Digital twins in medicine and engineering: How to tell a useful tool from an expensive mirage

Digital twins are having a moment because they sound like common sense: take a complex thing, build a virtual copy, feed it data, and use it to predict what happens next. In a wind turbine, that might mean spotting a gearbox problem before it becomes a breakdown. In a hospital, it might mean testing treatment options on a virtual heart before touching a real one.

The trouble is that “digital twin” has become a suitcase word. It can mean anything from a static computer model to a live, continuously calibrated system linked to sensors and operational context. NIST describes a digital twin as a dynamic virtual representation connected to its physical counterpart through data exchange, used for monitoring and analysis. By that definition, plenty of things sold as twins are closer to simulations plus dashboards.

Myths versus reality

Myth: A digital twin is a perfect copy.
Reality: A twin is a model with a purpose and a boundary. It mirrors what it is designed to mirror, not everything.

Myth: Real-time data makes a model trustworthy.
Reality: Bad data arrives faster too. Trust requires calibration, validation, and governance.

Myth: If a twin outputs a precise number, the system is precisely known.
Reality: Precision without uncertainty reporting is often a warning sign, especially in clinical contexts.

What makes a twin, not just a model

A simulation can be excellent without being a twin. The twin step is the link to the real world: the model is updated as the asset or patient changes, and the output is used to support decisions.

Healthcare authors sometimes classify twins as passive, semi-active, or active depending on whether they ingest real-time data and update behaviour. Industry standards are moving in the same direction: ISO 23247 lays out principles for manufacturing twin frameworks, aiming to bring consistency to what “twin” means in practice.

What feeds a twin, and why calibration is where truth lives

In engineering, twins typically drink from a firehose of sensors: vibration, temperature, power, pressure, strain. In medicine, they are built from imaging and patient records, then updated with physiological signals and lab results. The model is then calibrated, tuned so it matches the specific entity rather than an average.

Cardiovascular digital twin reviews describe a pipeline that makes this explicit: data collection, modelling, calibration, simulation and application. Calibration is the difference between “a heart model” and “this person’s heart model”. It is also where errors compound if the inputs are noisy or incomplete.

Where twins are genuinely useful

In cardiovascular care, a bounded, credible use case is planning or assessing interventions using patient-specific anatomy and haemodynamics. Reviews in major cardiovascular journals describe steady progress, alongside persistent barriers: data quality, workflow fit, regulatory credibility and the difficulty of validating predictions in complex patients.

In orthopaedics, digital twin language often means a patient-specific biomechanics workflow, sometimes combining imaging, motion capture and finite element simulation. A systematic review on arthroscopic knee surgery argues that a true twin should incorporate patient-specific data rather than generic simulators. A trauma workflow paper shows the more practical reality: build a model for a defined clinical question, then validate it for that use.

In intensive care, the attraction is dynamic state estimation: updating risk and response predictions as new observations arrive. Reviews in critical care describe potential in decision support and in silico research, while also warning that data integration, model drift and explainability are hard problems in a messy environment.

In engineering, wind turbine twins are a strong fit for condition monitoring and predictive maintenance, where telemetry is rich and failure modes are costly. A recent review frames twins as tools for monitoring and maintenance planning across components, while highlighting integration complexity and the need for reliable data and calibration. Manufacturing and factories are also pushing towards standardised frameworks such as ISO 23247, reflecting a desire to turn bespoke pilot projects into repeatable practice.

How twins go wrong

The classic failure is not a broken algorithm. It is a mismatch between how confident the output looks and how credible the evidence is.

• Bad data: sensors fail, clinical records are incomplete, imaging is inconsistent.
• Model drift: the asset wears, the patient changes, the model is not re-calibrated.
• Unvalidated assumptions: boundary conditions, physiology, load cases, and causal links are guessed, then forgotten.
• False precision: outputs are presented without uncertainty bounds.
• Security and trust gaps: a twin that centralises operational truth becomes a cybersecurity and governance risk.

Regulators have tried to bring sanity to this in medicine by insisting on credibility arguments tied to decision risk. The FDA’s guidance on computational modelling and simulation sets out a risk-informed framework and draws directly on ASME V&V 40 credibility concepts. The underlying message is simple: if a model will influence a high-stakes decision, the evidence burden rises.

How to evaluate a twin claim

Ask six questions.

1. What is the twin for, exactly? Maintenance scheduling, surgical planning, energy optimisation, diagnosis support.
2. What data feeds it, and what happens when data is missing?
3. How is it calibrated, and how often is re-calibration needed?
4. What validation exists for the intended decision? Look for a framework, not a testimonial.
5. How is uncertainty communicated? If there are no error bounds, be wary.
6. What are the governance and security controls? NIST highlights security and trust as core twin concerns.

What to watch next

Three trends matter more than marketing.

• Standards that narrow the definition gap: ISO 23247 in manufacturing and architectural work like ITU-T Y.3090 for networks are attempts to make twins more interoperable and comparable.
• Credibility-by-design in healthcare: risk-based credibility frameworks are becoming normal language for high-stakes modelling, which should reduce false precision.
• Better handling of drift and uncertainty: as more systems run continuously, the hard work shifts from building a model to operating it safely over time.

Digital twins are not an illusion. They are also not magic. The useful ones behave like good engineering: clear scope, honest uncertainty, and validation that matches the decision.

Fact-check list (claims, sources, confidence)

• NIST defines a digital twin as a dynamic virtual representation linked to a real-world entity via data exchange. Confidence: High
• NIST IR 8356 discusses characteristics, expected operational uses, and cybersecurity and trust considerations for digital twin technology. High
• ISO 23247-1 provides overview and general principles and requirements for a digital twin framework for manufacturing. High
• ITU-T Y.3090 specifies requirements and architecture for a “digital twin network”. High
• ASME V&V 40 provides a risk-based framework for assessing computational model credibility relative to how a model is relied upon in medical device decision-making. High
• FDA guidance (finalised 2023) provides a risk-informed framework for credibility assessment of computational modelling and simulation used in medical device submissions and uses concepts from ASME V&V 40. High
• Health digital twin literature distinguishes active versus passive or semi-active twins depending on real-time data ingestion and updating. High
• Cardiovascular digital twin reviews describe stages including data collection, modelling, calibration, simulation and applications, and highlight barriers such as validation and adoption. Medium-High (stage naming varies by author)
• A systematic review on arthroscopic knee surgery argues that a true digital twin should include patient-specific models, not only generic simulators. High
• A clinically applicable orthopaedic trauma workflow integrating imaging, motion capture, modelling and finite element simulation has been described in peer-reviewed literature. High
• Critical care reviews discuss digital twin technology for decision support and in silico research, alongside challenges in data integration and validation. High
• A wind turbine digital twin review describes use cases including monitoring and predictive maintenance across components and notes integration complexity. High
• Presenting highly precise outputs without uncertainty bounds is a recognised credibility risk in high-stakes modelling, addressed indirectly through risk-informed credibility frameworks. Medium (principle is widely accepted; phrasing is interpretive)