Forecasting disease<\/p>\n
After the training phase, the researchers could fine-tune the model to different tasks.<\/p>\n
First, they tested the model on standard sleep analysis tasks, such as classifying different stages of sleep and diagnosing the severity of sleep apnea. SleepFM performed as well as or better than state-of-the-art models used today.<\/p>\n
Then the researchers tackled a more ambitious goal: predicting future disease onset from sleep data. To identify which conditions could be forecast, they needed to pair the training polysomnography data with the long-term health outcomes of the same participants. Fortunately, they had access to more than half a century\u2019s worth of health records from a sleep clinic.<\/p>\n
The Stanford Sleep Medicine Center<\/a> was founded in 1970 by the late William Dement<\/a>, MD, PhD, widely considered the father of sleep medicine. The largest cohort of patients used to train SleepFM \u2014 some 35,000 patients ranging in age from 2 to 96 \u2014 had their polysomnography data recorded at the clinic between 1999 and 2024. The researchers paired these patients\u2019 polysomnography data with their electronic health records, which provided up to 25 years of follow-up for some patients.<\/p>\n (The clinic\u2019s polysomnography recordings go back even further, but only on paper, said Mignot, who directed the sleep center from 2010 to 2019.)<\/p>\n SleepFM analyzed more than 1,000 disease categories in the health records and found 130 that could be predicted with reasonable accuracy by a patient\u2019s sleep data. The model\u2019s predictions were particularly strong for cancers, pregnancy complications, circulatory conditions and mental disorders, achieving a C-index higher than 0.8.<\/p>\n The C-index, or concordance index, is a common measure of a model\u2019s predictive performance, specifically, its ability to predict which of any two individuals in a group will experience an event first.<\/p>\n \u201cFor all possible pairs of individuals, the model gives a ranking of who\u2019s more likely to experience an event \u2014 a heart attack, for instance \u2014 earlier. A C-index of 0.8 means that 80% of the time, the model\u2019s prediction is concordant with what actually happened,\u201d Zou said.<\/p>\n SleepFM excelled at predicting Parkinson\u2019s disease (C-index 0.89), dementia (0.85), hypertensive heart disease (0.84), heart attack (0.81), prostate cancer (0.89), breast cancer (0.87) and death (0.84).<\/p>\n \u201cWe were pleasantly surprised that for a pretty diverse set of conditions, the model is able to make informative predictions,\u201d Zou said.<\/p>\n Models of less accuracy, with C-indices around 0.7, such as those that predict a patient\u2019s response to different cancer treatments, have proven useful in clinical settings, he added.<\/p>\n Interpreting the model<\/p>\n The team is working on ways to further improve SleepFM\u2019s predictions, perhaps by adding data from wearables, and to understand exactly what the model is interpreting.<\/p>\n \u201cIt doesn\u2019t explain that to us in English,\u201d Zou said. \u201cBut we have developed different interpretation techniques to figure out what the model is looking at when it\u2019s making a specific disease prediction.\u201d<\/p>\n The researchers note that even though heart signals factor more prominently in heart disease predictions and brain signals factor more prominently in mental health predictions, it was the combination of all the data modalities that achieved the most accurate predictions.<\/p>\n \u201cThe most information we got for predicting disease was by contrasting the different channels,\u201d Mignot said. Body constituents that were out of sync \u2014 a brain that looks asleep but a heart that looks awake, for example \u2014 seemed to spell trouble.<\/p>\n