Digital Twins

FRCS Level

The Principle

A model becomes a twin when it is tied to one individual and updated as that individual changes.

A model built from a patient's scan is patient specific. It becomes a twin only when new data about that patient keeps changing it.

The difference is the update loop, and it is the part most often missing from things called digital twins.

At the Knee

The components exist separately. Patient specific geometry from imaging, a finite element or musculoskeletal model, and continuous data from wearables. 3.16.4 and 3.16.8 cover two of them.

What is missing is the loop that feeds the data back into the model and revises it.

A knee twin would predict how this patient's cartilage responds to this patient's activity, and update the prediction as the activity changes.

Nothing at the knee currently does that outside research demonstrations.

Industrial digital twins work because the physical system is manufactured to a specification. A knee is not, which is why biological twins are harder than the phrase suggests.

What Changes It

Availability of continuous individual data, which wearables are beginning to supply.

Computational speed, since a model that takes hours cannot update usefully. 3.16.6 covers the shortcut.

Material property estimation in a living person, which remains the hardest gap. 3.16.4 covers it.

Validation methods for an individualised prediction, which differ from population ones.

Clear definition, since the term is used for anything patient specific.

Clinical Relevance

Ask whether a claimed digital twin has an update loop, which is the definition.

Treat patient specific modelling as valuable and as a different thing from a twin.

Expect material properties to be the limiting gap rather than geometry or computation.

Note that nothing at the knee currently meets the definition outside research.

Fellowship Level

At the Knee

The plausible near term version is narrower than the phrase: a patient specific model of one tissue, updated by one stream of data, answering one question.

Cartilage response to cumulative loading is the obvious candidate, because the exposure is now measurable and the outcome is slow enough to track. 3.16.8 covers the exposure.

The hard part is not the model. It is estimating this person's material properties without a biopsy, and current practice assumes population values. 3.16.4 covers that assumption.

Quantitative magnetic resonance imaging offers a route to individual cartilage composition and lacks clinical thresholds. 5.11 covers cartilage imaging.

Without individual material properties, a twin is a population model wearing one patient's shape, which is a weaker claim than the term implies.

What Changes It

Whether individual material properties can be estimated non invasively.

Quality and continuity of the data stream feeding the update.

Model reduction and learned surrogates, which make the update computationally feasible.

The specific question, since a narrow twin is tractable and a general one is not.

Validation design, since predicting one person's future cannot be validated by a cohort.

How It Is Measured

Validation of an individualised prediction requires prospective follow up of that individual, which is a different study design from anything else in this chapter.

Group level validation tells you the method works on average, which is the thing a twin is supposed to improve on.

Research demonstrations report geometric and kinematic agreement rather than predictive accuracy.

Imaging based material property estimation, which is active research without clinical thresholds.

No knee digital twin has been prospectively validated to predict an individual's outcome, which is the claim the concept rests on.

Clinical Relevance

Judge a twin by its prediction of a future event in that individual rather than by its resemblance to them.

Expect material properties to be assumed from a population, and ask whether they were.

Treat narrow single tissue twins as the plausible version and general ones as marketing.

Note that group validation does not establish an individual prediction.

Consultant Perspective

At the Knee

The decision this page poses is about language before it is about technology: whether to let a term be used in a department when the thing it names does not yet exist there.

On one side, the concept is coherent, the components are real, and premature dismissal has a cost in missed early adoption. 3.16.8 covers the component that arrived.

On the other, the term is already being applied to ordinary patient specific models, and a borrowed word that implies prediction where there is only resemblance misleads colleagues and patients.

What moves the weight is whether an update loop and a prospective validation exist. Those two questions deflate most current claims and will eventually stop deflating them.

Both positions are defensible: insisting on the strict definition, or using the term loosely and qualifying it. What is not defensible is presenting a patient specific model to a patient as a prediction about them.

What Changes It

Whether an update loop exists, which is the definitional question.

Whether individual rather than population material properties were used.

Whether any prospective individual validation has been published.

Who is making the claim, since this term currently travels faster in industry than in literature.

The use to which the output is being put, since a planning aid and a prognosis are different.

How It Is Measured

Prospective individual prediction studies, which is the measurement that would settle it and which do not yet exist at the knee.

Component validation, which is what is currently published and is necessary rather than sufficient.

Agreement between the twin and the patient on a measurable present quantity, which is resemblance rather than prediction.

Regulatory evaluation, which is beginning to distinguish decision support from visualisation.

There is no agreed standard for what a clinical digital twin must demonstrate, which is why the term is doing work the evidence is not.

Clinical Relevance

Ask for the update loop and for prospective individual validation before accepting the term.

Describe a patient specific model to a patient as a model of their anatomy rather than as a prediction about them.

Hold a stated position on the terminology rather than letting it drift.

State that no agreed standard exists for what a clinical digital twin must demonstrate.

Weigh the patient's age and physiology, comorbidities, knee demand, psychological state and their own ideas, concerns and expectations before showing them a model of their own knee, because a picture of your own joint failing is a powerful thing to be given.