In engineering it means a live computational model of one particular machine, fed continuously by sensors, capable of prediction. Nobody has that for a human being — not a research group, not a company, not me. Here is what does exist, and what I am actually building.
The term comes from engineering, and it means something quite specific there.
NASA and Rolls-Royce build digital twins of jet engines. Not a diagram of an engine — a live computational model of that particular engine, fed continuously by sensors on the actual unit, running in parallel. You can ask it what happens if you increase the load, and it will tell you, because it knows the state of those specific components right now.
That is the standard. A model of one individual thing, updated by live data, capable of prediction.
Nobody has that for a human being. Not a research group, not a company, not me.
An engine has maybe a few thousand relevant parameters and known physics. Every interaction is describable in equations somebody has already written.
You have around thirty-seven trillion cells, a genome interacting with an environment, a microbiome with its own genome larger than yours, and a nervous system that changes the behaviour of everything else depending on what happened yesterday. Large parts of that are not merely unmeasured — the equations do not exist.
So a genuine human digital twin is not a matter of more data. It is a matter of a science that is not finished. Which does not mean the idea is useless — it means the honest version is a partial model, and the interesting question is which parts are worth modelling first.
Where the concept becomes practically useful is as an architecture. Four kinds of information, each answering something the others cannot.
Biochemical. What your blood, urine and stool show. Hormones, nutrients, inflammatory markers, metabolic by-products, microbial population. Most objective, slowest refresh rate.
Subjective. What you report — energy, mood, sleep quality, pain, digestion, cognition. Frequently dismissed as soft, and it should not be. It is the only layer capturing what you are actually experiencing.
Physiological, real-time. What wearables measure continuously. HRV, sleep staging, resting heart rate, temperature, continuous glucose. Enormous volume, moderate accuracy, updates by the minute.
Behavioural. What you actually did. Food, training, alcohol, sleep timing, adherence. The least glamorous layer and the one that most often explains the others.
The value is not in any single layer. It is in the relationships between them.
A wearable showing suppressed HRV is a signal without a cause. Blood chemistry showing a flattened cortisol curve is a cause without a timeline. Put them together with a training log and a sleep record, and you can see something neither would show alone — that the HRV drop began three weeks into a training block, alongside a change in sleep, in someone whose cortisol pattern says they had no capacity for it.
That is not a digital twin. But it is a considerable improvement on four separate systems that do not speak to each other, which is what most people currently have.
The biochemical layer is genuinely good and badly used. The tests exist and are affordable. What is usually missing is the longitudinal element — most people have snapshots taken years apart by different providers, in different units, sitting in unconnected systems. The data is not the constraint. The continuity is.
The wearable layer has volume without interpretation. Your watch will tell you your HRV was low and give you a readiness score. It has no idea whether that is iron depletion, thyroid downregulation, a virus, a bereavement or three glasses of wine. It is measuring one output of a system it cannot see into.
The subjective layer is collected inconsistently and analysed almost never — odd, given it is the one thing that determines whether an intervention worked.
And nothing joins them. That is the actual gap. Not more sensors, not more panels — a place where a cortisol result from March sits next to a sleep record from July and a note about how you felt in between, and where somebody looks at them together.
I want to describe this precisely, because this is the point where most people writing about digital twins start describing an aspiration in the present tense.
I have built a clinical knowledge base of a few hundred thousand words — test interpretation across the panels I use, movement assessment and prescription, contraindication logic, supplement evaluation, and the biomarker patterns that determine when an intervention is inappropriate. It runs as an AI clinical assistant. Clients can ask it about their results and it reasons across the markers rather than defining them one at a time.
That is not a digital twin of anybody. It is a model of clinical reasoning, applied to your data. It does not simulate you and it cannot predict what you will do next. When you ask it a question it is not consulting a model of your physiology — it is applying a body of clinical thinking to the numbers you have given it.
Which is a genuinely useful thing and a different thing. I would rather say so than borrow the more impressive term.
The layer I am actually working toward is narrower: the joining-up. A place where your results across years, your reported symptoms, and what you actually did all sit together and get read as one picture. That is a personal health record with clinical reasoning attached, and it is achievable now.
Calling it a digital twin would be marketing.
The direction of travel is real. Predictive models trained on large datasets are already flagging risk years before diagnosis. Continuous monitoring is getting cheaper and better. Multi-omics is moving from research into something affordable.
But when something is described to you as a digital twin, three questions sort the real from the marketing.
Does it model you, or people like you? A risk score derived from a population is a useful thing. It is not a model of you, and the difference is the whole concept.
Does it update? A twin is defined by the live data connection. A one-off report is a snapshot with a fashionable name.
Does it predict, and can the prediction be wrong? A model generating testable predictions — your ferritin should reach this level by twelve weeks; if it has not, absorption is the problem — is doing something. A framework that explains everything after the fact and is never falsified is doing something else.
The most valuable thing in this whole area is not a technology.
It is continuity. Somebody who has your results from three years ago and this year, knows what you changed in between, remembers what you reported at the time, and can tell you which of those things moved together.
That is not futuristic. It is what a good doctor did when people had a doctor for thirty years, and most of what the digital twin conversation is reaching for is an attempt to rebuild it with technology after we dismantled it.
The technology will get there, in part. The four layers will join up, and predictive models will get better and cheaper.
But if you want most of the benefit today, it is considerably less exciting than a twin: test the same markers with the same lab, at intervals, keep the results in one place, write down how you actually felt, and have someone read them together.
Nobody is selling that, because it does not sound like the future. It is also the part that works.
It is that nothing joins up — and joining it up is available now.
Ask AIdan → See the TDG Programme