Deep divefrom Episode 11Longevity9 November 202620 min

Can a Blood Clock Predict Dementia?

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: Biological age measured by DNA methylation clocks and neurodegenerative diseases: A systematic review and meta-analysis (medRxiv preprint)

MP3

Chapters

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.

Host 2Today is one paper: a systematic review and meta-analysis posted on the twenty-first of October by Zaylea Kua, Jian Hua Tay and colleagues, led by Andrea Maier at the National University of Singapore. It asks whether the so-called epigenetic clocks, blood tests that estimate how old your body is, can tell us anything about dementia, Parkinson's disease, multiple sclerosis and motor neurone disease.

Host 1Before we get into it, has this been checked by outside reviewers yet, or is it still a draft?

Host 2It is a preprint, so it has not been through peer review, and the authors say plainly it should not guide clinical practice. Their own bottom line is cautious: one clock showed a modest link with dementia, and the rest showed limited and inconsistent links.

A clock written on your DNA

Host 1Let's start at the beginning then: what exactly is an epigenetic clock, and what is it reading in my blood?

Host 2It reads DNA methylation: small chemical tags sitting on your DNA that help switch genes on and off, without changing the genetic letters themselves. At certain spots, where the letter C sits next to a G, the share of cells carrying a tag drifts in a remarkably regular way as we get older.

Host 1So the tags change with age so predictably that you could guess someone's age from a blood sample?

Host 2Yes, that is the idea, and a computer model does the guessing. Researchers feed it methylation readings from thousands of people, and it learns which few hundred spots, weighted how, best predict the number they were trained on.

Host 1How long have people been building these clocks, and who started it?

Host 2The paper starts its search in two thousand eleven, the year the first methylation clock was published. The wider picture in the field is that the big takeoff came in two thousand thirteen, when Steve Horvath published a clock that worked across many tissues and Gregory Hannum's group published one built on blood.

Host 1And when people talk about different generations of clocks, what changed from one generation to the next?

Host 2First-generation clocks, like Horvath's and Hannum's, were trained to predict plain calendar age. Second-generation clocks, like PhenoAge and GrimAge, were trained on things linked to illness and death, and third-generation clocks, like DunedinPACE, try to measure how fast you are ageing right now.

Host 1Why would a clock trained on illness and death be any better than one trained on your birthday?

Host 2Because a perfect birthday clock would be useless: it would just tell you your age. The interesting part is the gap, and clocks trained on health outcomes are built so that the gap means something about risk.

Host 1Tell me more about that gap, because I suspect it's the number everyone actually cares about.

Host 2It is, and the paper uses two versions of it. Age deviation is simply your clock age minus your real age, while epigenetic age acceleration is what is left over after a statistical model removes everything your real age explains.

Host 1Why would that difference matter, if both of them are just measuring how far off the clock is?

Host 2Because many clocks tend to underestimate old people and overestimate young ones, so the raw difference partly tracks age itself. The leftover version strips that out, and we will see later that this choice can create a result out of nothing.

What was known before

Host 1What did researchers already know about these clocks and brain diseases before this review came along?

Host 2The wider picture is that the faster clocks, especially GrimAge, have a solid track record of predicting death and heart disease in large cohorts. For brain diseases the evidence was scattered, and the authors note that earlier reviews found no consistent link with dementia, blamed on small studies and wildly different methods.

Host 1Why would anyone even hope a blood test could reveal something happening inside the brain?

Host 2Ageing is by far the biggest single risk factor for these diseases, and blood is cheap and easy to collect again and again. If a blood clock captured the body-wide ageing that makes a brain vulnerable, it could flag risk years ahead, which is what makes it appealing.

The open question

Host 1So what was the specific question nobody had properly answered yet?

Host 2Three questions, really: whether people who already have these diseases show faster clocks, whether a fast clock today predicts getting the disease later, and whether faster ageing could actually cause it. The authors say the many new clocks had never been pooled together with numbers, only described in words.

Host 1Why is predicting the disease later such a different thing from simply seeing faster clocks in people who already have it?

Host 2Because an illness can change your blood, your medicines and your lifestyle, which can push the clock rather than the other way round. Only a clock measured before diagnosis can be a warning sign.

How they did it

Host 1Walk me through how they actually did this, starting with how they found the studies.

Host 2They searched six large databases, including PubMed, Embase, Scopus and Web of Science, for everything from January two thousand eleven to the fifteenth of April twenty twenty-six, with help from a university librarian. They registered their plan in advance in a public registry for reviews, so they could not quietly change their goals afterwards.

Host 1And how big was that haystack they were searching through to find the right studies?

Host 2About forty thousand records, of which almost twenty-five thousand were duplicates and over fifteen thousand were ruled out from the title and abstract alone. Of one thousand six hundred and ninety full papers read, forty-four made it in: thirty-seven observational studies and seven of a genetic kind we will come back to.

Host 1Who decided which studies made it in, and how did they keep that from being one person's opinion?

Host 2Two reviewers screened every paper independently, and a third settled disagreements. Studies had to involve adults, measure the clock from something easy to collect like blood, saliva, cheek cells or skin, and diagnose disease using recognised clinical criteria.

Host 1How many people ended up being covered by those forty-four studies, all added together?

Host 2After removing people counted twice across overlapping cohorts, forty-six thousand three hundred and sixty-five individuals, sixty point five percent of them women. The typical study's average participant was about sixty-nine, and single studies ranged from thirty-two people to almost nineteen thousand.

Host 1Did they also check whether the studies themselves were any good, or did everything get equal trust?

Host 2They scored quality with a standard checklist for observational studies, and most came out well: sixteen of eighteen cohort studies and fifteen of nineteen case-control studies were rated high quality. All seven genetic studies were rated high quality against their own reporting standard.

Host 1Now the pooling part: how do you combine results when every study measures things a little differently?

Host 2They pooled only one kind of result: studies that followed people without dementia and reported a hazard ratio for later developing dementia of any cause. Every estimate was converted to the same unit, the change in risk for each one standard deviation higher age acceleration, and studies that did not give enough detail to convert were dropped.

Host 1Could you put a hazard ratio and a standard deviation into plain words for me?

Host 2A hazard ratio compares how often something happens in two groups over time: one point one means ten percent more often, one means no difference. A standard deviation is the typical spread around the average, so one standard deviation higher means noticeably faster than most people, not an extreme case.

Host 1How did they deal with one cohort turning up in several different papers?

Host 2That was a real issue, because some cohorts gave more than one estimate, and those are not independent. They used a model that lets true effects differ both between and within studies, and assumed the errors of estimates from the same cohort were correlated at zero point six.

Host 1Zero point six sounds like a guess, so did they check whether the answer depended on it?

Host 2They did: they reran everything with zero point three and zero point nine. They also dropped one study at a time, swapped which overlapping reports counted, used a correction that is stricter when there are few studies, and looked for signs that small studies with flashy results were skewing things.

Host 1And how many studies actually made it into that final pooled analysis in the end?

Host 2Eight cohorts, which is few. GrimAge had seven studies with ten estimates, Horvath and PhenoAge seven studies each, Hannum six and DunedinPACE five, and dementia was the only disease with enough comparable studies to pool at all.

A word from our sponsor

Host 1This episode is supported by the Proxy Hat, the first hat that never touches your head. It estimates whether you are wearing it from surrogate markers, like your coat and your posture.

Host 2Terms apply. The Proxy Hat was trained on thirty-seven heads, not all of them human, and every result is harmonised to one standard hat.

Host 1The Proxy Hat may report that you are wearing it while you are in the bath. This is a known within-cohort correlation, assumed to be zero point six.

Host 2I wear mine to bed. It has not been there for three years, and in fully adjusted models I have never been warmer.

Host 1Leave-one-out analysis shows the hat stays on even after the head is removed. The Proxy Hat: point estimates may vary.

What they found

Host 1Before the pooled numbers, what did the individual studies show for dementia?

Host 2Looking ahead in general populations, GrimAge was linked to later dementia in six of ten studies and its update GrimAge2 in three of three, while Horvath's original clock showed a positive association in one of nine, and Hannum's in five of six.

Host 1So if I have this right, the clocks trained on health did better than the ones trained on birthdays?

Host 2For dementia, roughly yes, but even the health-trained ones were mixed: PhenoAge linked in five of eleven studies and DunedinPACE in four of eight. And oddly, among people who already had Alzheimer's, GrimAge showed no difference from controls in all five studies that looked.

Host 1What about the other three diseases, Parkinson's, multiple sclerosis and motor neurone disease?

Host 2Too few and too varied to pool. A few studies of people already ill found faster clocks, for example a higher Hannum-based measure in Parkinson's in two of three studies and higher PhenoAge in one Swedish multiple sclerosis study, but looking forward the picture was mostly blank.

Host 1Was there a big forward-looking study in that group, or were they all small ones?

Host 2There was one, Generation Scotland, with eighteen thousand eight hundred and twenty-one people followed for ten years. In fully adjusted models, no clock predicted new multiple sclerosis; the closest, Hannum, had a hazard ratio of one point two six with a confidence interval running down to zero point nine nine, so not significant.

Host 1You promised earlier that the choice of age gap could create a result out of nothing, so where does that show up?

Host 2In Parkinson's. Two studies found that people with a higher raw age deviation on Horvath's clock got the disease younger, but when the same clinical cohort was analysed with the leftover version, the link vanished.

Host 1Why would that link disappear just by switching to the leftover version of the gap?

Host 2Because the raw gap partly encodes age itself: younger patients tend to get a bigger positive gap simply because of how the clock is calibrated. The authors flag it as an age bias, not a real sign that faster ageing brings the disease forward.

Host 1Now the main event: when they pooled the dementia studies, what came out?

Host 2One clock stood out. In the most fully adjusted models, each standard deviation higher GrimAge acceleration went with an eleven percent higher risk of dementia, a hazard ratio of one point one one, with a confidence interval from one point zero two to one point two.

Host 1Is eleven percent a big effect, or just a little one?

Host 2Small. The authors call it modest, and the wider picture in medicine is that a ten percent relative increase on top of a risk that is low in any given year changes very little for an individual.

Host 1Is there a way to put it in years of ageing, which is how these clocks are usually talked about?

Host 2Yes: each extra year of GrimAge acceleration went with about three percent higher risk, a hazard ratio of one point zero two nine. With fewer adjustments, the per standard deviation figure was a bit larger, one point one four five.

Host 1And all the other clocks, did any of them come anywhere close to that GrimAge result?

Host 2None reached significance. In the fully adjusted pooling, PhenoAge was one point zero eight, DunedinPACE one point zero seven, Horvath one point zero six, and Hannum zero point nine six, all with intervals that include no effect.

Host 1How confident should we be that the GrimAge result isn't a fluke of which studies happened to be included?

Host 2Mixed. The studies agreed with each other well, with heterogeneity measured at zero percent, and the funnel plot showed no sign of small studies skewing it; but when they dropped one study at a time, the estimate stayed between one point zero eight and one point one two while its statistical significance depended on which study was left out.

Host 1So the size of the effect holds steady, but whether it clears the bar for significance is wobbly?

Host 2Exactly that. A consistent small effect sitting close to the line, built on seven studies, is what an honest description looks like.

Host 1What were those seven genetic studies you mentioned, and what did they add?

Host 2They used Mendelian randomisation, which uses the genes you were born with as a natural experiment. If gene variants that nudge your clock faster also raise disease risk, that hints faster ageing might cause the disease, since your genes were fixed before any lifestyle or illness could interfere.

Host 1And did that genetic approach find any evidence that a fast clock actually causes these diseases?

Host 2Very little. GrimAge looked causal for Alzheimer's in one study, an odds ratio of one point zero two five, but it did not survive a standard robustness check, and that analysis rested on only seven gene variants with an estimated sixteen percent power.

Host 1Sixteen percent power sounds low, so what does that number actually mean?

Host 2It means that even if the effect were real, a study like that would detect it only about one time in six. The only result replicated in two studies was Hannum's clock and multiple sclerosis, and the authors read that cautiously as a hint about immune ageing.

Why it matters

Host 1Step back for me: what is the principle this paper teaches us, beyond the specific numbers?

Host 2That a clock is only as useful as what it was trained on. GrimAge was built from methylation stand-ins for smoking and blood proteins linked to inflammation and blood sugar, so it captures body-wide wear that also feeds dementia risk, while birthday clocks largely do not.

Host 1So is GrimAge really measuring brain ageing, or is it mostly picking up things like smoking and inflammation?

Host 2The paper cannot fully separate them. The authors point out that the result held in studies adjusting for smoking, weight, education and blood-cell mix, which makes ordinary risk factors less likely to explain it, but adjustment varied: only six of seven adjusted for smoking and four for cell composition.

Host 1For an ordinary person tempted by one of these biological age tests sold online, what does this mean?

Host 2The authors say directly that their findings do not show clinical use or justify screening the population. The wider picture is that consumer tests often use different clocks and labs, and nothing here says a result from one would tell you your own dementia risk.

Host 1And for researchers, what changes in how they should use these clocks from here?

Host 2GrimAge earns a place as a candidate to test in dementia prediction models, but it has to beat what is already there. The authors stress it must add information beyond the blood tests for Alzheimer's proteins now emerging, such as p-tau two seventeen.

Host 1What can we definitely not conclude from this study, even with the GrimAge result?

Host 2We cannot conclude that a fast clock causes dementia, that slowing your clock would protect your brain, or that any clock diagnoses Parkinson's, multiple sclerosis or motor neurone disease. It is an association, of modest size, for one disease, from a handful of studies.

Honest caveats

Host 1What are the biggest weaknesses the authors themselves admit to?

Host 2The studies differed so much in clocks, people, methods and diagnoses that only dementia could be pooled. Most participants were White, and the authors note some clocks behave inconsistently across ancestries, so the results may not carry over.

Host 1Is there an issue with measuring blood when the disease itself is happening in the brain?

Host 2Yes, and the authors treat it as central. Methylation in blood and brain match only at some spots, and they cite a previous review suggesting clocks measured in brain or spinal cord tissue show more consistent links than blood.

Host 1Did anything else jump out at you while reading the paper itself?

Host 2Two small things. The abstract says forty-two thousand records were screened while the main text says forty thousand, a typo-sized slip a reviewer should catch; and the senior author declares co-founding a private longevity clinic, which is disclosed in the paper.

Host 1Who is missing from these studies that we would want to see represented?

Host 2People in midlife, between forty and fifty-nine, which may be when the changes leading to these diseases begin. Sex differences were barely studied, even though one study hints the GrimAge link may differ between men and women, and vascular dementia, the second most common kind, was hardly covered.

What to watch next

Host 1So what should we be watching for next in this field?

Host 2Clocks built for specific diseases, trained on brain-related biology rather than lifespan, and then tested in people followed from before symptoms. The authors also point to clocks for single cell types, such as one for glial cells, the brain's support cells, which showed stronger age acceleration in Alzheimer's tissue.

Host 1What would a convincing result actually look like when one of those new clocks comes along?

Host 2A clock measured in midlife, repeated over time in a diverse population, that predicts who develops a specific disease and adds something beyond age, standard risk factors and the new blood biomarkers. Until then, these clocks are research tools.

Host 1So my one-line takeaway: GrimAge hints at slightly higher dementia risk, and nothing else holds up yet?

Sign-off

Host 2That is fair: a modest, consistent link for one clock and one disease, and no good evidence yet for the rest. And it is a preprint, so we'll see what review makes of it.

Host 1That's it for this deep dive. Every weekend this same feed carries the roundup, the week's new papers in one short episode.