Future capability, systems physiology
Recovery intelligence as a tool for predicting disease.
Concept in development, subject to clinical review and regulatory pathway
The finding
585,000 hours of physiology, 130 predictable conditions.
585,000
Hours of polysomnography used for training
65,000 participants
1,000+
Disease categories screened in health records
Stanford Sleep Medicine Center
130
Conditions predictable with reasonable accuracy
Nature Medicine, Jan 2026
25 yrs
Longest outcome follow-up in the cohort
Records from 1999 to 2024
Source: Stanford Medicine, "New AI model predicts disease risk while you sleep," January 6, 2026, reporting on a study published in Nature Medicine. Sleep Week does not operate SleepFM. This page describes a proposed future service layer.
How it reads a night
Six systems, one language.
Hover or select a channel to read what it contributes.
Foundation model output
A ranked risk horizon across 130 conditions, read from system relationships
The strongest signal is not any single channel. It is incoherence between channels, a brain that reads asleep while the heart reads awake.
Predictive strength
Where the signal is strongest.
C-index by condition, as reported in the Stanford Medicine study. 0.50 is chance. Bars are scaled from the 0.50 chance baseline, so differences between conditions are visible.
Strongest categories
- Cancers
- Circulatory conditions
- Pregnancy complications
- Mental health disorders
- Neurodegenerative disease
Risk horizon
Night 1
8 hours of physiology recorded
Minutes
5-second windows tokenized, the language of sleep
Same week
Clinician-reviewed risk ranking issued
1 to 25 years
Outcome window the model was validated against
Why it changes the deliverable
From a sleep score to a coherence-informed risk horizon.
Conventional sleep report
Predictive sleep analysis
Data used
A handful of summary metrics
Every channel, every five seconds, all night
Most of a polysomnogram is discarded today.
Question answered
How did you sleep last night
What is your physiology signalling about the next decade
Time frame
One night, retrospective
Validated against outcomes up to 25 years out
Modality handling
Channels read separately by a scorer
Channels contrasted against each other by the model
Disagreement between brain and heart carried the most information.
Output
Apnea index and stage percentages
Ranked, clinician-reviewed risk across condition families
Proposed Sleep Week pathway
How this would enter the program.
01
Consent and eligibility
Opt-in, separate from the core program, with a written scope of what is and is not being predicted.
02
Full polysomnography night
Lab-grade montage in the Sleep Suite rather than wearables alone.
03
Model-assisted analysis
Multimodal risk ranking produced as decision support, never as a diagnosis.
04
Physician review and referral
A clinician interprets, contextualises and routes anything actionable to real care.
Guardrails
Prediction is not diagnosis.
No unaccompanied results
Nothing is released without a physician reviewing it with the guest first.
No screening claims
The model is not a cleared screening test for cancer, dementia or cardiac disease.
Interpretability is partial
Researchers can see which channels drive a prediction, not the full reasoning.
Data stays yours
Recordings would not be used for model training without explicit separate consent.
Read the research context
SleepFM was developed by Stanford Medicine researchers with colleagues at the Technical University of Denmark, Copenhagen University Hospital, BioSerenity, University of Copenhagen and Harvard Medical School, with co-senior authors Emmanuel Mignot, M.D., Ph.D., and James Zou, Ph.D. The team is exploring whether wearable data can extend the model beyond the lab. Published in Nature Medicine on January 6, 2026.
Related