Medicine, Digital Twins and AI - Promising a predictive era of medicine?

In a hospital in Baltimore, doctors recently performed a heart procedure they had, in a sense, already carried out. Not on a similar patient, and not on a generic model, but on a digital copy of the very heart in front of them.

This copy had been created from the patient's own scans and then used to rehearse the precise fix. Only then did they treat the real heart, already knowing where to aim.

"We treat the twin before we treat the patient," is how the lead researcher, Professor Natalia Trayanova, put it. The copy is called a digital twin, a concept which could usher medicine into a new era of predictive practice.

The TL;DR?

• A digital twin is a virtual replica of a patient or one organ, built from their scans, labs, genetics and wearables, that updates as they change.
• In principle it lets a clinician test a treatment in simulation before giving it for real, reasoning from the individual rather than the population.

So twins are about to transform care? Not yet. The idea is powerful and parts of it already work, but we remain some way from digital twins used at scale:

• In one trial, doctors rehearsed a heart procedure on a virtual copy and eight of ten patients stayed arrhythmia-free at a year, against a usual rate near six in ten.
• But most twins model one organ, not a body, and the standards to prove one is accurate barely exist yet.
• AI helps build and run them, but it is only one of several things a twin needs, alongside data, computing power and trust.

The twin could change what medicine is able to anticipate, but much has to fall into place first. Read on for more detail.

Digital Twin Ancestors

Medicine has rehearsed on copies before. Surgeons plan complex operations on 3D-printed models of a patient's own heart or skull, built from their scans. In cancer care, radiotherapy has done it computationally for decades, building a model of a patient's anatomy from their imaging and simulating exactly where the beams will fall, in an attempt to spare healthy tissue. Both are ways of rehearsing on a patient copy before touching the real one.

But these copies are built once, for a specific treatment or procedure. The radiotherapy plan is made for a single course of treatment and does not change as the patient does.

A digital twin promises far more.

In principle it is a living copy, one that can be run forward in time, projecting how the patient might fare under one treatment, under another, or under none at all. And those projections would not be drawn from averages built on large datasets of people roughly like the patient. They would be built from the individual themselves, their own biology, their own history and their own choices.

A dashboard that displays a patient's numbers is not a twin. A risk score is not a twin. A twin simulates and predicts, updating with live data as the real patient changes. A clinician facing a difficult decision could test treatment options against the patient's twin, watch a condition unfold in simulation, and tailor interventions to this specific patient rather than to the statistical average.

Medicine would shift, at least in part, from reacting to what has already gone wrong toward anticipating it. That is the promise, and it is a genuinely exciting one.

It is also, for now, mostly a promise.

The Heart of Digital Twin Development

As of 2026, cardiology is the field furthest down the digital twin route, because to an extent the heart suits the method. It is an electrical and mechanical organ that obeys physical laws you can write as equations, and it can be imaged in fine detail.

The Baltimore operation is real. A team at Johns Hopkins, led by the biomedical engineer Natalia Trayanova and the cardiologist Jonathan Chrispin, built digital twins of ten arrhythmia patients’ hearts from their MRI scans, used them to pinpoint treatment, and rehearsed the ablation on the copy before carrying out the live procedure at the same spot.

Afterwards the arrhythmia could not be triggered in any of the ten patients, and eight stayed free of it at a mean of thirteen months, against the treatment's usual rate near sixty percent.

That is a genuine result, but still only a proof of concept.

Around it sit other early uses, all in the same young state. In heart-valve replacement, surgeons already build a model of a patient's aortic root from a scan and simulate fitting the new valve before the operation, in order to choose the size and predict problems.

The US Food and Drug Administration has spent years developing a validated virtual heart of its own, used to test how drugs affect cardiac rhythm without dosing a person. Beyond the heart, cancer twins model how a tumour might respond to chemotherapy and help adjust the treatment as the tumour changes. In diabetes, twins help personalise insulin dosing from continuous glucose readings.

AI is not the whole story

The gap between these early successes and a twin in every clinic is wide, and closing it depends on data science, validation and consensus as much as technical development.

Almost every popular account calls these "AI-powered" twins. That is only half right, and to see why, you have to look at how the model inside a twin is actually built.

There are two very different ways to do it.

The first method simulates real biology. The heart is an electrical organ, which has let physicists work out the equations describing how its electrical signal travels through muscle.

Feed in a scan of one patient's heart and you can compute how the signal moves through that specific heart, where it gets trapped, and so where the fault lies.

Nothing is learned from other patients. The model calculates what this organ will physically do, which is how the cardiac twin from the opening works.

Its strength is that every result traces back to a known law, so you can see why it reached its answer. Its weakness is that it is slow, and it only works for the parts of the body whose physics we can actually write down.

The second method uses AI. Instead of simulating the biology, it learns patterns from huge amounts of patient data, then predicts by matching a new patient to the patterns it has already seen.

A tool called Foresight, built at King's College London on the same kind of model as ChatGPT and trained on NHS records, does exactly this, forecasting where a patient's health may go next from the records of millions before them.

This approach is fast, and it works even where the biology is too complex to put into equations. But it cannot fully explain its reasoning, and is only as reliable as the data it learned from. It tells you what is likely, not why.

Perhaps the most interesting work combines the two approaches. The physicist Peter Coveney and the science writer Roger Highfield argue that the real power comes from fusing them, using AI to speed up the parts of a simulation that are too slow or too poorly understood to compute directly, while keeping the biology-based approach that lets a clinician see why a prediction holds.

But however the model is built, it is only one piece. A twin needs much more besides.

What a twin needs

A viable digital twin needs data, and a lot of it. It needs a rich, current picture of one person, drawn from scans, blood work, genetics and the steady stream from wearables. Crucially it needs that data over time, not once, since the whole point is a model that updates as the patient changes.

The twin then needs that data connected and interpreted. The useful sources currently sit in separate systems that were never built to talk to each other, hospital records, imaging archives, wearable apps. Getting them to share data in a usable form is a problem of data integration rather than science, and it remains one of the field's most stubborn barriers.

It is also demanding of computing power. A faithful organ model is a heavy simulation, and a whole-body one heavier still. The largest cardiac-twin work so far, modelling tens of thousands of hearts, was possible only with serious computing behind it. This is part of why twins are slow, and part of why AI is enlisted to approximate the most expensive steps.

And it needs validation. Before a clinician can trust a twin to guide a real decision, there has to be an agreed way to prove the model is accurate and to measure how far to rely on it. What makes this complex is that the two kinds of twin fail in different ways.

A physics-based twin hides nothing. Every result traces back to an equation you can point to, which ought to make it easy to trust. The catch is that transparency does not guarantee truth. The twin is built from that patient's own heart, its shape mapped precisely from their scan.

But a working model also needs things a scan cannot show, like how fast the electrical signal travels through each part of the muscle. Those have to be inferred, by tuning the model until its output matches the patient's readings.

The trouble is that the readings are usually sparse. It is a little like feeling a single weak pulse and trying to say why: the heart could be pumping too softly, or a valve could be leaking, or a vessel could be narrowed, and from that one reading you cannot tell which.

In the same way, when a patient's data is thin, several different combinations of hidden properties can produce the very same readings, each implying a slightly different problem inside the same heart. The model settles on one and presents it cleanly, and nothing in that clean answer tells you the data could just as easily have supported another.

An AI twin fails differently, and potentially more catastrophically, because the very thing that makes it useful is also what makes it dangerous.

AI models cope with missing information by filling the gap with the most plausible answer, which is how they handle cases at speed. But because of this, when an AI does not know, it does not stop. It generates a confident, plausible answer anyway, in the same fluent tone it uses when it is right.
In medicine this is not a quirk. An AI asked about a drug dose can return a precise figure that looks exactly like a real one, yet is dangerously wrong, and nothing on the surface tells the clinician which it is.

The danger in using an AI based twin is that it can fail very persuasively.

The frameworks for proving either kind trustworthy are only now being written, borrowed from engineering, adapted to medicine, and taken up by regulators like the FDA. Until they mature, the reach of digital twins stays deliberately limited.

What it means for the doctor

Run past all the obstacles for a moment. Imagine the data is connected, the models are validated, the regulation is settled, and digital twins work as well as anyone hopes.

Every problem in this piece, solved: In that world, is the doctor still needed?

The answer is yes, and understanding why is the whole point.

Even a perfect twin only tells you what is likely to happen to this body under this intervention. It does not tell you whether to do it. That judgement, weighing what the model predicts against what the patient wants, what they will tolerate, and what they value is not a calculation a twin can run.

The tool might become sharper. The hand must stay human.

The scan did not replace the radiologist, it gave her more to read. The blood test did not replace the physician, it gave him more to weigh. Digital twins may become the most personalised instrument medicine has ever built, but will remain an instrument requiring hands to use it and a minds to know when to put it down.

The conversation in Dubai

How twins move from a handful of watched cases into everyday practice is exactly the kind of question taken up at Next Generation Medicine 2026.

Across four days at Atlantis The Royal in Dubai, those  clinicians, technologists, regulators and investors working across longevity and regenerative medicine present the research and debate hard cases in the open.

Among them is Philippe Gerwill, a digital health futurist with nearly thirty years in global healthcare, including fifteen at Novartis, who now advises governments and organisations on AI-enabled care.

His session traces how a patient's data becomes an AI-powered digital twin, one that can predict where a disease is heading and test an intervention before it ever reaches the clinic. It is the question this whole piece has circled, taken up by someone who has spent a career at the point where technology meets the patient.

Next Generation Medicine 2026. 7 to 10 November. Atlantis The Royal, Dubai. Get tickets.