Political economy of healthcare digital twins: what a paradigm costs, what it costs not to have the others
A previous article established that the term digital twin covers at least four families of objects (organ, trajectory, territorial, care pathway), governed by four independent dimensions, and that contracting the concept to the high-fidelity organ twin alone produces a temporal framing error. That demonstration closed the terminological question. It left open the question that actually decides: the budget.
The thesis fits in one sentence. French healthcare digital twin programmes are investments in real options on a medico-economic benefit whose magnitude remains to be demonstrated, while several competing families operate on documented pools at modelling unit costs one to two orders of magnitude lower, without benefiting from any comparable national envelope. The opportunity cost does not lie in what is funded. It lies in what is not, and in the temporal window during which the unfunded families remain industrially behind.
The validity domain of this note is explicit. It does not conduct an exact cost-benefit evaluation, which would be indefensible given the state of public data. It compares economic profiles across classes of objects, with an assumed uncertainty margin of a factor of two to three.
MEDITWIN was announced in December 2023 for five years (2024-2029), led by Dassault Systèmes as industrial prime, associating seven University Hospital Institutes, Nantes University Hospital, eleven Inria project-teams and four startups (inHEART, Codoc, Qairnel, Neurometers), for seven virtual health products in cardiology, neurology and oncology. None of the official communications consulted states an amount. The phrasing is invariably that of State financial support within the France 2030 framework.
This opacity is not a fault of the consortium: it reflects standard France 2030 communication practice. It nonetheless has a direct consequence, and it is the only one this note draws from it. A public investment whose envelope is not publicly detailed cannot be publicly evaluated. The reasoning therefore proceeds by comparison with structuring programmes of the same format: 50 to 150 million euros all components combined, public subsidy and industrial self-funding cumulated, consistent with the 30 to 60 million euros per Priority Research Programme and Equipment on the public share, and with the 30 million euros obtained by Gustave Roussy for its IHU PRISM in the third France 2030 call.
No conclusion in this note depends on an amount attributed to MEDITWIN. It depends on cost profiles, which is precisely the methodological benefit of the constraint.
The cost of a high-fidelity organ twin decomposes into cumulative strata whose industrial determinants differ.
The programmatic stratum is that of the consortium. The marginal stratum is that of the patient: Bhagirath and colleagues (EP Europace, December 2024) document that a spatial resolution of 400 μm mobilises several CPU-days per patient on high-performance computing resources, and that doubling the resolution multiplies the cost by a factor close to ten. Adding engineer time for segmentation and meshing, clinician time, and the underlying imaging, the order of magnitude reaches a few thousand euros per patient, a surcharge of 15 to 30 % on the procedural cost of an atrial fibrillation ablation (a stay of 8,000 to 15,000 euros under the Homogeneous Stay Group classification). This surcharge is not problematic in itself. It becomes so if the expected benefit on hard outcomes is not demonstrated at the level that justifies it.
The third stratum is the one nobody budgets. An organ twin in a research environment is not industrially equivalent to an organ twin in regulated clinical deployment. The passage presupposes a redesign along hexagonal architecture lines (isolated input and output ports, dedicated reliability ports, continuous instrumentation of quality control) and a dynamic governance framework of the Predetermined Change Control Plan type, formalised by the FDA in December 2024, which this corpus describes elsewhere as a contractualised promotion port. Without that redesign, every modification calls for full re-validation and maintenance cost becomes prohibitive. Most prototypes documented in the literature are research artifacts intended to demonstrate feasibility, not systems intended to evolve under regulatory constraint. Moving from one to the other is not an improvement: it is a change of nature. This is the hidden cost of the paradigm.
Two registers of evidence must be separated, which public communication readily aggregates.
Documented benefits. The cohort of Hwang and colleagues (Circulation: Arrhythmia and Electrophysiology, 2024) validated the technical accuracy of the ventricular tachycardia twin on 18 patients (sensitivity 81.3 %, specificity 83.8 %, negative predictive value 98.8 %) without measuring the effect on survival. The TWIN-VT study published by Chrispin, Prakosa, Trayanova and colleagues in the New England Journal of Medicine of 2 April 2026 (10 post-infarction patients, FDA investigational device exemption trial NCT03536052) reports ventricular tachycardia non-inducible in 100 % of patients after twin-guided ablation, and 8 patients out of 10 free of recurrence without antiarrhythmic therapy at a mean follow-up of 13 months, against roughly 60 % success at one year for standard of care. What these figures prove: clinical feasibility of the paradigm is now supported on hard outcomes. What they do not prove: anything approaching a randomised multicentre demonstration, on a ten-patient cohort with no control arm and an industry co-author.
Projected benefits. An accurate electrophysiological twin could improve the success rate of persistent atrial fibrillation ablation by 5 to 15 points. What distinguishes these benefits from the former is not their prior probability, which may be high. It is their epistemic status. So long as the option is not exercised by a demonstration at scale, the investment remains a cost; it becomes a benefit only retrospectively. No randomised MEDITWIN trial on hard outcomes is published as of this note, and the associated prospective cohorts (including RHU-TALENT at Bordeaux University Hospital) will not deliver results before 2027 or 2028. The medico-economic study run by the University of Rotterdam within the European inEurHEART project (112 patients, 16 centres) will constitute the first published European reference on the cost-effectiveness ratio of an industrial cardiac twin in real-world conditions.
The scientific debate on feasibility is beginning to be settled. The allocative debate on comparative relevance remains entirely open. These are two distinct debates, and the first does not close the second.
Assessing comparative allocative relevance requires an object whose pool is already documented. Heart failure is one.
Santé Publique France records approximately 1,376,692 adult prevalent cases in 2022, or 2.6 % of the adult population and 23.7 % beyond age 85, and 181,178 adults hospitalised for acute heart failure in the same year, the leading cause of hospitalisation of the elderly. At the Cnam average of 8,338 euros per stay, the acute hospital line represents roughly 1.5 billion euros per year. A 10 % reduction in avoidable hospitalisations would represent on the order of 150 million euros per year, a 15 % reduction roughly 225 million.
These amounts are not acquired net savings, and the note does not present them as such. They define a gross addressable ceiling, whose conversion depends on deployment cost, patient adherence, coordination across ambulatory, hospital and medico-social sectors, the false-positive rate, and the share effectively capturable under the current tariff system. Several of these factors can turn a significant ceiling into a modest net saving. The distinction is not accounting, it is doctrinal: an unaddressed pool remains theoretical, an instrumented pool begins to become economic.
The order of magnitude of accessible performance is documented. Pavon and colleagues, on 150 post-discharge patients followed for one year, obtain with an LSTM recurrent network exploiting telemonitored vital signs and patient profile an area under the ROC curve of around 80 % for dynamic 30-day readmission prediction, maintained above 78 % when telemonitoring frequency is halved. One conceptual precision is required here, because it conditions the reading of the entire medico-economic literature in this field: a trajectory twin is not a model, it is a loop of prediction, decision, action, re-evaluation. The effect of an isolated predictive model is by construction limited; the effect of a loop mobilises four links, each of which can be the limiting factor. That meta-analyses on telemonitoring remain contrasted therefore does not disqualify the model: it measures the organisation into which it is inserted.
No dedicated France 2030 envelope exists for this family. The ETAPES scheme, instituted in 2018, is not backed by a structured trajectory twin programme.
The second case study demonstrates that the organ twin itself can target a high cost-benefit ratio, provided it is deployed in the right configuration.
The indication for an implantable cardioverter-defibrillator in post-infarction primary prevention rests largely on a left ventricular ejection fraction threshold of 35 % or below, validated by MADIT-II and SCD-HeFT, whose specificity is limited: a significant number of implanted patients never trigger the device. According to the September 2022 HAS evaluation, approximately 10,000 implantations are performed each year in France, at a ten-year cost of 25,000 to 40,000 euros per patient, an annual line of roughly 200 to 400 million euros.
The hypothesis must be stated with care. The 98.8 % negative predictive value reported by Hwang qualifies the identification of ablation targets in ischaemic scar ventricular tachycardia. It does not, as it stands, bear on the long-term sudden death risk that structures the primary indication. Bridging the two clinical objects in one step would be a shortcut, and this note does not take it. The operative hypothesis is more modest: an electrophysiological twin could become one component of an integrated de-indication score, articulated with the fibrotic substrate on late gadolinium enhancement MRI, the competing risk of non-rhythmic death, medical therapy optimisation and comorbidities, which presupposes a longitudinal cohort of several hundred to several thousand patients and a non-inferiority demonstration under de-indication intent. Under that reservation, identifying 10 to 15 % of current candidates as at sufficiently low risk would open a pool of 25 to 100 million euros per year in gross ceiling.
Two explicit architectural conditions govern the result. Compact form first: a twin intended for de-indication must be deployable beyond expert centres equipped with dedicated high-performance computing, which calls for a reduced-order model, either by hybrid eikonal approximation (computing arrival times rather than detailed potentials, moving from several CPU-days to a few minutes) or by a differentiable surrogate trained to reproduce the reference solver’s outputs. Indication regime second, rather than pure peri-interventional: the twin no longer serves to optimise a gesture, but to prevent it from taking place. The saving is no longer marginal, it is radical.
It remains to explain why the concentration occurs, even though the other families present higher expected ratios in order of magnitude. The answer is not a strategic intent. It is the administrative form of the fundable project.
The high-fidelity organ twin fits naturally into the large-programme format: identified industrial prime, unified computing platform, public-private alliance over five years, commercialisable deliverables, deployment on sovereign cloud. Each of these elements corresponds to an existing vehicle, Bpifrance i-Démo calls, IHU endowments, Inria agreements, CIFRE contracts.
Populational families present the inverse profile: cost dispersed across actors none of which alone carries the modelling, and data governance that simultaneously mobilises the National Health Insurance, the Regional Health Agencies, ambulatory medicine, the medico-social sector and territorial collectivities. Structuring investment finds no natural sponsor. It is neither a medical device, nor an industrial product, nor a research infrastructure. These families are fundable in theory and lightly funded in practice, because they fall into the interstices between vehicles. A conceptual framing problem compounds this: the corpora exist (PMSI, FRANCIM registries, the REIN database, the chained SNDS), but they are read as retrospective evaluation tools rather than as closed-loop decision infrastructures.
An allocation logic can be rational for each of the actors composing it and collectively suboptimal for the system that results. Without coordination, allocation stacks individually defensible decisions that produce a structural bias. There is, in 2026, no published and arbitrated French national strategy distinguishing the four families, comparing their expected ratios, and allocating accordingly.
First objection: comparing a pre-competitive research programme with a populational infrastructure compares objects of different maturity. That is correct, and it is precisely why the note retains the conservative bound for projected benefits and the median bound for costs. This asymmetry is deliberate: it shields the analysis against the charge of having maximised the benefits of counterfactual reasoning.
Second objection: the pools advanced are gross ceilings, not savings. That is the note’s thesis, not its blind spot. The argument does not bear on acquired savings. It bears on the order of magnitude of a pool that current allocation does not put in a position to be captured, and on the fact that this order of magnitude justifies conducting an explicit arbitration.
Third objection: this note attacks MEDITWIN. It does not demonstrate a misallocation. It demonstrates the absence of the public framework allowing one to know whether the current allocation is sound. The difference is not rhetorical: the first charge would be unverifiable given the state of public data, the second is verified by simple observation of absence.
A national digital health strategy needs several paradigms to be effective. A strategy that preferentially funds a single one may do so for sound industrial reasons. It must then name what it does not fund, and accept the associated opportunity cost.
The explicitation asked for is technically feasible. It presupposes a comparative medico-economic mapping exercise, distinguishing the four families, their validation regimes, their expected cost-benefit ratios and their reciprocal opportunity costs. The Haute Autorité de Santé, the National Health Insurance, Inserm or France 2030 itself each hold the legitimacy to pilot it.
Not to name this cost is to count by halves. To count by halves is to arbitrate without knowing it. To arbitrate without knowing it is to fund one objective while believing one funds another.
Doctrinal notes and explorations on AI in regulated systems. Once or twice a month. One-click unsubscribe.