Experiencing a sleepless night can lead to a sluggish mind the following day. However, recent studies indicate that even a single night’s sleep may hold significant insights regarding health issues that could remain asymptomatic for years.
In one notable experiment, an AI system analyzed overnight physiological signals to assess an individual’s risk for over 100 potential health disorders.
This innovative model, named SleepFM, was created by researchers from Stanford Medicine. It was trained using nearly 600,000 hours of polysomnography data collected from around 65,000 individuals, involving comprehensive overnight sleep studies that monitored various bodily functions including brain activity, heart rate, breathing, and movement.
Exploring Sleep Data with SleepFM
Traditionally, polysomnography has been utilized primarily as a clinical tool, focusing on scoring sleep stages and identifying sleep apnea. However, the research team contends that this merely scratches the surface of the invaluable data contained within these recordings.
“We collect a vast array of signals while studying sleep,” noted co-senior author Emmanuel Mignot, a professor of sleep medicine at Stanford. “It’s a rich source of general physiological data, captured over eight hours while the subject is still.”
Up until recently, human analysts and conventional software could only process a fraction of this complexity.
AI offers a solution by identifying patterns across thousands of nights and multiple physiological systems simultaneously.
Integrating AI in Sleep Studies
While AI has made significant advancements in fields such as radiology and cardiology, it has notably lagged behind in sleep science, despite its critical interconnections with brain function, metabolism, respiration, and cardiovascular health.
Study co-senior author James Zou, who serves as an associate professor of biomedical data science, remarked, “From an AI perspective, sleep is relatively unexplored. Much AI research addresses pathology or cardiology, while sleep remains undervalued, even though it is a fundamental aspect of our lives.”
This insight directed the team’s strategy. Rather than constructing a model for a singular task, they designed a foundational model to learn broad patterns first and make tailored predictions later.
SleepFM: Learning the Language of Sleep
SleepFM was trained similarly to a large language model, but instead of focusing on words, it learned from minute physiological segments.
The polysomnography data was divided into five-second intervals, allowing the model to interpret lengthy sleep sessions as sequential data and discern typical patterns of occurrence. “SleepFM essentially learns the language of sleep,” noted Zou.
The model simultaneously incorporated various channels, including electroencephalography (to measure brain activity), electrocardiography (for heart rhythms), electromyography (to assess muscle activity), along with pulse and airflow information.
Ensuring Reliable Model Training
The aim extended beyond merely interpreting each channel; it was about comprehending the relationships between them.
To achieve this, the researchers developed a training strategy wherein one data stream would be hidden, compelling the model to reconstruct it from the remaining data.
“A significant advancement we made was figuring out how to harmonize diverse data modalities so they could jointly learn the same language,” Zou stated.
Following training, the team fine-tuned SleepFM for conventional tasks in sleep medicine, such as classifying sleep stages and evaluating sleep apnea severity.
On these assessments, the model either matched or surpassed the performance of leading tools currently utilized in the field.
This outcome was crucial, as it indicated that the model wasn’t merely learning irrelevant noise and could reliably perform fundamental tasks before attempting more complex predictions.
Analyzing Sleep Data for Disease Risk
The pivotal phase involved forecasting potential diseases based solely on one night’s sleep. For this, researchers aligned sleep data with extensive medical outcomes, leveraging decades of records from a prominent sleep clinic.
Founded in 1970 by the esteemed William Dement, the Stanford Sleep Medicine Center provided a wealth of data for this project, encompassing around 35,000 patients aged 2 to 96, all of whom underwent polysomnography tests from 1999 to 2024.
The researchers matched these sleep studies with electronic health records, offering as much as 25 years of follow-up data for certain individuals.
SleepFM evaluated over 1,000 disease categories and successfully predicted 130 of them with notable accuracy based solely on sleep data.
Forecasting Diseases Years in Advance
The model achieved its most promising results in predicting cancers, complications related to pregnancy, circulatory disorders, and mental health issues, yielding a C-index exceeding 0.8 in these domains.
The C-index evaluates how effectively a model ranks risk among individuals. It does not indicate certainty for a single case but rather assesses whether the model typically ranks higher-risk individuals above those at lower risk.
“For every pair of individuals, the model ranks who is more likely to experience an event—like a heart attack—earlier. A C-index of 0.8 means that 80% of the time, the model’s prediction aligns with actual outcomes,” Zou explained.
Particularly strong results were observed for various conditions, including Parkinson’s disease, dementia, hypertensive heart disease, heart attack, prostate cancer, and breast cancer, as well as mortality.
“We were pleasantly surprised to find that the model was capable of making valuable predictions across a diverse array of health conditions,” Zou added.
Understanding the Insights from SleepFM
Despite these promising results, a critical question persists: what specific factors is SleepFM detecting? The researchers are currently developing tools to enhance interpretation and may also explore adding data from wearable devices to improve predictions.
“SleepFM doesn’t communicate its insights in plain language,” Zou pointed out. “However, we have created various interpretation techniques to determine what the model focuses on when it predicts a particular disease.”
One clear pattern has emerged: the most accurate predictions were linked not to individual channels, but to the analysis of multiple channels and identifying discrepancies.
“Our findings demonstrate that the most informative disease predictions arise from contrasting different channels,” Mignot emphasized.
This suggests that a mismatch in physiological signals may indicate underlying health issues. For instance, a brain that appears to be asleep while the heart shows signs of being “awake” could suggest that an internal problem is brewing.
The findings are detailed in the journal Nature Medicine.
—–
Enjoyed this article? Subscribe to our newsletter for exciting articles, exclusive content, and the latest news.
Explore more at EarthSnap, a complimentary app brought to you by Eric Ralls and Earth.com.
—–
