The AI that reads your sleep: what Stanford's SleepFM can and cannot do

What is SleepFM?

SleepFM comes from Stanford, from the teams of sleep physician Emmanuel Mignot and data scientist James Zou, working with partners including the Technical University of Denmark. The study, led by Rahul Thapa and Magnus Ruud Kjaer, was published on January 6, 2026, in Nature Medicine.

The model learned from about 585,000 hours of polysomnography, the full overnight sleep test, recorded in roughly 65,000 people. The largest group came from the Stanford Sleep Medicine Center: about 35,000 patients aged 2 to 96, recorded between 1999 and 2024. Their recordings were then linked to their health records, with up to 25 years of follow-up.

SleepFM is a foundation model. Like the language models behind chatbots, it first learns the general patterns in its data, here a night of sleep, and then reuses that knowledge for specific questions. Older tools were usually built to detect one condition at a time.

How does an AI read a night of sleep?

What a sleep lab records

Polysomnography records several signals at once: brain activity (EEG), eye movements (EOG), muscle tone (EMG), heart rhythm (ECG), breathing and blood oxygen. Together they map the stages of the night, including REM sleep, when the eyes move quickly, the muscles go slack and dreams tend to be most vivid.

These recordings have mostly served to diagnose sleep apnea or narcolepsy. The question behind SleepFM is simple: do they hold more information than doctors usually read?

One model, many questions

The model cuts each night into five-second segments. During training, it hides one type of signal, the heart for example, and learns to reconstruct it from the others. This step needs no diagnosis: the model only learns how brain, heart, muscles and breathing usually move together.

Only then is it linked to health records, to see which sleep patterns come before which diagnoses. According to Stanford, combining all the signals gave the best results. The researchers also noticed that channels out of step with each other, such as a brain that looks asleep while the heart looks awake, seemed to go with a higher risk.

What can SleepFM predict, and how reliably?

The team tested more than 1,000 disease categories. For 130 of them, the model reached a C-index of at least 0.75. This score measures ranking: take two people and check whether the model gives the higher risk to the one who falls ill first. A score of 0.5 is chance; 1 is a perfect ranking.

Neurodegenerative diseases

The most striking scores concern the brain: 0.89 for Parkinson's disease and 0.85 for dementia. That fits what sleep medicine already knew. In REM sleep behavior disorder, people act out their dreams because the usual muscle paralysis of REM sleep fails, and this disorder often comes years before Parkinson's disease or related conditions.

SleepFM does not rely on that single sign; it looks for subtler patterns across all the signals. A good score is still not a verdict. A model that ranks thousands of people well can be wrong about one person.

Cardiovascular diseases and cancers

Beyond neurology, Stanford reports a C-index of 0.81 for heart attack, 0.84 for death from any cause, 0.89 for prostate cancer and 0.87 for breast cancer. One hypothesis is that some diseases change the automatic regulation of heart rate and breathing during sleep long before symptoms appear.

These figures describe people studied in sleep clinics. They do not tell you your own risk, and the model is not available as a test you can book.

The DREAM database: brain activity meets dream reports

Dream research has a different problem: few labs wake enough sleepers to collect solid data. The DREAM database, presented in Nature Communications in August 2025, pools that work. Its first release brings together 20 datasets, 505 participants and 2,643 awakenings, each with the brain recording just before waking and a standardized account of what the person experienced.

The first analysis is modest, and telling. From the EEG alone, a model could guess better than chance whether the sleeper would report an experience, more clearly in REM sleep than in other stages. It says whether someone was probably dreaming, not what they dreamed.

The database is open to researchers. It is a step toward understanding why we dream, and a reminder that dream content still comes from the dreamer's own words.

What can AI do with your dream reports?

From hand coding to automated analysis

For decades, dream content was studied by hand. In the Hall and Van de Castle system, published in 1966, trained coders count the characters, interactions and emotions in each report. It is reliable, but slow.

In 2020, researchers built a tool that applies this grid automatically and ran it on about 24,000 reports from the DreamBank archive. Its results fit the continuity hypothesis: dreams tend to carry over the concerns and relationships of waking life. A 2026 study on what shapes dream content points the same way.

These tools describe large groups. They do not decide what a single dream means, and they are not used to diagnose anyone.

How Noctalia's AI reads your dream reports

Here the AI works on what you remember, not on brain signals. You record the dream by voice or in writing, it is transcribed, and the AI suggests a reading: symbols, emotions, possible links with your days, and questions to explore. That reading happens only if you have agreed to AI processing in the app. Each reading is a lead, not a verdict.

Fictional example: over three weeks, a journal collects a mirror showing a slightly different face, a key that no longer fits the front door, and a path that forks in the fog. Each entry also notes a mood: unease, then curiosity. Read together, these dreams raise a question rather than an answer: what is changing in my life right now, and how do I feel about it?

That is what a dream journal makes possible: seeing what comes back over time. Noctalia does not replace a therapist, and nightmares that keep disturbing your sleep deserve professional help.

What comes next, and where are the limits?

A health check-up in one night?

Not yet. SleepFM needs a full lab recording, with many sensors and a technician. The Stanford team wants to improve its predictions, possibly by adding data from wearables. Watches and rings measure far fewer signals, and their accuracy has limits, as our article on sleep trackers explains.

If tools like this reach the clinic one day, they would most likely serve as a first screen, followed by the usual medical tests.

Limitations and ethical questions

The authors list several limits themselves. Most patients had been referred to a sleep clinic, so they do not represent the general population. Performance dropped somewhat on more recent recordings. Outside validation, on the Sleep Heart Health Study, could only cover part of the diseases. And the model still cannot explain, case by case, what drives a prediction.

Ethical questions follow. Who gets to see these predictions? Could an insurer ask for them? How do you live with a false alarm about Parkinson's disease? The right not to know, long discussed in genetics, may need to cover sleep data too.

Research on dreams and mental health and on dreams and creativity is a reminder that sleep is more than a biomarker. It is an intimate experience, and monitoring it every night raises real questions of privacy and autonomy.

One practical point in the meantime: if you snore with pauses in your breathing, feel very sleepy during the day, or a partner tells you that you act out your dreams, see a doctor. A sleep study is the test designed for exactly that.