Every morning your watch hands you a number. Sleep score 82. Or 61. Or 91. And every morning a lot of us look at that number and think: that is not how I feel.
That gap is the point of a new paper out of Tel Aviv University. A team led by Omer Hausner, Eva Mishor, Maya Reiner and Erez Shmueli published a study in Communications Medicine on November 10 that tries to predict how well you felt you slept, not just what your body did overnight. The result is a model that can separate the nights you would call terrible from the nights you would call great with an area under the curve of 0.88.
The research didn't need a new sensor. It needed a different question.
What they actually did
The data came from the PerMed study, a long-running observational cohort in Israel. The numbers are big enough that they're hard to argue with:
- 4,794 participants
- 759,116 self-reported daily sleep questionnaires
- 1,765,264 days of smartwatch recordings
- Three years of collection
Each night, people rated their sleep. The model had access to the watch's own view of the night (sleep data, physical activity, heart rate, and stress inferred from heart rate variability), plus a set of individual characteristics: age and other sociodemographic details, personality traits, and health attributes.
Then it tried to guess the rating before seeing it.
The headline result is that it works best at the extremes. Telling a "very poor" night from a "very good" one hit that 0.88 figure. That's a genuinely strong number for a prediction built on wrist data and a questionnaire filled out once at setup.
The second finding matters just as much for anyone shipping a sleep score: watch data alone and personal characteristics alone each did a limited job, and the two together did considerably better. The authors are explicit that combining them beat either input on its own.
Why your current sleep score misses
Most sleep scores on the market do something subtly different from this. They judge the night itself, using population-level norms: how much deep sleep you got, how fragmented it was, how long you were in bed. Then they compare you against everyone else and print a number.
That approach has a structural problem. If 85 is your normal and you sleep the same way every night, an 85 tells you nothing. If you normally score 92 and drop to 78, the drop is the story, and a population norm buries it.
The paper's framework points at a fix: score the night against that person's own recent history, and fold in the day before. The authors frame it as a conceptual shift in how subjective sleep quality gets measured rather than as a better sensor.
There's a product read here that any of the big three could act on. Garmin already leans hard on sleep and recovery data for Training Readiness, Body Battery and Training Status, all metrics owners have no independent way to check. Apple and Samsung both push a nightly score with a lot of confidence behind it. A model like this would let a watch say "you slept about like you usually do, and you'll probably feel fine" instead of "82".
That is a much easier claim to believe, and belief is the currency. A score that matches your morning experience earns trust in every other number on the watch.
What this is not
Worth being clear about the limits, because the paper is careful about them too.
- It predicts a rating, not health. The target is how rested you felt, which is a real and useful signal, but it isn't a clinical measure of sleep quality. The authors position subjective and objective sleep quality as complementary dimensions, not substitutes.
- It's a model, not a feature. Nothing here is shipping on a watch. It's a research framework built on a specific cohort.
- The cohort is one population. Nearly 4,800 volunteers in one country over three years is substantial, but generalizing to a global wearable user base is a separate question.
- The paper is in early-access form. It's peer reviewed, citable and carries a permanent DOI, and the publisher notes that this version may still receive editorial revisions.
None of that weakens the central result, which is that the information needed to estimate how rested you feel is already being collected by the watch on your wrist. It just isn't being used that way.
The part that should annoy you
Here's the uncomfortable detail for the industry. If a model can estimate how you felt about a night's sleep using signals every mainstream watch already records, then the remaining barrier isn't hardware. It's product design.
Adding a handful of setup questions (how would you say you usually sleep, what's your baseline) and scoring each night relative to your own trailing history is not a sensor problem. It's a software decision about what a sleep score is for.
Some of this is already moving at the edges, with recovery features increasingly framed as "relative to your normal" rather than against one curve for everyone. This paper gives that direction a much stronger evidentiary floor: three years, 759,000 ratings, and a machine learning result that got better, not worse, when it stopped treating people as interchangeable.
If you're in the market for a watch that takes sleep and recovery seriously in the meantime, the health-focused end of the Garmin lineup is the most developed, and the Venu 4 is the sensible starting point. If you want to compare how the two biggest platforms handle a nightly score, our Garmin sleep score breakdown and the Galaxy Watch 9 vs Apple Watch Series 11 comparison cover the mechanics. For a look at how watch sensor data is being pushed into clinical prediction more broadly, the Nature hypertension study is the closest parallel. And if a sleep score alone isn't enough to justify an upgrade, the Forerunner 265 is the cheapest way into the same recovery metrics.
Bottom Line
The interesting thing about this study isn't the 0.88. It's that the fix for "my sleep score doesn't match how I feel" turns out to be personalization rather than more hardware.
Every major watch brand already has the raw data. What's missing is the willingness to score you against your own normal, and to admit that the night is only half the story once you factor in the day that came before it. A watch that gets that right would be a lot more useful than one that just prints a bigger number.
Sources: Hausner et al., Communications Medicine, November 2026, PerMed cohort study, Communications Medicine, The Guardian on smartwatch stress and sleep self-reports.
