Portuguese 
English 
Spanish 
5 min of reading 
0 comments 
Trained with 600 thousand hours of sleep records from 65 thousand people over 25 years, an artificial intelligence from Stanford can predict 130 diseases — including Parkinson’s, cancer, and heart attack — by analyzing a single night of sleep
While you sleep, your body tells a story that no doctor can read with the naked eye.
But an artificial intelligence developed at Stanford University has learned to decode this story — and what it found is disturbing.
The model called SleepFM, published in the journal Nature Medicine in 2026, analyzes data from a single night of sleep and can predict the risk of developing over 130 diseases, including Parkinson’s, breast cancer, dementia, and myocardial infarction — years before the first symptoms appear.
-
China hints in a video at its possible 4th aircraft carrier, which could be the first with nuclear propulsion, and promises to expand artificial islands to reinforce maritime power, pressure Taiwan, and harden disputes in the South China Sea.
-
160 scientists from over 20 countries are putting the planet on alert by revealing that Earth has already entered the danger zone of tipping points, with corals, polar ice, the Amazon, and ocean currents at risk of cascading collapses that could affect billions of people.
-
Scientists doped 105 salmon with cocaine and metabolites in Sweden, and exposed fish swam up to 14 km more in eight weeks, raising alarms about how drugs dumped into rivers can alter entire ecosystems.
-
A mysterious golden ball found 3,200 meters deep in the ocean intrigued scientists for two years until DNA sequencing revealed that it came from an unexpected giant animal.
600 thousand hours of sleep analyzed: how the artificial intelligence learned to “read” the sleeping body
SleepFM was trained with 600 thousand hours of polysomnography records from approximately 65 thousand participants, collected over 25 years at the Stanford Sleep Medicine Center.
This data was cross-referenced with over 50 years of medical records, including electronic health records and clinical follow-up of each patient.
The result is a colossal database that connects what happens during each person’s sleep with the diseases they developed — or did not develop — decades later.
“SleepFM is essentially learning the language of sleep,” summarizes James Zou, associate professor of biomedical data science at Stanford University and co-author of the study.
The artificial intelligence does not just analyze whether the person slept well or poorly.
It identifies patterns invisible to humans: how the brain, heart, and breathing synchronize — or fail to synchronize — during the different phases of sleep.
When the brain sleeps but the heart stays awake: the signal that warns of future diseases
The most surprising discovery of the study is that body systems desynchronized during sleep act as early warnings of serious diseases.
When the brain has already entered deep sleep but the heart continues to beat as if the person were awake, something is wrong.
This desynchronization is invisible to the patient themselves. They wake up thinking they slept normally.
But for the artificial intelligence, this mismatch is a risk signature that may indicate cardiac, neurological, or oncological problems that would only appear clinically years later.
Parkinson’s with 89% accuracy, prostate cancer with 90%: the accuracy that scares
The researchers measured the accuracy of SleepFM using the C-Index — a concordance index where values above 0.80 mean that the model is correct in at least 80% of cases.
The results impressed even the creators themselves.
- Prostate cancer: C-Index of 0.90 — correct in 90% of cases
- Breast cancer: C-Index of 0.87 to 0.90
- Parkinson’s disease: C-Index of 0.89
- Hypertensive heart disease: C-Index of 0.88
- Dementia: C-Index of 0.85
- All-cause mortality: C-Index of 0.84
- Myocardial infarction: C-Index of 0.81
- Heart failure: C-Index of 0.80
In total, the artificial intelligence identified 130 predictable conditions with an accuracy greater than 0.80, covering categories such as cancer, pregnancy complications, circulatory diseases, and mental disorders.
What changes in practice: medicine that acts before symptoms
The direct implication is a revolution in preventive medicine.
Today, most diagnoses of Parkinson’s, dementia, and various types of cancer occur when symptoms are already evident — and often when the most effective treatment is no longer possible.
If a sleep test can identify risks years in advance, doctors could act before the disease sets in.
For Parkinson’s, for example, studies show that the disease begins to affect the brain up to 20 years before the first tremors.
The difference between intervening at this pre-clinical stage and intervening when symptoms appear could be the difference between maintaining or losing autonomy for the rest of one’s life.
What this means for those who sleep poorly in Brazil
Brazil has over 73 million people with some degree of sleep disorder, according to data from the Brazilian Sleep Association.
Most live with insomnia, apnea, or fragmented sleep without ever undergoing a polysomnography — the test that SleepFM uses as a basis.
If Stanford’s artificial intelligence can work with data from smartwatches and wearables, the impact in Brazil would be immense.
Millions of Brazilians who already use smartwatches to monitor sleep could discover risks of Parkinson’s, dementia, or heart diseases years in advance — without needing a laboratory test that costs hundreds of reais.
The technology is not yet available for clinical use. But the existence of an AI with 90% accuracy for prostate cancer using sleep data is something no doctor predicted ten years ago.
The limitations that the Stanford team emphasizes
The study, despite being promising, has important limitations.
The accuracy was measured retrospectively — that is, the AI analyzed past data and compared it with existing diagnoses. Prospective tests, with new and diverse populations, are still needed.
The data was collected via polysomnography, a lab test with sensors on the body. It is not something that can be done at home with a smartwatch.
The team is working to make the model compatible with wearable devices, but they have not reached that point yet.
Moreover, the model functions as a “black box” — it identifies patterns but does not yet explain why certain desynchronizations lead to certain diseases.
The team is developing techniques to interpret these specific physiological patterns, which would make the tool not only predictive but also explanatory.
The future where your watch knows you will get sick before you do
If SleepFM is adapted to work with data from smartwatches and smart rings, the scenario changes radically.
Millions of people who already use devices to monitor sleep could receive early risk alerts — transforming a wellness gadget into a real medical prevention tool.
For now, the technology is restricted to clinical polysomnographies.
But the foundation is laid: 600 thousand hours of sleep, 65 thousand patients, 50 years of medical records, and an artificial intelligence that has learned to listen to what the body says when we think it is silent.
The question remains: when we know with 90% certainty that we will develop a disease in 10 years, what will we do with that information?
Be the first to react!