For years, doctors have relied on symptoms, check-ups, and medical history to assess a patient’s risk to disease. But what if the body’s signals during a single night’s sleep could provide those insights instead?
Researchers from Stanford Medicine in the United States have developed a new artificial intelligence (AI) model that is capable of predicting a person’s risk of developing more than 100 health conditions based on physiological recordings from just one night’s sleep.
The foundational AI model known as SleepFM has been trained on nearly 6,00,000 hours of sleep data collected from 65,000 participants. The dataset brings together a wide range of sleep signals, including brain activity, heart activity, respiratory signals, leg movements, eye movements, and other physiological data captured overnight from the participants using various sensors, according to Stanford Medicine’s blog post on January 6.
While using AI-powered tools to predict diseases is not entirely new, SleepFM appears to be one of the first AI models that analyses sleep data to achieve similar outcomes.
“From an AI perspective, sleep is relatively understudied. There’s a lot of other AI work that’s looking at pathology or cardiology, but relatively little looking at sleep, despite sleep being such an important part of life,” Dr James Zou, associate professor of biomedical data science and one of the co-authors of the study, was quoted as saying.
“We record an amazing number of signals when we study sleep. It’s a kind of general physiology that we study for eight hours in a subject who’s completely captive. It’s very data rich,” said Dr Emmanuel Mignot, another co-author of the study that is titled ‘A multimodal sleep foundation model for disease prediction’ and published in Nature Medicine.
The Stanford Medicine researchers’ findings come at a time when AI companies are increasingly pushing into healthcare by rolling out specialised tools and services. In the past few days, OpenAI launched ‘ChatGPT for Health’ and Anthropic announced a similar offering called ‘Claude for Healthcare’. However, the growing number of AI-powered healthcare tools has sparked privacy concerns while also highlighting the risk of ‘hallucinations’ which could lead to inaccurate or misleading medical information.
What is SleepFM? How was it developed?
In order to collect sleep data to train the AI model, the Stanford Medicine researchers said that they relied on a large cohort of 35,000 patients ranging in age from two to 96 years. These patients had had their polysomnography data recorded at Stanford’s sleep clinic between 1999 and 2024.
Around 5,85,000 hours of polysomnography data was collected and paired with the patient’s electronic health records. The sleep data was then split into five-second increments, analogous to words that large language models (LLMs) use to train on.
Overview of SleepFM framework. (Image: Nature Medicine)
After training the AI model, researchers further fine-tuned it to carry out different tasks.
How did the AI model perform?
In their testing, the researchers assessed the performance of SleepFM based on standard sleep analysis tasks such as classifying different stages of sleep and diagnosing the severity of sleep apnea.
Next, SleepFM was used to achieve the more ambitious goal of predicting future disease risk from sleep data. Upon analysing more than 1,000 types of diseases based on the participants’ health records and sleep data, SleepFM was found to be capable of predicting 130 diseases with reasonable accuracy.
For predicting diseases such as various types of cancers, pregnancy complications, circulatory conditions, and mental disorders, SleepFM obtained an average C-index higher than 0.8. Simply put, the concordance index (C-index) is used to measure an AI model’s ability to predict which of any two individuals in a group will experience an event first.
C-indices higher than 0.8 is significant since AI models with scores of around 0.7 have also proven useful in clinical settings, as per the researchers. SleepFM scored C-index 0.89 for predicting Parkinson’s disease, along with dementia (0.85), hypertensive heart disease (0.84), heart attack (0.81), prostate cancer (0.89), breast cancer (0.87), and death (0.84).
Moving forward, the researchers said that they are looking to improve the accuracy of SleepFM’s predictions by adding data from wearables to the training dataset.
More research is also needed to understand exactly what the model is interpreting. “It doesn’t explain that to us in English. But we have developed different interpretation techniques to figure out what the model is looking at when it’s making a specific disease prediction,” Zou said.
The AI model performed as well as or better than existing state-of-the-art models, as per the researchers. (File photo) 
Overview of SleepFM framework. (Image: Nature Medicine)
Share
