Deep divefrom Episode 7Longevity20 October 202623 min

Ninety-Four Million Check-Ups and the Uneven Pace of Ageing

A deep dive into one paper: what was known before, what the researchers did, what the numbers say and what they don't.

The paper: A longitudinal Health Atlas of age-associated multisystem physiology from population health examinations in China (Zhang et al., npj Aging, 2026)

MP3

Chapters

Also linked

Supplementary information (PDF)

Transcript

The script the two AI voices read from. The times are worked out from the script, not heard, so a line can start a few seconds early or late.

Read the whole episode

Cold open

Host 1You heard the trailer — this is the whole story. Welcome to the Hastrup Review of Science — where artificial intelligence reads real science, so you don't have to pretend you understood the paper. Today, one hundred and eighty-nine thousand people in eastern China, ninety-four million routine check-up results, and an attempt to map how different body systems drift with age.

Host 2The paper is called A longitudinal Health Atlas of age-associated multisystem physiology from population health examinations in China. It has eleven authors, with Shilong Zhang as first author and Cong Feng and Ming Chen as corresponding authors.

Host 1Where are they, and where was it published?

Host 2Mostly the Department of Bioinformatics at Zhejiang University in Hangzhou, plus the Zhejiang Provincial Center for Disease Control and Prevention and the geriatrics department of a Zhejiang University hospital. It is peer reviewed and appeared in npj Aging, a Nature Portfolio journal, on the twenty-first of September twenty twenty-six, as an early accepted version. Funding is Chinese national and provincial grants, and the authors declare no competing interests.

Host 1And what did we actually read?

Host 2One flag up front: the main text was not available to us, only the abstract and the thirty-eight-page supplementary file, which holds the tables and figure legends. The numbers here come from those, the reasoning in the main text we could not check, and where I add textbook background, I will say so.

Host 1The claim, in two sentences.

Host 2Ordinary check-up data, organised into ten body-system domains, show that systems drift from their young-adult values at different ages, and differently in women and men. A summary of how far someone's systems sit from typical for their age is linked to dying sooner, but modestly, and a prediction model built on these data forecasts seven common lab abnormalities about as well as standard off-the-shelf methods.

Why this field exists

Host 1Start with the basics. Why would anyone want to measure ageing, rather than just count birthdays?

Host 2Background, not from this paper. Two people of the same age can be in very different shape, and age alone predicts disease and death only coarsely. Since at least the nineteen-sixties, gerontologists have looked for so-called biomarkers of ageing: measurements that track how old the body behaves, rather than how old it is.

Host 1What did those early attempts look like?

Host 2Background again. The simplest version takes a handful of blood tests, blood pressure and lung function, fits a formula that predicts age, and calls the gap between predicted and real age your acceleration. A widely used refinement from two thousand six, the Klemera–Doubal method, does this more carefully, and this paper compares against an adapted version of it.

Host 1And more recently?

Host 2Background. From about two thousand thirteen, epigenetic clocks used chemical marks on DNA to estimate age closely, and later versions were trained to predict death rather than age. Another line of work, called homeostatic dysregulation, simply asks how statistically unusual a person's whole profile of lab values is, compared with healthy younger adults.

What was known before

Host 1So what was already known about different organs ageing differently?

Host 2Background. A small Stanford study around twenty twenty followed about a hundred people intensively and proposed ageotypes: some people seemed to age mainly along a metabolic axis, others along immune, liver or kidney axes. Since then, studies using blood proteins have built organ-specific clocks, and a twenty twenty-five paper built a methylation test scoring eleven physiological systems separately.

Host 1What is the weakness in those approaches?

Host 2Background. They tend to use either expensive measurements, like proteomics or methylation, in thousands of people, or a few cheap lab values in large surveys. Most are snapshots, or have only a few repeat visits, so they say little about how each person's systems change from year to year.

Host 1Where does a routine check-up come in?

Host 2Background, hedged. In China, regular health examinations are common, through employers and through a public programme that offers annual check-ups to people aged sixty-five and over, so health centres sit on years of repeated measurements for very large numbers of people. The authors' opening line is that these examinations capture longitudinal physiology but are seldom used to characterise ageing.

The open question

Host 1So what is the open question?

Host 2Whether messy, real-world check-up records, never designed for research, can be turned into a coherent map of how many body systems change with age. And if so, whether that map tells you anything about who dies, or who develops abnormalities, beyond what the plain lab values already tell a doctor.

Host 1That second part sounds like the hard one.

Host 2It is, and it is the part to keep in mind through the results. A new score is only useful if it adds something to the raw numbers it was built from.

What the paper did

Host 1Walk me through the data. Where do the ninety-four million numbers come from?

Host 2Health examination records from Zhejiang, a coastal province just south of Shanghai. The source covered three hundred and forty-five thousand five hundred and thirty-nine people with a valid exam; after cleaning, one hundred and eighty-nine thousand and ninety-five remained in the final atlas, contributing ninety-four million, fifty-five thousand, three hundred and twenty-six observations.

Host 1Who are these people?

Host 2Older than you might expect: in the final atlas the median age at first examination is sixty-two, about fifty-two percent are women, and the median first exam year is twenty nineteen. The median person has six visits and three years of follow-up, and about one hundred and fifty-seven thousand had more than one visit.

Host 1And how do you get from a check-up form to ninety-nine biomarkers?

Host 2The records contained eight hundred and sixty differently named items, which they standardised down to four hundred and two, and then selected ninety-nine for the atlas. Fifty-four are continuous measurements, like blood counts or creatinine; forty-five are qualitative findings, like a urine dipstick result, an ECG reading or an imaging finding, coded as normal or abnormal.

Host 1What are the ten domains?

Host 2Cardiovascular and vital signs with fifteen markers, haematologic-immune, essentially the blood count, with twenty-five, urinalysis with seventeen, physical examination with fifteen, and imaging and metabolic-adiposity with seven each. Then renal-purine, meaning kidney function and uric acid, with four, and three each for liver, ECG and a functional-sensory domain.

Host 1Who decided those groupings?

Host 2The authors, in advance, based on clinical meaning rather than on how the data happen to cluster. That is a reasonable choice for interpretability, and they checked it against a purely data-driven ten-cluster grouping, which we will come back to.

Host 1Then what is a trajectory, concretely?

Host 2For every biomarker, each sex and every single year of age from thirty to eighty-nine, they took the median value, then smoothed it with a five-year rolling average. Each age-and-sex cell needed at least twenty records, and the resulting curve shows how the typical value moves with age.

Host 1Typical in which population? Healthy people?

Host 2No, and this matters: the reference is the observed check-up population itself, including people with disease. The authors call these curves empirical age patterns, not disease-free norms, though a healthy-only reference gave similar curves, with a median correlation of zero point eight eight one.

Host 1How do they turn that into a score for one person?

Host 2For each measurement they compute how far the person sits from the typical value for their age and sex, scaled by the spread at that age, which gives a kind of z-score, clipped at five either way. Biomarkers are oriented so that the direction they move with age counts as positive, and each domain score is the median of its biomarkers' scores at that visit.

Host 1And the overall numbers?

Host 2Two main ones. Mean absolute domain burden averages how far, in either direction, a person's domains sit from typical. Signed domain dispersion is the spread between domains, meant to capture systems ageing out of step with each other.

Host 1So dispersion is the idea that your kidneys might be old while your blood is young.

Host 2That is the intuition behind it. Whether it measures that, or something more mundane, turns out to be one of the more instructive parts of the paper.

Host 1What about the abstract's thresholds crossed at sex-specific ages?

Host 2For each domain and sex they compare against thirty-to-thirty-nine-year-olds and ask in which decade the average deviation first exceeds a preset threshold, mainly zero point three standard deviations. They call these detectable clinical phenotype shift windows, and they stress that they are not biological onset ages.

Host 1Then the mortality analysis.

Host 2They linked participants to death records up to the end of twenty twenty-three. The main analysis has one hundred and forty-seven thousand, four hundred and ninety-two people and four thousand five hundred and ninety-nine deaths, over a median follow-up of about three and two-thirds years.

Host 1And the statistics?

Host 2Cox regression, the standard survival model, which estimates hazard ratios: how much the risk of dying at any given moment rises per unit of the score. The models were stratified by sex and adjusted for age, age squared, examination year, and recorded current and chronic disease.

Host 1And the prediction model, DynLifeNet.

Host 2A neural network that reads a person's sequence of raw check-up visits, plus a summary of latest values and trends, and predicts seven abnormalities one to three years later. They are kidney decline, high glucose, high blood pressure, anaemia, raised liver enzymes, abnormal blood lipids and high uric acid, each predicted only in people who did not already have it.

Host 1Where does the atlas come into the network?

Host 2As a correction on top. The network first learns a prediction from the raw data alone, called the raw anchor, then freezes it and trains a second branch on atlas features that can only nudge the answer by a bounded amount, starting from zero.

Host 1That is a cautious design.

Host 2It is, and it means the full model should land close to its raw anchor. They tested in two ways: on held-out people, and with a time split, training on twenty sixteen to twenty twenty, tuning on twenty twenty-one, and testing on twenty twenty-two.

Host 1And the outside checks?

Host 2Two of them. They carried the framework over to the China Health and Retirement Longitudinal Study, a national survey of middle-aged and older adults, and they compared some atlas axes with public molecular data: a single-cell blood atlas of six hundred and twenty-five donors, and gene expression and DNA methylation from the Genotype-Tissue Expression project.

A word from our sponsor

Host 1Before the results, a word from our sponsor. Do you go between check-ups looking slightly less fresh with every visit? The Hastrup Review of Science is brought to you by dry shampoo.

Host 2Your hair is not one of the ninety-nine biomarkers.

Host 1Dry shampoo reduces visible deviation from your own baseline, without a full wash. A few seconds of spray, and you are back near typical for your age.

Host 2Has that been validated in an independent cohort?

Host 1Not by Cox regression. Dry shampoo is available now, and our referral link is in the show notes.

Results

Host 1Back to the paper. What do the trajectories look like?

Host 2Heterogeneous is the authors' word, and it fits. Different biomarkers and domains rise, fall or stay flat with age along different shapes, and many differ between women and men.

Host 1Give me the windows.

Host 2At the zero point three threshold, women's metabolic-adiposity and blood-count domains first cross in their forties, while men's domains mostly cross in their fifties. Kidney, cardiovascular, liver and ECG domains cross in the fifties in both sexes, and the physical-examination domain never crosses in either.

Host 1How solid are those decades?

Host 2Less solid than a headline would suggest. Across the twenty domain-and-sex combinations, the point window stayed the same across thresholds of zero point two, zero point three and zero point five in only nine, and a more conservative estimate often pushes the window into the seventies or out of range altogether.

Host 1Any reason women would shift earlier on metabolism?

Host 2Background, and speculation that is mine, not the paper's: menopause typically falls around fifty and is known to change lipids and body fat, so a shift visible by the late forties is plausible. The supplement only reports the windows, so I cannot tell whether the main text tests that.

Host 1Now the mortality headline.

Host 2Each one standard deviation of higher mean absolute domain burden goes with a hazard ratio of one point zero five eight, with a ninety-five percent confidence interval of one point zero four five to one point zero seven one. In plain terms, about six percent higher risk of death at any point in follow-up, per standard deviation, after adjusting for age and recorded disease.

Host 1Is six percent a lot?

Host 2It is modest. As rough background, adult mortality rises by around nine or ten percent per year of age, so one standard deviation of burden is worth less than one extra birthday, in a model that already accounts for age.

Host 1Is it stable over time?

Host 2Not really. The model's proportional-hazards assumption fails, so they split the follow-up: in the first two years the hazard ratio is one point two zero five, and after two years it is one point zero three zero.

Host 1Meaning?

Host 2Most of the signal is short-term, which fits a score that partly picks up people who are already ill at the check-up. Dropping deaths in the first year only brings the estimate down to one point zero four seven, so it does not vanish.

Host 1Which causes of death?

Host 2Ischaemic heart disease shows the largest, at one point zero six eight, then cancers at one point zero four seven and stroke at one point zero four one, all surviving correction for multiple testing. Respiratory deaths were not clearly associated.

Host 1And individual domains?

Host 2The blood-count domain stands out, with a hazard ratio of one point two seven five per standard deviation, and the kidney and uric acid domain is next, at one point one one four. Oddly, the liver and imaging domains go the other way, with hazard ratios slightly below one, which the supplement does not explain and which I would not over-read.

Host 1Now the dispersion idea. Do systems ageing out of step predict death?

Host 2At first glance, yes: raw dispersion has a hazard ratio of one point zero three eight. But the authors then show that ninety-five point eight percent of the variation in raw dispersion is explained by how many domains a person had measured, and by their overall burden.

Host 1So it is mostly a measure of how much was measured.

Host 2Largely, plus how far off you are overall. Once that is removed, the hazard ratio is zero point nine nine five, essentially no association, and the authors say plainly that their results do not support dispersion as an independently harmful asynchronous ageing process.

Host 1That is unusually candid.

Host 2It is, and it deserves credit. Dispersion also barely correlates with the adapted Klemera–Doubal measure or with homeostatic dysregulation, and swapping in the data-driven domain grouping leaves the dispersion scores with a correlation of only about zero point one to the original.

Host 1The abstract also mentions instability over time.

Host 2A different idea: how erratically a person's domain profile changes between visits, measured with a one-year lag in held-out participants. The abstract links it to mortality but labels it exploratory, the supplement shows it only as a figure without numbers I can read, and the authors call it a prognostic marker, not evidence of cause.

Host 1What about systems moving together, as a network?

Host 2Weak. The median correlation between domains after adjustment is zero point zero two nine, and of two hundred and twenty-five domain-pair-and-age-band combinations, only nineteen gave stable links above zero point one.

Host 1The independent cohort. What held up?

Host 2In the China Health and Retirement Longitudinal Study, only four of the ten domains could be matched with overlapping measurements. The ranking of those domains was preserved, but the scales differed, so Zhejiang's reference values do not transfer directly to another population.

Host 1And the prediction model?

Host 2On held-out people, DynLifeNet's average area under the curve across the seven endpoints is zero point eight zero two, where zero point five is a coin toss and one is perfect. Age and sex alone manage zero point six zero seven.

Host 1Compared with simpler methods?

Host 2This is the key table. A logistic model using each person's latest values reaches zero point seven eight six, adding the trend gives zero point seven nine one, a standard gradient-boosting model on the raw data gives zero point eight zero zero, and XGBoost, another common off-the-shelf method, gives zero point eight zero five.

Host 1So XGBoost scored higher.

Host 2Numerically, by zero point zero zero three, which is not statistically distinguishable. The atlas branch improves on the network's own raw anchor by zero point zero zero two four, reliable but small, and the authors themselves write that the results are not evidence that the atlas branch alone outperforms raw gradient boosting.

Host 1Which endpoints are easy, and which are hard?

Host 2A new glucose abnormality is the easiest, around zero point nine four, which makes sense, since earlier glucose values are strongly predictive. Raised liver enzymes and abnormal lipids are hardest, at around zero point seven four to zero point seven six.

Host 1And the time-split test?

Host 2Predicting twenty twenty-two from earlier years, across six endpoints, DynLifeNet scores zero point seven seven five, essentially tied with XGBoost at zero point seven seven four. Every model drops when moving forward in time, which is the more realistic test.

Host 1And the molecular data?

Host 2For the blood, metabolic and kidney axes, the direction of check-up age effects lines up with age signals in the external single-cell, gene expression and methylation data. The authors frame this explicitly as concordance, not mechanism or causation.

Why it matters

Host 1Step back. What is the principle here?

Host 2That the routine numbers health systems already collect can be organised into a reproducible description of how body systems change with age, at a scale no dedicated study could afford. The atlas is mostly a descriptive resource, closer to a reference growth chart than to a new test.

Host 1What changes, and for whom?

Host 2For ageing researchers, it is a template and a set of benchmark curves, plus a worked warning about composite scores that turn out to track measurement coverage. For health services in China, it hints that check-up data could flag drifting systems, though nothing here shows that acting on such a flag helps anyone.

Host 1And what does it not mean?

Host 2It does not mean your organs have ages you can read off a check-up, and the windows are not when ageing starts. It does not show a new AI model beating standard tools, and it does not show that systems ageing out of sync is harmful; the paper's own analysis argues against that.

Host 1For a listener who gets an annual check-up?

Host 2Nothing changes yet. In this study, the individual lab values your doctor already looks at carry most of the predictive information.

Caveats

Host 1Now the caveats. Start with who is in the data.

Host 2One province, people who chose or were offered repeated check-ups, with a median age of sixty-two. The authors' own audit found repeat visitors almost perfectly distinguishable from one-time visitors, with an area under the curve of zero point nine nine three, so the longitudinal sample is a selected group.

Host 1Follow-up?

Host 2Under four years for most people, so the mortality results are mostly short-term, and the two-year split shows the score works best at spotting people who are already unwell. About four thousand six hundred deaths is enough for overall estimates, but thinner for single causes.

Host 1The reference curves?

Host 2They come from the same examined population, sick people included, and the domain-specific mortality analyses each cover only the subset with complete data, around fifty-two to fifty-five thousand people for the blood, kidney, metabolic and liver domains. Uneven measurement coverage across people is exactly what undermined the dispersion score, and it may affect other scores less visibly.

Host 1And the domains themselves?

Host 2The authors find their clinical groupings stable for most trajectories, but they note that exact global scores and dispersion effects depend on how the biomarkers are partitioned. And this is observational, so none of these associations show cause.

Host 1Anything about our own reading?

Host 2Yes: we had the abstract and supplement, not the main text, so we could not check the methods prose, the main figures, or how the authors discuss these limits. The supplement is unusually thorough with sensitivity analyses, which is reassuring.

What to watch next

Host 1What should we watch for?

Host 2Whether the full Zhejiang reference curves and code are released, so other regions can compare their own check-up data. And a proper external test with more than four matched domains, ideally outside China.

Host 1And on the prediction side?

Host 2Whether any atlas-based flag, used in a real clinic, changes outcomes, which would need a trial rather than a better area under the curve. And longer follow-up, to see whether the burden signal holds beyond the first couple of years.

Sign-off

Host 1That is the deep dive: A longitudinal Health Atlas of age-associated multisystem physiology from population health examinations in China, by Shilong Zhang, Cong Feng, Ming Chen and colleagues, in npj Aging. The link is in the show notes.

Host 2Short version: ninety-four million check-up results from one hundred and eighty-nine thousand people show body systems drifting at different, sex-specific ages, and a higher overall burden goes with a modestly higher death risk, mostly in the short term. The out-of-sync ageing idea did not survive adjustment, and the AI model ties with standard tools; we read the abstract and supplement, not the main text.

Host 1Every Friday this same feed carries the roundup episode: five new papers as trailers, and you pick which become deep dives. This has been the Hastrup Review of Science.