In a new paper published in npj Digital Medicine, a Nature portfolio journal, Johns Hopkins researchers Will Coon and Mattson Ogg asked a fundamental question: How much health-related information is lost when an entire night of sleep is summarized using only the conventional five sleep stages?
They trained Sleep2.0, a self-supervised AI foundation model, on 11,261 overnight sleep recordings without providing sleep-stage labels. The model independently rediscovered the familiar organization of wakefulness, REM sleep, and the non-REM stages, confirming that conventional staging captures important physiological structure. But it also preserved finer-grained patterns within those stages—particularly within non-REM stage N2 sleep—that conventional staging compresses. These patterns carried additional information associated with aging, metabolic health, and cardiopulmonary health, and differed in familiar physiological features such as sleep-spindle and slow-oscillation activity.
The results suggest that traditional sleep stages provide a valuable clinical foundation but not a complete account of the health information encoded in brain activity during sleep. AI may eventually allow researchers to complement the conventional sleep-stage chart, or hypnogram, with higher-resolution biomarkers for health screening, longitudinal monitoring, and scientific discovery.
“The question wasn’t simply whether a large AI model could predict health-related traits from sleep EEG. It was whether those predictions were just a sophisticated repackaging of the traditional hypnogram, or whether the model had uncovered meaningful physiology hidden within the stages. We found evidence for both: conventional sleep stages explain a great deal, but they do not tell the whole story,” says Coon.
“Two 30-second periods can both be labeled N2 sleep and still contain meaningfully different physiology. When we treat them as identical, we may average away information related to our health state and disease risk in the future,” Coon explains. “Sleep gives us hours of continuous information about the brain and body every night. Foundation models offer a way to use that information at much higher resolution, without requiring experts to label every pattern in advance.”
Coon and Ogg both teach Introduction to Brain-Computer Interfaces – 585.783, and Coon also teaches Frontiers in Neuroengineering – 585.781 through the Applied Biomedical Engineering program for Johns Hopkins Engineering for Professionals.