Digital twins

A digital twin is a computational replica of something real, fed with data from that real thing, allowing one to simulate what would happen if this or that were done. Aviation has used them for decades. Transferred to medicine, the idea has obvious power: build a model of the individual patient — their anatomy, their biomechanics, their physiology — and rehearse on it before operating. Let error cost computing time and never tissue.

It is also, today, one of the most inflated terms in medical vocabulary. A systematic review published in npj Digital Medicine found that the label covers eight categories of application as different from one another as hospital bed management and the simulation of a heart, and concludes that the field lacks standards and interoperability¹. Before discussing promises, then, three things need separating: what reached the clinic, what works in neurosurgery even if it goes by another name, and what is still only a promising idea.

The one that got through the door

The case that serves as a yardstick is not neurosurgical but cardiological. From a coronary CT scan, a three-dimensional model of the patient's arteries is reconstructed and blood flow is simulated on it using computational fluid dynamics, to estimate whether a lesion limits flow without the need for catheterisation. The US agency created a dedicated regulatory classification for it in 2013².

What makes that case exemplary is the complete body of evidence around it: a pivotal study measuring its accuracy against the invasive reference standard³, a randomised trial showing that its use substantially reduces catheterisations that end without findings⁴, and another randomised trial that failed to meet its primary cost objective⁵. Authorisation, measured accuracy, randomisation, and a published negative result. That is the standard of proof. No neurosurgical application comes close yet.

What does work in neurosurgery

It exists, and it is measured in millimetres. When the dura is opened, the brain shifts: the anatomy stops matching the preoperative scan on which the plan was based, and navigation loses accuracy just when it is most needed. A patient-specific biomechanical model, fed with surface images taken in the operating room, allows that deformation to be estimated and the images updated.

Work published in the Journal of Neurosurgery shows the procedure reduces target registration error from over six millimetres to under two, automatically and in under four minutes⁶. That is precisely a digital twin at work, even though the term does not appear. It is also a short, single-centre series measuring technical accuracy: nobody has yet shown it translates into more resection, less deficit or better survival. That is the exact boundary between the measured and the promised.

Where simulation fell short of the human eye

A failure is worth telling too, because it teaches more than a success. In cerebral aneurysms it has been argued for years that flow simulation can predict rupture risk. A collective experiment tested that claim as cleanly as possible: the same two aneurysms — one ruptured, one not, of similar geometry — were given to twenty-six simulation groups from fifteen countries and to forty-three neurosurgeons⁷.

The simulation groups identified the ruptured one in four out of five cases. The neurosurgeons, looking at the image, got it right nine times out of ten. And for the exact rupture site, the groups proposed six different locations and none found the correct one. The problem is not the physics: it is the boundary conditions, the segmentation and the wall properties each team chooses. When someone claims that simulation predicts rupture, the two correct questions are whether it predicts better than a trained observer and whether the laboratory next door gets the same answer. Today the documented answer to both is no.

The trials that run on a computer

They exist and they work, within narrow limits. The best-documented case was run by the US regulator itself: it generated almost three thousand synthetic patients, simulated image acquisition and reading, and reproduced with that virtual trial the conclusions of a real breast screening trial⁸.

It is a genuine achievement and also a ceiling. Simulating the physics of an X-ray detector and the detection of a lesion is a bounded problem. Simulating a tumour's response to a drug, or the functional outcome of a resection, is not. There is today no synthetic control arm supporting the approval of a treatment.

What the regulator asks of a simulation

There is a framework, and it is more mature than usually supposed. Since 2023 there has been guidance establishing how to assess the credibility of a computational model submitted in a marketing application: it requires declaring exactly what the model is used for and grading the evidence according to the risk of the decision the model influences⁹.

But it is written for physical models assessed once. A true digital twin — one that reassimilates new patient data, whose prediction today is not last week's — does not fit that scheme well. That gap is a verifiable fact, not an opinion.

What to take away

The commonest error in reading about this subject takes two forms. The first is treating "digital twin" as if it were a technology, when it is a family of things that barely resemble one another. The second, more damaging, is turning a feasibility series into a clinical result. In April 2026 the New England Journal of Medicine published a letter on arrhythmia ablation guided by a cardiac digital twin: ten patients, no control group, no randomisation, most free of recurrence at a year¹⁰. It is promising, it is in the best journal in the world, and it remains a ten-patient study without a comparator. All three are true at once.

The practical rule we use to read any work in this field is three questions: is there a comparator? was the model validated against an independent measurement in the same patient? has anyone reproduced the result at another centre? In neurosurgery, today, all three are rarely present.