From one night of sleep-clinic recording, Stanford’s model predicted 130 medical conditions, including heart failure and stroke, with a C-index of at least 0.75 [Nature Medicine]. That is the central finding from Stanford’s 『SleepFM』, a foundation AI model published in January 2026 in Nature Medicine. It was trained on roughly 585,000 hours of sleep-study recordings from about 65,000 participants [Nature Medicine]. The authors say it can complement existing risk assessment tools and help identify early signs of disease [Nature Medicine]. It is early work. The training data came mostly from people referred for sleep studies, and the disease predictions were tested on patients from the Stanford sleep clinic.
Sleep as a Biological Broadcast
Most of us think of sleep as downtime.
Photo by Recovery. A blank stretch between days. But sleep is anything but quiet, biologically speaking. During polysomnography, sensors record brain activity and eye movement, heart rhythm, muscle activity and breathing across many channels, and the recordings in this study could include up to 23 [Nature Medicine]. Each channel tells its own story about what’s happening inside the body.
What makes the 『SleepFM』 research notable is how it uses this data. The authors say earlier models mostly targeted individual outcomes, were trained on relatively small datasets, and often generalized poorly across cohorts [Nature Medicine]. They describe polysomnography as underutilized for questions beyond sleep disorders [Nature Medicine].
Annual bloodwork captures a single snapshot. A sleep recording offers hours of continuous data from several body systems.
What SleepFM Actually Found
Stanford’s model doesn’t just analyze sleep quality.
Photo by It estimates future disease risk from a single night’s recording [Nature Medicine]. The authors evaluated 1,041 conditions in a held-out test set of 5,019 Stanford sleep clinic patients, and 130 reached a C-index of at least 0.75 [Nature Medicine]. Risk was assessed for diagnoses that appeared after the sleep study, and the 6-year AUROC counts diseases that developed within six years [Nature Medicine]. Mortality prediction reached a C-index of 0.84, dementia prediction 0.85, and myocardial infarction 0.81 [Nature Medicine]. The C-index measures how well the model ranks people’s risk across a group. It is not the probability that any one person will develop the disease. Atrial fibrillation prediction reached an AUROC of 0.81, close to the 0.82 an earlier study reported for that condition, although direct comparisons are difficult because the samples differ [Nature Medicine].
Each prediction uses one night of recording as input. Under the hood, convolutional layers and a transformer that reads 5-minute windows produce embeddings for each signal type. A two-layer LSTM network then reads a standardized 9-hour night of those embeddings together with age and sex [Nature Medicine]. The model learns from brain, heart, muscle and breathing signals together.
What stands out is the breadth of the model. The authors note that earlier approaches typically focused on individual outcomes [Nature Medicine]. SleepFM produced predictions for 130 conditions that met the 0.75 threshold from the same recordings [Nature Medicine].
What This Could Mean for Preventive Health
Many people feel disconnected from preventive medicine.
Photo by The advice feels abstract, the tests infrequent. 『SleepFM』 points toward something different: a risk estimate built from a physiological recording rather than a general warning.
A dementia C-index of 0.85 does not mean a person has an 85% chance of developing dementia. The C-index is the proportion of correctly ranked risk pairs [Nature Medicine]. It shows that the model tends to rank people who go on to develop dementia above people who do not, within the test group. That is different from an individual’s probability.
If the scores hold up in wider groups, they could help clinicians identify higher-risk patients earlier. The authors suggest SleepFM could complement existing risk assessment tools [Nature Medicine], but its value in clinical care has not been shown.
Limits and What Comes Next
『SleepFM』 is a proof of concept, not a finished clinical product.
Photo by It’s worth being honest about the gaps.
Access is the most immediate barrier. The model was built and tested on polysomnography, the overnight recording made in a sleep lab with clinical-grade sensors, and not on data from consumer wrist trackers [Nature Medicine]. Someone outside a sleep clinic has no such recording to analyze.
Who was studied is another limit. The authors note that the dataset consists primarily of patients referred for sleep studies because of suspected sleep disorders or other conditions needing overnight monitoring. They say this selection bias means the cohort is not representative of the general population, and that people without sleep complaints or with limited access to sleep clinics are underrepresented [Nature Medicine]. They also report that performance degraded somewhat on a test set of later recordings, and that only a subset of conditions could be checked in an independent dataset. Further work is needed to understand which sleep patterns drive the predictions [Nature Medicine].
The authors write that as wearable sleep technologies advance, models like 『SleepFM』 may offer opportunities for noninvasive, real-time health monitoring [Nature Medicine]. That is a possibility, not a result, because the study itself used lab recordings.
For now, this research invites a gentle reconsideration: prioritizing sleep isn’t just about feeling rested. It may be one of the most information-rich things your body does.
Stanford’s 『SleepFM』 suggests that a night of sleep-clinic recording carries information about the risk of many diseases over the following six years. It is a research result from a population referred for sleep studies, and validation in wider groups will decide whether it can inform preventive care.
