For most of medical history, sleep has been treated as a passive state, something the body does while clinicians wait for symptoms to appear during waking hours. Two studies published in Nature Communications and Nature Medicine are challenging that assumption in a concrete way. The argument they make, taken together, is that sleep data already contain health signals that current clinical practice largely ignores.
Sleep Changes Before Diagnosis, Not After
The foundational insight behind both studies is not new, but it is underappreciated. Growing evidence, as summarized by Rachel K. Rowe of the Department of Integrative Physiology at the University of Colorado Boulder, suggests that changes in sleep patterns can emerge before a disease is formally diagnosed. This is the key detail that most coverage of sleep research misses.
The standard clinical model works in reverse: a patient develops symptoms, seeks care, receives a diagnosis, and then a clinician might ask about sleep quality as a secondary concern. What the research trajectory described by Rowe points toward is a different sequence entirely. Sleep disruption may not be a consequence of disease. It may be a precursor, a signal that something is shifting in the body before it becomes visible through conventional diagnostic criteria.
This reordering of cause and effect has significant implications. If sleep changes precede diagnosis, then sleep data become prospectively useful rather than retrospectively descriptive.
Two Studies, One Shared Claim
The study by Bilal et al., published in Nature Communications, and the study by Thapa et al., published in Nature Medicine, arrive at a shared conclusion through what appear to be complementary approaches. Both are described as foundation models, a term from machine learning referring to large models trained on broad datasets that can then be adapted to specific tasks.
The framing Rowe uses to describe their significance is precise and worth unpacking: sleep data hold substantially more information about health than clinicians and researchers currently extract from them. That is not a claim about what sleep data might theoretically contain. It is a claim about a gap between what is already present in the data and what current practice actually uses.
This gap is the real subject of both papers. Polysomnography, the standard clinical tool for measuring sleep, captures a range of physiological signals including brain activity, breathing, and movement. The data it generates are rich. But the way those data are typically analyzed, focused on identifying specific disorders like sleep apnea, leaves most of the signal on the table. Foundation models trained on sleep data are designed to recover some of what is being discarded.
The titles of the two papers make the ambition explicit: one concerns “sleep-based risk stratification and clinical outcome,” the other is described as “a multimodal sleep foundation model for disease prediction.” Risk stratification and disease prediction are not the same as diagnosis. They operate earlier in the timeline, which is precisely the point.
What This Means Beyond the Clinic
The broader significance of this research sits at the intersection of AI capability and medical infrastructure. The ability to train large models on physiological time-series data, and to extract predictive signals from them, is not limited to sleep. But sleep is a particularly useful domain for several reasons.
Sleep data are already being collected at scale. Wearable devices, including consumer smartwatches, generate continuous sleep-related measurements for millions of people. Clinical sleep studies produce detailed polysomnographic records. The data infrastructure exists. What has been missing is the analytical framework capable of treating that data as a source of early health signals rather than a record of the previous night.
Foundation models trained on sleep data could, in principle, augment the work of clinicians by surfacing patterns that would not be visible through standard interpretation. A clinician reviewing a sleep study for apnea events is doing one job. A model trained to identify early signals of a broad range of conditions is doing a different job, and the two are not in competition. The model handles the volume and complexity of signal extraction; the clinician interprets the result in the context of a specific patient.
This is also a reminder that AI’s most consequential applications in medicine may not be the dramatic ones, robotic surgery or real-time diagnosis, but the quieter ones: finding information that was always there, in data that was already being collected, that no one had the tools to read.
In Short
Sleep disruption can precede disease diagnosis, which means sleep data may contain early health signals that current clinical practice does not fully use. Two foundation models, one published in Nature Communications by Bilal et al. and one in Nature Medicine by Thapa et al., are designed to close that gap. The core claim, articulated by University of Colorado Boulder researcher Rachel K. Rowe, is straightforward: sleep data hold substantially more health information than is currently being extracted from them. AI is not replacing clinical judgment here. It is reading a signal that was always present but previously unreadable at scale.
Based on reporting from Nature: Machine Learning.