Genealogy of a contraction, epistemic asymmetry, and the distinction between peri-interventional, indication, and predictive regimes
Since 2023, the French agenda has installed a dominant representation of the digital twin: a three-dimensional simulated organ, animated by multi-physics equations, queryable before a specialised intervention. This representation is scientifically solid, clinically promising, industrially structuring. It does not, on its own, suffice to define the concept it embodies.
The term digital twin does not canonically designate a mechanistic organ model. It designates a broader class of numerical representations coupled to a real referent: object, system, organ, patient, trajectory, cohort, territory, or care pathway. Reducing the concept to the simulated organ is a late semantic contraction, made possible by the demonstrative power of a few use cases. Conceptual dilution does not lie in broadening. It lies in restriction.
The concept was not born in medicine. Glaessgen and Stargel (NASA, 2012) formalise it as integrated multi-physics, multi-scale, probabilistic simulation of a vehicle or system. Kritzinger et al. (2018) then distinguish three architectural regimes: digital model (no automated connection to the referent), digital shadow (unidirectional feed), and digital twin in the strict sense (bidirectional coupling). This taxonomy conditions infrastructure, validation, governance, and degree of autonomy. It implies that a significant number of systems labelled digital twins are, in reality, models or shadows.
Healthcare presents a radically different configuration from the sector where the concept was structured. Fragmented data, observation biases, High Dimension Low Sample Size regime, behavioural and social determinants that are neither stationary nor homogeneous across populations. The modelling exercise is more dependent on hypotheses than in a heavily instrumented industrial universe. This is not a methodological weakness; it is an epistemic constraint.
To this data asymmetry is added a computational asymmetry worth quantifying. The reference review by Bhagirath et al. (EP Europace, 2024) documents the point without ambiguity: at a discretisation of around 400 µm, already used in VT stratification, VT ablation, and AF ablation, computational costs reach the scale of several CPU-days per patient; doubling spatial resolution may multiply cost by a factor close to ten. What this figure proves: generalising the mechanistic paradigm to massive cohorts (tens of thousands of patients per year for a given indication, hundreds of thousands at the scale of a secondary prevention policy) is not an incremental engineering problem. It is an architectural one.
A digital twin is not defined by its visual form. It is defined by a quadruplet of independent dimensions. The nature of the referent (organ, longitudinal patient, cohort, territory, care pathway) determines the class of relevant objects. The coupling regime (unconnected, unidirectionally fed, looped) conditions regulatory qualification and accountability. The temporal scale (point intervention, months, years, decades) commands data architecture. The validation regime (mechanistic fidelity, predictive utility, decisional fidelity, fidelity to real-world data) follows the finality, not an abstract hierarchy of rigour.
Current confusion stems from reducing this quadruplet to a single dimension: the three-dimensional representation of an organ. This narrowing does not bring precision. It erases.
The word prediction today covers three distinct operations that need to be separated.
Strong prediction assumes an open temporal horizon, a not-yet-declared event, a real intervention window, and the possibility of modifying the trajectory through a present decision. It grounds populational predictive and preventive medicine.
Specialised indication concerns a patient already identified, often already ill, whose interventional decision remains open: should a defibrillator be implanted, an ablation proposed, surgery performed? A hybrid zone, strategically important.
The peri-interventional regime applies once the procedure is decided: the question becomes technical, anatomical, electrophysiological, on a horizon of days, hours, or minutes. This is computational prediction, not necessarily predictive medicine in the preventive sense.
Predicting the probable outcome of an imminent procedure is not the same operation as predicting a disease that has not yet declared. The strategic consequence is clear-cut. The peri-interventional optimises a societal cost already engaged. Strong prediction aims to avoid that a societal cost be engaged. A public policy that places these three regimes under the same banner without distinguishing them may fund essentially tertiary optimisation while believing it funds populational prevention.
The organ twin has as referent an organ or a physiological system. Its clinical coupling is most often unidirectional or mediated: the simulation informs the decision, it does not act directly on the biological referent. The myocardium is not an addressable actuator. The most advanced terrain in prospective validation today is scar-related post-infarction VT. Hwang et al. (Circulation: Arrhythmia and Electrophysiology, 2024), on 18 patients: sensitivity 81.3%, specificity 83.8%, negative predictive value 98.8%. What these figures prove: on a limited but prospective cohort, a model reconstructed from standard imaging can non-invasively designate targets whose agreement with critical sites mapped in the lab is high. What they do not prove: that the twin replaces invasive mapping. Validity perimeter: ischaemic scar substrate, interventional-grade LGE-MRI.
The patient trajectory twin aggregates clinical events, biomarkers, treatments, behaviours, exposures, social context. Its scale is months to years. Pavon et al. trained an LSTM on 150 post-discharge HF patients: AUROC around 80%, performance maintained above 78% AUROC even when telemonitoring frequency was halved. What this literature establishes: at a defensible performance level, on a real clinical cohort, an object of the trajectory family exists, is validated, is deployable on lightweight infrastructure. What it does not resolve: the net effect on hospitalisations in routine practice, which depends as much on care organisation as on the predictive model itself.
The territorial twin has as referent a population basin. Its validation is not anatomical, it is decisional: better resource allocation, weak-signal detection, hospital tension anticipation, prevention campaign targeting. It is closer to a populational decision support system than to an organ simulator.
The care pathway twin has as referent a pathology, a care episode, or a medico-organisational sequence traversed by a heterogeneous patient population. It identifies frictions, delays, ruptures of continuity, losses to follow-up. Particularly relevant in oncology, rare diseases, mental health, geriatrics, post-acute pathways.
These four families are not in competition. They do not cover the same needs, are not validated by the same criteria, do not have the vocation to live within the same industrial architectures. A coherent national strategy articulates them. A confused strategy lets one of them silently absorb the other three.
First scenario. An elderly, polypathological patient in a nursing home or medically underserved area: heart failure with preserved ejection fraction, diabetes, hypertension, atrial fibrillation, chronic bronchopathy. No intervention is planned. What changes week after week is not myocardial geometry: it is volume status, respiratory rate, adherence, renal function, weight, care environment. The right object is a compact trajectory twin, not a high-fidelity multi-physics solver whose validity domain in this context is not even defined. The geometry of a heart does not change in seven days. The vital prognosis, sometimes, does.
Second scenario. A post-infarction patient with reduced ejection fraction, in the grey zone of an implantable cardioverter-defibrillator indication. Here, the organ twin is highly relevant: scar geometry, electrophysiological substrate, in silico inducibility may enrich the decision beyond ejection fraction alone. The 98% NPV reported by Hwang, if confirmed on larger cohorts, opens the way to informed de-indication for patients whose twin does not generate clinically plausible VT inducibility. The stake ceases to be only scientific; it becomes industrial: moving from a costly high-fidelity model to a version robust, fast, and integrable enough to be used beyond a few expert centres. This is the work of reduced-order models, differentiable surrogates, hybrid architectures.
These two scenarios are not opposed to MEDITWIN. They trace another map.
This thesis is neither a scientific critique of the MEDITWIN consortium nor a contestation of the relevance of its use cases. It bears on the conceptual contraction that may install itself around it. Recognising the excellence of a project within its validity perimeter does not preclude recalling that this perimeter does not exhaust the conceptual category in which it inscribes itself.
To the objection “broadening the term dilutes it”: the opposite is true. We do not dilute, we restore. The taxonomy proposed here is closer to the historical formulations of Grieves, the aerospace definition of Glaessgen and Stargel, the observations of Negri, and the categorisation of Kritzinger, than the recent contraction of the concept to the simulated organ alone. Dilution lies in the capture of a concept by its most visible instance.
To the objection “this distinction changes nothing in practice”: on the contrary, it changes the financing criteria, evaluation grids, validation regimes, and industrial trajectories. A trajectory twin evaluated with the criteria of an organ twin fails structurally, not because it is mediocre, but because it is judged within a frame that is not its own.
To the objection “the distinction between strong prediction, specialised indication, and peri-interventional is excessive”: it is, on the contrary, minimal. A solver predicts a deformation. A cardiovascular risk model predicts an event at ten years. An indication model predicts the expected benefit of a procedure in an already-identified patient. The three use the same verb. They do not belong to the same decisional regime, are not financed in the same way, do not industrialise within the same architectures, and do not produce the same public value.
Conflating these regimes leads to judging an object with the promises of another. This produces structurally false arbitrations that nevertheless preserve an appearance of procedural rationality, that is, the most difficult form of framing defect to correct, because it presents itself as its opposite.
This note does not resolve the question of clinical validation proper to each family. It does not treat in detail the governance of data, which becomes central as soon as one leaves the strongly delimited tertiary frame. It does not prejudge the relative industrial maturity of the different families: the organ twin is today more advanced than the territorial twin, both in clinical validation and in industrial integration. The rapid evolution of reduced-order models (hybrid eikonal, physics-informed networks, differentiable surrogates, compact models coupled to longitudinal data) may, in the coming years, reduce part of the computational asymmetry between high-fidelity organ twins and lighter twins. This convergence will not annul the taxonomic distinction. It will displace some of its industrial consequences.
The digital twin is not a simulated organ. It is a class of representations coupled to a real referent, defined by four independent dimensions: referent, coupling, temporal scale, validation regime. In healthcare, this class declines into at least four families, across three distinct temporal regimes. In the absence of this distinction, public arbitrations may remain rational in form while being structurally ill-posed.
The high-fidelity organ twin holds its promises within its perimeter. The public promise of populational predictive and preventive medicine is of another order. A national strategy that rigorously distinguishes them serves both. A strategy that conflates them compromises both.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.