Your Biomarkers Improved. Did Your Healthspan?

Longevity medicine runs almost entirely on markers. They are fast, quantifiable, and satisfying to both parties in a way that "come back in five years and we'll see" is not. But a marker earns its place only if moving it moves the thing the patient actually came for, and in medicine that assumption has failed often enough, and badly enough, to deserve permanent suspicion.

This is not an argument against measurement. It is an argument for knowing what a measurement is entitled to tell you.

Marker, target, endpoint

Three different things get called a biomarker, and conflating them is where most of the trouble starts.

A prognostic marker predicts an outcome. High hs-CRP is associated with cardiovascular events. That association may exist because the marker reflects a process that causes the outcome, or merely because both track something else.

A response marker shows that an intervention did what it was designed to do. A rise in IGF-1 after a growth hormone secretagogue confirms the drug engaged its target. It says nothing about benefit.

A surrogate endpoint is a marker validated to stand in for a clinical outcome in trials. This is a much higher bar, and it requires more than correlation. The marker must sit on the causal pathway, and changes in it must capture essentially the whole effect of the treatment on the outcome. If the intervention affects the outcome through routes that bypass the marker, the marker will mislead.

Most of what longevity practice measures is category one or two being used as though it were category three.

Four times the marker moved and the patient got worse

The argument is not theoretical. Medicine has run these experiments, at scale, with mortality as the readout.

Ventricular ectopy. Frequent premature ventricular beats after myocardial infarction predict sudden death, and encainide and flecainide suppressed them reliably. The Cardiac Arrhythmia Suppression Trial tested whether suppression saved lives. It was stopped early because mortality in the treated patients was roughly double that of placebo. [1] The drugs did exactly what they were designed to do, to the marker that had justified them.

HDL cholesterol. Low HDL predicts coronary events, and torcetrapib raised it substantially while also lowering LDL. The ILLUMINATE trial found increased cardiovascular events and increased mortality, and development was halted. [2]

Bone mineral density. Low BMD predicts fracture, and sodium fluoride increased it. Fracture rates went up.

Haemoglobin. Anaemia in chronic kidney disease predicts poor outcomes, and erythropoiesis-stimulating agents correct it. Driving haemoglobin toward normal increased stroke and thrombotic events.

Each of these was a reasonable inference from a real association. Each was wrong in the specific way that matters: the drug reached the marker by a route that carried consequences the marker could not see.

The general lesson was formalised three decades ago and has not been improved upon: a surrogate can only substitute for an outcome when it captures the full effect of the intervention, and most do not. [3]

Why the failure mode is structural

An intervention almost never has one effect. It has a principal effect, which is the one being measured, and a collection of others that nobody is measuring because nobody predicted them.

A marker is a window into one room of a large building. When the intervention works through that room alone, the window is informative. When it also works, or does damage, in rooms with no windows, the view is not merely incomplete. It is actively misleading, because it shows improvement while the net effect is harm.

This is why the strength of the association between marker and outcome is not reassuring. Ventricular ectopy genuinely predicts sudden death. HDL genuinely predicts coronary events. The associations were real. They just weren't the mechanism the drugs operated through.

The longevity version

Epigenetic clocks are where this problem now lives most visibly, and the honest position is more uncertain than the marketing suggests.

The clocks were built by training algorithms to predict chronological age, or mortality and disease risk, from methylation patterns. They do that well, which is a genuine scientific achievement. But being trained to predict something is not the same as being a lever that controls it, and there is an active and unresolved debate about whether the clocks are causal markers of biological ageing or statistical proxies for accumulated physiological burden. A 2026 review of the field states plainly that the clocks have not been thoroughly validated as clinical surrogate endpoints. [4]

Two findings sharpen the point.

A 2025 randomised weight-loss trial measured both epigenetic ageing and cardiometabolic markers. Changes in epigenetic age did not mediate the cardiometabolic improvement, and the longitudinal associations between the two were not significant — the clock and the health benefit moved independently. Its authors concluded that using these measures as surrogate endpoints for intervention efficacy should remain exploratory. [5]

And in the caloric restriction analysis from CALERIE, DunedinPACE slowed meaningfully, GrimAge and DunedinPoAm did not reach significance, and several other clocks moved in the opposite direction. [6] One intervention, one cohort, and the surrogates disagreed with each other about whether anything good had happened.

Despite this, trials are increasingly using clocks as primary endpoints on the assumption that they are causal markers — an assumption that is being adopted faster than it is being tested.

The cleanest example in this field

There is a longevity intervention where the surrogate and the outcome were both measured, and the result should be better known than it is.

Growth hormone in healthy older adults reduces fat mass by roughly two kilograms and increases lean mass by roughly two kilograms. That is precisely the body composition change the therapy is sold on, and it is real. In the pooled trial data it came with increased rates of oedema, arthralgia, carpal tunnel symptoms and impaired fasting glucose, and no demonstrated improvement in function. [7]

The marker moved exactly as intended. The DEXA scan would have looked like a success. Nothing the patient came for changed.

What sits closer to healthspan

If markers are unreliable proxies for functional years, the alternative is to measure function, which is not difficult and is mostly free.

Gait speed, grip strength, chair-rise time, balance, and cardiorespiratory fitness all predict disability, hospitalisation and mortality — and unlike a methylation reading, they are not standing in for the outcome. They are close to being the outcome. A patient who walks faster, carries the shopping without stopping and recovers from a chest infection in days rather than weeks has healthspan, whatever their panel says.

The practical implication is not to stop ordering laboratory tests. It is to stop letting them be the only thing that gets followed over time. A clinic that measures ten biomarkers at every visit and no functional capacity at any of them has optimised for the thing it can show the patient rather than the thing the patient wants.

Three questions before acting on a marker

Has this been validated as a surrogate for an outcome — with this intervention, in this population? For almost everything in longevity practice the answer is no. That is not a reason never to measure, but it is a reason not to describe the result as improvement in health.

Would the result change management? A marker that will not alter a decision is being ordered for reassurance, and reassurance obtained from an unvalidated number is not reassurance.

What would falsify the intervention? Define this before starting. Without it, panels expand, something always looks better, and the programme continues indefinitely on the strength of noise. Regression to the mean guarantees that an abnormal value retested will tend to look better regardless of what was done in between — which is why the outlier a patient was most alarmed by is the one most likely to improve on its own.

What markers are actually for

None of this makes biomarkers worthless. They are how early-phase research gets done, how dose is titrated, how safety is monitored, and how a clinician confirms that a drug reached its target at all. Those are real functions and they matter.

The error is narrower and more specific: presenting a marker as though it were an outcome, to a patient who is paying for outcomes. Someone who came for more good years did not buy a better number, and the distinction is theirs to understand rather than ours to blur.

The honest version of the six-month conversation is that the marker moved, that we do not yet know whether that predicts anything for them personally, and that the functional measures are how we will find out. It is a less satisfying sentence than the alternative. It has the advantage of being true.

For the framing of healthspan as a clinical goal, see Dr. Hassan Bennani's Lifespan versus healthspan. For how to grade the evidence behind any intervention, see Evidence-based longevity medicine.