BlogInsight
How do you know a health score is right?
Lessons from validating health scores when there is no gold standard
aConvergence with several anchors
Health score
(any score under evaluation)
Survival
mortality risk beyond age and sex
↕ adds beyond age and sex?
Daily life
function, self-rated health
↕ reflects daily function?
Molecular markers
of biological aging
↕ reflects biological aging?
Clinical judgment
of practicing clinicians
↕ agrees with clinicians?
No single anchor settles it: do the anchors point the same way?
bBehavior as a measurement
Repeatable
similar result when a person is measured again
Robust
remains reliable when a test result is missing
Consistent
performs equally well across groups of people
Health scores are everywhere. The health apps from Apple, Google and Samsung turn sleep, heart rate and activity into scores: sleep scores, daily readiness and energy scores, cardio fitness levels. Biological-age algorithms such as PhenoAge estimate a person’s biological age from routine blood tests, and DNA methylation tests that offer the same are sold directly to consumers. Large language models can now be asked to rate a person’s health from their medical records.
Medicine’s established risk scores have a clear target. SCORE2, used across Europe, estimates a person’s chance of a heart attack or stroke over the next ten years (SCORE2 working group, European Heart Journal, 2021), and it can be checked against what actually happens. A score that claims to measure health itself has no such target. Health has no blood test of its own, and there is no gold-standard measurement of a person’s true health against which a score could be checked.
So how do you know a health score is right? For biomarkers of aging, one family of such measures, researchers have noted that there is no consensus on how they should be validated before they are used in the clinic (Moqri et al., Nature Medicine, 2024). We met this question while building Double Check, which combines routine clinical data into a single health score and won the ACUHIT 2026 hackathon. What follows is what that work taught us, and the questions we think every health score should face.
Why the usual check is not enough
A common way to validate a health score is to treat it like a risk score and ask whether it predicts death. That is a reasonable start. On its own, it has three limitations.
Age can do most of the work. Chronological age predicts death remarkably well by itself. A score that mostly re-expresses age can look excellent at predicting death while adding little to what age already tells us.
Predicting is not measuring. A score can predict death for reasons that have little to do with a person’s current health. Strong prediction alone does not show that a score measures health.
Measurement quality is a separate question. A good death prediction does not show whether a score gives the same answer twice, is robust to a missing test result, or works equally well across groups of people. These properties are checked far less often.
Results are also hard to compare across studies. Scores are tested on different populations, against different outcomes, with different definitions of success.
What building a score taught us
These limits are easy to state and hard to respect. Three lessons stood out.
Normal is personal. Reference ranges describe populations; what often matters more is how far a person has moved from their own usual value. A creatinine level that rises from 0.6 to 1.0 mg/dL stays inside the usual range, yet for that person it is a large change. Our score uses ranges that follow sex and age and penalizes a move into a worse range since the previous visit; reading every result against the person’s own baseline is the next step.
Every point should be traceable. A clinician cannot check a number they cannot take apart. Every point a person loses can be traced to the value or rule that caused it, and the reference range shown on screen is the one used in the calculation. Keeping that true as the score changes is harder than it sounds, so an automated test checks the explanation against the score itself.
Do not chase the headline number. When few people in a study experience the outcome, two versions of a score can differ by an amount small enough to arise by chance. Many of the changes we tested moved the headline number by less than that. So we show the uncertainty next to every comparison and decide on clinical grounds. Some changes we kept made the headline number slightly worse, because they made the score clinically sounder.
What to look for
These lessons lead to six principles, in three groups.
Compare against several kinds of evidence
- Many anchors, not one. No single outcome can stand in for health. Compare a score with several anchors, signals that any real measure of health should track: survival, how people feel and function day to day, molecular markers of aging, and the judgment of practicing clinicians. Then ask whether the picture holds together.
Look past age, and test the score as a measurement
- Beyond age, beyond prediction. Ask what a score adds beyond age and sex. Survival is one anchor among several, and predicting it well is not, on its own, evidence that a score measures health.
- A score must behave like a measurement. Check whether it gives the same answer when a person is measured again, whether it is robust to a missing test result, and whether it works equally well across groups of people.
Set the rules in advance, and let your own score fail
- Rules first. Write down every hypothesis, threshold and decision rule before the main analysis is run, and register them with a timestamp. Report any later change as a deviation.
- Test the test. Run your validation method on deliberately meaningless control scores, which it should reject, and on a widely studied score with well-documented links to health, which it should not. If it cannot tell them apart, it cannot tell you much about a new score either.
- Same rules for everyone, including yourself. Judge your own score by the same rules as the established ones, and set up the checks so that it can fail.
What we will not claim, and what comes next
We will not claim that any score is “true,” and no validation can show that. What it can show is how much evidence supports a score and where that evidence runs out. None of these scores, ours included, is a diagnostic test.
Next, we are applying these principles ourselves, and we will share what we find, including where we fall short.
There may never be a single number that proves a health score is right. But there can be an open, repeatable way to check. That is how we want every health score to be checked, starting with our own.