trained on PSG data that includes far more signals than current consumer wearables in the market, Ukwakwe believes the technology could analyze wearable sleep data in the future.
“I can see data collected from wearables powered by AI being used to support clinical decision making,” Ukwakwe said.
This project is only the latest example of how AI can be used to integrate multimodal physiological data and glean clinical insights from sleep, which is now being considered a window into not just our current but future health.
“Sleep contains so much physiological information that we are only beginning to tap into,” said Thapa.