AI and Sleep Data: New Study Suggests Disease Risk Predictions

A recent study indicates that advancements in artificial intelligence could enable the use of sleep data to assess disease risk. Researchers at Stanford Medicine have developed an AI model, named SleepFM, which is based on around 600,000 hours of sleep data gathered from over 60,000 individuals across various sleep clinics.

This model has the capacity to forecast a person’s likelihood of developing more than 100 different health conditions, as highlighted in a university announcement.

SleepFM was trained using polysomnography, a detailed method that captures brain and heart activity, breathing patterns, limb movements, and eye movements—often referred to as the “gold standard” in sleep research. Dr. James Zou, an associate professor of biomedical data science and one of the study’s co-authors, emphasized that “Sleep contains much more information about our future health than what we use today.”

“By learning the language of sleep, our AI model opens new doors for research in the science and medicine of sleep,” he added, noting that humans typically spend about a third of their lives asleep.

In the course of the study, the team integrated sleep data with electronic health records, which provided insights spanning up to 25 years. This approach allowed the model to analyze disease data across 1,000 categories, ultimately identifying 130 diseases that it could predict with “moderate accuracy.”

According to Zou, “By analyzing a night of sleep with powerful AI, we found that sleep patterns can predict risk for more than 100 different diseases years before diagnosis.” The model’s predictions were particularly strong for conditions such as cancer, cardiovascular disease, and mental health disorders.

However, some experts urge caution. Dr. Harvey Castro, a board-certified emergency physician, remarked that while SleepFM is groundbreaking, it’s not yet ready for clinical use. “A critical signal is not equivalent to a ready-to-use drug,” he stated. “Ranking risks is not the same as predicting outcomes.” He pointed out that while the tool can rank risks, it does not necessarily guarantee disease prediction or outcomes.

The researchers acknowledged limitations in their study. Zou mentioned, “There’s still a lot we don’t understand… most analyses focus on narrow tasks like sleep staging or apnea detection.” They also warned that this research does not provide specific medical advice, but rather underscores the importance of sleep.

Looking ahead, the research team aspires to expand their efforts by collecting data from wearables, which could lead to better insights about what the model interprets. As of now, this technology is still confined to a research setting and hasn’t yet become available for public use.