One Night of Sleep Can Predict 130 Diseases. What Happens Next?

In my last post, I wrote about how three years of sleep tracking bought me almost nothing. Six extra minutes per night. +2.4 points on my Oura score. Over 1,000 nights of data, and I couldn't move the needle on my actual sleep.

That's still true. Using sleep data to sleep better remains frustratingly hard.

But this week, Stanford published a striking study in Nature Medicine you’ve probably been seeing in your feeds that reframes some of the ways sleep data can be truly useful.

Researchers from Stanford built an AI model called SleepFM and trained it on 585,000 hours of polysomnography (sleep lab) data from over 65,000 participants across multiple sleep studies. They then linked a subset of roughly 35,000 Stanford Sleep Medicine Center patients (recorded between 1999 and 2024) to long-term health outcomes. PSG is the clinical gold standard: brain activity (EEG and EOG), heart rhythms (ECG), muscle activity (EMG), and respiratory signals. The full orchestra.

Then they looked at what actually happened to those people in the years after their night in the lab. Who developed dementia? Who had a heart attack? Who died?

From a single night of sleep, the model can rank who is more likely to develop certain conditions, years before diagnosis. The key metric is the C-index: a score of 0.85 means that if you picked two people (one who developed the disease and one who didn't) the model correctly identified the higher-risk person 85% of the time. For dementia, SleepFM hit 0.85. Death from any cause: 0.84. Heart attack: 0.81. Heart failure: 0.80. Stroke: 0.78.

We already knew sleep matters for health. What's new is how much predictive power is hiding in a single night of data, and that AI can learn to read it.

The Stanford team described their approach as "learning the language of sleep." They trained the model to find cross-system mismatches: a brain that looks asleep but a heart that stays alert, or other channels out of sync with each other. Much of the predictive power came from moments when different body systems seemed out of rhythm with each other — not failing individually, but falling out of sync. That's one place the predictive signal seems to show up.

Different signals predict different diseases. Brain activity flags neurological conditions. Heart signals flag cardiovascular disease. But all the signals together produced the best predictions across the board.

Here's what strikes me most: we've been collecting this data for decades. Sleep studies sitting in clinic databases. The information was always there. We just couldn't find the full meaning in it. Now we are layering intelligence on top of data we've had all along, finding patterns no human could see across hundreds of thousands of nights. And… these models are the least sophisticated they'll ever be...

Most sleep tracking today is retrospective. You wake up, check your app, see what happened. Maybe you slept poorly. Now what?

The current phase is making it actionable in the moment. What can you do today based on last night's data? That's where most consumer wearables are trying to go.

But this study points to something bigger: using sleep data to predict future health outcomes, years in advance. The question is whether consumer devices can get there to democratize these insights, and what we do with those predictions once we have them.

The signal gap: Stanford used clinical-grade PSG: brain waves, eye movements, muscle activity, heart rhythms, respiratory signals. Your Oura or WHOOP captures heart rate, heart rate variability, estimated respiratory rate, temperature, blood oxygen, and movement among other things. Useful signals, but no brain activity. PSG captures the full orchestra. Your wearable only hears the drums and bass.

Companies like Berkeley-based Somnee are working to close this gap. Co-founded by my (incredibly inspiring) Sleep Data Science professor Matt Walker and led by former Fitbit CMO Tim Rosa, Somnee uses lab-grade EEG sensors to capture brain signals directly, not infer them from heart rate and motion. But Somnee isn't just capturing better data — it's also intervening. Using gentle neurostimulation, the headband helps users fall asleep faster and stay asleep longer.

Full disclosure: I own a Somnee product, I've noticed my sleep latency decrease while using it, and I'm genuinely bullish on this company and their strong team. I'll be writing more about the emerging category of "electroceuticals" in a future post. The question isn't whether we'll capture richer signals at home. It's how fast.

The action gap: Even if you could predict disease risk from sleep, what would you do with that information? A scary graph doesn't help you sleep better. Prediction without action is just anxiety.

Eight Sleep is the clearest example of a company working on closing this gap. Another disclosure: I own an Eight Sleep, love the product, and am inspired by the company's mission.

They already have the action layer — the Pod adjusts temperature and elevation automatically while you sleep, responding to what your body is doing. You don't have to obsess over the data, you don’t have have to know what to do with it, and you don’t have to do the hard work of habit change on your own. The ambient intelligence of the autopilot system just... makes your sleep better.

Now they're building the prediction layer. They just raised $100 million to pursue building "digital twins" of their users: simulating thousands of scenarios per user, modeling health outcomes, and using that to intervene more intelligently. They have over a billion hours of sleep data to train on.

That's the unlock. Once you have both prediction and action, you're not just reacting to what happened last night. You're proactively optimizing sleep to set someone up for a healthier long-term path. Not showing them a scary graph about disease risk. Actually doing something about it, automatically, every night.

This is a very similar philosophy to what I'm aiming to build with Remi: using data to proactively adapt the environment around you, to support your sleep and circadian rhythm, rather than just giving you more information to worry about.

And even the best closed-loop sleep systems only address one risk factor. Sleep matters, but so do exercise, nutrition, stress, and genetics. The real promise of prediction isn't just "your mattress will help you sleep better." It's that early warning could prompt a whole cascade of interventions: lifestyle changes, screenings, maybe even preventive treatments. Sleep optimization is one piece of the puzzle. Prediction is what unlocks the rest.

If predictive sleep models go consumer without the action layer, if your watch starts estimating disease risk but can't do anything about it, that's a recipe for anxiety. Orthosomnia is already real: tracking that makes sleep worse, not better. Add disease predictions and false positives to the mix, and you've got healthy people obsessing and spiraling.

The answer isn't to stop building. It's to build thoughtfully. Systems that give people agency, not anxiety. That translate prediction into action, not just alerts.

Stanford showed us what's possible. One night of sleep, decoded properly, contains real signal about your future health.

Consumer hardware isn't quite there yet. But two gaps are closing: the signal gap (companies like Somnee bringing clinical-grade brain data home) and the action gap (companies like Eight Sleep building prediction plus automatic intervention).

And I can't help but feel a sense of wonder about where this goes. What else will we learn to read from sleep data as models get smarter and datasets get bigger? If we can predict disease years in advance, what happens when that prediction feeds into personalized prevention protocols, earlier screenings, or interventions we haven't even imagined yet?

The future of sleep tech isn't better dashboards. It's closed-loop systems that predict what your sleep reveals about your health and do something about it, quietly, automatically, every night.

Our sleep is talking to us. We're finally learning to listen. And we need to start dreaming bigger about how to respond.