A reference page on how a biological age is computed, why two readings of the same blood can land years apart, what the clocks do well at the level they were built for, and which measurements come with an action attached.
At the population level, epigenetic clocks are accurate enough to predict who dies sooner. For one person they are not yet accurate enough to act on. The reason is measurable: one blood sample assayed twice and scored by a conventional clock has returned ages up to nine years apart (Higgins-Chen et al., 2022), while the biological spread the clocks are meant to detect is only three to five years wide. That is the short answer. The rest of this page shows the arithmetic behind it, gives the clocks the credit they have earned, and ends with the numbers I would take instead.
This page stands outside the three series. It picks up where the wearables page left off, which sorted the numbers on a wrist into the few that are measured and the many that are modelled, and it applies the lactate page's rule: a figure earns its place only if some result would alter what you do next. The Substack essay makes the same case shorter and more personally; here every study carries its numbers and its limits, the caveats get a section of their own, and the sources sit at the end. Nothing is argued that the essay doesn't argue.
Key points
A biological age is the output of a regression trained on thousands of other people. The weights are population averages; the number is yours only in the sense that your sample went in.
Technical noise alone moved six conventional clocks by up to 3 to 9 years between replicates of the same sample, against a biological spread of 3 to 5 years (Higgins-Chen et al., 2022).
Even clocks that agree with themselves across replicates drift with meals, stress and surroundings when the same person is sampled again days apart, and the two kinds of reliability are unrelated, r = 0.02 (Sehgal et al., 2026).
In a meta-analysis of 13 cohorts (13,089 participants), each tested measure of epigenetic age acceleration predicted death independently of age, roughly 4 percent more hazard per year of acceleration (Chen et al., 2016).
The field's own consortium states that no surrogate biomarker of ageing has been validated yet; blood pressure is their example of one that has (Moqri et al., 2023).
Fitness, grip strength and sleep regularity each predict mortality in cohorts of 60,977 to 139,691 people, can be repeated on the same person, and point at an action.
Before paying for a test, write down what you would do at each possible result. If the columns match, skip the test. Measure the action instead.
This is an educational and strategic perspective, not personal medical advice.
The views are the author's own and not statements by Atlas Cove Lda.
Where the clocks are built
Every epigenetic clock starts from the same raw material. Along your DNA sit millions of positions called CpG sites, and at each one some share of your cells carries a small chemical tag, a methyl group. That share drifts with age, rising at some sites and falling at others. A laboratory array reads the tag level at several hundred thousand sites and reports each as a value between zero and one, and from there the work is statistics, and the statistics are published, which is the only reason the question in the title can be answered at all rather than guessed at. The recipe is open.
The first multi-tissue clock came from Steve Horvath in 2013. He assembled roughly 8,000 samples from 82 datasets covering 51 tissues and cell types, handed a penalised regression 21,369 candidate sites that appeared on two generations of array, and let the algorithm select. It settled on 353. Their weighted sum predicted chronological age in held-out data with a correlation of 0.96 and a median miss of 3.6 years (Horvath, 2013). Note the target. The first generation was trained to reproduce the date on your passport, so a flawless version would tell you nothing new. All the interest lies in the residual, the distance between predicted and true age, which the field calls age acceleration.
Levine and colleagues moved the target in 2018. Step one used no methylation at all. Using 9,926 adults from a US national survey followed for roughly two decades, a penalised survival model reduced 42 candidate blood measures to nine (albumin, creatinine, glucose, C-reactive protein, lymphocyte percentage, mean cell volume, red cell distribution width, alkaline phosphatase and white cell count) plus chronological age, and expressed the resulting mortality risk in years. I read that table twice before I believed how ordinary the nine are. Step two took that phenotypic age into a second cohort and regressed it on methylation, 20,169 candidate sites going in and 513 coming out of the elastic net (Levine et al., 2018). DNAm PhenoAge therefore forecasts a forecast: a methylation estimate of a blood-chemistry estimate of time to death, two regressions deep before the number reaches the page.
GrimAge went further. Rather than predicting age or a composite directly, it first built methylation-based stand-ins for seven plasma proteins and for smoking pack-years, then combined those surrogates with age and sex into a predictor of lifespan (Lu et al., 2019). The pace-of-ageing clocks took a different route again. DunedinPACE was trained on how quickly 19 organ-system markers declined across four measurement waves and two decades inside a single New Zealand birth cohort. That slope was then compressed into a single-timepoint blood test reading 173 sites, chosen only from probes that had already passed a test-retest reliability threshold (Belsky et al., 2022). Four clocks, four targets.
What unites these is the origin of the weights. All of it is borrowed. Each site's coefficient summarises how that site tracked the target, on average, across the thousands of people in the training data, so your score reports how people who shared your tag pattern tended to age, or to die, in those cohorts, and nothing about you that those cohorts didn't already contain. It is a legitimate quantity. It is a useful one. It is also, by construction, a population statistic with your name printed above it.
Four clocks, what each was trained to predict
| Clock | Trained to predict | Sites it reads |
|---|---|---|
| Horvath multi-tissue, 2013 | Chronological age across 51 tissues | 353 |
| DNAm PhenoAge, 2018 | A mortality score built from nine blood measures plus age | 513 |
| GrimAge, 2019 | Time to death, via surrogates for seven proteins and smoking | Seven surrogate scores plus age and sex |
| DunedinPACE, 2022 | The 20-year decline of 19 organ-system markers in one cohort | 173, reliability-filtered |
The decision sheet, before the noise
The order here is deliberate. The technical section that follows is persuasive in a way that can make the practical step feel unnecessary, so do the practical step first.
Take a sheet of paper and write three possible results across the top: a number about five years below your age, a number about five years above it, and a number close to it.
Under each heading, write what you would change in the coming week.
Add one line at the foot of the sheet: the result that would make you do less than you do now.
When the three columns carry the same entries (train, keep a bedtime, eat properly, book the appointment you keep postponing), the test has already been answered without being bought. Spend the money on the action, or on measuring it. When the columns differ, the two sections that follow examine whether a test can tell them apart, which turns out to be the harder question. And when nothing in the last line would ever make you ease off, the test can only push in one direction, which makes it an instrument of sales rather than of measurement, whoever is selling it.
Why a few drifting sites can move the whole number
The signal at any single clock site is small. Across Horvath's 353 sites, people over 55 differed from people under 35 by a mean absolute tag level of 0.032 on the zero-to-one scale (Horvath, 2013). Against that sit the ordinary sources of measurement error: how the sample was prepared, how many beads on the chip happened to read each site, the chemistry of the probe, and the batch the tube ran in (Higgins-Chen et al., 2022). A weighted sum hides none of it. The error goes straight through.
Higgins-Chen and colleagues quantified the consequence. They took a public dataset of 36 whole-blood samples, each measured twice, and pushed the pairs through six well-known clocks. Horvath's multi-tissue clock gave the two copies ages a median of 1.8 years apart, and 4.8 years in the worst pair. The other five clocks showed median gaps between 0.9 and 2.4 years and worst pairs between 4.5 and 8.6 (Higgins-Chen et al., 2022). Set those gaps against the standard deviation of age acceleration in a population, 3 to 5 years, and four of the six clocks could be moved by more than one standard deviation of the very quantity they exist to measure. By technical variation alone.
The most reliable of the six was GrimAge, with an intraclass correlation of 0.989. The paper traces that to its two-step design: a clock assembled from seven surrogate scores averages away noise that a clock built from single sites cannot (Higgins-Chen et al., 2022). The group's general fix works on the same principle. Instead of trusting individual sites, their principal-component clocks extract the shared ageing signal across many of them. The retrained versions agreed between replicates within 1 to 1.5 years in more than 90 percent of pairs, with intraclass correlations above 0.99 for epigenetic age and above 0.97 for acceleration. That fix was published in 2022, and whether a given commercial test applies it is something to ask before paying.
The problem the fix leaves untouched
A clock that reads the same sample identically twice can still read the same person differently on Tuesday and on Sunday, because the person is not the same sample on Tuesday and on Sunday, and no amount of laboratory precision repairs that. Sehgal and colleagues put 18 methylation-based ageing measures through two separate repeat designs. Technical replicates, the same extraction assayed again on EPIC and 450K arrays, agreed well for most clocks, although several shifted with slide position and with the DNA extraction protocol. Biological repeats, the same individuals sampled again within a short interval under changing conditions such as meals, stress and environmental exposures, agreed far less, and stability fell further once immune-cell composition was adjusted for (Sehgal et al., 2026).
For a buyer, one finding matters most. The two reliabilities are unrelated. How well a clock survived the technical repeat predicted nothing about how it survived the biological one: across the 18 measures the two reliabilities correlated at 0.02. Two widely used second-generation clocks landed in the worst corner, technically excellent and biologically unstable, while the principal-component clocks did comparatively better on both counts (Sehgal et al., 2026). The practical reading is plain. A report cannot tell you whether the sample it analysed was your blood after a quiet weekend or your blood after a bad week, and the two may not give the same age. The kit doesn't ask.
A third paper from the same laboratory asks what happens when a trial reports that an intervention moved one clock. Borrus and colleagues took previously published single-clock intervention results and re-ran them across additional clock models, and found that no further clock reached statistical significance. They propose three conditions before a result deserves trust: a reliable principal-component clock responds, a mortality-trained or pace-of-ageing clock responds, and several clocks move in the same direction (Borrus et al., 2026). Those were written for researchers, and they are a reasonable bar for anyone deciding whether a before-and-after pair of kits, bought six months apart and compared on a kitchen table, has shown them anything at all.
I have never bought one of these tests. So there is no reading of my own to set beside these figures, and the replicate pairs above are the closest thing on this page to a personal result, and they belong to other people.
What the clocks do well, and it is a great deal
None of the above makes the clocks bad science. At the level they were designed for, they perform. Chen and colleagues pooled 13 cohorts, 13,089 participants across three ethnic groups, and found that each acceleration measure they examined predicted all-cause mortality independently of chronological age, and still did so after adjusting for the standard risk factors (Chen et al., 2016). The size of the effect is modest per year and compounds: a hazard ratio of 1.040 for each year of extrinsic acceleration becomes 1.48, a 48 percent higher hazard, for someone running ten years ahead of the calendar, and falls to 0.82, an 18 percent lower hazard, for someone five years behind it.
Across thousands of people the technical wobble cancels out and a genuine gradient appears, and the research that gradient has made possible is among the most informative ageing biology of the past decade. The clocks have shown how behaviour, environment and adversity register in the epigenome, and how some interventions appear to alter that trajectory (Apsley et al., 2025). The group behind DunedinPACE built it precisely so that trials could measure a change in the pace of ageing within a person (Belsky et al., 2022). All of that is population science. All of it is sound.
The field itself has drawn the line between that use and personal use. The Biomarkers of Aging Consortium set out, in 2023, the stages a candidate biomarker must pass, from analytical validation through validation against outcomes in people to acceptance as a surrogate endpoint, and stated that no formally validated or likely surrogate endpoint biomarker of ageing exists yet, partly because the candidates are so new. Their example of a surrogate that has passed every stage is blood pressure: lowering it is accepted as a proxy for preventing strokes (Moqri et al., 2023). Keep that example in mind for the final section.
Apsley and colleagues then asked the individual-level question head on. Their answer: no, for the present. They name three barriers: a single reading is unreliable, a change between two readings cannot be interpreted in practice, and a score lacks the specificity to say what, if anything, is wrong. Underneath those sits a design point, which is that the clocks were engineered to capture correlations across a population, not accuracy for an individual, so they fall short of the standards an established laboratory test must meet, and treating one score as grounds for action means acting on a faint signal inside loud technical and biological noise (Apsley et al., 2025).
Which answers the question autocomplete offers as soon as someone types the first words. Is biological age real? Across populations the gradient is real, and so is the research built on it, and anyone dismissing the clocks as nonsense is arguing against the mortality meta-analyses. The report handed to one person is a forecast with an unprinted error bar, and that error bar is exactly what a cohort of thousands averages out and a single reader cannot.
Fitness: the number with no ceiling
Here is what I would measure in place of an age, starting with the one that carries the most evidence. Among 122,007 people at a single US centre who underwent symptom-limited treadmill testing, with 8.4 years of median follow-up, risk-adjusted mortality fell step by step with cardiorespiratory fitness. No upper limit of benefit appeared. Sitting below the age-and-sex average instead of above it meant a hazard ratio of 1.41, identical to the figure smoking produced in that model, and the gap between the least fit and the fittest was a hazard ratio of 5.04 (Mandsager et al., 2018). An earlier meta-analysis of 33 cohort studies with 102,980 healthy participants put a dose on it: a risk ratio of 0.87 for all-cause mortality per additional MET of capacity, roughly one kilometre per hour of running speed (Kodama et al., 2009).
Both are associations. Nobody was randomised to be fit, and the treadmill cohort was a referred population at one hospital, so neither proves cause. What distinguishes fitness from a clock is simpler than that. The repeat test is the same test, the instruction for raising it is not in doubt, and a second treadmill will show whether the instruction was followed.
Grip: ten seconds, one number
Grip strength does something no methylation array can, which is to stand in for whole-body strength and for how well the nervous system recruits muscle, in ten seconds and for the price of a dynamometer. In the PURE study, 139,691 adults across 17 countries were followed for a median of four years, and each 5 kilograms less of grip carried a hazard ratio of 1.16 for all-cause mortality, 1.17 for cardiovascular death and 1.17 for non-cardiovascular death. As a predictor of all-cause and cardiovascular death, grip outperformed systolic blood pressure (Leong et al., 2015). The strength-dose page covers what raises it and how little that takes.
One caveat belongs here rather than in the list below, because it bears on the day-six measurement later on this page, and I learned it with a spring dynamometer in September. A grip reading moves with hydration, sleep, caffeine, recent training and how familiar the person is with the device, and the first weeks of any strength gain come from the nervous system rather than from new muscle. A second grip reading taken days after the first mostly shows learning and rest, and that is true in our house as much as anywhere. That is still information, and worth having. New tissue it is not, and nobody should sell it as such.
Sleep regularity: the lever is a clock on the wall
The third measurement differs from the other two in what moves it. A bedtime, rather than a load. Windred and colleagues computed a sleep regularity index from more than ten million hours of wrist accelerometry in 60,977 UK Biobank participants, worn for seven days, and followed deaths for up to 7.8 years. The four more regular quintiles had 20 to 48 percent lower all-cause mortality than the least regular fifth. Cancer mortality was 16 to 39 percent lower, cardiometabolic mortality 22 to 57 percent lower, after adjustment for age, sex, ethnicity and a long list of lifestyle and health factors. Duration added nothing. In nested models, regularity predicted survival better than duration did (Windred et al., 2024). The sleep page explains why timing outranks hours.
A consistent bedtime is the entire intervention. It costs nothing, and a wrist device or a notebook can confirm within a month whether it happened.
What a second morning can and cannot show
At the September residency we recorded four things before breakfast on the first morning and again on the sixth: VO2 max, grip strength, resting blood pressure and sleep. No age estimate was taken. Those four were chosen because, among everything measurable on a healthy adult in a morning, they carry the best evidence for length and quality of life, and each is read against a published reference for the person's age or against that person's own first morning, never against another guest.
The same scepticism applied to the clocks applies to our second morning. Blood pressure and sleep timing can shift within a week, and those are the two readings worth watching across six days. VO2 max cannot remodel in that time. A higher grip on day six reflects a nervous system that now knows the device, plus a week of regular sleep, which is what the handout given to the group says. What the repeat provides is the same four readings taken after a week lived differently, so that the person can see which of them responded, and a change that can be explained is worth more than a verdict that cannot.
Blood pressure deserves the last word of this section because the consensus paper chose it as the example of a validated surrogate (Moqri et al., 2023), and because taking it twice costs nothing. One reading is noise. A series is a measurement, and the series is what the week is for.
What a single value could not tell me
I know the pull of one number because I spent two or three winters under it. Something was wrong in those months, and I went looking for the single blood value that would explain it. Every panel came back unremarkable, and each time I assumed the next marker would be the one that did not. What eventually worked would never have appeared on a report: no alcohol, a fixed bedtime, training intensity cut back, and recovery returning months before performance did. A value that reads fine while you feel anything but is a measurement of the wrong thing. In those winters I would have paid for a biological age without a second thought, and I would have had no idea what to do with it.
Which sets my own bar for one today. A result eight years above my passport age would not change a single repetition in my week until a repeat test showed the gap was more than technical drift or a poor night's sleep. And a report that cannot say whether the problem lies in metabolic health, in the liver or in recovery has not supplied a target. It has supplied a worry.
Caveats, and where the evidence is thin
Every limit I know of, because a reference page should show its weak points.
The replicate study is small. Higgins-Chen's headline deviations come from 36 blood samples measured twice, plus the group's own replicate sets; the direction is consistent across datasets, the exact maxima are not.
The nine-year figure is a maximum. Median gaps sat between 0.9 and 2.4 years. Most replicate pairs disagreed by less than the worst case, and the worst case is what a report cannot rule out.
Reliability has partly been fixed. Principal-component clocks and two-step designs agree within about 1.5 years between replicates. Whether a commercial test uses them is not stated on most reports.
Biological drift is measured over short intervals. Sehgal's repeats within days under varying conditions show instability, and how much of that is measurement noise and how much is a real short-term change in the epigenome, which would be interesting in its own right, is not settled by that paper or by any other I found.
The mortality evidence is observational. Chen's 13 cohorts show association independent of risk factors, not that lowering a clock score lowers risk.
The population critique is a perspective. Apsley and colleagues argue from the design of the clocks and from the standards an established laboratory test must meet; the paper is a reasoned position, not a trial.
The fitness cohorts are selected. Mandsager's 122,007 people were referred for a treadmill test at one hospital; Kodama's meta-analysis pools healthy cohorts with different fitness definitions.
Grip is noisy on a single squeeze. Hydration, sleep, caffeine, pain and familiarity move it by several percentile points; the published norms assume best-of-two or best-of-three.
Sleep regularity was measured once. Windred's participants wore the device for a single week, and the index was treated as a stable trait over up to 7.8 years of follow-up.
Where Atlas Cove fits
The thread through this page is the difference between a forecast and a measurement. A forecast tells you what people with your pattern tended to do. A measurement tells you what your body did this morning. A second one tells you whether something changed. That is the whole page.
An Atlas Cove week begins with readings of how your body is performing and ends, on the sixth morning, with the same readings again, so that what you take home is a comparison you can act on, read with you by the hosts who took both.
If you write the three columns and they match, you already know as much about your biological age as anyone can: it is the result of your next treadmill test, and that result you can move.
Common questions
Is biological age real?
Across populations, yes. People whose epigenetic age runs ahead of their calendar age die sooner on average, independently of age and the usual risk factors. What is not yet real, in the sense of validated for one person, is the number on an individual report, which carries a technical error and a day-to-day drift that together can exceed the biological signal.
How accurate are epigenetic age tests?
Against chronological age, the first-generation clocks miss by a median of about 3.6 years. Between two runs of the same sample, conventional clocks have differed by up to 9 years, and between two samples from the same person days apart they drift further. Newer principal-component versions agree within about 1.5 years on replicates; whether your test uses one is worth asking.
Can your biological age change in a few months?
A reported number can change in a few months. It can also change in a few days, which is the problem. Separating a real shift from technical drift, and from the state you were in when you spat in the tube, needs repeat samples and several clocks moving the same way, which is the bar researchers now propose for intervention studies and a fair one for a person comparing two kits.
What is the difference between first- and second-generation clocks?
First-generation clocks were trained to predict chronological age, so their only interesting output is the miss. Second-generation clocks were trained against mortality or a blood-chemistry score of risk, and pace-of-ageing clocks against how fast organ systems declined within a cohort over decades. All of them are regressions whose weights come from other people.
What should I measure instead of biological age?
Cardiorespiratory fitness, grip strength, blood pressure and the regularity of your sleep timing. Each predicts mortality in large cohorts, each can be repeated on you rather than on a chip, and each points at something you can do this week. Blood pressure is the only one of the four with a validated place as a surrogate endpoint.
Sources
Horvath, S. (2013). DNA methylation age of human tissues and cell types. Genome Biology, 14(10), R115. DOI: 10.1186/gb-2013-14-10-r115
Levine, M. E., Lu, A. T., Quach, A., Chen, B. H., Assimes, T. L., Bandinelli, S., Hou, L., Baccarelli, A. A., Stewart, J. D., Li, Y., Whitsel, E. A., Wilson, J. G., Reiner, A. P., Aviv, A., Lohman, K., Liu, Y., Ferrucci, L., & Horvath, S. (2018). An epigenetic biomarker of aging for lifespan and healthspan. Aging, 10(4), 573-591. DOI: 10.18632/aging.101414
Lu, A. T., Quach, A., Wilson, J. G., Reiner, A. P., Aviv, A., Raj, K., Hou, L., Baccarelli, A. A., Li, Y., Stewart, J. D., Whitsel, E. A., Assimes, T. L., Ferrucci, L., & Horvath, S. (2019). DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging, 11(2), 303-327. DOI: 10.18632/aging.101684
Belsky, D. W., Caspi, A., Corcoran, D. L., Sugden, K., Poulton, R., Arseneault, L., Baccarelli, A., Chamarti, K., Gao, X., Hannon, E., Harrington, H. L., Houts, R., Kothari, M., Kwon, D., Mill, J., Schwartz, J., Vokonas, P., Wang, C., Williams, B. S., & Moffitt, T. E. (2022). DunedinPACE, a DNA methylation biomarker of the pace of aging. eLife, 11, e73420. DOI: 10.7554/eLife.73420
Higgins-Chen, A. T., Thrush, K. L., Wang, Y., Minteer, C. J., Kuo, P. L., Wang, M., Niimi, P., Sturm, G., Lin, J., Moore, A. Z., Bandinelli, S., Vinkers, C. H., Vermetten, E., Rutten, B. P. F., Geuze, E., Okhuijsen-Pfeifer, C., van der Horst, M. Z., Schreiter, S., Gutwinski, S., Luykx, J. J., Picard, M., Ferrucci, L., Crimmins, E. M., Boks, M. P., Hägg, S., Hu-Seliger, T. T., & Levine, M. E. (2022). A computational solution for bolstering reliability of epigenetic clocks: Implications for clinical trials and longitudinal tracking. Nature Aging, 2(7), 644-661. DOI: 10.1038/s43587-022-00248-2
Sehgal, R., Borrus, D. S., Gonzalez, J., Markov, Y., & Higgins-Chen, A. (2026). Biological versus technical reliability of epigenetic clocks and implications for disease prognosis and intervention response. Aging Cell, 25(8), e70635. DOI: 10.1111/acel.70635
Borrus, D. S., Sehgal, R., Armstrong, J. F., Kasamoto, J., Gonzalez, J., & Higgins-Chen, A. (2026). When to trust epigenetic clocks: Avoiding false positives in aging interventions. Aging Cell, 25(10), e70756. DOI: 10.1111/acel.70756
Chen, B. H., Marioni, R. E., Colicino, E., Peters, M. J., Ward-Caviness, C. K., Tsai, P. C., Roetker, N. S., Just, A. C., Demerath, E. W., Guan, W., Bressler, J., Fornage, M., Studenski, S., Vandiver, A. R., Moore, A. Z., Tanaka, T., Kiel, D. P., Liang, L., Vokonas, P., Schwartz, J., Lunetta, K. L., Murabito, J. M., Bandinelli, S., Hernandez, D. G., Melzer, D., Nalls, M., Pilling, L. C., Price, T. R., Singleton, A. B., Gieger, C., Holle, R., Kretschmer, A., Kronenberg, F., Kunze, S., Linseisen, J., Meisinger, C., Rathmann, W., Waldenberger, M., Visscher, P. M., Shah, S., Wray, N. R., McRae, A. F., Franco, O. H., Hofman, A., Uitterlinden, A. G., Absher, D., Assimes, T., Levine, M. E., Lu, A. T., Tsao, P. S., Hou, L., Manson, J. E., Carty, C. L., LaCroix, A. Z., Reiner, A. P., Spector, T. D., Feinberg, A. P., Levy, D., Baccarelli, A., van Meurs, J., Bell, J. T., Peters, A., Deary, I. J., Pankow, J. S., Ferrucci, L., & Horvath, S. (2016). DNA methylation-based measures of biological age: meta-analysis predicting time to death. Aging, 8(9), 1844-1865. DOI: 10.18632/aging.101020
Moqri, M., Herzog, C., Poganik, J. R., Biomarkers of Aging Consortium, Justice, J., Belsky, D. W., Higgins-Chen, A., Moskalev, A., Fuellen, G., Cohen, A. A., Bautmans, I., Widschwendter, M., Ding, J., Fleming, A., Mannick, J., Han, J. J., Zhavoronkov, A., Barzilai, N., Kaeberlein, M., Cummings, S., Kennedy, B. K., Ferrucci, L., Horvath, S., Verdin, E., Maier, A. B., Snyder, M. P., Sebastiano, V., & Gladyshev, V. N. (2023). Biomarkers of aging for the identification and evaluation of longevity interventions. Cell, 186(18), 3758-3775. DOI: 10.1016/j.cell.2023.08.003
Apsley, A. T., Etzel, L., Ye, Q., & Shalev, I. (2025). From population science to the clinic? Limits of epigenetic clocks as personal biomarkers. Epigenomics, 17(18), 1447-1461. DOI: 10.1080/17501911.2025.2603880
Mandsager, K., Harb, S., Cremer, P., Phelan, D., Nissen, S. E., & Jaber, W. (2018). Association of cardiorespiratory fitness with long-term mortality among adults undergoing exercise treadmill testing. JAMA Network Open, 1(6), e183605. DOI: 10.1001/jamanetworkopen.2018.3605
Kodama, S., Saito, K., Tanaka, S., Maki, M., Yachi, Y., Asumi, M., Sugawara, A., Totsuka, K., Shimano, H., Ohashi, Y., Yamada, N., & Sone, H. (2009). Cardiorespiratory fitness as a quantitative predictor of all-cause mortality and cardiovascular events in healthy men and women: a meta-analysis. JAMA, 301(19), 2024-2035. DOI: 10.1001/jama.2009.681
Leong, D. P., Teo, K. K., Rangarajan, S., Lopez-Jaramillo, P., Avezum, A., Jr., Orlandini, A., Seron, P., Ahmed, S. H., Rosengren, A., Kelishadi, R., Rahman, O., Swaminathan, S., Iqbal, R., Gupta, R., Lear, S. A., Oguz, A., Yusoff, K., Zatonska, K., Chifamba, J., Igumbor, E., Mohan, V., Anjana, R. M., Gu, H., Li, W., & Yusuf, S. (2015). Prognostic value of grip strength: findings from the Prospective Urban Rural Epidemiology (PURE) study. The Lancet, 386(9990), 266-273. DOI: 10.1016/S0140-6736(14)62000-6
Windred, D. P., Burns, A. C., Lane, J. M., Saxena, R., Rutter, M. K., Cain, S. W., & Phillips, A. J. K. (2024). Sleep regularity is a stronger predictor of mortality risk than sleep duration: a prospective cohort study. Sleep, 47(1), zsad253. DOI: 10.1093/sleep/zsad253
This is an educational and strategic perspective, not personal medical advice.
The views are the author's own and not statements by Atlas Cove Lda.
Tom Wuerden · Co-Founder
Engineer turned Ironman
Read next
What does it take to host a longevity retreat? Nine lessons from our September residency
What hosting our September residency at Casas na Ferraria taught me: the breaks we forgot to plan, why one-on-ones matter most, the phone-free week our guests loved, and what we have changed.
Does heat training improve endurance? Sauna, heat acclimation and the evidence
Why heat lifts your heart rate at the same pace, what one to two weeks of heat change, how fast that fades, and what a sauna after training adds.
Does menopause change your metabolism? What happens to muscle, fat and blood sugar
In menopause the scale often stays on course while fat gain speeds up and muscle falls. What the evidence shows, what helps with sleepless nights, and what we saw in a week.