Could Your Digital Twin Help Choose a Treatment?

Before changing something important in the physical world, engineers often explore it in a model. The idea of doing something similar for a person is both compelling and easy to overstate.

Imagine a consultation in which a clinician can compare several treatment possibilities using a model informed by a patient's measurements. A change in those measurements updates the model. Its predictions help identify the next question to investigate.

That is a hypothetical scene, not a claim that a complete, clinically reliable virtual copy of each person is available. It makes the central issue easier to see: which part of a person needs to be represented accurately enough to help with a specific decision?

A useful twin has a purpose

In a September 2026 highlighted research topic, NIH described digital twins as virtual representations that are updated from their physical counterparts, make predictions, and inform decisions. Its research interests include validation, uncertainty, data integration, and responsible use. This is an invitation to advance research and prototypes, not evidence that all proposed applications are ready for care.[1]

The word “twin” can invite an image of completeness. A useful model may instead be deliberately narrow. It might represent a particular process or outcome while leaving much of a person's life outside its scope. That limitation can be sensible if the intended decision is equally specific and the model is evaluated for it.

A weather forecast does not need to reproduce every leaf in a forest to help someone decide whether to carry an umbrella. But a forecast designed for one place and time cannot be assumed to answer every question about the climate. The same discipline applies when a medical model moves between populations, settings, or purposes.

Medicine already uses models

FDA describes model-informed drug development as the use of biological and statistical models, drawing on preclinical and clinical data, to support development or regulatory decisions. Applications can include dose selection, trial design, and the investigation of safety or response. These established categories of modeling should not all be relabeled as patient digital twins; they provide context for the broader ambition.[2]

The next step in the imagined consultation would be to ask what the model adds. Does it improve a decision compared with the information the clinician already has? Can it identify when the available data are insufficient? Has its performance been evaluated beyond the observations used to construct it?

A vivid visualization can make the result easier to discuss, but visual resemblance and predictive usefulness are separate qualities. A simple display backed by appropriate evaluation may be more helpful than a detailed avatar whose accuracy is unclear.

When the model and the person disagree

Suppose our fictional patient reports feeling worse while the model predicts improvement. That disagreement should create a question, not erase the report. Perhaps an important measurement is missing. Perhaps the situation falls outside the model's intended use. Perhaps the chosen outcome does not capture the experience that matters to the patient.

A good workflow would make it possible to investigate and document the discrepancy. It would identify who reviews the output, who can challenge it, and how new information changes the decision. More data alone do not settle which information deserves weight.

There is also the question of whose history the model has learned from. A tool can perform well on average and still be poorly suited to some people. Evaluating the relevant population and the consequences of error is part of making “personalized” a meaningful description.

Trust includes the right to understand

A person should be able to ask what information is being used, what the prediction means, and what remains uncertain. The model's role in a decision should be understandable even when its mathematics are complex. Responsibility for interpreting and acting on its output needs to remain visible.

The most useful future for digital twins may arrive through modest, well-defined successes: a specific question answered more reliably, a measurement chosen more intelligently, or an uncertainty made easier to discuss. Each would be valuable without requiring a complete simulation of a human being.