Physiological “Disconnection” During Sleep: AI Traces Clues to Neurodegenerative Risk in 124,000 Studies

A single night’s sleep recording captures more than breathing and brain waves; it may also contain faint traces of health changes that emerge years later. A newly released preprint shows that researchers used a human-supervised multi-agent AI to analyze approximately 124,000 polysomnography studies and identified signals in the degree of coordination between the brain and other physiological systems that were associated with subsequent diagnoses of neurodegenerative diseases.

The environment, called AI Sleep Co-Scientist, did not conduct the research independently. Sleep experts first set the direction, after which different agents divided the work of generating hypotheses, designing signal-preprocessing workflows, and performing statistical analyses. Researchers reviewed intermediate results, while the formal selection of hypotheses and initiation of full-dataset analyses also required human confirmation. The system also linked numbers and figures in the paper back to executable code in an attempt to reduce the risk of large language models producing erroneous analyses or conclusions that cannot be traced.

The research covered four clinical and epidemiological databases comprising more than 50 TB of raw signals and used five case studies to examine disease risk, clinical phenotypes, and sleep regulation. The most prominent analysis focused on “physiological coupling”: whether activity among different brain regions, as well as among the brain, heart, muscles, and eye movements, maintains stable coordination during sleep. The researchers conducted the exploratory analysis using data from a Stanford sleep clinic and repeated the comparison using another Human Sleep Project dataset.

The results showed that, during stage 2 non-rapid eye movement sleep, each one-standard-deviation decrease in coupling between brain regions was associated with a risk ratio of 1.48 for a subsequently recorded diagnosis of Parkinson’s disease and 1.38 for Alzheimer’s disease. The respective 95% confidence intervals were 1.31 to 1.67 and 1.25 to 1.53. Lower brain–heart coupling was also associated with subsequent diagnoses of Parkinson’s disease, Alzheimer’s disease, and dementia. However, the overall pattern of effects showed only moderate consistency across databases, indicating that differences in healthcare settings, patient composition, and measurement conditions still affect the signals.

The AI also helped construct a “sleep age” model comprising seven domains, including sleep architecture, breathing and blood oxygen, autonomic function, and brain–heart interactions. An approach that retained physiological modules and then integrated their predictions was slightly more accurate in both databases than directly mixing all features. Its age residual was also associated with multiple cardiovascular, kidney, and respiratory diseases. Other case studies characterized arousal features in insomnia with sleep apnea, the regulation of rapid eye movement sleep, and differences in brain-wave oscillations in narcolepsy type 1.

These results cannot currently be interpreted as showing that reduced sleep coupling causes neurodegeneration, nor can a single sleep study be used as an individual screening tool. The study used a retrospective design, and the data largely came from referred patients or specific epidemiological populations. Disease endpoints were determined primarily from diagnosis codes in medical records, while unmeasured comorbidities, medications, and differences in healthcare utilization could all introduce confounding. The paper has also not yet undergone peer review. At this stage, these signals are better regarded as hypotheses for validation in prospective studies rather than mature clinical biomarkers.

The real advance of this work may lie not only in identifying a set of risk associations, but also in demonstrating how AI can participate in scientific exploration involving vast physiological datasets. A design that makes code traceable, preserves expert decisions, and checks results across databases is more pragmatic than fully automated research. The next step will still require independent teams to reproduce the analyses and confirm, in prespecified, representative follow-up studies, whether their incremental value exceeds that of age, medical history, and established sleep measures.