Adding life to years,
not just years to life

Lifespan, healthspan and what aging is

Two numbers matter, and confusing them is the most common error in this field.

• Lifespan — how many years someone lives.
• Healthspan — how many of those years are free of serious chronic disease and disability.

Across the twentieth century, medicine got very good at extending the first number. It was less good at extending the second, which is why the gap between them widened: more years survived, a growing share of them spent managing disease. Every serious longevity programme now targets the gap, not the endpoint. The goal is compression of morbidity — pushing the onset of decline back faster than you push back death.

Two different bets

That framing splits the field cleanly in two:

• Disease-by-disease medicine. Cure or manage one condition at a time. It works, and it has a ceiling: eliminating any single major disease adds surprisingly few years of population life expectancy, because the other age-related diseases are queued behind it.
• Geroscience. Treat the aging process itself as the shared upstream driver of those diseases. Slow it, and you delay the whole queue at once. This is the bet that makes longevity a distinct field rather than a rebranding of general medicine.

The hallmarks of aging

Geroscience is organised around a well-known set of interlinked cellular processes — the hallmarks of aging. They give AI a set of measurable targets rather than a vague notion of "getting old". The main ones:

The data layer: what AI actually learns from

No model outperforms its inputs. Longevity is a data problem before it is an algorithm problem, and the field runs on a handful of distinct data types with very different strengths.

Why biobanks changed the field

Longevity research has a structural problem: the outcome takes decades to arrive. Large prospective cohorts — national biobanks linking genomes, imaging, blood panels, lifestyle questionnaires and, crucially, long-term follow-up on who actually got sick — are what make supervised learning possible at all. Without linked outcomes you can build a model that describes aging; with them you can build one that predicts it.

Multimodal beats single-modal

The consistent finding across the field is that combining modalities beats any one of them. Methylation captures cumulative cellular wear; blood panels capture current inflammatory and metabolic state; wearables capture behaviour and recovery; imaging captures accumulated structural damage. They are correlated but not redundant, and models that fuse them predict outcomes better than the best single input.

Aging clocks and biological age

Two people are both 60. One runs half-marathons; the other has been managing three chronic conditions for a decade. Chronological age says they are identical. Biological age is the attempt to put a number on the difference.

An aging clock is a machine learning model trained to predict something age-related from biological measurements. The first generation predicted chronological age itself; the useful part turned out to be the error term.

If the model says 68 for a 60-year-old, that eight-year residual is the signal. Consistently, across cohorts, a positive residual is associated with higher rates of age-related disease and mortality; a negative one with the opposite. The clock's "mistake" carries the information.

Three generations of clock

1. First generation — trained to predict chronological age from DNA methylation at a few hundred sites. Impressively accurate, but a model trained to predict calendar age is, by construction, penalised for noticing that someone is aging unusually fast.
2. Second generation — trained on outcomes instead: time to death, disease onset, or a composite of clinical markers. These predict health outcomes considerably better, because that is what they were asked to do.
3. Third generation and beyond — trained on the rate of change measured across repeated visits, and extended to other data types entirely: retinal images, ECGs, brain MRI, voice, gait, and organ-specific clocks built from circulating proteins.

What clocks are good for — and what they are not

• Good for: comparing groups in research, as a fast surrogate endpoint in trials, and for testing whether an intervention shifts a population's residual over time.
• Weak for: telling one individual their "true age". Test-retest variability between samples can be a meaningful fraction of the effect people are trying to detect, different clocks disagree with each other on the same sample, and the number is a correlation, not a diagnosis.
• Unproven for: confirming that a specific product worked on you. A clock reading that moves is not the same as disease avoided — that link is exactly what the field is still establishing.

AI in drug discovery for aging

Bringing a drug to market is famously slow and expensive, and most of that cost is failure — candidates that die in late clinical trials after years of investment. AI is being applied at nearly every stage, though not evenly, and the honest summary is that it has compressed the early stages far more than the late ones.

Where AI enters the pipeline

• Target identification. Mining omics, literature and knowledge graphs to propose which biological mechanism to attack. For aging this is unusually hard: the hallmarks are interlinked, so hitting one perturbs others.
• Structure prediction. Predicting a protein's 3D shape from its sequence used to take months of laboratory work; it now takes minutes computationally, at accuracy competitive with experiment for many proteins. This changed what "we don't know the structure" means.
• Generative chemistry. Designing novel molecules conditioned on a target, then filtering computationally for binding, toxicity, and whether the thing can actually be synthesised.
• Repurposing. Screening approved drugs for unexpected geroprotective effects. Attractive because safety data already exists — the origin of the long-running interest in certain widely used metabolic and immune drugs as candidate aging interventions.
• Trial design. Identifying which patients are likeliest to benefit, and predicting outcomes to reduce the number of participants and years required.

The bottleneck nobody has removed

Computational design has genuinely produced clinical candidates in a fraction of the historical discovery timeline. But discovery was never the expensive part. Clinical trials are — and they are gated by human biology and calendar time, not compute. A trial that needs three years of follow-up needs three years whether the molecule was designed by a chemist or a model.

For aging specifically there is an extra problem. To show that a drug slows aging, you need an endpoint. "Fewer deaths over thirty years" is unaffordable and unrunnable, which pushes the field toward surrogate endpoints — clock readings, biomarker panels, incidence of a bundle of age-related diseases. Whether regulators accept those surrogates is the single biggest open question in the field, and it is the subject of module 7.

Early detection: the maths that decides everything

Catching disease earlier is the most immediately credible way AI extends healthy life. Deep learning models read medical images at or near specialist level in several narrow tasks, and — more interestingly — extract signals humans were not looking for: cardiovascular risk from retinal photographs, structural heart disease from an ECG, cancers on scans years before they were flagged.

Where it works today

• Imaging triage — flagging suspicious findings for a radiologist and reordering the worklist so urgent cases surface first.
• Screening support — second-reader systems in mammography and diabetic retinopathy programmes, several deployed at national scale.
• Opportunistic screening — mining scans taken for another reason for incidental signals such as coronary calcium or low bone density.
• Risk stratification — combining records, labs and wearables to predict who will deteriorate, so scarce clinical attention goes to them first.

The number that gets left out of the press release

A model announced at "94% sensitivity and 94% specificity" sounds decisive. Whether it is useful depends on a number that has nothing to do with the model: how common the disease is in the people being tested. This is the base rate, and ignoring it is the most expensive mistake in screening.

What false positives actually cost

In a screening programme the harm is not abstract. A false positive means anxiety, a biopsy or repeat imaging, occasional complications from the follow-up procedure, cost, and clinical time taken from someone who needed it. It also produces overdiagnosis: finding real abnormalities that would never have caused harm in that person's lifetime, then treating them anyway. A more sensitive model can increase this harm while looking better on paper.

Personalised prevention and digital twins

The interventions with the strongest evidence for extending healthy years are not exotic: physical activity, sleep, nutrition, not smoking, managing blood pressure and metabolic health, maintaining muscle mass, and staying socially connected. None of that is new, and none of it needs AI.

What AI can plausibly add is not a new lever but better use of the existing ones — personalisation, timing, and adherence.

• Heterogeneous response. The same diet or exercise prescription produces very different metabolic responses across individuals. Models trained on continuous glucose, microbiome and response data can predict some of that variation, moving advice from population averages toward the individual.
• Continuous rather than episodic signal. An annual check-up is one sample per year. Wearables produce a stream, which changes what is detectable: deviations from a person's own baseline, drift in resting heart rate or HRV, sleep fragmentation, and declining activity that precedes clinical decline.
• Adherence, which is the real bottleneck. Knowing what to do has never been the hard part. Well-designed nudges, coaching and feedback loops beat better recommendations that nobody follows.
• Deprescribing and interaction checking. In older adults, drug interactions and inappropriate polypharmacy do measurable harm. Automated review is unglamorous and one of the clearest wins available.

Digital twins

A digital twin is a computational model of an individual physiology, updated from that person's data, used to simulate interventions before trying them. In cardiology and diabetes, patient-specific models are already used to test scenarios — a mechanistic simulation of a specific heart, tuned to a specific patient's imaging.

Applied to whole-body aging it remains aspirational. The honest state of the art is that partial twins of well-characterised subsystems work; a twin faithful enough to answer "what will this intervention do to my aging over twenty years" does not exist. A related, more practical use is the synthetic control arm — modelling what a trial's placebo group would likely have done, to reduce how many people must be assigned to placebo.

Proving it works: evidence and regulation

Longevity is unusually vulnerable to weak evidence, because the outcome that matters arrives decades after the intervention and almost everything in the meantime is a proxy. Knowing how to rank evidence is the most portable skill in this course.

The surrogate endpoint problem

Because you cannot run a thirty-year mortality trial, the field needs an accepted stand-in. A valid surrogate has to do more than correlate with the outcome — it has to be on the causal path, such that moving it reliably moves the outcome. Medicine has a long history of surrogates that failed this test: drugs that improved the marker beautifully and harmed patients.

Two structural problems follow. Regulators in most jurisdictions do not recognise aging itself as an indication, so a "longevity drug" must be developed against a specific disease. And no aging biomarker is yet accepted as a validated surrogate endpoint, which is why proposed trials tend to use composite endpoints — the incidence of a bundle of age-related diseases — rather than a clock reading.

How AI tools themselves get regulated

• Software as a medical device. A model that diagnoses, screens or guides treatment is regulated as a device in most major markets; wellness and general-information tools sit outside that perimeter, which is where most consumer longevity products live.
• Locked versus adaptive models. Regulators have historically approved fixed models. A model that keeps learning after deployment changes the thing that was approved, which is why predetermined change-control plans have become central.
• External validation. Performance on the development dataset means little. What matters is prospective performance at sites that had no part in training — where most published models lose a substantial chunk of their accuracy.
• Post-market surveillance. Populations, scanners and clinical practice drift. A model that was accurate at approval degrades silently unless monitored.

Bias, equity, privacy and hype

The risks here are not hypothetical, and several of them make health outcomes worse for exactly the people already doing worst.

Bias that becomes clinical harm

• Unrepresentative training data. Models built mostly on one ancestry group, one country's care patterns, or one manufacturer's scanners degrade on everyone else. Genomic reference databases have historically over-represented European ancestry, which limits how well polygenic risk scores transfer.
• Proxy variables that encode inequity. The best-known case in health AI used prior healthcare spending as a proxy for health need. Because less money had historically been spent on Black patients at the same level of illness, the model systematically underestimated their need. The algorithm was working exactly as designed — the label was the problem.
• Access asymmetry. Longevity technology arrives priced for the wealthy. A field that adds years mostly to the people who already have the most of them widens a gap it claims to close.

Privacy, when the data is your biology

Genomic data is identifying, unchangeable, and partly shared with relatives who never consented. Continuous wearable streams reveal pregnancy, illness, depression and sleep patterns. The relevant questions are concrete: who holds it, what secondary uses are permitted, what happens on acquisition or bankruptcy, whether it can be deleted, and whether it can reach an insurer or employer. Consumer health platforms frequently sit outside the strict clinical data regimes people assume protect them.

Reading the hype

Longevity attracts overstatement because the promise is enormous, the evidence is slow, and the products are sold direct to consumers. Recurring patterns worth recognising:

• Mouse or worm results reported as if they were human results.
• A biomarker or clock reading presented as an outcome.
• Correlational cohort findings described in causal language.
• Model accuracy quoted from internal validation, with no external test site.
• Testimonials and n-of-1 self-experiments used in place of controls.
• A real mechanism used to imply a product that has never been tested in humans.

Best practices and common pitfalls

If you are building in this space

• Define the outcome before the model. "Predicts biological age" is not an outcome; "identifies people who will develop condition X within five years, earlier than current practice" is.
• Validate externally, on sites and populations that contributed nothing to training, and publish the drop.
• Report performance by subgroup — ancestry, sex, age band, site, device — not just in aggregate. Aggregate metrics hide exactly the failures that matter.
• Design for the base rate of the population you will actually deploy in, and report predictive values, not only sensitivity and specificity.
• Interrogate your labels. Ask what the label really measures and whose behaviour generated it, since most bias enters here rather than through the algorithm.
• Decide the deployment question early: does this replace a clinician, assist one, or triage a queue? The evidence bar and the regulatory path differ sharply.
• Monitor after launch for drift, and set a threshold at which the model is pulled rather than patched.
• Be explicit about what you are not: a wellness product should not use clinical language, and a clinical product should be regulated as one.

If you are evaluating claims

• Ask which band moves — lifespan or healthspan. Extending one without the other is not the win it sounds like.
• Check the species. Then the sample size. Then whether there was a control group.
• Separate the marker from the outcome. A clock that moved is a hypothesis about an outcome, not the outcome.
• Check whether the model was validated anywhere other than where it was built.
• Establish who profits from the reading, especially when the same company sells the test and the intervention.
• Weigh against the boring baseline. Sleep, activity, blood pressure control and not smoking have decades of evidence behind them and are the comparison any new intervention must beat.

Pitfalls that catch capable people

• Optimising a proxy until the proxy stops tracking the thing you cared about.
• Reporting accuracy on an imbalanced dataset — 99% accuracy on a 1% prevalence condition can be achieved by answering "no" every time.
• Leaking data across the split — the same patient in both training and test sets, or normalising before splitting, producing results that evaporate in the real world.
• Confusing precision with accuracy — a number reported to two decimal places from an assay with wide test-retest variability.
• Ignoring the deployment context — a model that performs beautifully and changes no decision, or that adds work to an already saturated clinic.
• Treating regulatory clearance as proof of benefit — clearance often establishes substantial equivalence, not that outcomes improve.

Test yourself — 10 questions

Pick one answer per question, then score. Explanations appear with your result.